A two-stage integrated fault diagnosis model based on deep sparse autoencoder

Abstract

The application discloses a two-stage integrated fault diagnosis model based on deep sparse auto-encoder, which comprises the following steps: 1) first, different types of activation functions are adopted to train sparse auto-encoders with different learning behaviors, and then the sparse auto-encoders are stacked to form multiple different deep sparse auto-encoders; 2) the deep sparse auto-encoders are integrated in two stages, the method is that a boosting integration strategy is adopted to obtain multiple integrated models based on deep sparse auto-encoders, and then a voting method is used to integrate the multiple integrated models to obtain a final two-stage fault diagnosis model based on deep sparse auto-encoders. The application utilizes different activation functions to reflect the learning behaviors of the sparse auto-encoders, then a hierarchical pre-training method adopted by the deep sparse auto-encoders can learn more useful features in the original data, and finally a two-stage integration is adopted to fully utilize the advantages of the integrated learning and the deep learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis and computer artificial intelligence, and particularly relates to a two-stage integrated fault diagnosis model based on a deep sparse autoencoder. BACKGROUND

[0002] Rolling bearings, as important components of rotating machinery and electrical equipment, are known as the "industrial joint", and in order to ensure the production efficiency and safe and reliable operation of the equipment, the state monitoring of the rolling bearings is very important. Since the vibration signal of the rolling bearing when running is rich in information reflecting the running state, a large number of sensors are usually arranged in some large industrial equipment such as an aero-engine and a gas turbine to obtain the monitoring data of the rolling bearing, but how to obtain the running information of the rolling bearing from a large number of vibration signals and discover faults early has become a new problem faced by the field of bearing fault diagnosis.

[0003] In recent years, various signal processing techniques such as time domain analysis, frequency domain analysis, wavelet transform, empirical mode decomposition, etc. have been widely studied and applied for diagnosing bearing faults by extracting signal features such as mean value, peak-to-peak value, root mean square value, etc. from original vibration signals.

[0004] However, the fault feature extraction method based on signal processing technology requires the staff to master and be proficient in certain signal processing knowledge, and at the same time, since it needs to be operated manually, a large amount of manpower is needed to extract and analyze the distinguishing fault features. Since it is manually selected, it is difficult to determine the most suitable features to extract, resulting in unstable diagnosis results. In addition, the number of parameters of the existing features is huge, which can easily cause the problem of model overfitting. Therefore, how to design a method to automatically learn rich fault features from the vibration signal of the rolling bearing is particularly important.

[0005] In addition, the existing fault diagnosis model is a single deep learning model, and since the vibration signal of the vibration signal has the characteristics of mass, diversity and complexity, the normal samples and the fault samples are usually unbalanced, which leads to low fault prediction generalization performance. SUMMARY

[0006] The purpose of the present application is to provide a two-stage integrated fault diagnosis model based on a deep sparse autoencoder, which uses different activation functions to reflect the learning behavior of the sparse autoencoder, and a hierarchical pre-training method used by the deep sparse autoencoder can learn more useful features in the original data. Finally, the two-stage integrated fault diagnosis model can avoid the problems of manual feature extraction, overfitting and low generalization performance of a single deep learning model in the prior art.

[0007] The above technical purposes of the present application are achieved by the following technical solutions:

[0008] A two-stage integrated fault diagnosis model based on deep sparse autoencoder, comprising the following steps:

[0009] 1) First, different types of activation functions are used to train sparse autoencoders with different learning behaviors, and then the sparse autoencoders are stacked into multiple different deep sparse autoencoders;

[0010] 2) Then, the deep sparse autoencoders are integrated in two stages. The method is to obtain multiple integrated models based on deep sparse autoencoders by using the boosting integration strategy, and then to obtain the final two-stage fault diagnosis model based on deep sparse autoencoders by using the voting method to integrate multiple integrated models.

[0011] Further, the different types of activation functions include exponential type activation functions and non-exponential type activation functions.

[0012] Further, the sparse autoencoder uses a regularized cross-entropy loss function to ensure that the sparse autoencoder can learn different transformations.

