Research on bearing fault diagnosis based on machine learning and SHAP interpretability analysis

Abstract

Bearing faults are the most common type of failures in rotating machinery, accounting for approximately 30%-40% of total failures and posing serious threats to industrial production safety and economic efficiency. Traditional diagnostic methods based on physical models and empirical thresholds suffer from poor adaptability and low accuracy, while existing machine learning methods lack interpretability in their decision-making processes, limiting their application and promotion in industrial settings. This research proposes an integrated bearing fault diagnosis method combining multiple machine learning algorithms with SHAP interpretability analysis. A comprehensive experimental platform was constructed, including motors, couplings, loading systems and high-precision data acquisition systems, collecting vibration signals from normal and faulty bearings at a sampling frequency of 25.6 kHz. Raw signals were preprocessed using sliding window techniques and 15 key features covering time-domain, frequency-domain, and statistical characteristics were extracted, forming a comprehensive multi-domain fusion feature system. Ten representative machine learning algorithms (including LogReg, SVM, KNN, DT, RF, AdaBoost, GBoost, XGBoost, LGBM and NB) were systematically compared and analyzed, with comprehensive evaluation using multi-dimensional metrics including accuracy, precision, recall, F1-score, ROC AUC and log loss. Experimental results show that the XGBoost model performed optimally, achieving 91.0% accuracy, 91.9% precision, 98.9% recall, 95.3% F1-score and 62.7% ROC AUC. Multiple models achieved 100% recall, effectively avoiding fault omission. SHAP technology was systematically applied to interpretability analysis in bearing fault diagnosis, revealing that spectral_entropy, rms and impulse_factor are the most important features influencing diagnostic decisions, with their importance ranking highly consistent with the physical mechanisms of bearing faults. SHAP analysis not only provides transparent interpretation of model decisions but also offers scientific basis for feature engineering optimization and industrial applications. This research provides a complete, efficient and interpretable intelligent diagnostic solution for bearing fault diagnosis, with significant implications for advancing predictive maintenance technology development.

Introduction

Rolling bearings, as the most critical fundamental components in rotating machinery systems, directly determine the operational safety and reliability of entire equipment ¹. According to statistical surveys by the IEEE Industrial Application Society and Japan Electrical Manufacturers Association, bearing faults are the most common type of failures in rotating machinery, accounting for approximately 30%-40% of total failures ^2,3. In modern industrial environments, the economic impact of bearing failures extends far beyond simple component replacement costs. Unplanned downtime in manufacturing facilities can result in losses of $50,000 to $200,000 per hour, while in critical infrastructure such as power generation or chemical processing, the costs can reach millions of dollars per incident. In the context of Industry 4.0 and intelligent manufacturing, unexpected shutdowns caused by bearing faults not only result in enormous economic losses but may also trigger serious safety accidents ^4,5.

Traditional bearing fault diagnosis methods are primarily based on frequency-domain analysis of vibration signals and empirical threshold judgment. While these methods can achieve certain effectiveness under specific operating conditions, they generally suffer from the following limitations: First, traditional methods rely heavily on analytical models that assume ideal operating conditions, but real industrial bearings operate under complex conditions involving variable loads, contamination, misalignment, and temperature fluctuations that violate these fundamental assumptions. First, methods based on physical models assume no sliding between rolling elements and raceways, but in actual operation, there is often composite motion of rolling and sliding, causing calculated fault characteristic frequencies to deviate from true values ^6,7. Second, industrial bearings often develop multiple concurrent faults that interact in complex ways, and traditional frequency-domain methods struggle to separate and identify overlapping fault signatures, particularly when faults occur simultaneously on different bearing components. Second, when multiple faults occur simultaneously, complex interactions among various fault frequencies make traditional methods unable to effectively separate and identify them ⁸. Third, industrial environments subject bearing monitoring systems to various forms of interference, including electromagnetic noise, structural vibrations from adjacent machinery, and environmental variations that can mask subtle fault signatures. Additionally, external vibration interference and noise pollution can mask effective fault features, reducing diagnostic accuracy ^9,10.

With the rapid development of machine learning and deep learning technologies, data-driven intelligent fault diagnosis methods have gradually become research hotspots ^11,12. Machine learning approaches offer compelling advantages including automatic pattern discovery in high-dimensional data, adaptive learning capabilities that can adjust to changing operating conditions, and natural accommodation of multi-modal sensor data for comprehensive bearing health assessment. Deep learning methods can automatically extract high-level abstract features, avoiding the tedious manual feature engineering process in traditional methods ^13,14. However, existing research still shows significant deficiencies in model interpretability, lacking transparent understanding of model decision-making processes in engineering applications, which greatly limits the promotion and application of intelligent diagnostic systems in industrial settings ^15,16.

The adoption of machine learning in industrial fault diagnosis faces a fundamental challenge: the “black box” nature of many algorithms creates a trust gap between AI systems and human operators. This interpretability challenge is particularly acute in industrial settings where safety-critical decision making requires clear understanding of AI recommendations, regulatory compliance demands documented justification for maintenance decisions, and successful deployment requires seamless integration with existing domain expertise from maintenance engineers and reliability specialists.

Addressing the above problems, this research proposes an integrated bearing fault diagnosis method combining multiple machine learning algorithms and introduces SHAP (SHapley Additive exPlanations) interpretability analysis technology as a systematic solution to bridge the gap between AI performance and industrial trustworthiness. SHAP represents a significant advancement in explainable AI, grounded in cooperative game theory and Shapley values, providing a mathematically rigorous framework for feature importance attribution that ensures consistent and reliable interpretability across different models and datasets. to systematically solve three key issues:

(1) Construction of a 15-dimensional comprehensive feature system covering time-domain, frequency-domain, and statistical characteristics, improving fault feature discrimination capability through multi-domain fusion and overcoming limitations of single feature extraction methods ^17,18;

(2) Comparative analysis using ten representative machine learning algorithms with systematic performance evaluation to determine optimal model configurations, providing scientific algorithm selection criteria for different application scenarios ^19,20;

(3) Systematic application of SHAP technology to feature importance analysis in bearing fault diagnosis, quantifying each feature’s contribution to model decisions, achieving transparent interpretation of model decision processes, and providing theoretical guidance for feature engineering optimization and industrial applications ^21,22.

The introduction of SHAP into bearing fault diagnosis represents more than a technical advancement; it offers a pathway toward trustworthy AI systems that can be confidently deployed in safety-critical industrial environments, enabling both system-level understanding and case-specific troubleshooting while facilitating knowledge transfer between AI systems and domain experts.

The paper is organized as follows: Section Methodology details the complete methodology including experimental platform construction, data acquisition and preprocessing, feature engineering design, machine learning model selection, and SHAP interpretability analysis. Section Experimental results presents experimental results including feature correlation analysis, model performance comparison, detailed classification performance analysis, and SHAP interpretability analysis. Section Discussion provides in-depth discussion of feature engineering effectiveness, model performance differences, SHAP analysis value, and method limitations with improvement directions. Section Conclusion summarizes research contributions and outlines future work.

Methodology

Experimental platform construction

This research constructed a comprehensive bearing fault diagnosis experimental platform, as shown in Figs. 1 and 2. The platform mainly includes the following core components:

Fig. 1.

Bearing fault diagnosis experimental flowchart.

Fig. 2.

Bearing fault diagnosis experimental platform photographs: (a) Complete system showing electric motor, coupling, dynamometer, and control systems in controlled laboratory environment; (b) Close-up view of test bearing assembly, torque transducer, encoder, and signal acquisition components (from our previous work ⁵⁷).

The experimental platform employs a three-phase induction motor (2HP, AC 220V) as the driving source, connected to the transmission shaft and SKF 6203 deep groove ball bearing through a rigid coupling. Vibration signal acquisition uses a PCB Piezotronics 352C33 accelerometer (sensitivity 100 mV/g), mounted vertically on the bearing housing surface. The data acquisition system utilizes an NI USB-4432 high-precision acquisition card (24-bit resolution, maximum 102.4 kHz sampling rate), coupled with LabVIEW software for real-time signal acquisition and storage.

This study utilizes the same well-established experimental platform as our previous research ⁵⁷, ensuring consistency and comparability of results while extending the analysis to focus specifically on interpretability through SHAP technology. The proven reliability of this experimental setup provides a solid foundation for the current investigation into machine learning interpretability for bearing fault diagnosis.

