Skip to main content
ACS Omega logoLink to ACS Omega
. 2026 Jun 23;11(26):39252–39262. doi: 10.1021/acsomega.6c03627

A Novel Dynamic Principal Component Analysis Based on Block Dynamic Mode Decomposition for Chemical Process Fault Detection

Pu Yang 1, Zongquan Xie 1,*, Shuqi Sheng 1
PMCID: PMC13347635  PMID: 42428826

Abstract

Dynamic principal component analysis (DPCA) has been widely adopted for process monitoring and fault detection in dynamic systems. However, its application in chemical process industries remains constrained by several limitations, including vulnerability to intermittent disturbances caused by equipment vibrations, inability to effectively capture long-term time dependence, and lack of consideration for subsystem coupling. To overcome these limitations, we propose an enhanced methodological framework. First, a frequency-based clustering technique is introduced to remove intermittent noise caused by equipment vibrations. Second, a novel matrix augmentation strategy is introduced to improve DPCA’s capacity to capture long-term time dependence. Third, block matrix dynamic mode decomposition is introduced to model the subsystem coupling. Experimental results demonstrate that the proposed method achieves competitive performance under moderate lag settings and superior fault detection performance under large lag settings, with advantages in detecting correlated faults across multiple subsystems. In addition, ablation experiments demonstrate that each improvement plays a positive role, and their integration leads to performance gains in complex industrial environments.


graphic file with name ao6c03627_0006.jpg


graphic file with name ao6c03627_0004.jpg

1. Introduction

With the rapid development of information technology and automation, modern process systems are increasingly equipped with extensive sensor networks, enabling the continuous collection of process data. Such data can provide valuable insights for process monitoring and fault detection. Multivariate Statistical Process Control (MSPC) has thus attracted significant attention as a data-driven approach for monitoring and diagnosing complex industrial processes. By analyzing the statistical correlations and patterns among process variables, MSPC facilitates real-time fault detection, thereby helping to maintain the stability and safety of system. −

Principal Component Analysis (PCA) has long been regarded as an important technique in multivariate statistical analysis and process monitoring. PCA was originally developed to identify orthogonal directions of maximum variance in high-dimensional data and reduces dimensionality. It projects sample data onto a lower-dimensional subspace while retaining as much variability as possible. This data-driven characteristic has made PCA particularly valuable in modern process systems, where large-scale data sets are collected. Over the past few decades, numerous studies have demonstrated the effectiveness of PCA and its variants in fault detection and process monitoring across a wide range of industrial sectors including the chemical industry − and power industry. , Its ability to identify differences in the statistical distribution of process variables has been well established. However, in modern process systems, which are characterized by nonlinearity and time dependence, these assumptions are frequently violated, thereby undermining the diagnostic effectiveness of PCA.

To address these challenges, many extensions and modifications to PCA have been developed. Ku proposed Dynamic PCA (DPCA) to enable the model to capture time dependence by augmenting the input matrix. However, given that the principal components extracted by DPCA may not be inherently dynamic, Dong introduced Dynamic Inner PCA (DiPCA) to extract a set of dynamic latent variables,ensuring the derived principal components are dynamic. Nonlinear extensions of PCA, such as Kernel PCA (KPCA), have also been developed to address nonlinearities in process data. Furthermore, Ren proposed an autoencoder-based KPCA to improve the computational efficiency of traditional KPCA. Inspired by deep learning, Deng developed Deep PCA (DePCA) to enhance fault detection performance via multilayer feature extraction. Notably, these methods for nonlinearity all leverage kernel tricks or neural network-based representations to map input data into higher-dimensional feature spaces, where linear separability is more achievable. Beyond PCA variants, researchers have explored other multivariate methods for specific scenarios: Partial Least Squares (PLS) is employed for cases involving both input and output variables, , while Independent Component Analysis (ICA) is used for fault detection in systems with non-Gaussian data. , In addition, deep learning methods such as convolutional neural networks have been applied to process fault detection by automatically extracting hierarchical features from time-series data. ,

Although numerous new methods have been developed to address PCA’s shortcomings, critical challenges remain in chemical process monitoring. In chemical production processes, the operation of rotating equipment or fluctuations in flow velocity induce the microvibrations of equipment. While small, such vibrations interfere with sensor measurements, thereby misleading the process monitoring models to make a wrong alarm. Although existing denoising algorithms effectively mitigate noise from continuous vibrations, they cannot suppress intermittent, decaying vibrations triggered by flow velocity fluctuations. Moreover, industrial chemical systems typically consist of multiple interdependent subsystems, yet existing approaches rarely account for the coupling between these subsystems, which limits model accuracy and robustness. Additionally, while the sampling period of sensors in chemical process systems operates on the order of seconds or even milliseconds, some process variables such as temperature and pressure change slowly. Given this condition, most approaches need to set the lag to a large value, which leads to a disproportionate increase in the number of model parameters and a decrease in both model accuracy and computational speed.

To address the aforementioned challenges, this paper proposes a novel framework termed DMDPCA. Unlike conventional DPCA-based methods that treat temporal dependence, noise suppression, and subsystem coupling independently, the proposed method provides a unified solution targeting three key limitations in industrial process monitoring.

The main contributions of this work are summarized as follows:

  • (1)

    A frequency-based clustering (FBC) method is developed to suppress intermittent and nonstationary vibration noise, which cannot be effectively handled by conventional denoising techniques.

  • (2)

    A transfer entropy-based augmentation strategy is proposed to selectively incorporate informative time-lagged variables, enabling efficient modeling of long-term dependencies while avoiding excessive dimensionality.

