A Bearing Fault Diagnosis Method and System Based on Multi-Source Domain Deep Transfer Learning

Abstract

The present invention relates to the technical field of fault diagnosis, and discloses a bearing fault diagnosis method and system based on multi-source domain deep transfer learning, including: collecting vibration acceleration data signals in typical fault states of bearing components under different working conditions; building the backbone network of the diagnosis framework, and during training, calculating the feature distribution distance between each source domain and the target domain and constructing a multi-source domain loss function, and setting a multi-source domain loss dynamic adjustment mechanism according to the Wasserstein distance; obtaining unlabeled signal data of the target domain to be diagnosed, preprocessing it and then inputting it into the trained diagnosis model, outputting the diagnosis result for saving and visually displaying the features. By reducing the attention of the model to the classification effects of multiple source domains, the influence of negative transfer is effectively reduced, providing support for further fault handling and ensuring that the system's fault response is more accurate and safe.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and specifically to a bearing fault diagnosis method and system based on multi-source domain deep transfer learning. Background Art

[0002] In today's highly automated living environment, the safe and stable operation of rotating components such as bearings is crucial for maintaining the continuity and efficiency of mechanical equipment. With the rise of the new energy vehicle industry, the maintenance and fault diagnosis of its electromechanical transmission system have attracted more and more attention and favor. Currently, the new energy vehicle industry has placed the testing work of key components at the core. Among them, the evaluation of safety performance and its possible degradation are particularly highly regarded by automobile manufacturers. Compared with the traditional transmission system of a transmission, a reducer is known for its streamlined components, simple structure, and no need for gear shifting. For electric vehicles, the operating characteristics of an electric motor are very different from those of a conventional internal combustion engine, showing characteristics of high-speed operation and high torque output. At the same time, the operating environment where the reducer is located is relatively harsh. Once a fault occurs in the reducer, it will not only cause economic losses but also pose potential safety hazards.

[0003] However, as an important part of the electromechanical transmission system, the predictive detection and fault diagnosis of the operating state of bearings play an important role in improving the safety and reliability of vehicles. Bearing intelligent fault diagnosis technology, with its high efficiency and accuracy, is becoming a powerful tool to solve this problem. It can not only detect equipment faults in a timely manner but also provide a scientific basis for the preventive maintenance of equipment, and has important social value and application prospects. In recent years, more and more scholars and researchers have widely applied the deep transfer learning theory to the field of intelligent diagnosis. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are: in the training process of the existing fault diagnosis model based on multi-source domain deep transfer learning, the problem that the losses of each source domain do not dynamically allocate weights as the model parameters are updated; in addition, the negative transfer problem caused by the large proportion of the multi-source domain classification loss in the total loss in the later stage of model training.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a bearing fault diagnosis method based on multi-source domain deep transfer learning, including:

[0007] Collect the vibration acceleration signals of the bearing in typical fault states by installing vibration acceleration sensors;

[0008] Preprocess the obtained source domain data;

[0009] Construct an initialized TCN feature extractor and a joint classifier for each source domain to form a backbone network;

[0010] Set an independent joint classifier C for each source domain i and add a Droupout module to it;

[0011] During training, calculate the feature distribution distance between each source domain and the target domain and construct a multi-source domain loss function. Set a multi-source domain loss dynamic adjustment mechanism according to the Wasserstein distance to align the feature distributions of the target domain and multiple source domains;

[0012] Obtain the unlabeled signal data of the target domain to be diagnosed. After preprocessing, input it into the trained diagnostic model, save the output diagnostic results and visually display the features;

[0013] Among them, C i represents the joint classifier of the i-th source domain.

[0014] As a preferred solution of the bearing fault diagnosis method based on multi-source domain deep transfer learning described in the present invention, wherein: the vibration acceleration signals in the typical fault states include the vibration acceleration signals in four typical fault states of normal bearings, inner ring fault bearings, rolling element fault bearings and outer ring fault bearings;

[0015] Construct a multi-source domain training data set with corresponding fault category labels and a target domain training data set without fault labels according to the vibration acceleration signals in the typical fault states;

[0016] The multi-source domain training data set includes taking each working condition collected as a source domain and collecting source domains of multiple working conditions to form a data set for training;

[0017] The target domain training data set includes the collected data set without fault labels.

