Cross-bearing migration diagnosis method based on deep separable transfer learning network
By adopting a deep-segmentable transfer learning network method in bearing fault diagnosis, using IJDA domain adaptation loss and I-Softmax classification loss, the problem of insufficient learning ability of segmentable characteristics under bearing signal noise pollution in the prior art is solved, and the accuracy and generalization ability of transfer diagnosis are improved.
Patent Information
- Application Number
- CN202211230003.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-08
- Publication Date
- 2025-03-18
- Estimated Expiration
- 2042-10-08
AI Technical Summary
The existing intelligent migration fault diagnosis method has poor feature learning ability when the bearing signal is contaminated by noise. It is limited to migration diagnosis of different loads, different speeds and different measurement points of the same machine. The distribution measurement distance does not take into account the distribution characteristics of the real signal, and there are a large number of approximate estimates, which affects the theoretical reliability.
A cross-bearing migration diagnosis method based on deep splittable transfer learning network (DDTLN) is adopted, and the bearing vibration signals are collected through acceleration sensors, and the sample data set is expanded using sliding sampling technology, and DDTLN based on IJDA domain adaptation loss and I-Softmax classification loss is constructed, and iterative update training is carried out to improve the accuracy and generalization ability of the migration model.
The accuracy and generalization ability of the intelligent migration diagnosis method are improved, and the problem of insufficient separable feature learning ability of the Softmax loss function in the case of noise pollution is solved. It is adapted to the Gaussian distribution characteristics of the bearing signal, which enhances the reliability of migration diagnosis.
Smart Images
Figure CN115479775B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and relates to a cross-bearing migration diagnosis method based on a deep separable transfer learning network. Background Art
[0002] As a basic general mechanical part, bearings are widely used in important fields such as rail transportation, energy and power, and aerospace. Once a bearing fails, it will cause unbearable economic losses. Therefore, in recent years, the intelligent fault diagnosis of bearings has attracted a large number of experts and scholars to conduct research. In particular, the intelligent migration fault diagnosis method of bearings represented by the transfer learning technology that considers data distribution differences has achieved great success. However, the existing intelligent migration fault diagnosis methods still have the following limitations: 1) The classification loss function used in the model is Softmax loss. Although this loss can achieve good results in most cases, its separable feature learning ability is poor when the bearing signal is heavily polluted by noise; 2) It is limited to the migration diagnosis task of different loads, different speeds and different measurement points of the same machine; 3) The selected distribution metric distance does not take into account the distribution characteristics of the real signal; 4) The existing transfer learning technology based on joint distribution alignment has a large number of approximate estimates, and its theoretical reliability cannot be guaranteed. Summary of the invention
[0003] In view of this, the purpose of the present invention is to provide a cross-bearing migration diagnosis method based on a deep separable transfer learning network to improve the accuracy and generalization ability of existing intelligent migration diagnosis methods.
[0004] In order to achieve the above object, the present invention provides the following technical solutions:
[0005] A cross-bearing migration diagnosis method based on a deep separable transfer learning network comprises the following steps:
[0006] S1: The original vibration signal of the bearing on the mechanical equipment is collected through the acceleration sensor to facilitate the construction of the subsequent migration diagnosis task. Then, the collected original vibration signal is used to expand the sample data set using the sliding sampling technology;
[0007] S2: Construct DDTLN based on IJDA domain adaptation loss and I-Softamx classification loss, where IJDA stands for improved joint distribution alignment and DDTLN stands for deep separable transfer learning network;
[0008] S3: Input the divided training samples into the DDTLN constructed in step S2, and use the I-Softamx classification loss function L of the labeled samples in the source domain y and the IJDA domain adaptation loss function L between the source domain and the target domain IJDAIteratively update and train the migration model;
[0009] S4: After multiple iterations of training, the error curve tends to be stable, and the migration model training is completed. The trained migration model will be used for cross-bearing fault migration diagnosis.
