One-class classification model for intelligent fault diagnosis of mine ventilation systems

Abstract

To address the problem of fault branch recognition in mine ventilation systems, a one-class classification algorithm is introduced to construct the MC-OCSVM ventilation system fault diagnosis model, which is integrated with multiple OCSVMs. This model adopts uniform hyperparameters and transforms the ventilation system fault diagnosis problem into a maximum decision distance problem, to realize the effective use of mine monitoring wind speed data. The experiments on public KEEL datasets verify that the one-class classification integration model can solve the multiclassification problem and that the MC-OCSVM model has better generalizability than other one-class classification integration models do. The experiment is carried out in the Buertai coal mine, and the results show that the proposed algorithm can identify fault branches quickly and accurately, with an accuracy of 93.2% and a single fault diagnosis time is 1.2 s, highlighting its strong robustness.

Introduction

The topological relationships of underground ventilation networks are complicated. When the wind resistance in one branch of the ventilation system changes, it impacts both the air volume in that branch and in other branches. Therefore, data monitored by wind speed sensors can only reflect the changes in air volume at the sensor’s location, making it difficult to determine if the specific branch with the sensor is faulty. This problem has become difficult to solve urgently in coal mines¹.

With the development of big data, the industrial internet, artificial intelligence and other technologies, fault diagnosis technology has been successfully applied in different industrial fields, providing a good reference for the application of this technology in mine ventilation systems^2–12. In 2018, Liu et al.^13,14 introduced a method using air volume as an input feature, obtained a ventilation fault sample set through an intelligent ventilation simulation system, and applied a support vector machine algorithm to predict the fault location and magnitude of the ventilation system. This method pioneered the application of the intelligent AI algorithm to fault mine ventilation systems. In subsequent research, Zhou¹⁵ used an improved genetic algorithm to optimize the penalty coefficient and kernel function coefficient of a support vector machine. On the basis of a coefficient-optimized support vector mechanism, he built a resistance-type fault diagnosis model for mine ventilation systems that improved the accuracy of fault diagnosis to a certain extent. Huang et al.^16–18 analyzed the problem of ventilation system fault diagnosis from two perspectives: with and without sample sets. They built an unsupervised learning model with a hybrid coding adaptive evolutionary strategy for mine ventilation systems, verified the performance advantages of the algorithm through experiments, and realized the integrated prediction of fault location and fault quantity. Zhao et al.^19–21 applied an improved SVM algorithm to achieve fault diagnosis of a ventilation system in the Daming Coal Mine. They constructed a fault tunnel range database and proposed an online fault diagnosis model for the mine ventilation system to improve the fault diagnosis efficiency. Zhang et al.²² built a pipeline experimental platform and established three fault diagnosis machine learning models for mine ventilation systems, namely, support vector machines, neural networks and random forests. Through model experiments and mine production simulation experiments with different ventilation modes, it was verified that the neural network model under the characteristics of air volume–wind pressure generated better fault diagnosis results among the three models. Li et al.²³ studied a fault diagnosis method for ventilation systems based on monitoring air volume, analyzed the change characteristics of air volume, and built a mathematical model of fault diagnosis on this basis, providing a new method for fault diagnosis on the basis of online monitoring. Wang et al.²⁴ studied an identification algorithm when multiple branches of a mine ventilation system fail at the same time and built a machine learning model based on multilabel K-nearest neighbors to solve this problem. Liu et al.^25,26 applied four kinds of machine learning algorithms to fully evaluate the performance of a fault diagnosis model for mine ventilation systems and demonstrated the superiority of the K-nearest neighbor and decision tree models. Moreover, a support vector machine was used as the classification model. The influence of four factors, sample dispersion, sample number, input feature and feature number, on the generalizability of the fault diagnosis model was analyzed, providing a reference for the establishment of a machine learning model for fault diagnosis of ventilation systems.