[0018] As a preferred solution of the bearing fault diagnosis method based on multi-source domain deep transfer learning described in the present invention, wherein: the preprocessing includes intercepting the signal, with an intercept window size of 1024, a step size of 128, and a batch size batch of 128;

[0019] Perform mean-std normalization processing on the data.

[0020] As a preferred solution of the bearing fault diagnosis method based on multi-source domain deep transfer learning according to the present invention, wherein: the obtained multi-source domain data and target domain data are respectively input into the constructed TCN feature extractor in batches, and the deep features of each source domain and target domain are obtained and mapped into a renewable Hilbert space, and the feature distribution differences between the two source domains and the target domain are respectively calculated:

[0021]

[0022] where n si and n T are the total number of samples in the source domain si and the target domain T, Z Si represents the sample set of the i-th source domain, Z T represents the target domain sample set, z i and z j respectively represent samples in the source domain and target domain sample sets, φ() represents the mapping transformation function, and H K is the multi-kernel reproducing kernel Hilbert space;

[0023] The deep features of the bearings in the source domain and target domain extracted by the TCN feature extractor are input into the constructed joint classifier, and the predicted labels of each source domain and target domain are obtained simultaneously.

[0024] As a preferred solution of the bearing fault diagnosis method based on multi-source domain deep transfer learning according to the present invention, wherein: the TCN feature extractor includes a feature extraction based on a temporal convolutional network;

[0025] The calculation formula of the classification loss of the joint classifier of each source domain can be expressed as:

[0026]

[0027] where si represents the i-th source domain, and n i represents the total number of samples in the i-th source domain, and respectively represent the predicted outputs of the fully connected classifier C linear and the LSTM classifier C lstm for the j-th sample in the i-th source domain, represents the true label of the j-th sample in the i-th source domain;

[0028] An independent classifier C i is set for each source domain, and a Droupout module is added thereto;

[0029] Meanwhile, the diagnostic accuracy acc i of each source domain is calculated through the source domain labels predicted by the joint classifier of each source domain.

[0030] As a preferred solution of the bearing fault diagnosis method based on multi-source domain deep transfer learning according to the present invention, wherein: the multi-source domain loss dynamic adjustment mechanism aligns the feature distributions of the target domain and multiple source domains in the form of weight coefficients;

[0031] The weight coefficient formulas for the feature alignment losses of different source domains are as follows:

[0032]

[0033] wherein, Π(P Si , P T ) represents the set of all possible joint distribution sets, γ represents one in the joint distribution set, E (x,y)～γ represents the mathematical expectation under the joint distribution, P si and P T are the probability distributions of the predicted labels of the i-th source domain and the target domain respectively, is the Wasserstein distance between the i-th source domain si and the target domain T, k i represents the weight coefficient of the difference loss of the i-th source domain, N is the number of source domains, and ε is a real number that is infinitesimal and non-zero;

[0034] The feature alignment loss between each source domain and the target domain consists of two parts: the minimum class confusion loss and the multi-kernel maximum mean discrepancy.

[0035] As a preferred solution of the bearing fault diagnosis method based on multi-source domain deep transfer learning according to the present invention, wherein: a negative transfer suppression mechanism is introduced to reduce its attention to the source domain classification loss;

[0036] The negative transfer suppression mechanism includes two parts: the lowest accuracy setting and the weakening of the classification loss; a reduction factor is used to reduce the proportion of the classification loss of each source domain in the total loss, and at the same time, the focus is placed on the feature transfer alignment loss;

[0037] When calculating the total loss Loss i of the i-th source domain, the specific expressions of the two mechanisms are:

[0038]

[0039] wherein, represents the minimum class confusion loss, acc si represents the model diagnosis accuracy of the i-th source domain in the current iteration, a represents the multi-source domain lowest diagnosis accuracy value set before training, b represents the reduction factor, and is used to reduce the order of magnitude of the classification loss of each source domain ;

[0040] The total loss Loss updated in the reverse direction and the final optimization update process are expressed as:

[0041]

[0042] Among them, θ F and θ Ci represent the learning parameters of the feature extractor and the i-th source domain joint classifier, and are the optimal values obtained through training. η F and η C are the learning rates of the feature extractor and the joint classifier respectively. η F = 0.5η C = 0.001.