[0010] Further, in step S2, the DDTLN uses a convolutional neural network as the backbone network to adaptively extract features. The network consists of five "Conv1D" convolutional blocks, a global average pooling layer and two fully connected layers; each "Conv1D" convolutional block consists of a convolutional layer, a batch normalization layer and a maximum pooling layer; among them, the global average pooling layer and the batch normalization layer can accelerate the network convergence and effectively suppress the overfitting phenomenon of the network. In order to save computing resources, the IJDA domain adaptation loss is added to the last fully connected layer; similarly, the I-Softmax classification loss also acts on the last fully connected layer.
[0011] Furthermore, in step S2, the I-Softamx classification loss function L y The expression is:
[0012]
[0013] Among them, F i Represents the sample X i The output vector of the feature extractor; c and j are the corresponding feature vectors F i The element index of ; n represents the number of feature vectors, that is, the size of the current input batch; k ≥ 0 and m ≥ 1 are defined as two hyperparameters to control the decision boundary. At the same time, when m = 1 and k = 0, the I-Softmax classification loss function is equivalent to the original Softmax function.
[0014] Furthermore, in step S2, the IJDA domain adaptation loss function L IJDA The expression is:
[0015]
[0016] Among them, the distribution difference index DDM(A,B)=MMD(A,B)+CORAL(A,B), where A and B represent the feature representation of the source domain and the target domain respectively, MMD(·) and CORAL(·) are the maximum mean difference and covariance difference respectively; X S and X T Represent the feature samples of the source domain and the target domain respectively, Y S and Y TRepresent the labels of the feature samples in the source domain and the target domain respectively, T(·) is the feature representation, c and C represent the cth category and the number of categories respectively, P(·) represents the distribution probability, and (T(X)|Y=c) represents the feature vector of the cth category.
[0017] Further, in step S3, the IJDA domain adaptation loss function L IJDA Optimize the network parameters Θ by gradient descent fe , its gradient expression is:
[0018]
[0019] in, n S and n T Represents the sample batch size of the source domain and the target domain, F S and F T Represent the feature vectors of the source domain and the target domain respectively.
[0020] The beneficial effects of the present invention are as follows: the present invention improves the accuracy and generalization ability of the existing intelligent migration diagnosis method as a whole, which is specifically reflected in the following two aspects:
[0021] 1) The present invention designs an I-Softmax classification loss function to solve the problem of poor separability of the diagnostic model caused by the Softmax function commonly used in existing migration diagnosis methods.
[0022] 2) The present invention designs an improved joint distribution alignment (IJDA) mechanism and designs a new distribution difference metric (DDM) to implement the IJDA mechanism, so as to better adapt to the Gaussian distribution characteristics of mechanical vibration signals.
[0023] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0025] Figure 1 This is a schematic diagram of the structure of the Deep Separable Transfer Learning Network (DDTLN) proposed in the present invention;
[0026] Figure 2 is the probability density function of the four bearing health states;
[0027] Figure 3 It is the principle diagram of formula (2);
[0028] Figure 4 This is a schematic diagram explaining the principle of I-Softmax;
[0029] Figure 5 This is a schematic diagram of the CWRU test bench;
[0030] Figure 6 This is a schematic diagram of the RDS test bench;
[0031] Figure 7 Schematic diagram of the SWJTU test bench. DETAILED DESCRIPTION
[0032] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0033] See also Figure 1 to Figure 7 The present invention designs a cross-bearing migration diagnosis method based on a deep separable transfer learning network, which specifically includes the following steps:
[0034] Step 1: The original vibration signal of the bearing on the mechanical equipment is collected by the acceleration sensor to facilitate the construction of the subsequent migration diagnosis task. Then, the collected original vibration signal is used to expand the sample data set using the sliding sampling technology.
[0035] Step 2: Construct a deep separable transfer learning network (DDTLN) based on the IJDA mechanism and I-Softamx classification loss.