  • (3)

    A block dynamic mode decomposition (Block DMD) framework is introduced to explicitly model subsystem coupling, improving the detection performance for correlated faults.

The remainder of this paper is organized as follows. Section presents the proposed algorithm. The results and discussion on the asphalt production process are presented in Section to demonstrate the effectiveness of the proposed algorithm. The final section gives conclusions.

2. Methodology

To effectively address the challenges of intermittent noise, long-term time dependence, and subsystem coupling in chemical processes, the proposed framework integrates three dedicated modules. The FBC module is designed to handle nonstationary vibration noise, the transfer entropy-based augmentation module aims to capture long-term dynamic dependencies with controlled model complexity, and the Block DMD module is introduced to explicitly characterize subsystem interactions. Unlike traditional approaches that address these issues separately, the proposed method combines them in a unified framework, where each module targets a specific limitation while complementing the others.

2.1. Denoise with Frequency-Based Cluster

Noise induced by equipment vibrations in chemical processes is typically composed of two components: constant noise and intermittent, decaying noise. The former is linked to the sustained operation of rotating equipment, while the latter arises from abrupt changes in flow velocity, such as those triggered by valve adjustments. Most conventional denoising techniques assume noise is constant and follows a normal distribution, rendering them ineffective at suppressing the intermittent, decaying noise. To address this limitation, we design FBC to discriminate between stable intervals (i.e., those without intermittent noise) and disturbed intervals (i.e., those with intermittent noise). A key issue here is that applying clustering directly to raw data typically yields only a few identified outliers, rather than meaningful segmentation of stable and disturbed regions. We think the primary distinction between these interval types resides in their frequency-domain characteristics. Accordingly, we partition the data sequence into multiple intervals and perform clustering on their frequency-domain representations to distinguish between the two. Compared with conventional denoising methods that assume stationary or Gaussian noise, the proposed FBC approach leverages frequency-domain characteristics to distinguish intermittent disturbances, making it more suitable for nonstationary vibration noise commonly observed in industrial processes.

Prior to applying FBC, the raw data are standardized using z-score normalization (mean-centering and scaling to unit variance) to eliminate differences in magnitude and unit across variables. Since the constant noise component appears uniformly across all intervals, it does not affect the frequency-domain clustering and is implicitly handled by the following compensation step. Therefore, no additional denoising is required prior to FBC. The standardized data are then partitioned into p nonoverlapping intervals, each containing q consecutive samples, as illustrated in eq

[x1x2⋯xqxq+1xq+2⋯x2q⋯⋯⋯⋯x(p−1)q+1x(p−1)q+2⋯xpq] 1

where p represents the total number of intervals, q represents the number of samples within each interval and x i represents the i-th sample composed of process variables.

Subsequently, a Fast Fourier Transform (FFT) is applied to each row of the data matrix, converting the time-domain signals into their corresponding frequency-domain representations. Due to the inherent symmetry of the FFT for real-valued signals, only the first q/2 frequency components that contain unique information are reserved. To keep the difference between different intervals while reducing dimensionality, the magnitudes of these components are doubled and used as input for clustering. In the following study, the interval length q is set to 100 samples, providing a frequency resolution of approximately 0.016 Hz at the 1 Hz sampling rate. This value balances the need for sufficient frequency resolution against the requirement for an adequate number of intervals for reliable clustering. The number of clusters C in FCM is set to 7, corresponding to disturbed intervals from 6 process units plus one stable intervals. The fuzziness exponent is set to 2.0, and the algorithm terminates when the change in the objective function falls below 1 × 10–3 or the maximum number of iterations (100) is reached.

Finally, the Fuzzy C-means (FCM) clustering is employed for clustering. FCM partitions the data set into C clusters, allowing each sample to belong to multiple clusters simultaneously. This feature makes it well-suited for handling transitional data segments that lie between stable and disturbed intervals. The cluster containing the largest number of segments is then identified as corresponding to the stable intervals and others are disturbed intervals. The stable intervals are directly transformed into time-domain data through Inverse FFT­(IFFT), while the interference intervals are the same after noise compensation. Let F s and F d denote the cluster centroids of the stable and disturbed classes in the frequency domain, respectively. For each disturbed interval with frequency representation F i, the compensated frequency representation F c is computed as follows:

Fc=Fi−(Fd−Fs) 2

This shifts the frequency profile of the disturbed interval toward that of the stable class, suppressing the intermittent noise component. The compensated interval is then converted back to the time domain via IFFT.

2.2. Data Matrix Augmentation

2.2.1. DPCA

In chemical production, some process variables change slowly, such as temperature and pressure. These process variables thus exhibit long-term time dependence, which presents a need for handle time dependence. To address this issue, DPCA augments the input data matrix and conducts PCA analysis on the augmented matrix to handle time dependence. Given a data matrix X as eq

X=[x1x2⋯xn]T 3

Where x i is a l × m vector representing a sample, and each column of x i represents a monitored variable, n is the total number of samples. The common PCA extracts principal components from X through the eq

X=TPT+Ẽ 4

Where matrix P represents the load matrix and its column vectors are called load vectors, each column of T is the projection of X in the direction of load vectors, E represents the residual matrix. Equation can be solved by singular value decomposition.

While in DPCA, to handle time dependence, time-lagged samples are simply appended to the original sample, forming a new sample that contain time dependence. And the data matrix X aug is augmented as follows

Xaug=[x̂1x̂2⋯x̂n−l]Tx̂i=[xixi+1⋯xi+l−1] 5

Where X aug ∈ R (n–l)×ml is the data input in DPCA and l represents the max lag of time dependence. This augmentation strategy leads to a disproportionate increase in the dimension of the data matrix when dealing with long-term time dependence, as it takes all time-lagged variables into account. Especially when the sample frequency of sensors is rather high, the max lag must be increased to capture the long-term time dependence. This disadvantage brings about both performance degradation and heavy computational burden in DPCA when addressing long-term time dependence.