Many important research achievements have been made in the resistance fault diagnosis of mine ventilation systems, but the following problems remain: (1) Unsupervised learning fault diagnosis models consume too many computing resources, have high hardware requirements, and have a slow fault diagnosis speed, which are not conducive to industrial applications. (2) Supervised learning models require a considerable number of data samples. Given that it is difficult to obtain a sufficient number of fault samples due to the constraints of current production mine conditions, in previous studies, sample data have been generated through simulations. There must be a certain gap between the simulated sample and the real sample, so how to diagnose the fault through the limited normal historical data is highly important. Fault sample identification is a problem of anomaly detection (fault data) of known classes (normal data), which can be solved via a single classification (OCC) algorithm. In view of this, in this paper, several OCSVM models were integrated to build the MC-OCSVM algorithm. The main contributions of this study are as follows: (1) The proposed model only requires normal sample data for training, which solves the problem of machine learning modeling of ventilation system fault diagnosis under the condition of insufficient fault samples. (2) The model exhibits strong robustness and generalizability and can meet the requirements of timeliness of ventilation system fault diagnosis. This study is closely aligned with the development of intelligent mines and is the basis for realizing intelligent maintenance and management of mine ventilation systems.

Single-classification support vector machine

The essence of SVM in the OCC problem is to obtain a nonlinear mapping based on a normal sample training set of mine ventilation system monitoring data , map the samples to a high-dimensional space, and construct a hyperplane so that the target class samples in the mine ventilation system monitoring dataset are separated from the outlier class samples to the greatest extent. Therefore, single-class support vector machines (OCSVMs) can be formalized as:

1

where is the normal vector of the hyperplane; is the number of training samples; and is the hyperplane intercept. Since it is impossible to completely separate all the target class samples from the outlier class samples, the optimal solution cannot be obtained. To obtain the suboptimal solution, relaxation variables and parameters are introduced. indicates the rejection ratio.

By introducing kernel functions and Lagrange multipliers , the parameters can be computed for any sample as follows²⁷:

2

The decision function of the OCSVM model is as follows:

3

When the OCSVM model is used for fault diagnosis, the model uses the entire training set as target class samples for learning. During the test phase, the statistical test quantity and threshold value need to be set. If the statistical test quantity of the test sample exceeds the threshold value, it is classified as a fault sample; otherwise, it is classified as a target class sample. On the basis of the test quantity and threshold settings presented in reference¹², in the OCSVM, the statistical test quantity is as follows:

4

Assuming that the model training is complete and that the statistical tests corresponding to each training sample are sorted as , the decision threshold value is as shown in Eq. (5), and the function floor(x) represents the maximum integer value that does not exceed x.

5

where v is the rejection ratio and N is the number of samples.

The decision threshold minus the threshold from the sample to the classification boundary is used as the classification decision value:

6

According to Formulas (5) and (6), the result of the OCSVM model for a new sample xi can be obtained as follows. When xi is a normal sample, the output is 1, and xi is a normal sample, the output is -1.

7

MC-OCSVM model

On the basis of the relevant theories of the OCSVM, multiple OCSVMs were integrated to construct an MC-OCSVM model suitable for multiple classifications of mine ventilation systems, as shown in Fig. 1. Unlike the "one-to-many" or "one-to-one" decomposition strategies that are often used to extend binary classifiers to multiple classifiers, the MC-OCSVM is trained by building a separate OCSVM model based on each class. After the test sample is input to the MC-OCSVM, each OCSVM outputs a predicted value, which represents the distance of the sample from the normal sample classification decision vector, and the branch corresponding to the maximum predicted value is the final predicted fault branch of the MC-OCSVM. Since each OCSVM model is trained separately and the mapping space of each model is different, the output results cannot be directly compared. To avoid this problem, a unified hyperparameter for training the OCSVM model was selected, and a scale factor was applied to uniformly process the output of the OCSVM model. The problem of ventilation system fault diagnosis is transformed into the problem of maximum decision distance.The pseudo-code of MC-OCSVM algorithm is shown in Table 1.

Fig. 1.

Principle of MC-OCSVM.

Table 1.

MC-OCSVM algorithm.

Algorithm: MC-OCSVM algorithm for resistance fault diagnosis in mine ventilation systems
Input: Xtrain, Xtest, Kernel type: Kernel_type, Nuclear parameter: kernel_parameters, Number of iterations: c, Class tag: T
Output: , scale_factor,
1) Train
For i = 1 to c
Htrain{i} = Kernel_function(Xtrain(i),Kernel_type, kernel_parameters)
The output of each OCSVM: ,f = [f1,f2,…fn]
Determine the decision threshold for each OCSVM:
scale_factor(i) = mean(exp((θ(i)-ptrain{i}/0.01))
end
2) Test
For i = 1 to c
Htest{i} = Kernel_function(Xtest(i),Kernel_type, kernel_parameters)
Calculate the predicted value of each OCSVM model: ptest{i} = T-f{i}
Calculate the final decision value: final_ptest{i} = ptest{i}/scale_factor(i)
end
The branch corresponding to the calculated maximum decision value is the final predicted fault branch: l = find(final_ptest = max(final_ptest))

The scale factor of a single OCSVM model is as follows:

8

where p(i) is the classification threshold of a single OCSVM.