[0043] A bearing fault diagnosis system based on multi-source domain deep transfer learning adopting the method as described in any one of the present invention, characterized in that: a sensor acquisition module, a data preprocessing module, a model training and storage module, a target domain diagnosis and fault warning module, and a visualization presentation and response prompt module;

[0044] The sensor acquisition module is used to acquire 1D acceleration vibration signals during the operation of rotating components;

[0045] The data preprocessing module uses a program to intercept and batch normalize the acquired data;

[0046] The model training and storage module is used to provide support for model training and store the model learning parameters meeting the training requirements into the diagnosis module;

[0047] The target domain diagnosis and fault warning module provides application support for the trained model, provides reliable diagnosis results for the unlabeled data in the target domain, and gives fault warnings to the operators for the convenience of users' observation and analysis;

[0048] The visualization presentation and response prompt module presents the deeply extracted features to the user in a visual form for the convenience of subsequent analysis and maintenance.

[0049] A computer device, comprising: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, the steps of the method as described in any one of the present invention are implemented.

[0050] A computer-readable storage medium, on which a computer program is stored, wherein: when the computer program is executed by a processor, the steps of the method as described in any one of the present invention are implemented.

[0051] Advantages of the present invention: The bearing fault diagnosis method based on multi-source domain deep transfer learning provided by the present invention adopts a new combined classifier structure, which combines a classifier based on a long short-term memory network (LSTM) and a fully connected layer classifier in parallel, and incorporates an enhanced temporal convolutional network (TCN) feature extractor. This design significantly improves the ability of the network structure to identify fault features in time series vibration signals; a novel multi-source domain loss dynamic adjustment mechanism is proposed. This mechanism uses the Wasserstein distance between the target domain and different source domains as weights to dynamically adjust the contribution degrees of the feature alignment losses of each source domain, so as to obtain a diagnostic model that better fits the fault feature distribution of the target domain data; a negative transfer suppression mechanism is proposed. In the later stage of model training, this mechanism effectively reduces the impact of negative transfer by reducing the attention of the model to the classification effects of multiple source domains, and finally constructs a diagnostic framework that focuses more on the transfer effect of the target domain; relying on hardware facilities and functional modules, the diagnostic prediction and visualization presentation of unlabeled data in the target domain are established, so as to provide support for further fault handling and ensure that the system's fault response is more accurate and safe. The present invention achieves better results in terms of safety, accuracy and applicability. Description of the Drawings

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings according to these drawings without creative efforts.

[0053] Figure 1 It is a schematic diagram of a target-oriented multi-source domain deep transfer learning method provided by the first embodiment of the present invention.

[0054] Figure 2 It is a training flowchart of a target-oriented multi-source domain deep transfer learning method provided by the first embodiment of the present invention.

[0055] Figure 3 It is a visualization diagram of the transfer of source domains H and I to target domain J in a target-oriented multi-source domain deep transfer learning method provided by the first embodiment of the present invention;

[0056] Figure 4 It is a visualization diagram of the transfer of source domains I and J to target domain H in a target-oriented multi-source domain deep transfer learning method provided by the first embodiment of the present invention;

[0057] Figure 5A visualization graph showing the transfer of source domains H and J to target domain I in a target - oriented multi - source domain deep transfer learning method provided by the first embodiment of the present invention;

[0058] Figure 6 A graph showing the total loss and accuracy changes during the training process in a target - oriented multi - source domain deep transfer learning method provided by the first embodiment of the present invention

[0059] Figure 7 A schematic diagram of the working process of a target - oriented multi - source domain deep transfer learning system provided by the second embodiment of the present invention. Detailed implementation manners