2.2.2. Augmentation Based on Transfer Entropy

Given the weakness of the existing augmentation strategy, we propose a novel strategy for data matrix augmentation. This strategy controls the dimension growth of the data matrix and the number of parameters by selectively retaining time-lagged variables with a high likelihood of long-term dependence, while discarding those with low. It begins by calculating transfer entropy (TE) between every pair of variables. As a directional, information-theoretic measure, TE effectively captures causal interactions and dynamic dependence and is inherently robust to noise. It is employed to identify the variables and their corresponding lags that show strong causal influence on the current variables, contributing to the construction of a more compact and informative augmented matrix. The formal expression for TE is provided in eq .

Tx→y(τ)=∑xt,yt,xt−τp(xt,yt,xt−τ)log⁡p(yt|xt,xt−τ)p(yt|xt) 6

Where T x→t represents the TE from x to y with time lag τ, x t–τ represents the x at time t – τ; p(x t , y t , x t–τ) represents the joint probability of X at t and t – τ, and y at t; p(y t | x t , x t–τ) represents the conditional probability of y t given both x t and x t–τ, p(y t | x t ) represents the conditional probability of y t given only x t without time-lagged X. In practice, the joint and conditional probability distributions in eq are estimated using k-nearest neighbor (KNN) probability estimation, a nonparametric approach well-suited for large-scale industrial time-series data. The number of nearest neighbors is set to 10 to balance estimation accuracy and computational efficiency. For each pair of variables (x, y), the TE is computed across a range of candidate lags from 1 to the largest lag. The lag that maximizes the TE is selected as the candidate lag for that variable pair. If the maximum TE surpasses the threshold, the time-lagged x t–τ variable is appended into the data matrix as an additional column. Otherwise, no augmentation is performed for that variable pair. Notably, the threshold is defined using a surrogate test. For each variable pair, the source time series is randomly shuffled 50 times to generate a set of surrogate TE values. The 95th percentile of these surrogate TE values is taken as the threshold to screen the original TE results. The augmentation strategy is defined in eq

XTE=[XV] 7
V=[vτv(τ+1)⋯v(τ+n)]T 8

Where v i represents the i-th sample of the selected variable, τ is the lag that maximize the TE. If multiple variables are selected, the augmentation proceeds in the same way. By selectively appending time-lagged variables, we can effectively control the dimension of the data matrix while maintaining the performance of the model. Unlike traditional lag-based augmentation that blindly includes all time-lagged variables, the proposed strategy selectively retains only those with strong causal influence, thereby significantly reducing redundant information and improving computational efficiency while preserving essential dynamic characteristics.

2.3. Block DMD

In previous studies, researchers treated the entire process system as a unified entity, applied PCA or its variants for process monitoring under this assumption. This simplification of model complexity enabled effective monitoring performance in certain scenarios. However, in some chemical process systems that consist of multiple interdependent subsystems, this assumption leads to a deterioration in model performance.

Given the weakness above, we introduce Block DMD to model the subsystem coupling. Block DMD is a structured extension of Dynamic Mode Decomposition (DMD), which explicitly models the coupling relationships between subsystems by partitioning high-dimensional data matrices into multiple low-dimensional block matrices based on physical modules, temporal scales, or functional correlations. The core concept involves decomposing the global dynamics of complex systems into the superposition of local modes and coupling modes, thereby preserving the linear approximation advantages of DMD while enhancing the capability to characterize complex systems with multisubsystems. Compared with conventional PCA-based methods that ignore subsystem structure, the Block DMD framework explicitly models intersubsystem interactions, enabling the detection of correlated faults that may be overlooked due to the cancellation effects in global modeling approaches.

Given a input data matrix X TE augmented by eq , each row of X TE represents a sample of monitored variables. The first step involves rearranging the column based on the structure of subsystem, such that variables belonging to the same subsystem are grouped into the same block matrix. The rearranged matrix X̂ is represented as eq

X̂=[S1,S2,···,Sk] 9

where S k (k ≤ m) represents block matrix composed of k-th subsystems’ monitored variables. The assignment of variables to subsystem blocks is based on the physical structure of the production process. Variables measured within the same process unit are grouped into the same block. In the asphalt production case studied in Section , six blocks are defined corresponding to the heating furnace, electric desalting, initial distillation, atmospheric distillation, vacuum distillation, and oxidation blending units. Variables appended through the TE-based augmentation are assigned to the same block as their original (nonlagged) variable.

To find the dynamic nature through which the current state is most predictable by its past, Block DMD gives the following objective function

minJ(A)=min||X̂′−X̂A||F2 10
A=[A11A12⋯A1kA21A22⋯A2k⋮⋮⋱⋮Ak1Ak2⋯Akk] 11

where X̂ ′ is the data matrix that is one step later than X̂ and A ij represents the influence of the i-th subsystem on the j-th subsystem.

The objective of the above function is to identify a block matrix A that most accurately characterizes the system’s state transitions. The diagonal blocks of A capture the modes within individual subsystems and reflect the internal characteristic of each subsystem. On the other hand, the off-diagonal blocks introduce new mode that do not present in conventional PCA frameworks, enabling the explicit model of coupling characteristic between subsystems. Simultaneously, the use of the augmented and rearranged data matrix X̂, which incorporates time-lagged variables with strong causal influence on the current state, contributes to the model’s ability to capture both intrasubsystem and intersubsystem dynamics effectively.