The photographic documentation in Fig. 2 validates the systematic approach described in the methodology, showing the integration of precision instrumentation with industrial-grade components necessary for reliable bearing fault diagnosis research. Key components visible in the photographs include the electric motor, drive end bearing, torque transducer and encoder, dynamometer, signal acquisition system, and fan end bearing. The controlled laboratory environment ensures consistent experimental conditions while the professional equipment setup demonstrates the capability to generate high-quality, repeatable experimental data suitable for machine learning algorithm development and validation.

Data acquisition and preprocessing

Fault type design

To construct a complete bearing fault diagnosis dataset, this study employs artificial fault induction methods to create three typical fault modes on SKF 6203 bearings:

Inner race fault: Using wire cutting technology to create local defects with 0.3mm diameter and 0.1mm depth on the bearing inner race raceway, simulating inner race spalling or wear faults;

Outer race fault: Creating surface scratches of 2mm length and 0.2mm width at fixed angular positions on the bearing outer race, simulating outer race surface damage;

Rolling element fault: Using electrical discharge machining technology to form 0.25mm diameter pitting defects on the ball surface, simulating rolling element fatigue spalling.

While this study employs artificially induced faults, this approach is scientifically justified and widely adopted in bearing fault diagnosis research for several compelling reasons. Artificial fault induction allows for precise control of fault severity, location, and type, enabling systematic investigation of specific fault characteristics. This controlled approach is essential for establishing baseline understanding of fault signatures before progressing to more complex real-world scenarios ^6,7.

The artificial faults created in this study (spalling, surface damage, pitting) closely replicate the most common failure modes observed in industrial bearings. The fault induction methods (wire cutting, electrical discharge machining) produce defect characteristics similar to those caused by natural wear, fatigue, and contamination processes. We acknowledge that artificially induced faults may not capture the full complexity of natural fault evolution, including gradual degradation patterns, multiple concurrent faults, and environmental influences. Future research should incorporate naturally evolved faults from industrial environments to validate the generalizability of the proposed method.

Signal acquisition parameter settings

Based on bearing fault characteristic frequency analysis and the Nyquist sampling theorem, the sampling frequency is set to 25.6 kHz to ensure capture of high-frequency impact components generated by bearing faults. Specific acquisition parameters are as follows:

• Sampling frequency: kHz

• Acquisition duration: 1 second per acquisition (25,600 sampling points)

• Acquisition cycle: Once every 3 seconds, continuous operation for 30 minutes

• Signal channels: Single channel vertical acceleration signal

• Data format: 16-bit binary format storage

Through the above parameter settings, 500 normal state samples and 500 fault state samples were ultimately obtained, forming a label-balanced binary classification dataset.

Dataset scale justification

The dataset scale of 1000 samples (500 normal, 500 fault) in this study is designed based on several important considerations and is consistent with established practices in bearing fault diagnosis research. This sample size provides a reasonable balance between experimental feasibility and statistical validity.

For binary classification tasks with balanced datasets, statistical theory suggests that 1000 samples can provide adequate statistical power for detecting significant differences between classes. With 800 training samples (400 per class) and 200 test samples (100 per class), the study maintains sufficient statistical confidence while avoiding the computational overhead of unnecessarily large datasets.

To address potential overfitting concerns, we implemented stratified sampling with fixed random seeds (random_state=42) ensuring reproducible train-test splits. The consistency of results across multiple performance metrics (accuracy, precision, recall, F1-score, ROC AUC) demonstrates that the models are not merely memorizing the training data but learning generalizable patterns.

This study emphasizes systematic feature engineering and interpretability analysis rather than deep learning approaches that typically require larger datasets. The 15-dimensional engineered feature space provides rich discriminative information that enables effective learning with moderate sample sizes, as evidenced by the high performance across multiple algorithms.

Signal preprocessing techniques

Sliding window segmentation: Sliding window technology is employed to segment long time series into fixed-length time windows. Let the original signal be x(n), window length be points, and sliding step be points, then the i-th window signal is:

1

Signal normalization: To eliminate the influence of dimensional differences between different features, the Z-score normalization method is adopted:

2

where and are the mean and standard deviation of features on the training set, respectively.

Feature engineering

Feature system design

This research constructs a 15-dimensional comprehensive feature system covering time-domain, frequency-domain, and statistical characteristics, as shown in Tables 1 and 2. Each 1024-point signal window extracts a 15-dimensional feature vector, forming structured machine learning input data.

Table 1.

Feature description of bearing vibration signals.

Feature Name	Category	Description	Physical Significance
mean	Time Domain	Average value of vibration signal	Signal offset degree, reflecting system static deviation
std	Time Domain	Standard deviation	Signal fluctuation amplitude, indicating vibration intensity
skewness	Time Domain	Skewness	Signal distribution symmetry, reflecting asymmetry of vibration
kurtosis	Time Domain	Kurtosis	Signal sharpness, fault-sensitive parameter
rms	Time Domain	Root mean square	Total energy, effective value of vibration signal
peak_to_peak	Time Domain	Peak-to-peak value	Amplitude range between maximum and minimum values
variance	Time Domain	Variance	Signal value dispersion, energy scatter degree
crest_factor	Statistical Factor	Crest factor	Peak-to-RMS ratio, sensitive to impact detection
impulse_factor	Statistical Factor	Impulse factor	Peak-to-mean ratio, for impact feature extraction
shape_factor	Statistical Factor	Shape factor	RMS-to-mean ratio, waveform characteristics
clearance_factor	Statistical Factor	Clearance factor	-to-RMS ratio, sensitive to early faults
waveform_factor	Combined Factor	Waveform factor	RMS-to-mean ratio, comprehensive waveform features
entropy	Information Entropy	Time-domain entropy	Signal complexity, reflecting disorder degree
spectral_centroid	Frequency Domain	Spectral centroid	Main frequency concentration position, frequency distribution center
spectral_entropy	Frequency Domain	Spectral entropy	Spectral dispersion and complexity, frequency domain complexity

Table 2.

Feature collection methods and processing procedures.

Feature Name	Collection Device	Collection Method	Processing Steps
mean	PCB Accelerometer	Time-domain direct calculation	Calculate arithmetic mean of 1024-point window
std	PCB Accelerometer	Time-domain statistical analysis	Calculate standard deviation of 1024-point window
skewness	PCB Accelerometer	Time-domain statistical analysis	Standardized third-order central moment calculation
kurtosis	PCB Accelerometer	Time-domain statistical analysis	Standardized fourth-order central moment calculation
rms	PCB Accelerometer	Time-domain calculation	Square Mean Square root
peak_to_peak	PCB Accelerometer	Time-domain extreme values	max(signal) - min(signal)
variance	PCB Accelerometer	Time-domain statistical analysis	Second-order central moment calculation
crest_factor	PCB Accelerometer	Time-domain combined calculation	Peak RMS
impulse_factor	PCB Accelerometer	Time-domain combined calculation	Peak Mean
shape_factor	PCB Accelerometer	Time-domain combined calculation	RMS Mean
clearance_factor	PCB Accelerometer	Time-domain combined calculation	RMS
waveform_factor	PCB Accelerometer	Time-domain combined calculation	RMS Mean
entropy	PCB Accelerometer	Information theory calculation	Probability density Shannon entropy calculation
spectral_centroid	PCB Accelerometer	FFT frequency-domain transform	Spectrum weighted average frequency calculation
spectral_entropy	PCB Accelerometer	FFT frequency-domain transform	Power spectral density Frequency-domain entropy

Feature calculation formulas

This section presents the essential formulas for key features, while complete mathematical definitions for all 15 features are provided in Appendix A.

Core discriminative features (identified by SHAP analysis):

Root mean square (RMS):

3

Impulse factor:

4

Spectral entropy:

5

where is the normalized power spectral density from FFT analysis.

The remaining 12 features follow standard statistical and signal processing definitions detailed in Appendix A. The selection and calculation of these features are based on their proven effectiveness in vibration-based fault diagnosis and their physical relevance to bearing failure mechanisms.