[0036] like Figure 1As shown in Figure 1, the DDTLN network uses a convolutional neural network as the backbone network to adaptively extract features. The detailed parameters are shown in Table 1. The network consists of five "Conv1D" convolution blocks, a global average pooling layer and two fully connected layers. Each "Conv1D" convolution block consists of a convolution layer, a batch normalization layer and a maximum pooling layer. Among them, the global average pooling layer and the batch normalization layer can accelerate the convergence of the network and effectively suppress the overfitting phenomenon of the network. In order to save computing resources, the IJDA loss proposed in the present invention is added to the last fully connected layer; similarly, the I-Softmax classification loss function also acts on the last fully connected layer. The IJDA mechanism and the I-Softmax classification loss function in the deep separable transfer learning network will be introduced in detail below.
[0037] Table 1 DDTLN network detailed parameters
[0038]
[0039] 1) I-Softmax classification loss
[0040] like Figure 2 As shown in Figure 2, the probability density functions of the original vibration signals from different bearing faults are usually overlapped and mixed due to noise interference. Figure 2 In the equation, NC, IF, BF, and OF represent normal bearing, inner ring fault, rolling element fault, and outer ring fault, respectively. Therefore, more separable classification features are required, but the traditional intelligent fault diagnosis method based on the Softmax function has limited separable feature learning ability and cannot meet the classification diagnosis of mixed data of multiple bearing faults. Therefore, based on the Softmax function, the present invention proposes a more separable classification function, namely the I-softmax classification loss function, which has two hyperparameters (m, k) to control its separable learning performance, as shown in the following formula (1):
[0041]
[0042] Among them, F i represents the output vector of this feature extractor (usually a convolutional neural network); c and j are the vectors corresponding to sample X i The label index of n represents the number of feature vectors, i.e., the size of the current input batch; k≥0 and m≥1 are defined as two hyperparameters to control the decision boundary. When m=1 and k=0, the I-Softmax classification loss function we invented is equivalent to the original Softmax function. Therefore, the I-Softmax function proposed in this invention is a more universal classification loss function than the original Softmax function.
[0043] In order to analyze the mechanism of I-Softmax function more clearly, we will analyze it from the perspective of angle. Figure 3 As shown, this feature vector can be decomposed into angle and amplitude through cosine similarity, as shown in the following formula (2):
[0044]
[0045] Among them, Z i Defined as the output vector of the second-to-last fully connected layer; W c is the cth weight vector from the last two fully connected layers, and θ c ∈[-π,π] is the vector W c and Z i The cosine angle between .
[0046] Next, the present invention analyzes I-Softmax using a binary classification example. Assuming a sample feature Z from class 1, the original Softmax must satisfy the following inequality in order to accurately classify the sample feature Z:
[0047]
[0048] However, the original Softmax still cannot have a strict decision boundary, which affects the intra-class compression and inter-class separability of feature learning. Therefore, in order to achieve a more satisfactory decision boundary and achieve this appropriate classification, for I-Softmax, the following formula must be satisfied:
[0049]
[0050] Obviously, the I-Softmax proposed in the present invention has a stricter decision boundary than the original Softmax. In essence, the superiority of I-Softmax over the original Softmax is still reflected in the angle. For example, it can be proved that:
[0051]
[0052] in, represents the equivalent edge angle, that is, k and m in the control equation (1) are equivalent to the control angle θ c .
[0053] like Figure 4 As shown, the present invention uses ||W1||=||W2|| and As an example, the effectiveness of I-Softmax is demonstrated intuitively. Also consider a feature vector Z belonging to class 1. In order to accurately classify, the original Softmax will force For I-Softmax, it will strictly force Therefore, it learns a more separable decision boundary.