The following outlines the procedure for calculating the block matrix A. Since X̂ ′ and X̂ are block matrices, and the Frobenius norm is additive for block matrices, the J(A) can be expressed as eq

J(A)=min∑j=1k||Ŝi′−∑i=1kŜjÂij||F2=min∑j=1kJj(Aj) 12

where A j represents the j-th column of A. J i (A i ) and J j (A j )­(i ≠ j) are independent of each other, the overall objective function J reaches its minimum when all J j attain their respective minimum values. By taking the partial derivatives with respect to all A ij as eq and setting them to zero, we obtain eq .

∂J∂Aij=−2ŜiT(Ŝj′−∑e=1kŜeAej) 13
ŜiTŜj′=∑e=1kŜiTŜeAej(∀i,j) 14

Define the matrix C ∈ R m × m , partitioned into blocks such that the (i,e)-th block is the m i × m e matrix C ie = Ŝ i Ŝ e . Similarly, define L j as a column of blocks, where the i-th block is the m i × m j matrix Ŝ i Ŝ j

C=[Cie]=[ŜiTŜe] 15
Lj=[Ŝ1TŜj′Ŝ2TŜj′⋯ŜkTŜj′]T 16

the eq can be simplified as follows

Lj=CAj 17

Where A j represents the j-th column of A. Thus, the A can be solved through eq

Aj=C+Lj 18

where C + is the Moore-Penrose pseudoinverse of C.

Inspired by the, here we operate the PCA on the incremental matrix X △, and the X △ is defined as follows

X̂△=X̂′−X̂A 19

By arranging the eigenvalue of the covariance matrix X̂ △ X̂ △ in descending order, we extract h eigenvectors corresponding to the largest h eigenvalues as the principal components, the value of h is determined by the cumulative percent variance (CPV). CPV is a principle that is widely used to determine the value of h. These components form a matrix B ∈ R n × h and matrix B is the score matrix in PCA framework. According to the formula in PCA framework which is eq , we can get the residual matrix R.

R=X̂△−BΨTB=X̂△Ψ 20

On the online monitor stage, we use T 2 and squared prediction error (SPE) as indicator to monitor the system. The T 2 and SPE for i-th sample are defined as follows

x̂i△=x̂i△−x̂(i−1)△A 21
Ti2=xi△Ψcov(B)−1ΨTxi△T 22
SPEi=RRT 23

where x i is a single sample derived from the test data set or the data obtained from online operation test, cov(B) is the covariance matrix of the B.

The upper control limit (UCL) is the threshold used to determine whether a fault has occurred in PCA framework. It was calculated by the past normal data, denoting the upper limit range of normal process variation. Typically, once either of the two indicators exceed the UCL corresponding to the 99% confidence level, it is considered that a fault has occurred. According to PCA framework, the T 2 statistic normally follow the F distribution, and the upper control limit (UCL) is

UCLT2=h(n−1)n−hFα(h,n−h) 24

where the α is confidence coefficient, h is the number of principal components, n is the number of samples. And the SPE statistics normally follow the chi-squared distribution, its UCL is

UCLSPE=v2mχα2(2m2v) 25

where m and v are the mean and variance of SPE. By monitoring both indicators, we can detect the fault. The whole flowchart is shown in Figure . As illustrated, all model components, including the FBC cluster centroids, the selected time-lagged variables and their lags, the block transition matrix A, the loading matrix Ψ, and the control limits, are determined entirely in the offline training stage and remain fixed during online operation. Online monitoring proceeds with a sliding window of 100 samples: each incoming sample is standardized, denoised, augmented, and evaluated through (eqs 21–23) using the precomputed parameters. These precomputed parameters remain fixed during online monitoring without incremental updates, which avoids model drift and ensures stable alarm performance. The measured average processing time per sample is 6.3 ms on a workstation equipped with an Intel Core i7–10700 CPU (2.90 GHz) and 16 GB RAM, well below the 1 s sampling interval, confirming that the framework meets the real-time requirement of the monitoring task.

1.

1

Flowchart of DMDPCA for fault detection. The left branch shows the offline training stage, and the right branch shows the online monitoring stage. Arrows between branches indicate parameters computed offline and applied online without updating.

3. Results and Discussion

3.1. Asphalt Production Process and Experimental Setup

Asphalt is a crucial material for road construction and waterproof engineering, and its quality directly dictates the durability, high-temperature stability, and low-temperature crack resistance of the resulting infrastructure. Since asphalt quality is closely tied to the operational stability of production equipment, it is necessary to develop methods for detecting equipment faults. The asphalt production process involves multiple disciplines including chemical reaction engineering, thermal process engineering and automatic detection technology.

In the case we study, the asphalt production equipment comprises heating furnace units, electric desalting units, initial distillation units, atmospheric distillation units, vacuum distillation units, and oxidation blending units. These units are susceptible to various faults, which can be categorized into two types: some are local faults, typically mitigated by the unit’s own backup or adjustment mechanisms and posing few impacts on other units. Others are correlation faults, which often affect multiple units and can ultimately lead to production halts. The data set used in this paper includes both fault types, encompassing 18 distinct fault categories. The data set comprises approximately 30,000 normal samples and 3000 fault samples, collected at a 1 Hz sampling rate. Typically, each fault sequence starts with a normal operating segment, followed by the onset and evolution of the fault. Each sample contains 3001 time steps of process data, corresponding to 3000 s. For model development, 25,000 normal samples plus half of the fault samples are used for training, validation, and parameter tuning. The remaining normal samples and fault samples are reserved for testing to evaluate detection performance. The fault set includes 18 distinct fault categories: 8 local faults and 10 correlated faults. The number of test samples per fault category ranges from 104 to 312. A total of 226 process variables are monitored, distributed across 6 subsystem blocks corresponding to the six production units mentioned above. The 226 monitored variables are distributed across the six subsystem blocks as follows: heating furnace (45 variables), electric desalting (25 variables), initial distillation (35 variables), atmospheric distillation (48 variables), vacuum distillation (43 variables), and oxidation blending (30 variables). Part of these faults and variables are listed in Tables and .