On the basis of Eq. (8), the scale unification formula of each OCSVM in the MC-OCSVM is determined as follows:

9

Fault diagnosis of the mine ventilation system MC-OCSVM

Fault diagnosis process

The steps of applying the MC-OCSVM algorithm to the resistance fault diagnosis of a mine ventilation system are as follows:

1. The wind speed (air volume) is used as the input feature of the fault diagnosis model, and the model is trained with the fault-free sample data of the mine ventilation system.

2. The parameters of a single classification model are initialized, including the kernel function K, nuclear parameter g and penalty factor c, the single classification model is calculated according to Formula (2), the classification hyperplane is calculated according to Formula (3), and the threshold value of a single classification model is calculated according to Formulas (4) and (5).

3. The test samples are input into the trained MC-OCSVM model, the scale factors are calculated according to Eqs. (8) and (9), and the predicted values of each OCSVM model are calculated.

4. The prediction branch number corresponding to the final maximum predicted value is output.

Evaluation indices

For multiclassification problems with unbalanced samples, the accuracy rate (ACC) index may not fully capture the performance of the classification results. Therefore, the recall rate (Re), accuracy rate (Pr), G-mean, and F1 score are also used in this study to comprehensively evaluate the ventilation fault diagnosis model¹⁹.

10

11

12

13

14

where N is the number of branches of the mine ventilation network in the input mode. T_Pi is the true example of the i-th fault branch, T_Ni is the true negative example of the i-th fault branch, F_Pi is the false-positive example of the i-_th fault branch, and F_Ni is a false-negative example of the i-th faulty branch.

Experimental analysis

To verify the generalizability of the MC-OCSVM and the feasibility of its application in ventilation system fault diagnosis, the proposed algorithm was applied to the KEEL dataset and mine ventilation system examples. The classification performance of the multiclassification algorithms (MC-OCELM, MC-SVDD) constructed from other commonly used single classification algorithms (OCELM, SVDD) is compared with that of MC-OCSVM. These multiclassification algorithms adopt the same construction strategy as MC-OCSVM. The test environment of all the experiments in this paper is as follows: Windows 10 operating system, Intel Core i7 processor, 16 GB RAM, running software MATLAB R2017b. The dd_tool program package is used to implement the MC-SVDD algorithm, the LibELM program package is used to implement the MC-OCELM algorithm, and the LibSVM program package is used to implement the MC-OCSVM algorithm.

KEEL dataset validation

The KEEL dataset is a commonly used open test dataset in the field of machine learning. Five unbalanced datasets in the KEEL open dataset are selected as research objects in this paper. These datasets have different unbalanced proportions and contain binary or multiclassification characteristics. The basic attributes are shown in Table 2. Each dataset is randomly divided into a training set and a test set at a ratio of 7:3. To avoid contingency and randomness of the results of a single training operation, for each algorithm, 10 repeated experiments are conducted on each dataset. The classification evaluation index values of each model are shown in Table 3 (the optimal results are shown in bold), and the results are the average values of the 10 experiments.

Table 2.

Basic properties of the KEEL dataset.

Dataset	Quantity feature	Quantity category	Category distribution	Imbalance ratio
glass1	214	9	138/76	1.82
N-thyroid	215	5	150/35/30	5
balance	625	4	288/288/49	5.88
autos	159	25	48/46/42/20/3	16
ecoli	340	7	143/81/52/35/20/5/2/2	71.5

Table 3.

Experimental results of the algorithms on the KEEL dataset.