[0054] In addition, I-Softmax does not require additional training techniques when learning features. It can optimize the network parameters Θ through stochastic gradient descent like the original Softmax. fe Define the I-Softmax output vector as Z, the one-hot encoding as Y, and the gradient of I-Softmax can be expressed as follows:
[0055]
[0056] In the formula for:
[0057]
[0058] 2) Improved joint distribution alignment (IJDA)
[0059] The general joint distribution alignment domain adaptation mechanism (Joint distribution alignment, JDA) is defined as consisting of marginal distribution alignment (Marginal distribution alignment, MDA) and class conditional distribution alignment (Class conditional distribution alignment, CCDA), as shown in formula (8):
[0060]
[0061] The first term on the right side of the equation represents MDA and the second term represents CCDA; X S and X T Represent the feature samples of the source domain and the target domain respectively, Y S and Y T Represent the labels of the source domain and target domain feature samples, T(·) is the feature representation, and E P (·) represents the expected value under P distribution, c and C represent the cth class and the number of categories, respectively.
[0062] There is an assumption in formula (8), that is, the class conditional distribution is equal to the conditional distribution, which will reduce the theoretical reliability and universality of JDA. Therefore, the present invention proposes an IJDA domain adaptation mechanism to improve the above assumption, as shown in the following formula:
[0063]
[0064] Formula (9) improves the assumptions in formula (8) through Bayesian theory, namely:
[0065]
[0066] At the same time, in order to better measure the marginal distribution difference (the first term on the right side of the equation) and the conditional distribution difference (the second term on the right side of the equation) in equation (9), and to better adapt to the Gaussian distribution characteristics of the mechanical vibration signal, the present invention defines a new distribution discrepancy metric (DDM), as shown below:
[0067] DDM(A,B)=MMD(A,B)+CORAL(A,B) (11)
[0068] Among them, A and B represent the feature representation of the source domain and the target domain respectively, MMD(·) and CORAL(·) are the maximum mean difference and covariance difference respectively, and the specific form is as follows:
[0069]
[0070]
[0071] In formula (11), n S and n T Represent the sample batch sizes of the source domain and the target domain respectively, and represents the source domain and the target domain, ||·|| H represents the regenerated Hilbert space, Φ(·) represents the mapping function in the regenerated Hilbert space. In formula (13), d is the sample dimension size, Cov S and Cov T Represent the covariance matrix of source domain and target domain samples respectively, and their formulas are as follows;
[0072]
[0073] Here, I represents a row vector whose elements are all 1.
[0074] Combined with the DDM distribution distance, the IJDA domain adaptation mechanism expressed in equation (9) can be restated as:
[0075]
[0076] The above IJDA domain adaptation loss function can optimize the network parameters Θ through gradient descent fe , whose gradient is shown as follows:
[0077]
[0078]
[0079] by and For example, the detailed calculation formula is as follows:
[0080]
[0081]
[0082] Step 3: Input the divided training samples into the constructed deep separable transfer learning network (DDTLN) and use the source domain labeled sample classification loss function L y (Equation (1)) and the IJDA domain adaptation loss function L between the source domain and the target domain IJDA (Equation (15)) Iteratively update the migration model for training.
[0083] Step 4: After multiple iterations of training, the error curve tends to be stable, the model training is completed, and the trained migration model will be used for cross-bearing fault migration diagnosis.
[0084] The above is the process of the cross-bearing migration diagnosis method based on the deep separable transfer learning network proposed in the present invention. The following experimental results have demonstrated the effectiveness of this method.
[0085] Verification experiment: The original vibration signals of the bearings collected in this experiment come from three test benches, namely the CWRU standard bearing dataset published by Case Western Reserve University in the United States. The schematic diagram of the CWRU bearing dataset is shown in the figure. Figure 5 As shown in Figure 1, it consists of a motor, bearings at both ends of the motor, a torque sensor and a power meter; and a self-tested RDS bearing data set. The schematic diagram of the RDS test bench is shown in Figure 1. Figure 6 As shown in Figure 1, it consists of a servo motor, a coupling, a rotor, and bearings and bearing seats at both ends. The SWJTU bearing dataset from Southwest Jiaotong University, the schematic diagram of the test bench is shown in Figure 1. Figure 7 As shown in Figure 2, it consists of a motor, a support bearing, a faulty bearing used for testing, and a loading system. Each bearing dataset includes information such as its fault type, fault size, and speed as shown in Table 2. With the above three bearing datasets, we can establish six cross-bearing migration diagnosis tasks, namely A→B, B→A, A→C, C→A, C→B, and B→C.