1. Some Faults of Asphalt Production Equipment.

fault category type numbers of example description
F1 Correlation 104 Coking of the furnace tubes
F4 Local 302 Valve drift
F6 Local 245 Vacuum pump failure in distillation units
F10 Correlation 148 Flooding in distillation units
F15 Correlation 203 Tube leakage caused by corrosion of the flange gasket

2. Some Monitored Variables of Asphalt Production Equipment.

variable code subsystems unit description
TIC1001 Heating furnace °C The temperature in heating vessel
PI1005 Initial distillation MPa The pressure of pipeline
TIC2005 Heating furnace °C The temperature of heating furnace flue gas
SI510 Oxidation blending rpm Stirring motor speed
LIC2009 Vacuum distillation m The liquid level of distillation unit

3.2. Comparative Experiments on the Performance of Different Methods

The approaches involved in the comparison experiments include dynamic reconstruction PCA (DRPCA), Dynamic inner PCA (DiPCA), deep convolutional neural network (DCNN), orthogonal multiblock dynamic PLS (OMDPLS) and the proposed DMDPCA. Since DCNN operates as a classifier that outputs fault probabilities rather than generating statistical monitoring indicators such as T 2 and SPE, it is evaluated only through the FAR and FDR metrics in Tables and and is not included in the monitoring curve comparisons in Figures and . The hyper-parameters of the comparison experiments are shown in Table . The lag represents the maximum lag considered in each approach and it should be set according to the actual situation. Here it is determined to be 20 as the performance of all approaches is quite good under this parameter. The l of DiPCA represents the number of dynamic latent variables and it is determined via the parameter determination method described in. The h represents number of principal components and is determined via the principle of CPV. We take the CPV up to 0.95 as the criterion for select h. All model parameters are determined from the training set. During testing, each test sample is processed individually using the offline-trained parameters, and no information from future samples is used at any step.

4. Average FAR and FDR on Test Dataset.

  DRPCA DiPCA DCNN OMDPLS DMDPCA
Average FAR 0.0926 0.0862 0.0941 0.0878 0.0812
Average FDR 0.8913 0.8960 0.9038 0.8952 0.9025

5. Average Performance Under Different lag.

    lag = 2 lag = 5 lag = 10 lag = 20 lag = 40 lag = 100
FAR DRPCA 0.3503 0.2673 0.1497 0.0926 0.0689 0.0523
DiPCA 0.3487 0.2253 0.1241 0.0862 0.0612 0.0548
DCNN 0.3012 0.2156 0.1289 0.0941 0.0723 0.0695
OMDPLS 0.3498 0.2314 0.1267 0.0878 0.0633 0.0567
DMDPCA 0.3661 0.2731 0.1345 0.0812 0.0594 0.0482
FDR DRPCA 0.5314 0.6793 0.8573 0.8913 0.8699 0.8358
DiPCA 0.5646 0.6684 0.8476 0.8960 0.8762 0.8411
DCNN 0.5923 0.7108 0.8634 0.9038 0.8912 0.8724
OMDPLS 0.5587 0.6701 0.8489 0.8952 0.8798 0.8467
DMDPCA 0.5289 0.6815 0.8453 0.9025 0.9321 0.9528

2.

2

Monitoring curves of four MSPC methods on the same test sample sequence. DCNN is excluded as it does not produce monitoring curves.

3.

3

Normalized T2/0.99UCL curves of four MSPC methods on local fault (a) and correlated fault (b) sequences.

3. Hyper-Parameters of the Approaches Involved in the Experiment.

  DRPCA DiPCA DCNN OMDPLS DMDPCA
parameters lag = 20 lag = 20 lag = 20 Kernel Size = 3 × 3 lag = 20
h = 11 Batch Size = 64
h = 8 l = 18 h = 6 Epochs = 150 h = 9
Learning Rate = 0.003

Table shows the average false alarm rate (FAR) and average fault detection rate (FDR) of the five methods on all test sample sequences. Under the parameter lag = 20, DMDPCA achieves the lowest FAR among all methods while its FDR is comparable to that of DCNN. DCNN achieves the highest FDR due to its nonlinear feature extraction capacity, but its FAR is also the highest, likely because the classifier lacks an explicit mechanism for suppressing intermittent vibration noise. OMDPLS performs comparably to DiPCA, reflecting their shared latent-variable modeling foundation. The advantage of DMDPCA becomes more pronounced at larger lag values, as shown in Table .