Dataset	Algorithm	Re (%)	Pr (%)	F1 (%)	G-mean (%)
glass1	MC-OCSVM	82.7	83.1	83.5	82.6
	MC-OCELM	78.1	78.2	77.9	78.2
	MC-SVDD	77.4	74.3	76.5	77.5
N-thyroid	MC-OCSVM	91.2	91.2	90.5	90.4
	MC-OCELM	80.5	81.4	82.3	81.2
	MC-SVDD	77.8	79.2	79.5	80.3
balance	MC-OCSVM	90.2	90.2	90.2	90.2
	MC-OCELM	80.5	80.6	80.6	80.6
	MC-SVDD	78.5	78.5	78.5	78.5
autos	MC-OCSVM	93.5	93.1	92.3	91.2
	MC-OCELM	82.5	80.7	81.5	84.5
	MC-SVDD	87.6	81.6	87.6	88.5
ecoli	MC-OCSVM	88.6	86.1	86.1	85.5
	MC-OCELM	76.5	77.5	77.5	79.6
	MC-SVDD	80.5	81.3	81.3	81.5

As shown in Table 3:

1. The three single-classification integration models have all completed the multiclassification test of the KEEL datasets, and the average Re, Pr, F1 and G-mean index values of MC-OCSVM, MC-SVDD and MC-OCELM on the five public datasets are as follows: at 83.1%, 82.5%, 83.1%, and 83.4%, respectively, the evaluation index values of all datasets exceeded 80%. Therefore, the single-classification integration model can effectively address the classification of unbalanced data.

2. Compared with the MC-OCELM and MC-SVDD models, the MC-OCSVM classification model proposed in this paper has advantages in terms of classification results on the 5 public datasets, and the Re, Pr, F1 scores and G-mean index values are all optimal, indicating that the model has the best generalizability.

Production mine example verification

Data preparation

The Buertai coal mine was chosen as an example to carry out MC-OCSVM fault diagnosis experiments on mine ventilation systems. The mine has a diagonal ventilation network, as shown in Fig. 2. The number of branches is 98, and the number of nodes is 94. The training sample set was obtained via real-time network solution and monitoring data export in the Intelligent Security Control Platform (VentAnaly6.0) developed by Coal Science and Technology Research Institute. The method of obtaining the test sample set is as follows. Ten branches with regulating air windows and dampers that are prone to faults in the ventilation system, e₆₇, e₅,e₂₂, e₈₁, e₃₂, e₇₁, e₂₀, e₂₁, e₅₃, and e₉, were selected to simulate the fault experiment. The initial parameters of the corresponding branches are shown in Table 4. According to the initial conditions of the change in the wind resistance, the branch fault was simulated by changing the wind resistance value with a VentAnaly 6.0 step length. The resistance increase simulation was performed on branches e₂₁, e_3, and e₃₀, where the air window is located, and the resistance reduction simulation was performed on branches e₇, e₂₅, e₁₈, e₅₆, e₄₈, e_78, and e₁₂, where the air door is located. Considering the large span of these wind resistance changes, the maximum value is 196.552, and the minimum value is 0.1215. The wind resistance change operation was performed on the branches to ensure the normal calculation of the ventilation network by determining different drag increase or decrease coefficients. The drag increase or decrease coefficients are shown in Table 5, and the wind resistance decreases linearly according to u times the original wind resistance, that is, each branch generates 30 fault samples, and a total of 300 fault samples are generated to constitute the test set. To verify the superiority of the MC-OCSVM in fault diagnosis for production mines, the MC-OCSVM, MC-OCELM and MC-SVDD models were established for comparative analysis.

Fig. 2.

Ventilation network of the Buertai coal mine.

Table 4.

Branch initial parameters in the Buertai coal mine.

Branch number	Sectional area/m2	Air volume/q0 (m3 s−1)	Wind resistance r0/(N s2 m−8)
e67	7.5	1.65	196.522
e5	8.3	2.65	95.2612
e22	10.6	2.95	50.655
e81	7.9	1.85	100.621
e32	7.6	2.02	60.232
e71	8.5	4.265	20.325
e20	8.3	1.365	30.325
e21	7.5	45.621	0.1542
e53	8.0	42.325	0.1215
e9	8.8	44.265	0.1325

Table 5.

Coefficient of resistance variation.