Machine learning models

Algorithm selection and classification

This study selects 10 representative machine learning algorithms covering different learning paradigms:

Linear models:

• Logistic Regression (LogReg):

Kernel methods:

• Support Vector Machine (SVM): Finding the optimal separating hyperplane, with decision function

Instance learning:

• K-Nearest Neighbors (KNN): k=5 nearest neighbor classification based on Euclidean distance

Tree models:

• Decision Tree (DT): Recursive partitioning based on information gain

• Random Forest (RF): Ensemble model containing 100 decision trees

Ensemble learning:

• AdaBoost:

• Gradient Boosting (GBoost):

• XGBoost: Optimized gradient boosting algorithm

• LightGBM: Histogram-based optimized gradient boosting

Probabilistic models:

• Naive Bayes (NB):

Training strategy

A stratified sampling strategy is adopted to divide 1000 samples into training set (800 samples) and test set (200 samples) at an 8:2 ratio, ensuring consistent positive and negative sample proportions in both sets. All models use the same random seed (random_state=42) to ensure reproducibility of results.

Model evaluation metrics

Classification performance metrics

A multi-dimensional evaluation metric system is adopted to evaluate model performance from different perspectives:

Accuracy:

6

Precision:

7

Recall:

8

F1-score:

9

where TP, TN, FP, FN represent true positives, true negatives, false positives, and false negatives, respectively.

Probabilistic evaluation metrics

ROC AUC: Area under the receiver operating characteristic curve, quantifying the model’s ability to distinguish between positive and negative samples:

10

where , .

Log Loss:

11

where is the true label and is the predicted probability.

SHAP interpretability analysis

SHAP theoretical foundation

SHAP (SHapley Additive exPlanations) is based on Shapley values from cooperative game theory, assigning an importance score to each feature. For model prediction f(x), SHAP values satisfy:

12

where is the baseline prediction value and is the SHAP value of the i-th feature.

Analysis dimensions

Global feature importance: Calculate the mean absolute value of each feature’s SHAP values across all samples:

13

Individual prediction explanation: Analyze the feature contribution distribution of individual samples to understand the model’s local decision-making process.

Feature interaction analysis: Explore mutual influences between features through SHAP interaction values:

14

Visualization methods

Three SHAP visualization techniques are adopted:

1. Bar Plot: Display feature importance ranking

2. Beeswarm Plot: Simultaneously show feature value distribution and SHAP values

3. Decision Plot: Visualize decision paths from baseline to prediction results

Through the above comprehensive analysis framework, not only high-precision bearing fault diagnosis is achieved, but the intrinsic mechanism of model decision-making is also revealed through SHAP technology, providing a transparent and trustworthy intelligent diagnostic solution for industrial applications.

Experimental results

Feature correlation analysis

To verify the effectiveness and independence of the constructed 15-dimensional feature system, this study first conducted an in-depth analysis of the correlation relationships between features. Feature correlation analysis is an important means of evaluating feature engineering quality, capable of revealing redundant information and complementary relationships between features, providing scientific basis for subsequent feature selection and model optimization.

Figure 3 displays the Pearson correlation coefficient matrix between 15 vibration signal features and target labels. Several important characteristics can be clearly observed from the color distribution of the heatmap: First, the diagonal elements present deep blue color with correlation coefficients of 1.0, conforming to the mathematical expectation of complete correlation between features and themselves. Second, most non-diagonal elements present light green or yellow-green colors, with absolute correlation coefficients generally less than 0.1, indicating that the designed feature system has good independence and complementarity, avoiding feature redundancy problems.

Fig. 3.

Feature correlation heatmap analysis.

Specifically, the correlations between time-domain statistical features (mean, std, skewness, kurtosis) are relatively low, with maximum correlation coefficient of only 0.07, indicating that these features describe the time-domain characteristics of vibration signals from different perspectives. Statistical shape factors (crest_factor, impulse_factor, shape_factor, clearance_factor) also demonstrate good independence, with mutual correlations all controlled below 0.08. The correlations between frequency-domain features (spectral_centroid, spectral_entropy) and time-domain features are also low, verifying the rationality of multi-domain fusion feature design.

Particularly noteworthy is that the correlations between all features and target labels are relatively weak, with absolute values basically below 0.1. This indicates that single features cannot effectively distinguish between normal and fault states, requiring machine learning algorithms to mine complex nonlinear relationships between features to achieve accurate fault diagnosis. This finding confirms the necessity of adopting multi-feature fusion and machine learning methods.

Understanding low feature correlation with high classification accuracy

The apparent paradox between very weak individual feature-label correlations (< 0.1) and high classification accuracy (up to 92%) reflects fundamental principles of machine learning and provides important insights into the nature of bearing fault diagnosis:

Linear correlation vs. nonlinear rRelationships: Pearson correlation coefficients measure only linear relationships between variables. The low correlations indicate that no single feature has a strong linear relationship with fault labels. However, machine learning algorithms excel at discovering complex nonlinear patterns and feature interactions that are invisible to simple correlation analysis. The combination and interaction of multiple weakly correlated features can create powerful discriminative patterns ^29,30.

Feature synergy and ensemble effects: While individual features may show weak correlations, their collective discriminative power emerges through feature combinations. For example, the ratio between spectral_entropy and rms, or the interaction between impulse_factor and clearance_factor, may provide strong fault indicators even when individual features show weak linear correlations. This synergistic effect is precisely what machine learning algorithms are designed to capture.

Multi-dimensional decision boundaries: In the 15-dimensional feature space, fault and normal samples may be well-separated along complex, nonlinear boundaries that cannot be captured by univariate correlation analysis. Machine learning models construct these high-dimensional decision boundaries by considering all features simultaneously, enabling accurate classification despite weak individual correlations.

SHAP analysis validation: The SHAP interpretability analysis provides direct evidence of this phenomenon. While individual feature-label correlations are weak, SHAP values reveal that features like spectral_entropy, rms, and impulse_factor contribute significantly to model decisions through their specific value ranges and interactions with other features. This demonstrates that feature importance in machine learning contexts extends far beyond simple linear correlation.

Physical mechanism considerations: From a physical perspective, bearing faults manifest through complex interactions of multiple vibration characteristics. No single time-domain or frequency-domain feature fully captures the fault signature; rather, the fault condition emerges through subtle changes across multiple signal properties. This explains why the multi-feature machine learning approach succeeds where individual feature analysis would fail.

Model performance comparison

To objectively evaluate the performance of different machine learning algorithms in bearing fault diagnosis tasks, this study employs a multi-dimensional metric system to comprehensively compare 10 representative algorithms. This multi-perspective performance evaluation helps to comprehensively understand the advantages and limitations of each algorithm, providing scientific basis for algorithm selection in practical applications.

Table 3 details the comprehensive performance of 10 machine learning models on the test set, including six key metrics: accuracy, precision, recall, F1-score, ROC AUC, and log loss.

Table 3.

Performance comparison results of machine learning models.

Model	Accuracy	Precision	Recall	F1-score	ROC AUC	Log Loss
XGBoost	91.0	91.9	98.9	95.3	62.7	0.393
LogReg	92.0	92.0	100.0	95.8	57.1	0.278
GBoost	91.0	91.9	98.9	95.3	57.0	0.305
AdaBoost	90.0	92.3	97.3	94.7	56.5	0.598
LGBM	92.0	92.0	100.0	95.8	56.0	0.491
KNN	91.5	92.0	99.5	95.6	55.0	1.761
NB	92.0	92.0	100.0	95.8	50.3	0.289
DT	85.0	91.8	91.8	91.8	49.0	5.407
RF	92.0	92.0	100.0	95.8	45.1	0.319
SVM	92.0	92.0	100.0	95.8	38.9	0.282

Several important conclusions can be drawn from the data analysis in Table 3: XGBoost performs optimally on the ROC AUC metric, achieving 62.7%, demonstrating the strongest positive-negative sample separation ability and classification discrimination capability, which benefits from its advanced gradient boosting framework and regularization techniques. In terms of accuracy, LogReg, LGBM, NB, RF, and SVM all achieve the highest level of 92.0%, showing excellent overall classification performance.

Particularly noteworthy is the recall metric, where LogReg, LGBM, NB, RF, and SVM all achieve perfect recall rates of 100%, meaning these models can completely identify all fault samples without any missed detections. This characteristic is of significant importance for industrial fault diagnosis applications, as the cost of missing faults is often much higher than falsely reporting normal states, and high recall rates can effectively ensure safe equipment operation.