Figure presents the monitoring curves of the four MSPC methods (DRPCA, DiPCA, OMDPLS, and DMDPCA) on a sample sequence from the test data set. DCNN is excluded from this figure because it operates as a classifier and does not produce continuous monitoring statistics comparable to T 2 and SPE. This sequence exhibits a microvibration around the 500th sample and a F4 fault (local faults) around the 1100th sample. Notably, the factory-provided labels for the data set are not completely accurate; as a result, the time points may earlier than the actual timing of vibration and fault onset. All four models performed well during both the normal stage (1–∼500) and the fault phase (∼1100–2000). During the normal phase, monitoring indicators generally remained below the upper control limit (UCL) at the 99% confidence level (0.99UCL), resulting in few false alarms. In the fault stage, these indicators rose rapidly and quickly exceeded the 0.99UCL, with sustained deviations thereafter. By contrast, during the microvibration stage (∼500–∼1100), the monitoring indicators of DRPCA and DiPCA exhibited larger fluctuations, resulting in numerous false alarms. This arose because equipment microvibrations induced intermittent noise in the data from some contact sensors, and conventional denoising techniques failed to effectively mitigate. Among these, the Q r indicator of DiPCA showed particularly significant fluctuations. This is likely because the Q r indicator is specifically designed to monitor changes in the system’s dynamic components, and data variability was exceptionally high during the microvibration stage. Although the indicators of DMDPCA also exhibited larger fluctuations, they rarely exceeded 0.99UCL, and thus there were fewer false alarms. Similarly, the indicators of OMDPLS showed fluctuations during the microvibration stage that were larger than those of DMDPCA but comparable to DRPCA and DiPCA, as OMDPLS also lacks the FBC-based noise suppression mechanism.

We also examined the impact of lag changes on model performance. Table shows the average FAR, FDR and test time of five models under the different lag. When the lag is 2 or 5, due to the lack of capturing dynamic characteristics, all models perform poorly on FAR and FDR. When the lag is 10, models can capture the dynamic characteristics of rapid processes, and thus the performance has significantly improved, but there are still some dynamic characteristics of medium and slow processes that have not been captured. When lag reaches 20, although the dynamic characteristics of slow processes (such as temperature) are not contained, the performance of DiPCA and DRPCA reaches the best. After that, as the lag increases, the dimension of the data matrix is too large due to the nonlinear growth of matrix dimensions, and DiPCA and DRPCA experience a certain degree of overfitting, with UCL becoming too large, resulting in a decrease in FDR. However, DMDPCA, by selectively augmenting the data matrix, captures the dynamic characteristics of slow processes while avoiding overfitting, and FDR reaches 95.28%. DCNN shows relatively stable performance across lag values due to its learned feature representations, but its FDR plateaus beyond lag = 20 and does not benefit from further increases. OMDPLS exhibits a performance trajectory like DiPCA, with FDR declining at large lag values due to the same dimensionality-related overfitting problem.

Figure presents the performance of the four MSPC methods on both types of faults. DCNN is excluded from this comparison for the same reason as in Figure . Note that the UCL values of each model are different, the vertical axis has been modified to T 2/0.99UCL for a clear show. For the local fault, all models demonstrated similar performance in terms of the time required to detect faults. As for the correlation fault, DiPCA and DRPCA are significantly slower than DMDPCA. As DRPCA and DiPCA do not explicitly model subsystem coupling, when a correlation fault occurs, different subsystems may exert opposing influences on the same principal component and restrain the magnitude of change in that component. This ultimately prevents statistical indicators from detecting the fault-induced changes in the principal component, leading to missed faults in early time. Only when the fault’s impact reaches certain magnitude do the principal components exhibit significant changes, allowing the indicators to detect the fault. OMDPLS, despite having a multiblock structure, does not explicitly model the dynamic coupling between subsystems via DMD and thus shows similar detection latency to DRPCA and DiPCA on correlated faults. Meanwhile, DMDPCA captures the subsystem coupling (including self-coupling) via DMD, constructs principal components on the incremental matrix and directly monitors the changes in the data itself, thereby it can detect correlation faults earlier than DRPCA and DiPCA.

To provide quantitative support beyond the individual monitoring curves in Figures , , and Table reports the average detection latency on correlation fault at lag = 100. The detection latency is defined as the number of samples from the labeled fault onset to the first 10 consecutive exceedance of the UCL. As shown in Table , at small lag values (lag = 2), all methods exhibit high detection latency due to insufficient capture of dynamic characteristics. At moderate lag (lag = 20), the detection latencies of all methods are comparable, with DiPCA and DMDPCA achieving the lowest values. However, as the lag increases beyond 20, the detection latency of DRPCA, DiPCA, and OMDPLS begins to rise due to the long computation time caused by excessive dimensions, while DMDPCA continues to decrease, reaching 64 samples at lag = 100. DCNN maintains relatively stable latency across lag values owing to its learned feature representations, but it does not achieve the low latency of DMDPCA at large lag settings. These results confirm that the selective augmentation strategy and Block DMD structure enable DMDPCA to detect correlated faults substantially earlier than other methods under large lag conditions.

6. Average Detection Latency on the Sequence of Correlation Fault.

    lag = 2 lag = 5 lag = 10 lag = 20 lag = 40 lag = 100
Detection latency DRPCA 198 156 118 95 108 131
  DiPCA 189 148 112 89 102 124
  DCNN 175 132 98 82 86 91
  OMDPLS 192 151 110 91 99 119
  DMDPCA 201 153 109 87 76 64

Table reports the average FDR on local and correlated faults separately at lag = 100. All five methods achieve similar FDR on local faults (0.87–0.94), while the difference on correlated faults is more pronounced: DRPCA, DiPCA, and OMDPLS achieve correlated FDR around 0.81, DCNN reaches 0.86 through nonlinear feature extraction, and DMDPCA achieves 0.96 by explicitly modeling subsystem coupling via Block DMD.