[0060] To make the above - mentioned objects, features and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] Example 1, referring to Figures 1-6 , which is an embodiment of the present invention. This example applies a bearing fault diagnosis method based on multi - source domain deep transfer learning and is implemented based on a target - oriented multi - source domain deep transfer learning method, including:

[0062] Step 1: Collect multi - source domains with corresponding fault category labels of bearings under a variety of operating conditions ( Figure 2 operating conditions 1 and 2 in it) by installing vibration acceleration sensors, and obtain target domain data without fault labels under other operating conditions different from the known operating conditions ( Figure 2 operating condition 3 in it). Among them, the operating condition refers to information such as the rotational speed change and load when the bearing is running (shown in Table 1). Since the collection method is the same, the variable - speed bearing dataset of the University of Ottawa in Canada is directly used here.

[0063] Table 1 Example bearing data parameter table

[0064]

[0065]

[0066] Furthermore, the collected vibration acceleration signals of the bearing include vibration acceleration signals in four typical fault states: normal bearing (NC), inner - race fault bearing (IF), rolling - element fault bearing (BF), and outer - race fault bearing (OF), and the labels are assigned as 0, 1, 2, and 3 respectively.

[0067] Step 2: Intercept and preprocess the source domain data obtained in Step 1, as shown in the training flow chart ( Figure 2 ).

[0068] Construct a multi-source domain training data set with corresponding fault category labels and a target domain training data set without fault labels based on the vibration acceleration signals in the typical fault states; the multi-source domain training data set includes, in the already collected data sets (such as CWRU, PU, etc.), each working condition represents a source domain, and multiple working conditions represent multiple source domains. The target domain training data set includes the collected unlabeled data set, which is used for further fault prediction and diagnosis, and then to detect and solve actual faults.

[0069] Among them, the data of source domain 1, source domain 2, and the target domain need to be processed into training sample sets respectively. The window size for intercepting all data samples in the source domain and the target domain is 1024, the step size is 128, and the batch size batch is 128. In addition, usually, the data is processed by "mean-std" standardization. In this example, in order to reflect the "end-to-end" diagnosis effect, 1D sample data is directly intercepted for training and recognition.

[0070] Step 3: As Figure 2 shown, construct an initialized TCN feature extractor network and a joint classifier network for each source domain to form a backbone network, and the parameters of each layer of its model are shown in Table 1:

[0071] Table 1 Parameter table of the backbone network model

[0072]

[0073]

[0074] Furthermore, input the obtained multi-source domain data and target domain data into the constructed TCN feature extractor batch by batch to obtain the deep features of each source domain and the target domain and map them into a renewable Hilbert space. When the features are mapped into the space, the fault feature distributions are different between different domains due to the large working condition span. Therefore, by continuously reducing the difference distance between the fault feature distributions of different domains, the transfer generalization performance of the training model is realized. The present invention uses the MK-MMD distance to calculate the difference and calculates the difference distance of the feature distributions between the i-th source domain and the target domain

[0075]

[0076] where, n si and n T are the total numbers of samples in the source domain si and the target domain T, Z Si represents the sample set of the i-th source domain, Z TDenote the target domain sample set as \(z\). i and \(z\) j represent samples in the source domain and target domain sample sets respectively, \(\varphi()\) represents the mapping transformation function, and \(H\) K is a multi - kernel reproducing kernel Hilbert space.

[0077] Furthermore, input the deep features of the bearings in the source domain and target domain extracted by the TCN feature extractor into the constructed joint classifier, and obtain the predicted labels of each source domain and target domain simultaneously. The joint classifier \(C\) i includes a classifier with a fully - connected layer network and a long - short - term memory network in parallel as shown in Table 1. Average the classification prediction outputs of the two networks and sum them with weights to obtain the prediction result of the joint classifier.