Initial wind resistance r0/(N s2 m−8)	r0 < 0.2	0.2 < r0 < 20	20 < r0 < 100	r0 > 100
increase or decrease drag coefficient u	0.5	0.005	0.0005	0.00005

Parameter selection

With different datasets, the model performance of different kernel functions has advantages and disadvantages. The kernel functions commonly used in the OCSVM model are the linear kernel function, polynomial kernel function, sigmoid kernel function, and radial basis function (RBF) kernel function. On the basis of the ventilation database of the Buertai coal mine, the performance of the OCSVM model was tested on four different kernel functions with 30 resistivity fault samples of the e₁₃ branch. With the other parameters fixed, the classification results of the four different kernel functions are shown in Fig. 3. In the classification test of the linear kernel function, only 16 samples out of 30 samples are classified correctly, and the prediction accuracy is only 53.3%. In the polynomial function classification test, 15 samples are predicted correctly, and the discrimination result is lower than that of the linear kernel function, with an accuracy of 50%. In the RBF kernel function classification test, 22 samples are predicted correctly, with an accuracy of 73.3%. In the sigmoid kernel function classification test, 24 samples are predicted correctly, with an accuracy of 80%. After a comprehensive comparison of the classification results of the above four kernel functions, the sigmoid kernel function has the best effect on this numerical fit. Therefore, the sigmoid kernel function is adopted in the OCSVM model for the Buertai coal mine fault diagnosis experiment. The other optimal hyperparameters of each single classification model were determined via fivefold cross-validation, as shown in Table 6.

Fig. 3.

Classification results of different kernel functions.

Table 6.

Parameters of the fault diagnosis model for the Buertai coal mine.

Single classification model	Parameter setting
OCSVM	v = 0.2, g = 0.145, K = Sigmoid
OCELM	v = 0.01, g = 0.023, K = RBF
SVDD	v = 0.05, g = 0.356, K = Sigmoid

Experimental results

Under the optimal parameter set, the fault diagnosis location index integrated by different single classification algorithms in the Buertai coal mine is shown in Fig. 4. The results of the predicted fault branches and real fault branches indicate that the fault diagnosis results are correct.

Fig. 4.

Fault branch diagnosis results for the Buertai coal mine under different single-classification integrated algorithms.

As shown in Fig. 4, the resistance fault branch diagnosis accuracies of the MC-OCSVM, MC-OCELM and MC-SVDD models are 93.2%, 87.7% and 91.6%, respectively, and the fault diagnosis accuracy of the MC-OCSVM model is the highest, increasing by 3.55% compared with the average values of the other two models. Compared with the traditional SVM ventilation system resistance fault diagnosis algorithm proposed in Reference⁵ (78.1%), the accuracy rate increases by 15.1%, and compared with the unsupervised ventilation system resistance fault diagnosis model proposed in Reference⁸ (92.5%), the accuracy rate increases by 0.7%. In the classification of the MC-OCSVM, MC-OCELM and MC-SVDD models, the proportions of fault branches misclassified as normal branches are 1.3%, 2.3%, and 5%, respectively. The MC-OCSVM model is the lowest, followed by the MC-OCELM model. The MC-OCSVM and MC-OCELM models maintain a low rate of misidentification of faulty branches as normal branches, which has obvious advantages.

The scores of the evaluation indicators of each model are shown in Table 7 (the mean standard deviation of each branch index). As shown in Table 7, the evaluation index values of the MC-OCSVM model of the Buertai coal mine ventilation system are as follows: Pr = Re = F1 = G-mean = 0.932; Pr = F1 = G-mean = 0.877 for the whole MC-OCELM model; and Pr = Re = F1 = G-mean = 0.916 for the MC-SVDD model as a whole, which indicates that for high-dimensional data of ventilation systems in large production mines, the single-classification integrated model has strong robustness and universality. Compared with those of the MC-OCELM and MC-SVDD models, the Pr, Re, F1 scores, and G-mean of the MC-OCSVM fault diagnosis model are increased by 5.5% and 1.6%, respectively, and a good fault diagnosis effect is achieved.

Table 7.

Evaluation index values of different single-classification integrated models in the Buertai coal mine.

Classification model	Pr	Re	F1	G-mean
MC-OCSVM	0.932 0.025	0.932 0.36	0.932 0.054	0.932 0.142
MC-OCELM	0.877 0.142	0.877 0.34	0.877 0.652	0.877 0.45
MC-SVDD	0.916 0.54	0.916 0.54	0.916 0.24	0.916 0.35

The training time of each single classification integration algorithm is shown in Table 8, the cumulative test time consumption of 30 fault diagnosis experiments of the e₈₆ branch is shown in Fig. 5a, and the learning curve of each algorithm is shown in Fig. 5b.

Table 8.

Comparison of the training times of the single classification integration algorithms.

Algorithm	MC-OCSVM	MC-OCELM	MC-SVDD
Training time/s	52.3	32.1	78.2

Fig. 5.

Comparison of single-classification integration algorithm time consumption and convergence rate. (a) Time consumption of each model. (b) Each model learning curve.