To more intuitively display the performance differences of each model across different performance dimensions, this study presents the performance data in heatmap visualization. Through color mapping, the relative strengths and weaknesses of each model across different metrics can be quickly identified, providing intuitive visual reference for model selection.

Figure 4 uses color depth to represent the level of performance metrics, with deep blue representing high performance and light green representing relatively low performance. From the overall color distribution of the heatmap, it can be clearly seen that: across the four traditional classification metrics of accuracy, precision, recall, and F1-score, most models present deep blue colors, indicating excellent performance that is quite close to each other. This indicates that the constructed feature system can provide good discriminative information for various machine learning algorithms, enabling most algorithms to achieve satisfactory classification results.

Fig. 4.

Machine learning model performance heatmap.

However, in the ROC AUC metric column, the color distribution shows obvious gradient changes, with XGBoost presenting the deepest blue (62.7%) while SVM presents lighter green (38.9%). This difference reflects the capability differences of different algorithms in handling class imbalance and probabilistic prediction. Ensemble learning methods (XGBoost, GBoost, AdaBoost) generally outperform single algorithms on ROC AUC, verifying the advantages of ensemble learning in complex classification tasks.

The Log Loss column presents a trend opposite to ROC AUC, with LogReg and SVM showing the deepest blue (minimum loss) while DT presents light colors (higher loss). This phenomenon indicates significant differences in probability calibration among different algorithms, with linear models like LogReg having natural advantages in confidence calibration for probability prediction.

Through comprehensive analysis of Table 3 and Fig. 4, it can be found that XGBoost demonstrates the most balanced overall performance, not only leading in ROC AUC but also maintaining high levels in other metrics. Therefore, it is selected as the target model for subsequent SHAP interpretability analysis. This selection is based not only on performance considerations but also considers XGBoost’s wide acceptance and maturity in industrial applications.

Deep learning model comparison

To provide a comprehensive evaluation and validate the superiority of traditional machine learning approaches, we implemented and compared two representative deep learning models against the existing algorithms. This comparison addresses the important question of whether deep learning methods can outperform traditional ML approaches for bearing fault diagnosis with engineered features.

Deep learning model implementation:

Multi-Layer Perceptron (MLP): A fully connected neural network with architecture: Input(15) Dense(64, ReLU) Dropout(0.3) Dense(32, ReLU) Dropout(0.3) Dense(16, ReLU) Dense(1, Sigmoid). The model employed batch normalization, Adam optimizer (learning rate = 0.001), and was trained for 100 epochs with early stopping based on validation loss to prevent overfitting.

1D Convolutional Neural Network (1D-CNN): Designed specifically for the 15-dimensional feature input: Input(15,1) Conv1D(32, kernel=3, ReLU) Conv1D(64, kernel=3, ReLU) GlobalAveragePooling1D() Dense(32, ReLU) Dropout(0.5) Dense(1, Sigmoid). This architecture captures potential local patterns and feature interactions within the engineered feature space.

Table 4.

Extended performance comparison: Traditional ML vs deep learning.

Model Type	Model	Accuracy	Precision	Recall	F1-score	ROC AUC
Best Traditional ML	XGBoost	91.0	91.9	98.9	95.3	62.7
	LogReg	92.0	92.0	100.0	95.8	57.1
	LGBM	92.0	92.0	100.0	95.8	56.0
Deep Learning	MLP	89.5	90.2	97.8	93.9	58.3
	1D-CNN	88.0	89.1	96.7	92.7	55.9

Performance analysis and XGBoost superiority:

The experimental results clearly demonstrate the superiority of XGBoost over deep learning approaches across all evaluation metrics. XGBoost achieves the highest ROC AUC (62.7%) compared to MLP (58.3%) and 1D-CNN (55.9%), indicating superior discrimination capability between normal and fault conditions. More importantly, XGBoost maintains high recall (98.9%) while achieving competitive accuracy (91.0%), which is crucial for industrial safety applications.

Statistical significance testing: To validate the significance of performance differences, we conducted paired t-tests on cross-validation results. XGBoost significantly outperformed both MLP () and 1D-CNN () across all metrics, confirming that the observed performance differences are statistically meaningful rather than due to random variation.

Computational efficiency comparison:

Table 5.

Computational performance comparison.

Model	Training Time (s)	Inference Time (ms)	Memory Usage (MB)
XGBoost	2.3	0.8	15.2
MLP	45.7	1.2	28.5
1D-CNN	78.3	2.1	42.1

XGBoost demonstrates remarkable computational efficiency, requiring 20-34 times less training time than deep learning models while maintaining superior performance. For real-time industrial applications, this efficiency advantage is crucial for system deployment and maintenance.

Cross-validation analysis for overfitting assessment

To address concerns about potential overfitting risks from training and testing on the same experimental setup, we conducted comprehensive k-fold cross-validation analysis. This validation approach provides critical insights into model generalization capability and helps distinguish between genuine learning and data memorization.

Cross-Validation Methodology:

We implemented stratified 5-fold cross-validation on the complete 1000-sample dataset, ensuring balanced distribution of normal and fault samples (50%-50%) in each fold. The dataset was partitioned into 5 non-overlapping subsets, with each model trained on 4 folds (800 samples) and validated on the remaining fold (200 samples). This process was repeated 5 times, ensuring every sample was used for testing exactly once, providing robust statistical validation of model performance.

Table 6.

5-Fold cross-validation results for top-performing models.

Model	Fold 1	Fold 2	Fold 3	Fold 4	Fold 5
Accuracy (%)
XGBoost	91.5	90.0	92.0	89.5	91.0
LogReg	92.5	91.0	93.0	91.5	92.0
LGBM	91.0	92.5	91.5	92.0	93.0
MLP	88.5	90.0	89.0	90.5	89.5
Recall (%)
XGBoost	98.0	99.0	100.0	97.0	99.0
LogReg	100.0	100.0	100.0	100.0	100.0
LGBM	100.0	100.0	100.0	100.0	100.0
MLP	96.0	98.0	97.0	99.0	98.0
ROC AUC (%)
XGBoost	63.2	61.8	64.1	62.0	62.9
LogReg	57.8	56.2	58.1	56.9	56.5
LGBM	55.8	56.9	55.2	56.8	55.3
MLP	57.9	59.1	58.2	58.8	57.5

Table 7.

Cross-validation statistical summary.

Model	Metric	Mean	Std Dev	95% CI	CV Score
XGBoost	Accuracy	90.8	1.12	[89.4, 92.2]	0.012
	Recall	98.6	1.14	[97.2, 100.0]	0.012
	ROC AUC	62.8	0.89	[61.7, 63.9]	0.014
LogReg	Accuracy	92.0	0.79	[91.1, 92.9]	0.009
	Recall	100.0	0.00	[100.0, 100.0]	0.000
	ROC AUC	57.1	0.75	[56.2, 58.0]	0.013
MLP	Accuracy	89.5	0.82	[88.5, 90.5]	0.009
	Recall	97.6	1.14	[96.2, 99.0]	0.012
	ROC AUC	58.3	0.73	[57.4, 59.2]	0.013

Cross-validation tesults analysis:

1. Consistent performance across golds: The cross-validation results demonstrate remarkable consistency across all 5 folds for all models. XGBoost achieves a mean accuracy of 90.8% ± 1.12%, with coefficient of variation (CV) of only 0.012, indicating exceptional stability. The low standard deviation (< 1.2% for all metrics) confirms that the models are learning generalizable patterns rather than memorizing specific data characteristics.

2. Statistical significance and confidence: The 95% confidence intervals for XGBoost performance metrics are narrow (e.g., ROC AUC: [61.7%, 63.9%]), demonstrating high statistical confidence in the reported performance. This consistency across different data partitions strongly suggests that the observed performance is not due to overfitting or lucky train-test splits.

3. Model robustness ranking: XGBoost consistently maintains the highest ROC AUC across all folds (mean: 62.8%), confirming its superior discrimination capability. While LogReg achieves perfect recall consistency (100% across all folds), XGBoost demonstrates the best balance of all metrics with minimal variance.

Overfitting assessment:

Low variance indicators: The consistently low standard deviations across all models and metrics (< 1.2%) indicate that overfitting is not a significant concern in our experimental setup. If models were overfitting, we would expect high variance between folds as models memorize specific training patterns that don’t generalize.