[0078] Step 4: To preserve more multi - domain feature information, set an independent classifier \(C\) i for each source domain to increase the robustness of the network. And add a Dropout module in it, which not only speeds up the convergence of the loss function but also reduces overfitting. The calculation formula of its classification loss can be expressed as:

[0079]

[0080] where \(s_i\) represents the \(i\) - th source domain, \(n\) i represents the total number of samples in the \(i\) - th source domain, and represent the predicted labels of the \(j\) - th sample in the \(i\) - th source domain by the fully - connected classifier \(C\) linear and the LSTM classifier \(C\) lstm respectively, and \(y_{ij}\) represents the true label of the \(j\) - th sample in the \(i\) - th source domain.

[0081] Furthermore, calculate the diagnostic accuracy rate \(acc\) i of each source domain through the source domain labels predicted by the joint classifier of each source domain. This accuracy rate will play the role of a precision threshold in the following negative transfer suppression mechanism.

[0082] Step 5: The Wasserstein distance between each source domain and the target domain actually reflects the minimum "workload" required to transfer all the mass of the target domain distribution to a source domain distribution. Here, the workload is calculated by the distance of moving a unit weight. In other words, it is by finding the lowest - cost transportation plan that can make the two distributions consistent. Therefore, the Wasserstein distance between each source domain and the target domain reflects the difference metric value between the target domain and each source domain at the output layer of the domain joint classifier \(C\) i This metric value is used to align the feature distributions of the target domain and multiple source domains through the multi - source domain loss dynamic adjustment mechanism in the form of a weight coefficient.

[0083] The weight coefficient formula for the cross - domain feature alignment loss is as follows:

[0084]

[0085] Among them, Π(P Si , P T ) represents the set of all possible joint distributions, γ represents one in the joint distribution set, E (x,y)～γ represents the mathematical expectation under the joint distribution, P si and P T are the probability distributions of the predicted labels of the i - th source domain and the target domain respectively, is the Wasserstein distance between the i - th source domain si and the target domain T, k i represents the weight coefficient of the difference loss of the i - th source domain, N is the number of source domains, and ε is a non - zero real number that is infinitesimal.

[0086] Furthermore, the feature alignment loss between each source domain and the target domain consists of two parts: the minimum class confusion (MCC) loss and the multi - kernel maximum mean discrepancy (MK - MMD). Among them, the multi - kernel maximum mean discrepancy (MK - MMD) has been obtained in step 3. The class confusion loss uses information entropy as the metric, and the calculation process is as follows:

[0087]

[0088]

[0089] Among them, |C| is the number of data classes, Y' ij represents the probability that the i - th sample belongs to class j, H(y' i ) represents the information entropy metric of the i - th row of samples in a batch, the symbol y j ' and y j' ' represent the predicted probability outputs of the classifier for each batch of samples belonging to classes j and j' respectively, B is the batch size, T is used as an abbreviation for the matrix transpose operation, and the class correlation C jj' reflects the tendency of the classifier to classify samples as class j and class j' simultaneously. W represents the diagonal matrix composed of the weights of a batch of samples, and W ii is the probability that measures the importance of the i - th sample for class confusion evaluation.

[0090] Step 6: Introduce a negative transfer suppression mechanism in the later stage of model training to reduce its attention to the source - domain classification loss, and thus pay more attention to the improvement of the transfer effect on the target domain.

[0091] The negative transfer suppression mechanism mainly consists of two parts: minimum accuracy setting and classification loss reduction. That is, the reduction factor is used to reduce the proportion of classification loss of each source domain in the total loss, while focusing on feature transfer alignment loss.

[0092] Furthermore, when calculating the i-th source domain loss, the specific expressions of the above two mechanisms are as shown:

[0093]

[0094] in, represents the minimum class confusion loss, acc si represents the model diagnosis accuracy of the i-th source domain in the current iteration, a represents the minimum diagnosis accuracy value of the multi-source domain set before training, and b represents the reduction factor, which is used to reduce the classification loss of each source domain. The total loss Loss of the reverse update and the final optimization update process are expressed as:

[0095]

[0096]

[0097] Among them, θ F and represents the learning parameters of the feature extractor and the i-th source domain joint classifier, and is the best value obtained through training, η F and η C are the learning rates of the feature extractor and the joint classifier, respectively, F =0.5η C =0.001.