As shown in Table 8 and Fig. 5, the training time of the MC-OCELM model is the shortest, and the training time of the MC-SVDD model is the longest, followed by the MC-OCSVM model. The time consumption of the training of the MC-OCSVM model proposed in this paper is not high, but the fault diagnosis model is trained offline. Training time consumption is not the main evaluation index of fault diagnosis model performance. The cumulative time consumption of each model shows that the fault diagnosis test time consumption is positively correlated with the number of tests. The slopes of MC-OCSVM, MC-OCELM, and MC-SVDD are 1.2, 6.1, and 3.2, respectively. The single fault diagnosis times of the MC-OCSVM, MC-OCELM and MC-SVDD models are 1.2 s, 6.1 s, and 3.2 s, respectively. The test time of the MC-OCSVM model is the shortest, the test time of the MC-SVDD model is the longest, and the test time of the MC-OCELM model is second longest. The learning curves of the three algorithms indicate that the MC-OCELM model converges the fastest, followed by the MC-OCSVM, and that MC-SVDD converges the slowest. Therefore, considering the training time and convergence speed, in the operating environment of this experiment, the time consumed by the MC-OCSVM fault diagnosis model is relatively stable. The single fault diagnosis time is short, which can meet the needs of ventilation system resistance fault diagnosis.

To verify the feasibility of the application of the MC-OCSVM model on the basis of measured data, field tests were conducted in the Buertai coal mine. Owing to the requirements of safe field production and the limitations of actual conditions, the field tests were conducted by adjusting the damper or wind window. When the ventilation system of the Buertai coal mine has no fault, the damper is closed, and the wind window is open. The resistive fault state was artificially simulated by opening the damper or closing the wind window. To ensure the safe production of the Buertai coal mine, the air doors of the upper left 1 return air lane of 6D in the west mining area and upper right 1 lane of 3# in the west mining area were opened, and the air windows of the lower left 2 lane of 7# in the fifth mining area and lower right 2 lane of 6D in the west mining area were closed. The average value of the monitoring data of 15 wind speed sensors within 10 min after the simulated fault occurred was used as the test sample (the fault diagnosis test was based on the number and position of sensors actually installed in the mine, without considering the optimization of the number and position of sensors), and the trained MC-OCSVM model was applied for fault branch diagnosis. In the field test of 4 fault branches, the MC-OCSVM model can correctly identify the fault branches in each test, and the fault diagnosis accuracy reaches 100%. The field application of the intelligent MC-OCSVM algorithm for resistance fault diagnosis of mine ventilation systems is feasible.

Conclusion

In this paper, a new classification algorithm is designed to solve the problem of fault diagnosis in mine ventilation systems with unbalanced sample data and effectively improve the accuracy of fault diagnosis in ventilation systems.

1. By introducing a single classification algorithm, a fault diagnosis model of a ventilation system based on the integration of multiple single classification algorithms (MC-OCSVM) is proposed to realize modeling without fault sample participation in training. The output of the OCSVM model is uniformly processed in combination with scale factors, thus transforming the fault diagnosis problem of the ventilation system into a maximum decision distance problem.

2. The experimental results of the KEEL dataset show that the single-classification integration model can effectively address the classification of unbalanced data and that the MC-OCSVM model has the best generalizability.

3. The fault diagnosis accuracy rate of the MC-OCSVM algorithm for mine ventilation systems is 93.2%, the index values of Pr, Re, F1 and the G-mean are all 93.2%, and the single fault diagnosis time is 1.2 s, which indicates strong robustness and generalizability. Compared with the single classification integration model of MC-OCELM and MC-SVDD, the fault diagnosis time of the MC-OCSVM algorithm is 1.2 s. The MC-OCSVM has advantages in terms of accuracy, recall rate, accuracy rate, F1 score, G-mean index, missed diagnosis rate, robustness, time overhead and convergence speed.

Owing to the complexity of mine ventilation networks and the restrictions of intelligent mine processes, this study still has certain limitations. For example, the timeliness of the model proposed in this paper makes it difficult to meet the requirements of online diagnosis. The focus of further studies will be on further shortening the time of fault diagnosis and improving the fault diagnosis rate.

Author contributions

Luo wrote the main manuscript text and Zhao prepared Figs. 1 and 2. All authors reviewed the manuscript.

Data availability

The data for this paper is available from the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.