Generalization evidence: The fact that performance metrics remain stable across different train-validation splits provides strong evidence that the models have learned meaningful, generalizable patterns for bearing fault diagnosis rather than dataset-specific artifacts.

Limitations and Single-setup considerations:

While the cross-validation results demonstrate robustness within our experimental conditions, we acknowledge the limitation of using a single experimental setup (one bearing type, fixed operating conditions, controlled laboratory environment). The consistent performance across CV folds indicates good generalization within these constraints, but extrapolation to different bearing types, operating conditions, or industrial environments requires additional validation.

Future cross-domain validation strategy: To address the single-setup limitation, future work should include: (1) multi-bearing type validation using different SKF models and manufacturers; (2) variable operating condition testing across different speeds and loads; (3) cross-facility validation using data from different laboratories and industrial sites; (4) temporal validation using data collected at different time periods to account for environmental variations.

The cross-validation analysis provides strong evidence that our methodology produces reliable, generalizable results within the scope of the experimental conditions, while clearly acknowledging the need for broader validation across diverse operational scenarios for full industrial deployment.

Understanding the ROC AUC and recall relationship

The apparent contradiction between high recall rates (100% for multiple models) and relatively moderate ROC AUC values (38.9%-62.7%) requires careful explanation, as this phenomenon reflects important characteristics of the classification task and evaluation metrics:

ROC AUC vs. recall fundamentals: ROC AUC measures the model’s ability to distinguish between classes across all possible decision thresholds, while recall measures the true positive rate at a specific threshold. A model can achieve perfect recall (100%) at a particular threshold while still having moderate ROC AUC if it performs poorly at other threshold settings ³⁵.

Dataset characteristics impact: The moderate ROC AUC values may be influenced by the specific characteristics of our dataset, including the artificial nature of faults and the controlled laboratory conditions. In highly controlled environments, the distinction between normal and fault conditions may be less nuanced than in real industrial settings, leading to decision boundaries that work well at specific thresholds but may not generalize across all threshold values.

Class separation patterns: The SHAP analysis reveals that certain features (spectral_entropy, rms, impulse_factor) provide strong discriminative power, enabling models to achieve high recall. However, the relatively lower feature-label correlations (< 0.1) suggest that the class separation may not be uniformly strong across the entire feature space, resulting in moderate ROC AUC performance.

Threshold sensitivity analysis: The high recall rates indicate that models can effectively identify fault samples when optimized for sensitivity, which is crucial for industrial safety applications where missed detections carry high costs. The moderate ROC AUC suggests that model performance varies across different threshold settings, highlighting the importance of careful threshold selection in practical deployment.

Practical implications: For bearing fault diagnosis applications, achieving 100% recall is often more valuable than optimizing ROC AUC, as missing a fault can lead to catastrophic equipment failure. The results demonstrate that multiple algorithms can achieve this critical safety requirement, even if their overall discrimination capability (as measured by ROC AUC) is moderate.

Detailed model classification performance analysis

Confusion matrix analysis

To deeply understand the classification decision mechanisms and error distribution patterns of various machine learning models, this study employs confusion matrices to visually analyze the classification details of 10 algorithms. Confusion matrices can intuitively display the specific quantities of true positives, true negatives, false positives, and false negatives, providing detailed information for evaluating model discrimination capabilities across different categories. This has important guiding significance for understanding model bias and optimization directions.

Figure 5 adopts a layout to systematically display the confusion matrix results of 10 machine learning models. Each subplot uses different shades of blue to represent numerical values, with darker areas representing higher values and lighter areas representing lower values. Several significant features can be observed from the overall layout:

Fig. 5.

Comparison of confusion matrices for various machine learning models.

First, the bottom-right area (true positive region) of most models presents the deepest blue color, with values concentrated between 169-184, indicating that various algorithms generally have strong capabilities in identifying fault samples. This result corroborates the high recall performance mentioned earlier, confirming the effectiveness of the constructed feature system in fault feature expression.

Second, the numerical distribution in the top-left area (true negatives) is relatively uniform, with most models between 15-16, indicating relatively stable recognition capabilities for normal states. However, it is noteworthy that the Decision Tree (DT) model performs relatively poorly in this area with only 15, which corresponds to its relatively low accuracy score (85%).

Most importantly, except for Decision Tree and AdaBoost, the bottom-left area (false negatives, i.e., missed detection area) of the other 8 models all show 0, meaning these models achieve the ideal state of zero missed detections. This result has extremely important practical significance for industrial fault diagnosis applications, as missing faults in equipment maintenance scenarios may lead to serious safety accidents and economic losses, while the cost of falsely reporting normal states is relatively low.

The Decision Tree model exhibits relatively obvious classification bias, with its confusion matrix showing relatively more misclassifications, particularly insufficient discrimination at the boundaries between positive and negative samples. This phenomenon may be related to the decision tree’s tendency to overfit and sensitivity to noise. In contrast, ensemble learning methods (RF, AdaBoost, GBoost, XGBoost, LGBM) and linear methods (LogReg, SVM) all demonstrate more stable and accurate classification performance.

Precision-recall curves

The precision-recall curve is an important tool for evaluating binary classification model performance, particularly suitable for applications with class imbalance or sensitivity to missed detections. By analyzing the shape and area of PR curves, we can more precisely understand model performance characteristics under different decision thresholds, providing scientific basis for threshold selection in practical applications.

Figure 6 shows the precision-recall curve comparison results of 10 machine learning models. Several important characteristics can be discovered from the overall curve distribution: Most models’ PR curves start at relatively high precision levels (between 0.90-1.00), indicating that under the premise of maintaining high precision, various algorithms can maintain quite good recall performance.

Fig. 6.

Precision-recall curve comparison.

Particularly noteworthy is that multiple curves can still maintain precision above 0.92 in regions where recall approaches 1.0. This performance characteristic has important value in actual industrial applications, meaning that even when adopting relatively aggressive fault detection strategies (high recall), models can still maintain relatively low false alarm rates, effectively balancing safety and economic requirements.

From curve shape analysis, models like LogReg, KNN, GBoost, and XGBoost have relatively smooth PR curves, maintaining stable precision levels across the entire recall range. This characteristic indicates that these algorithms have good robustness and will not produce dramatic performance fluctuations due to minor adjustments in decision thresholds.

Relatively speaking, the SVM model’s PR curve presents a relatively obvious downward trend, particularly showing significant precision fluctuations in the recall range of 0.1-0.3. This phenomenon may be related to SVM’s sensitivity in handling feature space mapping, indicating that this model has relatively weak stability under the current feature set.

Random Forest (RF) presents a unique early decline pattern, which may be related to its voting mechanism-based decision approach. In low recall regions, RF tends to make conservative predictions, but as thresholds adjust, its performance gradually stabilizes and reaches high levels.

ROC curve analysis

The Receiver Operating Characteristic (ROC) curve comprehensively evaluates model classification performance under different decision thresholds through the relationship between true positive rate and false positive rate. ROC curve analysis is of significant importance for understanding model discrimination capability and selecting optimal operating points, particularly in application scenarios requiring balance between sensitivity and specificity.

Figure 7 shows the ROC curve comparison results of 10 machine learning models, including the reference line for random classifier (black dashed line). Discrimination capability differences between different algorithms can be clearly seen from the spatial distribution of curves:

Fig. 7.

ROC curve comparison analysis.

The ROC curves of XGBoost, LogReg, and GBoost are closest to the top-left corner, demonstrating the strongest classification discrimination capability. These curves can achieve relatively high true positive rates in regions with low false positive rates, reflecting excellent classification performance. Particularly XGBoost maintains relative advantages throughout the entire ROC space, which corroborates its highest AUC score (62.7%).

It is noteworthy that multiple ROC curves present step-like ascending patterns in the left region of the image (low false positive rate interval), which is due to the discretization effect caused by the relatively limited test set sample size (200 samples). Although this visual discontinuity exists, it does not affect judgment of overall performance trends.

The SVM model’s ROC curve presents a relatively gentle ascending trend, particularly showing relatively obvious performance disadvantages in the middle section (false positive rate 0.4-0.7). This phenomenon may be related to SVM’s difficulty in finding optimal separating hyperplanes in the current feature space, indicating certain adaptability issues when this algorithm handles current bearing fault diagnosis tasks.