[0098] Step 7. Finally, the training model is obtained, and the processed unlabeled data of the target domain is input into the diagnosis model. The fault status category of the target domain sample to be tested is obtained through the integration of multiple classifier results and a fault warning is issued to achieve the classification goal. Figure 3 and Figure 6 As shown in Figure 2, the labeled data of the source domain H,I is preprocessed and input into the training model. After iterative learning, the most suitable migration model for the target domain J is obtained. Then, the target domain data is input into the trained model to obtain the fault feature visualization diagram of the target domain J ( Figure 3 ), thereby obtaining a clear classification boundary and realizing fault classification and diagnosis in the unlabeled target domain. Figure 4 and Figure 5 The principle is the same as above, except that the source domain and target domain are different.

[0099] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0100] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device.

[0101] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or processing it in other suitable ways if necessary, and then storing it in a computer memory.

[0102] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0103] Embodiment 2, referring to Figure 7 , is the second embodiment of the present invention, which provides a framework and system for the proposed target-oriented multi-source domain deep transfer learning method, including a sensor acquisition module, a data preprocessing module, a model training and storage module, a target domain diagnosis and fault warning module, and a visualization presentation and response prompt module.

[0104] Among them, the sensor acquisition module includes a vibration acceleration sensor and an acquisition device, which are used to acquire the vibration acceleration data of rotating components in different states. The data preprocessing module is used to intercept and batch normalize the acquired 1D signal data to obtain multiple source domain training data sets with fault labels and an unlabeled target domain data set respectively. The model training and storage module inputs the training data into the training model of the target-oriented multi-source domain deep transfer learning method proposed by the present invention according to the batch size and batches respectively, and stores the trained model in the target domain diagnosis hardware after training. The target domain diagnosis and fault warning module is used to diagnose the target domain test data, then output the corresponding prediction labels, and display the fault warning. The visualization presentation and response prompt module is used to introduce the T-SNE feature visualization method and the accumulated knowledge of fault handling experience for response prompts, helping operators quickly respond to and solve fault problems and facilitating subsequent further work.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A bearing fault diagnosis method based on multi-source domain deep transfer learning, characterized in that, Including: By installing vibration acceleration sensors, collect the vibration acceleration signals of the bearing under typical fault conditions; Preprocess the obtained source domain data; Construct an initialized TCN feature extractor and a joint classifier for each source domain to form a backbone network; The TCN feature extractor includes feature extraction based on a temporal convolutional network; The calculation formula for the classification loss of the joint classifier for each source domain can be expressed as: Among them, si represents the i-th source domain, and n i represents the total number of samples in the i-th source domain, and respectively represent the prediction outputs of the fully connected classifier C linear and the LSTM classifier C lstm for the j-th sample in the i-th source domain, represents the true label of the j-th sample in the i-th source domain; Set up an independent classifier C for each source domain i , and add a Dropout module to it; Meanwhile, the diagnostic accuracy acc of each source domain is calculated by the source domain labels predicted by the joint classifier of each source domain i ; During training, calculate the feature distribution distance between each source domain and the target domain and construct a multi-source domain loss function. Set a multi-source domain loss dynamic adjustment mechanism according to the Wasserstein distance to align the feature distributions of the target domain and multiple source domains; Obtain the unlabeled signal data of the target domain to be diagnosed. After preprocessing, input it into the trained diagnostic model, save the output diagnostic results, and visually display the features; Among them, C i represents the joint classifier of the i-th source domain.

2. The bearing fault diagnosis method based on multi-source domain deep transfer learning according to claim 1, characterized in that: The vibration acceleration signals under the typical fault conditions include the vibration acceleration signals of a normal bearing, an inner race fault bearing, a rolling element fault bearing, and an outer race fault bearing in four typical fault conditions; According to the vibration acceleration signals under the typical fault conditions, construct a multi-source domain training dataset with corresponding fault category labels and a target domain training dataset without fault labels; The multi-source domain training dataset includes taking each working condition collected as a source domain, and collecting source domains of multiple working conditions to form a training dataset; The target domain training dataset includes the collected dataset without fault labels.

