1. Introduction
The engine is the core power source of important equipment. It is particularly important for the normal operation of the equipment to detect and deal with the faults of the starting machine in time, which can greatly increase the safety of life and reduce the property loss.
Fault diagnosis methods based on artificial intelligence do not rely on the expert knowledge and experience and can achieve good results, so the paper uses artificial intelligence methods for fault diagnosis. As monitoring data are mostly time series signals, extracting the time series information from them is especially important for fault diagnosis, RNN (recurrent neural network) is widely used in fault diagnosis because of its excellent ability of time domain information mining. Although RNN is more effective in processing time series data, it faces the problem of vanishing gradient, which is alleviated by GRU to some extent. Gao [
1] used GRU to diagnose the collected sample set and compared it with BP (back propagation) network, and the results showed that GRU had higher accuracy. Shi [
2] used bidirectional GRU for fault diagnosis and showed high diagnostic accuracy. In Zhang [
3], CNN (convolutional neural network) and GRU are simultaneously used for feature extraction to complete fault diagnosis. GRU has high accuracy, fewer parameters, and high computational efficiency, so GRU is selected as the selection network in this paper.
Extraction of effective features can improve fault diagnosis accuracy, and various methods are used to extract effective features. For example, Zhou [
4] obtained the product of the square of the instantaneous amplitude and the instantaneous frequency of the vibration by using the Teager operator, effectively extracted the instantaneous energy of the cylinder vibration signal, and used it as a feature for the fault diagnosis of the cylinder. When selecting sensor signals for fault diagnosis, sensors are often selected according to the physical location of the research target. However, a sensor that is consistent with the physical location of the research target does not mean that it carries the most fault-related information. At the same time, many methods use VMD (variational mode decomposition), wavelet transform, and other algorithms to deeply mine the frequency domain information of the sensor signal after selecting the sensor signal and extracting the frequency domain information for fault diagnosis. For example, Wang [
5] uses wavelet transform to extract the information of instantaneous speed signals in normal state and fault state and compares them to obtain effective fault information. Jiang [
6] also obtained the effective frequency domain information of cylinder misfire fault by using wavelet transform. Bi [
7], Zhang [
8], and Qiao [
9] use VMD to extract information from multiple modes and use it for fault diagnosis. For example, Yang [
10] used correlation analysis to obtain characteristic correlation coefficients in the time domain, frequency domain, and time–frequency domain under different working conditions to obtain sensitivity features and used the normal state as the benchmark for fault diagnosis. Xu [
11,
12] used wavelet packet transform to extract the energy entropy of each node and Hilbert transform to extract the marginal spectrum energy entropy and sent it as a feature to SVM (support vector machine) for fault diagnosis. However, the performance of algorithms such as VMD and wavelet transform depends on the selected parameters to a large extent. At the same time, there is no complete theory about the selection of these parameters to select reasonable parameters for fault diagnosis. Therefore, it is difficult to extract frequency information for fault diagnosis. The time domain feature is used for fault diagnosis in this paper. To select the signals and features carrying more fault information, this paper proposes to use the Pearson correlation coefficient combined with fault state to extract time domain features for fault diagnosis.