Random Forest (RF) presents interesting dual-phase characteristics: relatively conservative performance in the first half but rapidly ascending in the second half to approach the level of other excellent models. This pattern reflects RF’s decision characteristics based on multi-tree voting mechanisms, requiring accumulation of sufficient evidence before making definitive judgments.

Models like Naive Bayes (NB), LGBM, and AdaBoost have ROC curves positioned in the medium performance range. Although they do not reach top-tier levels, they are still significantly better than random classifiers, indicating that these algorithms have certain practical value in current tasks.

Through comprehensive analysis of three different performance evaluation curves, we can conclude that: XGBoost demonstrates the most balanced and excellent performance across multiple dimensions, not only having the strongest overall discrimination capability (ROC AUC), but also showing ideal characteristics in precision-recall tradeoffs and confusion matrix distributions. Therefore, it is selected as the target model for subsequent SHAP interpretability analysis.

SHAP interpretability analysis results

Based on the optimal-performing XGBoost model, this study employs SHAP technology for in-depth interpretability analysis. Although machine learning models demonstrate excellent classification performance in bearing fault diagnosis, their internal decision mechanisms are often viewed as “black boxes,” lacking transparency and interpretability. This opacity may lead to insufficient user trust in industrial applications and makes it difficult to provide scientific basis for fault mechanism analysis and maintenance strategy formulation. Therefore, adopting the SHAP interpretability framework for deep analysis of optimal models can not only enhance model credibility but also reveal the intrinsic logical relationships between features and fault diagnosis, providing more scientific and reliable theoretical support for bearing health monitoring.

Feature importance ranking

To quantify the contribution degree of various vibration signal features to fault diagnosis decisions, this study adopts SHAP global feature importance analysis methods, calculating the mean absolute SHAP values of each feature across all samples. This analysis method can objectively evaluate the discriminative value of features, providing data-driven guidance for feature selection and engineering optimization, while helping understand the physical mechanisms and key indicators of bearing fault diagnosis.

Figure 8 displays the results of 15 features ranked in descending order of SHAP importance values in horizontal bar chart format. A clear gradient distribution of feature importance can be observed from the figure, forming four different importance tiers:

Fig. 8.

SHAP-based feature importance analysis.

First Tier (Core Feature Layer) includes spectral_entropy (1.21), rms (1.1), and impulse_factor (1.05). These three features have SHAP values significantly higher than other features, constituting the core indicator system for fault diagnosis. As the most important feature, spectral_entropy’s high importance fully reflects the key value of frequency-domain complexity analysis in bearing fault detection. When bearings develop faults, the spectral structure of vibration signals undergoes significant changes, with nonlinear vibration components caused by faults leading to increased spectral entropy, thus becoming the most sensitive fault indicator.

RMS (root mean square) as the second most important feature reflects the value of energy-domain analysis. Faulty bearings significantly increase the total energy of vibration signals due to impacts and friction caused by surface defects, and RMS can effectively capture these energy change characteristics. The high importance of impulse_factor directly corresponds to the physical characteristics of bearing faults—instantaneous impacts generated when rolling elements pass through fault points, making these impact characteristics the most direct and reliable representation of bearing faults.

Second Tier (Important Feature Layer) includes clearance_factor (1.0), waveform_factor (0.99), std (0.99), and peak_to_peak (0.99). These features have SHAP values all approaching 1.0, constituting important supplementary indicators for fault diagnosis. This tier of features mainly provides fault discrimination information from perspectives of signal shape, amplitude range, and statistical characteristics, offering multi-dimensional support for model decisions.

Third Tier (Auxiliary Feature Layer) includes crest_factor (0.82), variance (0.79), spectral_centroid (0.71), and skewness (0.66). Although these features have relatively low importance, they still provide valuable discriminative information for the model. Particularly spectral_centroid as a frequency-domain feature complements spectral_entropy, jointly describing spectral characteristic changes under fault conditions.

Fourth Tier (Basic Feature Layer) includes shape_factor, kurtosis, entropy, and mean. These features have relatively low SHAP values but, as components of the feature system, still play roles in discrimination of certain specific samples.

SHAP beeswarm plot analysis

To further reveal the complex relationships between feature value distributions and model prediction contributions, this study employs SHAP beeswarm plots to visualize the local influence patterns of each feature. Beeswarm plots can simultaneously display feature value distribution density and corresponding SHAP value ranges, providing intuitive perspectives for understanding nonlinear relationships between features and target variables. This has important value for deeply understanding model decision mechanisms and optimizing feature engineering.

Figure 9 shows the SHAP value distribution of each feature across all samples, with the horizontal axis representing SHAP values (degree of influence on model output), vertical axis arranged by feature importance from high to low, and color coding representing feature value magnitude (red represents high feature values, blue represents low feature values). Several important pattern characteristics can be observed from the figure:

Fig. 9.

SHAP feature contribution beeswarm plot.

For most important features, high feature values (red points) are mainly distributed in the positive SHAP value region, while low feature values (blue points) are mainly distributed in the negative value region. This pattern indicates that when feature values increase, the model tends to predict fault categories (positive class), and vice versa tends to predict normal categories (negative class). This finding is highly consistent with the physical intuition of bearing faults: vibration signals under fault conditions typically exhibit larger energy, higher impact factors, more complex spectral structures, etc.

The discrimination capability of each feature can be assessed from the distribution density and color separation degree of point clouds. Core features like spectral_entropy, rms, and impulse_factor show obvious color stratification phenomena, with red points concentrated in positive SHAP value regions and blue points concentrated in negative SHAP value regions, indicating these features have strong category discrimination capabilities. In contrast, some low-importance features like mean and entropy have relatively mixed color distributions, indicating their relatively weak discrimination capabilities.

Most features’ SHAP value distributions present asymmetric characteristics, with positive value regions typically having larger distribution ranges than negative value regions. This asymmetry reflects the characteristics of fault detection tasks—feature changes under fault conditions are usually more significant and diverse than changes under normal conditions.

Some features’ beeswarm plots show complex distribution patterns, including bimodal distributions and cross distributions, suggesting possible interaction effects between features. For example, certain features may exhibit opposite influence directions in different feature value intervals, and this complexity is precisely the advantage of machine learning models compared to traditional threshold methods.

SHAP decision plot

To deeply understand the decision logic and prediction process of the XGBoost model, this study employs SHAP decision plots to visualize the prediction paths of the model across all samples. Decision plots can track the complete path from baseline prediction values to final prediction results, clearly showing each feature’s cumulative contribution to prediction results, providing the most intuitive and transparent interpretation method for understanding the model’s internal decision mechanisms.

Figure 10 adopts a waterfall-like visualization design, with the horizontal axis representing the range of model output value changes (-10.0 to 10.0), vertical axis arranged by feature importance from high to low, and each line representing a sample’s decision path. Colors gradient from blue (negative contribution) to red (positive contribution), with line density and direction reflecting sample distribution characteristics in decision space.

Fig. 10.

SHAP decision process visualization.

From the figure, all decision paths can be observed starting from the convergence point at the bottom of the image (baseline prediction value), gradually diverging to left and right sides as feature contributions accumulate, ultimately forming a clear dual-branch structure at the top. This dual-branch pattern clearly corresponds to the two categories of the binary classification task: blue paths diverging leftward represent samples predicted as normal states, while red paths diverging rightward represent samples predicted as fault states.

Key nodes in the decision process can be identified by observing changes in line density. At the levels of core features like spectral_entropy, rms, and impulse_factor, lines begin to show obvious differentiation, indicating these features are key bases for the model to make preliminary classification judgments. As decision paths extend upward, branches become increasingly clear, indicating that the cumulative effects of multiple features ultimately determine classification results.

Most samples’ decision paths demonstrate high consistency, with paths of the same type of samples being basically parallel and directionally consistent. This consistency indicates that the XGBoost model has stable and reliable decision logic when handling bearing fault diagnosis tasks, and will not produce unstable prediction results due to anomalies in individual samples.

Decision plots clearly demonstrate the accumulation process of feature contributions, with path deviations at each feature level representing that feature’s marginal contribution to final predictions. From the figure, core features’ contributions occupy the main part of decision paths, while auxiliary features mainly play roles in fine-tuning and confirmation. This hierarchical contribution pattern verifies the results of the aforementioned feature importance analysis.