3. The bearing fault diagnosis method based on multi-source domain deep transfer learning according to claim 2, wherein: The preprocessing includes intercepting the signal with an intercept window size of 1024, a step size of 128, and a batch size batch of 128; Perform mean-std normalization processing on the data.

4. The bearing fault diagnosis method based on multi-source domain deep transfer learning according to claim 3, characterized in that: Input the obtained multi-source domain data and target domain data into the constructed TCN feature extractor batch by batch to obtain the deep features of each source domain and the target domain and map them into a reproducible Hilbert space, and calculate the feature distribution differences between the two source domains and the target domain respectively: where n si and n T are the total numbers of samples in the source domain si and the target domain T, Z Si denotes the sample set of the i-th source domain, Z T denotes the target domain sample set, z i and z j represent the samples in the source domain and target domain sample sets respectively, φ() represents the mapping transformation function, H K is the multi-kernel reproducing kernel Hilbert space; Input the deep features of the bearings in the source domain and the target domain extracted by the TCN feature extractor into the constructed joint classifier to obtain the predicted labels of each source domain and the target domain at the same time.

5. The bearing fault diagnosis method based on multi-source domain deep transfer learning according to claim 4, characterized in that: The multi-source domain loss dynamic adjustment mechanism aligns the feature distributions of the target domain and multiple source domains in the form of weight coefficients; The weight coefficient formula for the feature alignment loss of different source domains is as follows: Among them, Π(P Si , P T ) represents the set of all possible joint distributions, γ represents one in the set of joint distributions, E (x,y)～γ represents the mathematical expectation under the joint distribution, P si and P T are the probability distributions of the predicted labels of the i-th source domain and the target domain respectively, is the Wasserstein distance between the i-th source domain si and the target domain T, k i represents the weight coefficient of the difference loss of the i-th source domain, N is the number of source domains, and ε is a real number that is infinitesimal and non-zero; The feature alignment loss between each source domain and the target domain consists of two parts: the minimum class confusion loss and the multi-kernel maximum mean difference.

6. The bearing fault diagnosis method based on multi-source domain deep transfer learning according to claim 5, characterized in that: Introduce a negative transfer suppression mechanism to reduce its attention to the source domain classification loss; The negative transfer suppression mechanism consists of two parts: the lowest accuracy setting and the weakening of the classification loss; use a reduction factor to reduce the proportion of the classification loss of each source domain in the total loss, and at the same time focus on the feature transfer alignment loss; When calculating the total loss Loss of the i-th source domain i the specific expressions of the two mechanisms are as follows: Among them, represents the minimum class confusion loss, and acc si represents the model diagnosis accuracy of the i-th source domain in the current iteration, a represents the multi-source domain minimum diagnosis accuracy value set before training, and b represents the reduction factor used to reduce the classification loss of each source domain order of magnitude; The backward updated total loss Loss and the final optimization update process are expressed as: where, θ F and represent the learning parameters of the feature extractor and the i-th source domain joint classifier, and are the best values obtained through training, η F and η C are the learning rates of the feature extractor and the joint classifier respectively, η F = 0.5η C = 0.

001.

7. A bearing fault diagnosis system based on multi-source domain deep transfer learning using the method according to any one of claims 1-6, characterized in that: A sensor acquisition module, a data preprocessing module, a model training and storage module, a target domain diagnosis and fault warning module, and a visualization presentation and response prompt module; The sensor acquisition module is used to collect 1D acceleration vibration signals during the operation of rotating components; The data preprocessing module uses a program to intercept and batch-normalize the collected data; The model training and storage module provides support for model training and stores the model learning parameters that meet the training requirements in the diagnosis module; The target domain diagnosis and fault warning module provides application support for the trained model, provides reliable diagnosis results for the unlabeled data in the target domain, and gives fault warnings to the operator for convenient observation and analysis by the user; The visualization presentation and response prompt module presents the deeply extracted features to the user in a visual form for convenient subsequent analysis and maintenance.

8. A computer device, comprising: A memory and a processor; The memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the method according to any one of claims 1-6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the method according to any one of claims 1-6 are implemented.

Images