The fault type and severity diagnosis of the engine are particularly important for taking reasonable maintenance measures. However, most of the existing methods only diagnose the fault type without further classification of the fault severity. These classification methods are not conducive to taking reasonable maintenance measures. Some scholars have taken some methods to analyze the severity of faults in the fields of bearings and gears. Xiao [
12] uses wavelet transform and curve fitting to obtain warning value and alarm value and judges the fault degree based on the fault indication value obtained by fitting. This method is highly dependent on the selection of wavelet function and needs to determine the threshold manually, which is highly dependent on expert knowledge and experience. Pan [
13,
14] first performed feature selection and dimensionality reduction on the data. Then, he fed the selected features into a monotonic decision tree for fault severity diagnosis. The accuracy of his method reached 93.58%. JHA [
14] completed the fault severity diagnosis of the rolling bearing in two stages by using a multi-classification support vector machine. Machine learning methods such as support vector machines can only extract shallow features. Therefore, when the amount of data is large, the generalization ability of machine learning methods is poor, and the data information cannot be mined well. Almounajjed [
15] completed the fault severity diagnosis of the stator gate short circuit of the induction motor by using the mathematical model. Hang [
16] completed the detection of fault severity of high resistance wiring of the synchronous motor by establishing a mathematical model. Taha [
17] combined dissolved gas analysis and neural pattern recognition to complete the fault severity diagnosis of the power transformer. The accuracy of his method reached 93.58%. The model-based fault severity diagnosis method is difficult to implement in many cases because it requires professional knowledge and rich experience to establish the model, and the generalization ability of the model-based method is generally weak. Yang [
18] completed the fault severity diagnosis of the rolling bearing by integrated learning, with a training accuracy of 98.57%, verification accuracy of 100%, and test accuracy of 96.51%. Gai [
19] completed the fault severity diagnosis of rolling bearings using DBN (deep belief network), and the average detection accuracy reached 96%. The fault severity analysis of the above method is aimed at fewer fault severity types, and its ability to solve complex problems is not high. Sun [
20] completed the diagnosis of six fault types and their corresponding four fault severity levels by using a multi-head attention network combined with deep learning. However, the existing method requires environmental and control variables. These data requirements cannot be realized in many practical situations. Dibaj [
21] completed the fault severity diagnosis of bearings and gears by using VMD and CNN. However, this method needs to set the parameters of VMD and select the threshold. The reasonableness or otherwise of the parameters has an impact on the fault diagnosis results. The fault diagnosis results can only be obtained by further analyzing the results obtained by the network.
Most of the existing fault severity diagnosis methods require more demanding raw data and final results require further analysis of the model diagnosis results or the setting of thresholds based on experience. The selection of data processing methods and their parameters greatly depends on experience and professional theoretical knowledge. In order to select features containing more fault information, a feature selection method combining the Pearson correlation coefficient and faults is proposed in the paper to select features for fault severity diagnosis. First, sensors containing more fault information and the corresponding time domain features are selected twice using Pearson correlation coefficients [
22]. Then, the time domain features are sent to the GRU as input for fault diagnosis. Since multi-fault diagnosis is more difficult and requires more state information, we select the time domain features selected above from all vibration signals and concatenate them as input to GRU for multi-fault fault diagnosis.
3. Results
In this section, the single-fault diagnosis of four engine faults is carried out, and then the multi-fault diagnosis is carried out to directly obtain the fault severity of the corresponding fault type. To reduce the influence of random factors, five random experiments are carried out for each experiment, and the final experimental results are the average of the five random experiments.
3.1. Single-Fault Diagnosis
Four fault severity diagnosis models were trained for the four fault types. The effective features of four fault types (exhaust valve, cylinder, piston ring, and intake valve) are extracted and enhanced to obtain model input. Then, the training data set of four fault types are sent to four network models designed in this paper for training to obtain four single-fault diagnosis models.
Figure 7 are t-SNE dimension reduction diagrams of the 32-dimensional fully connected layer eigenvector of the single-fault diagnosis test set of four faults (exhaust valve, cylinder, piston ring, and intake valve). From the above feature reduction diagram, we can see that after combining the fault state with Pearson correlation coefficient and GRU, we can extract features with high resolution and distinguish between different fault severity. There are obvious differences between different fault severity, and the diagnosis results are obvious. The fault diagnosis method in this paper is effective.
In the four single-fault diagnosis models, the average test accuracy dimension of only five rounds of cylinders is 99.99%. The average test accuracy of the other three fault states reached 100%, and the generalization ability of the selected features was verified.
3.2. Multi-Fault Diagnosis
By training a multi-fault diagnosis model with the data of multiple fault types, different fault severity results of multiple fault types can be obtained. A dataset of 13 fault severities of four fault types (three wear degrees of the exhaust valve, four wear degrees of the cylinder, three wear degrees of the piston ring, and three wear degrees of the intake valve) is extracted by features to obtain the input of the multi-fault diagnosis model, which is then fed into the multi-fault diagnosis model for training.
The dimension reduction results of t-SNE before the full connection classification layer of multi-fault diagnosis are shown in
Figure 8.