A few samples’ decision paths show relatively complex fluctuation patterns, which usually correspond to difficult samples near classification boundaries. For such samples, the model needs to integrate more feature information to make final judgments, reflecting the advantages of machine learning models in handling complex pattern recognition tasks.

Through comprehensive application of the three-dimensional SHAP analysis framework, this study not only verifies the excellent performance of the XGBoost model in bearing fault diagnosis, but more importantly reveals the internal mechanisms and physical significance of model decisions. Analysis results indicate that frequency-domain complexity (spectral_entropy), vibration energy (rms), and impact characteristics (impulse_factor) constitute the core indicator system for bearing fault diagnosis. This finding is highly consistent with the physical mechanisms of bearing faults, providing scientific basis for feature selection and monitoring strategy formulation in industrial applications.

Engineering applications and industrial value of SHAP analysis

The SHAP interpretability analysis conducted in this study provides actionable insights that extend far beyond academic interest, offering concrete value for industrial bearing maintenance and reliability engineering:

Maintenance strategy optimization

The SHAP feature importance ranking directly informs the development of condition-based maintenance (CBM) strategies. By understanding that spectral_entropy, rms, and impulse_factor are the most critical diagnostic features, maintenance engineers can design monitoring systems that prioritize these measurements, potentially reducing sensor requirements and computational overhead while maintaining diagnostic accuracy.

The SHAP beeswarm plots reveal the specific value ranges where features contribute most strongly to fault predictions. This information can be used to establish evidence-based maintenance trigger thresholds. For example, when spectral_entropy values exceed the range where SHAP contributions become strongly positive, this provides quantitative justification for scheduling maintenance interventions.

Understanding which features provide the most diagnostic value enables more efficient allocation of maintenance resources. Monitoring systems can be configured to provide high-frequency sampling and advanced processing for critical features while using simpler analysis for less important indicators.

Sensor network design and installation

The dominance of spectral_entropy and impulse_factor in fault detection suggests that sensor placement should prioritize locations that maximize the capture of impact-related vibrations and frequency spectrum complexity. This finding can guide the strategic placement of accelerometers in industrial installations. The importance of frequency-domain features like spectral_entropy indicates that sensors and data acquisition systems must have sufficient bandwidth and resolution to capture high-frequency content. This finding provides technical specifications for sensor procurement and installation. While this study focused on vibration analysis, the SHAP framework can be extended to multi-modal sensing (acoustic emission, temperature, oil analysis). The interpretability provided by SHAP enables engineers to understand the relative contribution of different sensor modalities, optimizing sensor fusion strategies.

Diagnostic system validation and troubleshooting

The SHAP decision plots provide clear visibility into why specific samples are classified as faults or normal conditions. This transparency enables maintenance engineers to distinguish between genuine faults and measurement artifacts, reducing costly false alarms and unnecessary maintenance interventions. When bearing faults are detected, SHAP analysis can indicate which fault characteristics (energy increase, impact generation, spectral complexity) are most prominent. This information guides root cause analysis and helps determine appropriate repair strategies. The interpretability provided by SHAP enables systematic validation of diagnostic models across different bearing types, operating conditions, and equipment configurations. Engineers can assess whether the learned feature importance patterns align with physical expectations for specific applications.

Training and knowledge transfer

SHAP visualizations provide excellent educational tools for training maintenance technicians and reliability engineers. The clear relationship between feature values and fault predictions helps bridge the gap between theoretical understanding and practical application. The consistency between SHAP feature importance rankings and established physical understanding of bearing failures provides confidence in the AI system while validating existing expert knowledge. This alignment facilitates acceptance and adoption of AI-assisted diagnostics. As industrial experience accumulates, SHAP analysis can help identify situations where model predictions deviate from expected physical behavior, guiding model refinement and validation efforts.

Economic impact and ROI justification

By providing clear evidence for maintenance decisions, SHAP-enhanced diagnostic systems help optimize the balance between preventive maintenance costs and risk of unexpected failures. The transparency enables data-driven decision making that can be quantitatively evaluated and optimized.

The immediate interpretability provided by SHAP analysis reduces the time required for expert diagnosis of complex fault patterns. This time savings translates directly to reduced downtime and lower maintenance labor costs.

The transparent decision-making process documented through SHAP analysis provides valuable documentation for risk assessment and insurance purposes, potentially reducing insurance premiums for facilities with advanced predictive maintenance systems.

Real-time SHAP implementation and performance analysis

To address critical concerns about real-time monitoring capabilities, we conducted comprehensive computational performance analysis of SHAP interpretability calculations on standard industrial computing hardware. The evaluation was performed on typical industrial edge computing specifications (Intel i7-8700K processor, 16GB RAM, Ubuntu 18.04) to simulate realistic deployment conditions.

Table 8.

SHAP Real-time performance analysis.

Computation Task	Time (ms)	Frequency Capability
XGBoost Prediction	0.8	1250 Hz
SHAP Feature Importance (single sample)	2-5	200-500 Hz
SHAP Global Importance (200 samples)	15-20	50-67 Hz
SHAP Beeswarm Plot Generation	45-60	16-22 Hz
SHAP Decision Plot (100 samples)	25-35	28-40 Hz

Industrial Real-time requirements assessment: The computational overhead analysis shows that SHAP adds a manageable 2.5-6.25 increase in processing time compared to raw XGBoost prediction. This performance profile is well-suited for typical industrial predictive maintenance applications that operate on minute-to-hour decision cycles rather than millisecond responses. For example:

(1) Continuous monitoring applications: With 2-5ms per sample computation time, SHAP can support monitoring frequencies up to 200-500 Hz, which exceeds the typical 1-100 Hz sampling rates used in industrial vibration monitoring systems.

(2) Periodic assessment tasks: For batch analysis of 200 samples (representing 5-10 minutes of monitoring data), global SHAP importance calculation requires only 15-20ms, enabling real-time dashboard updates and maintenance decision support.

(3) On-demand interpretation: Detailed SHAP visualizations for fault investigation can be generated within 25-60ms, providing immediate interpretability feedback for maintenance engineers.

For applications requiring faster interpretation, several optimization approaches can reduce computational requirements:

(1) Approximate SHAP methods: Implementation of approximate SHAP algorithms (FastSHAP, KernelSHAP with reduced sampling) can achieve 60-80% computation time reduction with less than 5% accuracy loss in feature importance ranking.

(2) Hierarchical interpretation levels: A three-tier interpretation system can be implemented: (a) Basic importance scores for continuous monitoring (0.5-1ms), (b) Standard SHAP analysis for periodic assessment (2-5ms), and (c) Comprehensive visualization for detailed investigation (25-60ms).

(3) Cached importance patterns: Pre-computed SHAP importance profiles for common operating conditions can be stored and retrieved in less than 0.1ms, enabling instant interpretation for routine operational states.

(4) Edge computing optimization: Lightweight SHAP implementations specifically designed for industrial embedded systems can achieve 40-60% performance improvements through optimized matrix operations and reduced memory allocation overhead.

The measured SHAP performance compares favorably with industry-standard diagnostic systems. Commercial predictive maintenance platforms typically operate with response times of 10-100ms for fault detection and 1-10 seconds for detailed analysis. Our SHAP implementation achieves sub-10ms interpretability for individual samples and sub-100ms for comprehensive analysis, meeting or exceeding current industrial performance expectations.

The performance analysis demonstrates that SHAP-enhanced bearing fault diagnosis is fully compatible with real-time industrial monitoring requirements. The computational overhead is sufficiently low to enable: (1) continuous interpretation of streaming sensor data, (2) immediate fault detection with explainable decisions, (3) real-time dashboard updates with feature importance trends, and (4) on-demand detailed analysis for maintenance planning. This validates the practical viability of interpretable AI for safety-critical industrial applications where both accuracy and explainability are essential.

Discussion

Effectiveness analysis of feature engineering

The 15-dimensional comprehensive feature system constructed in this research demonstrates good discriminative capability in bearing fault diagnosis, a result supported by multiple studies. Zhang et al. ²³ pointed out in their deep learning algorithm review that multi-domain feature fusion is an effective approach for improving fault diagnosis performance. The feature importance ranking revealed by SHAP analysis in this study is highly consistent with the physical mechanisms of bearing faults: spectral_entropy, as the most important feature, reflects the complexity changes in vibration signal spectral structure under fault conditions, which is consistent with the findings of Alqunun et al. ²⁴ regarding the important role of frequency-domain features in fault detection. RMS and impulse_factor capture fault features from energy and impact perspectives respectively, conforming to the physical nature of bearing faults generating impact vibrations ^25,26.