[0013] Further, the specific process of step 1) is as follows:

[0014] S1. Use exponential type activation functions and non-exponential type activation functions as the activation functions of the sparse autoencoder, use regularized cross-entropy to construct the loss function of the sparse autoencoder, and train the sparse autoencoder based on bearing fault data;

[0015] S2. Use the same type of sparse autoencoder to construct several sparse autoencoders with a depth of 3 layers, and set the highest layer of each deep sparse autoencoder as a softmax layer for fault classification; train each deep sparse autoencoder using the original bearing vibration signal, and finally obtain deep sparse autoencoders of exponential type and non-exponential type;

[0016] S3. Select a fault classification accuracy threshold ε∈[0.80, 0.95], check the performance of each deep sparse autoencoder on the test set, and form a set S of all deep sparse autoencoders higher than the threshold ε; calculate the error rate e of each deep sparse autoencoder in the set S on the test set, arrange the deep sparse autoencoders in ascending order of error rate, and form an ordered set R;

[0017] S4. Select the first classifier f1 from the ordered set R, and use The weight of the training sample is modified, the weight of the sample with classification error is modified to d times of the original, and the weight of the sample with correct classification is modified to 1 / d times of the original; to ensure that a deep sparse automatic encoder with complementary learning behavior can be learned, a different classifier f2 from f1 is selected from the set S, and f2 is trained with the modified training sample; the above steps are repeated T times to obtain T deep sparse automatic encoders, and the weight of each classifier

[0018] Further, the specific process of step 2) is as follows:

[0019] S5. Deep sparse automatic encoder-based boosting integrated model Wherein, t = 1, 2, …, T represents the number of different deep sparse automatic encoders generated in the integrated learning process, and T represents the total number of deep sparse automatic encoders;

[0020] S6. Select the classifier with the second lowest error rate on the test set from the ordered set R as the first classifier, repeat steps 4 and 5 to obtain multiple different deep sparse automatic encoder-based boosting integrated models;

[0021] S7. The voting method is used to integrate multiple different deep sparse automatic encoder-based boosting integrated models to form the final two-stage deep sparse automatic encoder-based integrated model.

[0022] In summary, the present application has the following beneficial effects:

[0023] 1) The present application adopts a deep sparse automatic encoder-based boosting integrated model, which avoids the problem of unstable diagnostic results caused by manual feature extraction in the prior art.

[0024] 2) The present application avoids the overfitting problem of the prior art by using a sparse automatic encoder.

[0025] 3) The two-stage integrated deep learning model generated by the present application avoids the problem of low generalization performance of a single deep learning model in the case of massive, complex and high-dimensional original fault signals. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a schematic diagram of the sparse automatic encoder described in the present application.

[0027] Figure 2 is a schematic diagram of different types of deep sparse automatic encoders described in the present application.

[0028] Figure 3 is a two-stage integrated model based on deep sparse automatic encoders described in the present application. DETAILED DESCRIPTION

[0029] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application is further described below in combination with the drawings and specific embodiments.

[0030] As shown in Figure 1 , Figure 2 and Figure 3 , the present application proposes a two-stage integrated fault diagnosis model based on deep sparse autoencoder, including the following steps:

[0031] 1) First, different types of activation functions are used to train sparse autoencoders with different learning behaviors, and then the sparse autoencoders are stacked into multiple different deep sparse autoencoders;

[0032] 2) Then, the deep sparse autoencoders are integrated in two stages. The method is to first use the boosting integration strategy to obtain multiple integrated models based on deep sparse autoencoders, and then use the voting method to integrate multiple integrated models to obtain the final two-stage fault diagnosis model based on deep sparse autoencoders.

[0033] The specific process of step 1) is as follows:

[0034] S1. Use exponential type activation function and non-exponential type activation function as the activation function of sparse autoencoder, use regularized cross entropy to construct the loss function of sparse autoencoder, and train the sparse autoencoder based on bearing fault data;

[0035] S2. Use the same type of sparse autoencoder to construct several sparse autoencoders with a depth of 3 layers, and set the highest layer of each deep sparse autoencoder as a softmax layer for fault classification; train each deep sparse autoencoder with the original bearing vibration signal to obtain the deep sparse autoencoders of exponential type and non-exponential type;

[0036] S3. Select a fault classification accuracy threshold ε∈[0.80, 0.95], check the performance of each deep sparse autoencoder on the test set, and form a set S of all deep sparse autoencoders higher than the threshold ε; calculate the error rate e of each deep sparse autoencoder in the set S on the test set, arrange the deep sparse autoencoders in ascending order of error rate to form an ordered set R;

[0037] S4. Select the first classifier f1 from the ordered set R, and use The weight of the training sample is modified, the weight of the sample with classification error is modified to d times of the original, and the weight of the sample with correct classification is modified to 1 / d times of the original; in order to ensure that the deep sparse autoencoder with complementary learning behavior can be learned, a different classifier f2 from f1 is selected from the set S, and f2 is trained with the modified training sample; the above steps are repeated T times to obtain T deep sparse autoencoders, and the weight of each classifier

[0038] The specific process of the step 2) is as follows:

[0039] S5. Deep sparse autoencoder-based boosting integrated model Wherein, t = 1, 2, …, T represents the number of different deep sparse autoencoders generated in the integrated learning process, and T represents the total number of deep sparse autoencoders;

[0040] S6. The classifier with the second lowest error rate on the test set is selected from the ordered set R as the first classifier, and the steps 4 and 5 are repeated to obtain multiple different deep sparse autoencoder-based boosting integrated models;

[0041] S7. The multiple different deep sparse autoencoder-based boosting integrated models are integrated by using the voting method to form the final deep sparse autoencoder-based two-stage integrated model.