7. Average FDR on Local and Correlated Faults (lag = 100).

  DRPCA DiPCA DCNN OMDPLS DMDPCA
Local faults FDR 0.8690 0.8795 0.8891 0.8826 0.9413
Correlated faults FDR 0.8092 0.8104 0.8590 0.8180 0.9620

In summary, the case study on the asphalt production process demonstrates that DMDPCA offers clear advantages in two specific scenarios: when the lag is large, where the selective augmentation avoids the overfitting that degrades the performance of conventional methods and OMDPLS (Table ), and when correlated faults span multiple subsystems, where DMDPCA detects faults earlier than DRPCA, DiPCA, and OMDPLS (Figure ). DCNN achieves competitive FDR through nonlinear feature extraction but does not match DMDPCA in FAR reduction due to the lack of an explicit noise suppression mechanism. Under moderate lag settings, DMDPCA achieves the lowest FAR while maintaining comparable FDR to all baseline methods.

3.3. Ablation Experiments

To further validate the effectiveness of each component in the proposed DMDPCA framework, ablation experiments are conducted by systematically removing or modifying individual modules. Specifically, four variants are considered: (1) the baseline DPCA model, (2) DPCA with frequency-based clustering (FBC), (3) DPCA with FBC and transfer entropy-based augmentation (TE), and (4) the full proposed DMDPCA model incorporating Block DMD. All models are evaluated under the same experimental conditions using the asphalt production data set with lag = 100, and performance is assessed in terms of fault detection rate (FDR) and false alarm rate (FAR).

The results of the ablation study are summarized in Table . It can be observed that the baseline DPCA model achieves an FDR of 0.8863 with an FAR of 0.0675. Due to the presence of intermittent vibration noise and limited capability in capturing long-term dependencies, the baseline model exhibits relatively high false alarms and insufficient sensitivity to early faults.

8. Average FAR and FDR in Ablation Experiments.

number FBC TE-based augmentation BlockDMD FAR FDR
1       0.0675 0.8863
2 √     0.0532 0.9153
3 √ √   0.0536 0.9424
4 √ √ √ 0.0482 0.9528

After introducing the FBC module, the monitoring performance is improved to an FDR of 0.9153 and an FAR of 0.0532. This improvement can be attributed to the ability of FBC to suppress intermittent and nonstationary vibration noise in the frequency domain. By reducing noise contamination in the measurements, the model becomes more robust, leading to a noticeable decrease in false alarm rate.

When the transfer entropy-based augmentation is further incorporated, the FDR is further improved to 0.9424 while maintaining a relatively low FAR of 0.0536. This indicates that the TE-based strategy effectively captures informative long-term dependencies by selecting causally relevant lag variables, rather than blindly increasing the lag dimension as in conventional DPCA.

Finally, the complete DMDPCA model achieves the best overall performance among all variants. This demonstrates the effectiveness of the Block DMD module in explicitly modeling subsystem coupling. By capturing interactions among different subsystems, the proposed model can detect correlated faults, which are often difficult to identify using global modeling approaches due to the cancellation effects among variables.

Overall, the ablation results confirm that each module in the proposed framework plays a important role, and their integration leads to significant performance gains in complex industrial environments. The integration of these modules leads to a more effective monitoring framework for complex industrial processes.

4. Conclusion

In this paper, we present DMDPCA, a novel approach for chemical process monitoring and fault detection. To reduce the impact of microvibration of the equipment, we perform cluster based on the frequency of the sensor data and compensate for the noise based on the differences of the cluster centers. To prevent excessive lag from slowing down the calculation speed and accuracy of the model, we selectively append the time-lagged variables into the data matrix through TE. Finally, we model the coupling of the subsystems using dynamic mode decomposition and construct principal components on the incremental matrix, enabling DMDPCA to detect correlation faults earlier. To verify the effectiveness of the proposed approach, we perform comparison experiments on the asphalt production process. The experimental results demonstrate that DMDPCA outperforms DRPCA, DiPCA, DCNN, OMDPLS in scenarios involving large lag values and correlated faults across multiple subsystems. Under moderate lag settings, the proposed method achieves the lowest false alarm rate while maintaining a fault detection rate comparable to existing methods. The ablation study further confirms that each module contributes to the improvement of overall performance, validating the effectiveness of the proposed framework design. Although the proposed method shows promising results, further investigation is required to evaluate its generalization ability across different industrial processes. In addition, the selection of model parameters still relies on empirical settings, which may limit its adaptability in practical applications. Future work will focus on extending the proposed framework to more diverse industrial scenarios and developing adaptive parameter selection strategies to improve robustness and applicability in real-time monitoring systems.

Acknowledgments

This work is supported by the National Key Research and Development Program of China (2024YFB3311401) and the National Natural Science Foundation of China (62273176).

The authors declare no competing financial interest.