[0086] Table 2 Detailed information of three bearing datasets
[0087]
[0088]
[0089] Comparative experiment:
[0090] In order to demonstrate the superiority of the migration diagnosis method based on the deep separable transfer learning network (DDTLN) of the present invention, the current typical marginal distribution alignment network models (DDC, DCORAL, DANN) and the classic joint distribution alignment (JDA) are compared, and the experimental results are shown in Table 3.
[0091] Table 3 Experimental results
[0092]
[0093] From Table 3, we can conclude that the deep separable transfer learning network proposed in this invention has higher transfer diagnosis accuracy and stronger generalization ability. In particular, the average diagnostic accuracy of DDTLN reaches more than 90%, which is about 30% higher than other transfer diagnosis methods.
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A cross-bearing migration diagnosis method based on a deep separable transfer learning network, characterized in that: The method comprises the following steps: S1: Collect the original vibration signals of the bearings on the mechanical equipment, and then use the sliding sampling technology to expand the sample data set; S2: Construct DDTLN based on IJDA domain adaptation loss and I-Softamx classification loss, where IJDA stands for improved joint distribution alignment and DDTLN stands for deep separable transfer learning network; I-Softamx classification loss function L y The expression is: Among them, F i Represents the sample X i The output vector of the feature extractor; c and j are the corresponding feature vectors F i The element index of ; n represents the number of feature vectors, that is, the size of the current input batch; k ≥ 0 and m ≥ 1 are defined as two hyperparameters to control the decision boundary; IJDA domain adaptation loss function L IJDA The expression is: Among them, the distribution difference index DDM(A,B)=MMD(A,B)+CORAL(A,B), where A and B represent the feature representation of the source domain and the target domain respectively, MMD(·) and CORAL(·) are the maximum mean difference and covariance difference respectively; X S and X T Represent the feature samples of the source domain and the target domain respectively, Y S and Y T Represent the labels of the source domain and target domain feature samples respectively, T() is the feature representation, c and C represent the cth class and the number of categories respectively, P(·) represents the distribution probability, and (T(X)Y=c) represents the feature vector of the cth class; S3: Input the divided training samples into the DDTLN constructed in step S2, and use the I-Softamx classification loss function L of the labeled samples in the source domain y and the IJDA domain adaptation loss function L between the source domain and the target domain IJDA Iteratively update and train the migration model; S4: After multiple iterations of training, the error curve tends to be stable, and the migration model training is completed. The trained migration model will be used for cross-bearing fault migration diagnosis.
2. The cross-bearing migration diagnosis method according to claim 1, characterized in that: In step S2, the DDTLN uses a convolutional neural network as the backbone network to adaptively extract features. The network consists of five "Conv1D" convolutional blocks, a global average pooling layer, and two fully connected layers; each "Conv1D" convolutional block consists of a convolutional layer, a batch normalization layer, and a maximum pooling layer; the IJDA domain adaptation loss is added to the last fully connected layer; Similarly, the I-Softmax classification loss is also applied to the last fully connected layer.
3. The cross-bearing migration diagnosis method according to claim 1, characterized in that: In step S3, the IJDA domain adaptation loss function L IJDA Optimize the network parameters Θ by gradient descent fe , its gradient expression is: Among them, n S and n T Represents the sample batch size of the source domain and the target domain, F S and F T Represent the feature vectors of the source domain and the target domain respectively, and represents the source domain and the target domain, Φ(·) represents the mapping function in the regenerated Hilbert space, d is the sample dimension size, Cov S and Cov T Represent the covariance matrices of source domain and target domain samples respectively, and I represents a row vector whose elements are all 1.