It can be seen from the feature reduction diagram that the difference between different severity of different faults is obvious, the extracted features have high expressive power, and there is no crossover between different fault severity features. The classification result is excellent.
Figure 9 is the confusion matrix of diagnosis results of the test set of five rounds of random experiments with thirteen faults.
From the confusion matrix of five rounds of random experiments, the average test accuracy of five rounds of random experiments can reach 99.97%, and the generalization ability and diagnostic accuracy are excellent.
3.3. Analysis of Influencing Factors
3.3.1. Dropout Rate
To prevent overfitting, dropout is used in the network. However, a too large dropout rate will lead to underfitting, and a too small dropout rate will lead to overfitting. Therefore, choosing a suitable dropout rate is very important for the performance of the model. Five rounds of crossover experiments were carried out with the dropout rates of 0.2, 0.3, 0.4, and 0.5. The results of five rounds of crossover experiments with different dropout rates are shown in
Table 6.
According to the results in
Table 6, when the dropout rate is 0.2, the average test accuracy and average verification accuracy of the multi-fault model are the highest, and the results verify the superiority of the dropout rate selected by the model.
3.3.2. Number of Vibration Signals Selected
To verify the necessity of selecting the features of the three vibration signals as inputs, this subsection only uses one and two sensor signals (1 × 128 and 2 × 128) as inputs to the network for fault diagnosis to verify the necessity of the three sensor signals for multi-fault diagnosis.
To avoid the influence of random factors, the final results are taken as the average of five rounds of random experiments. The results of five rounds of crossover experiments with different numbers of vibration signals are shown in
Table 7.
From the results in
Table 7, we can conclude that the average test accuracy and the average validation accuracy of the multi-fault diagnosis model are the highest when the number of sensors is 3. Therefore, it is necessary to use the three sensor signals for multi-fault diagnosis.
3.4. Comparison of Different Feature Extraction Methods
To verify the effectiveness of the proposed feature extraction method, two feature extraction methods, PCA (principal component analysis) and CNN, are used to extract features from vibration signals. The extracted features are input to GRU for multi-fault diagnosis.
To avoid the influence of random factors, the final results are taken as the average of five rounds of random experiments. The results of five rounds of crossover experiments with different feature extraction methods are shown in
Table 8.
From the results in
Table 8, we can conclude that the average test accuracy and average verification accuracy of the proposed feature extraction method in the paper are the highest. The effectiveness of the proposed method is verified.
3.5. Comparison of Different Networks
To verify the effectiveness of the network model (GRU) used, two network models, LSTM and RNN, were used for comparison.
To verify the effectiveness of the proposed feature extraction method, two feature extraction methods, PCA (principal component analysis) and CNN, are used to extract features from vibration signals. The extracted features are input to GRU for multi-fault diagnosis.
To avoid the influence of random factors, the final results are taken as the average of five rounds of random experiments. The results of five rounds of crossover experiments with different networks are shown in
Table 9.
From the results in
Table 9, we can conclude that the network used in the article, GRU, has the highest average test accuracy and average verification accuracy. The network used in the paper is reasonable.
4. Conclusions
As the power source of important equipment, the normal operation of the engine is very important to ensure the safety of life and property. Therefore, we must use reasonable engine fault diagnosis [
28,
29] technology to diagnose the engine status.
Most of the previous methods only diagnose the fault type of the engine without further analyzing the fault severity of the engine fault type, which is not conducive to taking reasonable maintenance measures to ensure the safety of life and property.
To select the sensor signal with the most fault information and the time domain features, the variation coefficient of the cylinder head surface vibration signal is selected as the feature by using the double Pearson correlation coefficient and combined with GRU for fault diagnosis. The experimental results of single-fault diagnosis prove that the double Pearson coefficient can extract effective features. At the same time, the complexity of multi-fault diagnosis is higher than that of single-fault diagnosis, and more fault state information is required. Therefore, by stacking the variation coefficients of three vibration signals as inputs and sending them to the fault diagnosis network for fault diagnosis, the fault type and severity of the engine can be accurately diagnosed. The effectiveness of this method is verified, and the average test accuracy of the test set reaches 99.97%. The generalization ability of the method is also verified.