[0042] Figure 1 It is explained that the standard sparse autoencoder. According to the result of Figure 1 The deep sparse autoencoder can be constructed by using the hierarchical pre-training method, and Figure 2 According to the result of Figure 2 The two-stage integrated method can be used to obtain the deep sparse autoencoder-based two-stage integrated model.

[0043] The present application utilizes different activation functions to reflect the learning behavior of the sparse autoencoder, then the hierarchical pre-training method used in the construction of the deep sparse autoencoder can learn more useful features in the original data, and finally the two-stage integration fully utilizes the advantages of integrated learning and deep learning.

[0044] The present application can be applied to the predictive maintenance of manufacturing enterprises, which can help enterprises understand and solve the causes of equipment failure and the performance differences between equipment and other invisible problems, help enterprises to determine uncertain information, save costs and improve efficiency.

[0045] In this document, the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", "vertical", "horizontal", and the like, are used as terms of convenience to describe orientations or positional relationships based on the orientations or positional relationships shown in the figures, and are not to be construed as limiting the present application.

[0046] In this document, the terms "comprise", "comprising", or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0047] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A two-stage integrated fault diagnosis model based on a deep sparse autoencoder, characterized in that, Includes the following steps: 1) First, sparse autoencoders with different learning behaviors are trained using different types of activation functions. Then, the sparse autoencoders are stacked to build multiple different deep sparse autoencoders. The specific process of step 1) is as follows: S1. Use exponential and non-exponential activation functions as activation functions for the sparse autoencoder, construct the loss function of the sparse autoencoder using regularized cross-entropy, and train the sparse autoencoder based on bearing fault data. S2. Construct several sparse autoencoders with a depth of 3 layers using sparse autoencoders of the same type, and set the highest layer of each sparse autoencoder as a softmax layer to achieve fault classification; train each sparse autoencoder with the original bearing vibration signal to finally obtain exponential and non-exponential sparse autoencoders. S3. Select a fault classification accuracy threshold ε∈[0.80,0.95], check the performance of each deep sparse autoencoder on the test set, and form a set S of all deep sparse autoencoders that are higher than the threshold ε; calculate the error rate e of each deep sparse autoencoder in set S on the test set, and sort the deep sparse autoencoders in ascending order of error rate to form an ordered set R. S4. Select the first classifier f1 from the ordered set R, and use... Modify the weights of the training samples: increase the weights of misclassified samples by d times and the weights of correctly classified samples by 1 / d times. To ensure the learning of deep sparse autoencoders with complementary learning behaviors, randomly select a classifier f2 from set S that is different from f1, and train f2 using the modified training samples. Repeat the above steps T times to obtain T deep sparse autoencoders, each with its own weights. 2) Then, a two-stage integration of the deep sparse autoencoder is performed. The method is to first use the boosting integration strategy to obtain multiple integrated models based on the deep sparse autoencoder, and then use the voting method to integrate the multiple integrated models to obtain the final two-stage fault diagnosis model based on the deep sparse autoencoder.

2. The two-stage integrated fault diagnosis model based on a deep sparse autoencoder according to claim 1, characterized in that, The different types of activation functions include exponential activation functions and non-exponential activation functions.

3. The two-stage integrated fault diagnosis model based on a deep sparse autoencoder according to claim 1, characterized in that, The sparse autoencoder employs a regularized cross-entropy loss function to ensure that the sparse autoencoder can learn different transformations.

4. The two-stage integrated fault diagnosis model based on a deep sparse autoencoder according to claim 1, characterized in that, The specific process of step 2) is as follows: S5. Boosting Ensemble Model Based on Deep Sparse Autoencoder Where t = 1, 2, ..., T represents the number of different depth sparse autoencoders generated iteratively during the ensemble learning process, and T represents the total number of depth sparse autoencoders; S6. Select the classifier with the second lowest error rate on the test set from the ordered set R as the first classifier, and repeat steps 4 and 5 to obtain multiple different boosting ensemble models based on deep sparse autoencoders. S7. Multiple different boosting ensemble models based on deep sparse autoencoders are integrated using a voting method to form the final two-stage ensemble model based on deep sparse autoencoders.

Images