References

  1. Chen Y., Zhang X., Li D.. et al. A new class of fault detection and diagnosis methods by fusion of spatially distributed and time-dependent features. J. Process Control. 2025;146:103372. doi: 10.1016/j.jprocont.2024.103372. [DOI] [Google Scholar]
  2. Mason, R. L. ; Young, J. C. . Multivariate Statistical Process Control with Industrial Applications. Society for Industrial and Applied Mathematics 2002.
  3. Ge Z.. Review on data-driven modeling and monitoring for plant-wide industrial processes. Chemom. Intell. Lab. Syst. 2017;171:16–25. doi: 10.1016/j.chemolab.2017.09.021. [DOI] [Google Scholar]
  4. Jiang Q., Yan X., Huang B.. Review and perspectives of data-driven distributed monitoring for industrial plant-wide processes. Ind. Eng. Chem. Res. 2019;58(29):12899–12912. doi: 10.1021/acs.iecr.9b02391. [DOI] [Google Scholar]
  5. Melo A., Câmara M. M., Pinto J. C.. Data-driven process monitoring and fault diagnosis: A comprehensive survey. Processes. 2024;12(2):251. doi: 10.3390/pr12020251. [DOI] [Google Scholar]
  6. Li H., Jia M., Mao Z.. Dynamic reconstruction principal component analysis for process monitoring and fault detection in the cold rolling industry. J. Process Control. 2023;128:103010. doi: 10.1016/j.jprocont.2023.103010. [DOI] [Google Scholar]
  7. Khan A., Nahar R., Chen H.. et al. FaultExplainer: Leveraging large language models for interpretable fault detection and diagnosis. Comput. Chem. Eng. 2025;199:109152. doi: 10.1016/j.compchemeng.2025.109152. [DOI] [Google Scholar]
  8. Jiang Q., Yan X., Zhao W.. Fault detection and diagnosis in chemical processes using sensitive principal component analysis. Ind. Eng. Chem. Res. 2013;52(4):1635–1644. doi: 10.1021/ie3017016. [DOI] [Google Scholar]
  9. Yang G., Zhao Y., Gu X.. A novel Bayesian framework with enhanced principal component analysis for chemical fault diagnosis. IEEE Trans. Instrum. Meas. 2020;70:1–9. doi: 10.1109/tim.2020.3034975. [DOI] [Google Scholar]
  10. Sun C. Y., Yin Y. Z., Kang H. B.. et al. A quality-related fault detection method based on the dynamic data-driven algorithm for industrial systems. IEEE Trans. Autom. Sci. Eng. 2022;19(4):3942–3952. doi: 10.1109/TASE.2021.3139766. [DOI] [Google Scholar]
  11. Feng L., Zhao C.. Fault description based attribute transfer for zero-sample industrial fault diagnosis. IEEE Trans. Ind. Inf. 2021;17(3):1852–1862. doi: 10.1109/TII.2020.2988208. [DOI] [Google Scholar]
  12. Li W., Peng M., Wang Q.. Fault identification in PCA method during sensor condition monitoring in a nuclear power plant. Ann. Nucl. Energy. 2018;121:135–145. doi: 10.1016/j.anucene.2018.07.027. [DOI] [Google Scholar]
  13. Mohr F., Arnese-Feffin E., Barolo M.. et al. Smart process analytics for process monitoring. Comput. Chem. Eng. 2025;194:108918. doi: 10.1016/j.compchemeng.2024.108918. [DOI] [Google Scholar]
  14. Ku W., Storer R. H., Georgakis C.. Disturbance detection and isolation by dynamic principal component analysis. Chemom. Intell. Lab. Syst. 1995;30(1):179–196. doi: 10.1016/0169-7439(95)00076-3. [DOI] [Google Scholar]
  15. Dong Y., Qin S. J.. A novel dynamic PCA algorithm for dynamic data modeling and process monitoring. J. Process Control. 2018;67:1–11. doi: 10.1016/j.jprocont.2017.05.002. [DOI] [Google Scholar]
  16. Zhang Qi., Li P., Lang X.. et al. Improved dynamic kernel principal component analysis for fault detection. Measurement. 2020;158:107738. doi: 10.1016/j.measurement.2020.107738. [DOI] [Google Scholar]
  17. Ren Z., Jiang Y., Yang X.. et al. Learnable faster kernel-PCA for nonlinear fault detection: Deep autoencoder-based realization. J. Ind. Inf. Integr. 2024;40:100622. doi: 10.1016/j.jii.2024.100622. [DOI] [Google Scholar]
  18. Deng X., Tian X., Chen S.. et al. Deep principal component analysis based on layerwise feature extraction and its application to nonlinear process monitoring. IEEE Trans. Control Syst. Technol. 2019;27(6):2526–2540. doi: 10.1109/TCST.2018.2865413. [DOI] [Google Scholar]
  19. Jia Q., Zhang Y.. Quality-related fault detection approach based on dynamic kernel partial least squares. Chem. Eng. Res. Des. 2016;106:242–252. doi: 10.1016/j.cherd.2015.12.015. [DOI] [Google Scholar]
  20. Yi J., Huang D., He H.. et al. A novel framework for fault diagnosis using kernel partial least squares based on an optimal preference matrix. IEEE Trans. Ind. Electron. 2017;64(5):4315–4324. doi: 10.1109/TIE.2017.2668986. [DOI] [Google Scholar]
  21. Garcia-Bracamonte J. E., Ramirez-Cortes J. M., de Jesus Rangel-Magdaleno J.. et al. An approach on MCSA-based fault detection using independent component analysis and neural networks. IEEE Trans. Instrum. Meas. 2019;68(5):1353–1361. doi: 10.1109/TIM.2019.2900143. [DOI] [Google Scholar]
  22. Yang N.-C., Ismail H.. Robust intelligent learning algorithm using random forest and modified-independent component analysis for PV fault detection: In case of imbalanced data. IEEE Access. 2022;10:41119–41130. doi: 10.1109/ACCESS.2022.3166477. [DOI] [Google Scholar]
  23. Wu H., Zhao J.. Deep convolutional neural network model based chemical process fault diagnosis. Comput. Chem. Eng. 2018;115:185–197. doi: 10.1016/j.compchemeng.2018.04.009. [DOI] [Google Scholar]
  24. Yu J., Zhang Y.. Challenges and opportunities of deep learning-based process fault detection and diagnosis: a review. Neural Comput. Appl. 2023;35(1):211–252. doi: 10.1007/s00521-022-08017-3. [DOI] [Google Scholar]
  25. Hu C., Luo J., Kong X.. et al. Orthogonal multi-block dynamic PLS for quality-related process monitoring. IEEE Trans. Autom. Sci. Eng. 2024;21(3):3421–3434. doi: 10.1109/TASE.2023.3279575. [DOI] [Google Scholar]

Articles from ACS Omega are provided here courtesy of American Chemical Society

RESOURCES