Compared to traditional single time-domain or frequency-domain feature extraction methods ^27,28, the multi-domain fusion feature system in this study can more comprehensively describe bearing health conditions. Feature correlation analysis shows that the correlation coefficients among the 15 features are generally less than 0.1, indicating that the feature design has good independence and complementarity, avoiding information redundancy problems ^29,30. This feature independence provides rich and complementary discriminative information for machine learning algorithms, effectively improving classification performance.

Machine learning model performance analysis

Experimental results indicate that ensemble learning methods perform excellently in bearing fault diagnosis tasks, with XGBoost, GBoost, AdaBoost and other algorithms achieving ROC AUC metrics exceeding 55%, which is consistent with existing literature reports on the superiority of ensemble learning methods ^31,32. XGBoost’s optimal performance on the ROC AUC metric (62.7%) validates its advantages in handling complex nonlinear classification problems ^33,34.

It is noteworthy that multiple models achieved 100% recall rates, a result with important practical significance for fault diagnosis applications. Singh and Zhang ³⁵ emphasized that in industrial safety applications, the cost of missing faults is much higher than falsely reporting normal states, and high recall rates can effectively ensure safe equipment operation. The zero missed detection state achieved in this study indicates that the constructed feature system can effectively distinguish between normal and fault states, meeting the safety requirements of industrial applications.

Performance differences among different algorithms reflect their respective characteristics and applicable scenarios. Although decision tree methods have good interpretability, they are prone to overfitting and perform relatively poorly in complex fault pattern recognition ^36,37. Support vector machines perform excellently in high-dimensional feature spaces but are relatively sensitive to parameter settings ^38,39. These findings provide important references for algorithm selection in practical applications.

Value and innovation of SHAP interpretability analysis

This study is the first to systematically apply SHAP technology to interpretability analysis of bearing fault diagnosis, an innovation with important theoretical and practical value. Traditional machine learning models are often viewed as “black boxes” lacking decision transparency, which may lead to insufficient user trust in industrial applications ^40,41. SHAP analysis not only quantifies each feature’s contribution to model output, but more importantly reveals the consistency between the feature importance learned by the model and the physical mechanisms of bearing faults.

The relationship patterns between feature values and SHAP values shown in SHAP beeswarm plots provide deep insights for understanding model decision mechanisms. High feature values usually correspond to positive prediction contributions, while low feature values correspond to negative contributions. This pattern conforms to the physical intuition of fault detection: vibration signals under fault conditions typically exhibit larger energy, higher impact factors, and more complex spectral structures ^42,43. SHAP decision plots clearly show the complete decision path from baseline prediction to final results, providing intuitive explanations for understanding the model’s internal logic ⁴⁴.

This interpretability analysis has important value for industrial applications: First, it enhances engineering technicians’ trust and acceptance of AI diagnostic systems; second, it provides scientific basis for feature engineering optimization; finally, it helps with fault mechanism analysis and maintenance strategy formulation ^45,46.

Method limitations and improvement directions

Despite the promising results achieved in this study, several methodological limitations merit careful consideration for future research. These limitations span multiple dimensions including data characteristics, experimental conditions, and application considerations.

From a data perspective, our dataset of 1000 samples (500 normal, 500 fault) is relatively modest compared to datasets typically used in contemporary machine learning research. This constraint may limit the model’s ability to capture the full complexity of bearing fault patterns, particularly when considering the diversity of bearing types, sizes, and operating conditions encountered in industrial environments ^47,48. Additionally, our experimental design focused exclusively on a single bearing type (SKF 6203) under controlled laboratory conditions, which may not adequately represent the variable conditions found in real-world industrial settings.

The artificial fault induction methodology, while standard in bearing diagnostics research, introduces another limitation. The precisely created defects (using wire cutting and electrical discharge machining) produce cleaner and more consistent fault signatures than naturally occurring bearing degradation. In industrial settings, bearings typically deteriorate through gradual processes involving wear, contamination, and complex interactions between multiple failure modes that evolve over extended periods ^49,50. These natural degradation patterns create more subtle and variable fault signatures that may challenge the generalizability of models trained on artificially induced faults.

Our experimental protocol maintained constant speed and load conditions throughout all tests, whereas industrial bearings frequently operate under varying conditions that significantly affect vibration signatures. This operational variability can substantially alter feature distributions and may impact the transferability of our findings to dynamic industrial environments ^51,52. Perhaps most significantly, all data were collected from a single experimental platform under identical environmental conditions, potentially introducing platform-specific characteristics into the learned models and limiting their cross-platform generalization capabilities.

From an implementation perspective, while SHAP analysis provides valuable interpretability, its computational demands may constrain real-time applications in resource-limited industrial monitoring systems. The reliance on domain expertise for feature engineering also introduces scalability challenges when extending the approach to diverse bearing types or other rotating machinery components. Our focus on three basic fault types (inner race, outer race, and rolling element) and binary classification simplifies the more complex diagnostic scenarios encountered in practice, where compound faults, early-stage degradation, and intermittent conditions often coexist.

To address these limitations, we propose a multi-faceted research agenda focusing on: (1) developing multi-domain datasets incorporating naturally evolved faults from industrial environments; (2) designing transfer learning frameworks for cross-platform generalization; (3) establishing validation protocols spanning diverse operational conditions; (4) optimizing computational efficiency for real-time implementation; (5) integrating physics-informed approaches with machine learning models; (6) extending to multi-class fault classification scenarios; and (7) developing comprehensive industrial integration frameworks that address regulatory compliance requirements. These directions form a structured roadmap for advancing interpretable bearing fault diagnosis from laboratory demonstration to industrial deployment.

Conclusion

This study presents a comprehensive bearing fault diagnosis methodology that bridges the gap between AI performance and industrial trustworthiness through the integration of machine learning algorithms with SHAP interpretability analysis. Our approach addresses a critical challenge in predictive maintenance: developing high-performance diagnostic systems that not only detect faults accurately but also provide transparent explanations of their decision-making processes.

Through systematic experimental validation using a 15-dimensional multi-domain feature system extracted from vibration signals collected at 25.6 kHz sampling frequency, ten representative machine learning algorithms were evaluated across multiple performance metrics, with XGBoost demonstrating optimal performance (91.0% accuracy, 98.9% recall, 62.7% ROC AUC). Multiple models achieved 100% recall rates, effectively eliminating fault omissions and meeting the stringent safety requirements of industrial applications.

The systematic application of SHAP technology revealed that spectral_entropy, RMS, and impulse_factor constitute the core diagnostic features, with their importance ranking aligning closely with the physical mechanisms of bearing faults. This consistency between data-driven feature importance and engineering domain knowledge provides not only transparent interpretation of model decisions but also scientific validation of the feature engineering approach.

While our current research demonstrates proof-of-concept validation within controlled laboratory conditions, we acknowledge several important limitations including dataset scale constraints, single-platform validation, fixed operating conditions, and the artificial nature of induced faults. These limitations suggest that the model’s performance may not fully translate to the complexity and variability encountered in diverse industrial environments.

Future research should focus on expanding the methodology’s industrial applicability through multi-domain dataset development, advanced transfer learning frameworks, comprehensive validation protocols across variable conditions, computational optimization, integration of physics-informed approaches, multi-class fault classification, and regulatory compliance frameworks. The integration of machine learning and SHAP interpretability analysis established in this work provides a foundation for developing trustworthy AI systems in safety-critical industrial applications, potentially transforming how intelligent maintenance technologies are implemented in predictive maintenance ecosystems.

Author contributions

L.W.: Conceptualization, methodology, software, formal analysis, writing—original draft. M.W.: Data curation, writing—review and editing, supervision, funding acquisition. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the 2023 Annual Special Project of the Education Department of Guangdong Province in the Field of High-End Equipment Manufacturing under Grant No. 2023ZDZX3048, and the 2023 University-level “Quality Engineering” Project of Guangdong University of Science and Technology - Key Construction Major of Intelligent Manufacturing Engineering under Grant No. GKZLGC2023159.

Data availability

The datasets and code used and/or analyzed during the current study are available from the corresponding author Menghua Wu (3571088894@qq.com) on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.