A fault diagnosis method of decoupling inter-data causal representation and causal strength embedding

Abstract

The application relates to the technical field of fault diagnosis, in particular to a fault diagnosis method for decoupling and embedding of causal representation and causal strength between data. First, bearing historical fault data of a multi-source domain collected by multiple data collectors is acquired; then, the bearing historical fault data is subjected to decoupling processing of causal representation, and causal features and acausal features of each data collector are decoupled; then, all the causal features are used for embedding processing of causal strength to obtain comprehensive causal features; finally, the comprehensive causal features are subjected to fault prediction to predict corresponding bearing fault types. Thus, the application can effectively improve the precision of cross-domain bearing fault diagnosis by separating the causal features related to faults from the acausal features of environmental interference and fusing the causal information of multiple data collectors.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically a fault diagnosis method that decouples causal representations between data and embeds causal strength. Background Technology

[0002] As a core transmission component in industrial equipment (such as motors, fans, and machine tools), the operating status of bearings directly affects the safety and reliability of the equipment. During the operation of industrial equipment, bearing fault diagnosis is a crucial step in ensuring safe and stable operation and reducing maintenance costs. With the development of data acquisition technology and the Internet of Things (IoT), monitoring data from multiple sources and data acquisition devices in industrial scenarios (such as bearing vibration signals and bearing temperature signals) are widely collected, providing a rich data foundation for bearing fault diagnosis. However, in practical applications, the operating environment of equipment is complex and variable (such as different loads, speeds, and operating conditions). This results in the collected data containing not only features directly related to bearing faults (such as bearing fault type, size, and location) but also a large amount of domain-related non-causal interference information (such as environmental noise and differences in operating conditions). This interference information severely affects the generalization ability of bearing fault diagnosis models, especially in cross-domain scenarios (i.e., training data and test data come from different operating conditions or equipment), where diagnostic accuracy often drops significantly.

[0003] Traditional bearing fault diagnosis methods mainly rely on data-driven machine learning or deep learning techniques, such as extracting features and performing classification based on models like convolutional neural networks (CNNs) and recurrent neural networks (RNNs). However, these methods have the following limitations:

[0004] 1. Traditional models typically learn the overall features of the data directly, failing to effectively separate fault-related causal features from environment-related non-causal features. When the test domain differs from the training domain, changes in non-causal features can lead to model misjudgments, reducing diagnostic accuracy.

[0005] 2. In industrial scenarios, multiple data collectors are often used for collaborative monitoring. However, existing fusion methods do not consider the causal dependencies between data collectors, which leads to irrelevant information being mixed into the fused features, affecting the diagnostic accuracy.

[0006] 3. The performance of traditional models is highly dependent on the consistency of the distribution of training and testing data. When faced with new working conditions or equipment, the model is difficult to adapt quickly and requires a large amount of labeled data to be retrained. However, high-quality labeled data is often scarce in industrial scenarios, resulting in low diagnostic accuracy of the retrained model, which is difficult to meet the actual production applications in industry.

[0007] Therefore, it is evident that traditional bearing fault diagnosis methods still need further improvement at this stage. Summary of the Invention

[0008] To avoid and overcome the technical problems existing in the prior art, this invention provides a fault diagnosis method based on the decoupling of causal representations between data and the embedding of causal strength. By separating fault-related causal features from non-causal features of environmental interference and fusing causal information from multiple data acquisition devices, this invention can effectively improve the accuracy of cross-domain bearing fault diagnosis.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A fault diagnosis method for decoupling causal representations and embedding causal strength between data includes the following steps:

[0011] S1. Acquire historical bearing fault data from multiple data acquisition devices across multiple data sources;

[0012] S2. Perform causal characterization and decoupling processing on the historical bearing fault data to decouple the causal and non-causal features of each data acquisition device.

[0013] S3. Perform causal strength embedding on causal and non-causal features to obtain comprehensive causal features;

[0014] S4. Perform fault prediction based on comprehensive causal characteristics to predict the corresponding bearing fault type.

[0015] As a further aspect of the present invention, the sub-steps of step S1 are as follows:

[0016] S11. Select the type of bearing to be monitored and the industrial equipment using the bearing, and clarify the operating conditions of the bearing;

[0017] S12. Multiple data acquisition devices are sequentially arranged at various predetermined positions on the bearing;

[0018] S13. Under each working condition, the bearings with different fault types are tested. The original signals are collected synchronously by the data acquisition device. A set number of samples are collected for each type of bearing under each working condition. The original signals of each group of samples are continuous time-domain signals, and all the original signals under a single working condition constitute a source domain.

[0019] S14. Label the collected raw signals to obtain bearing historical fault data. The label content includes fault-related information and domain-related information. The fault-related information includes fault type, fault size and fault location. The domain-related information includes operating conditions and data acquisition device number.

[0020] As a further aspect of the present invention, the sub-steps of step S2 are as follows:

[0021] S21. Establish a decoupled model that includes a domain encoder, a causal encoder, a decoder, and a classifier. The domain encoder is used to extract non-causal features containing domain-related information, the causal encoder is used to extract causal features containing fault-related information, the decoder is used to reconstruct the signal, and the classifier is used to output fault classification results based on causal features.

[0022] S22. Use historical bearing fault data to jointly train the domain encoder, causal encoder, decoder and classifier in the decoupled model, optimize the model parameters through a multi-objective loss function, and stop training when the set conditions are met.

[0023] S23. Using the trained decoupled model, the historical bearing fault data is processed by the causal encoder and decoder respectively, and the causal and non-causal features of each data acquisition device are decoupled.

[0024] As a further aspect of the present invention, the training process of the decoupling model is as follows:

[0025] S221. Input the original bearing signal into the domain encoder and the causal encoder to obtain the non-causal features and causal features respectively, and then concatenate them to obtain the concatenated vector;

[0026] S222. Input the spliced ​​vector into the decoder to obtain the reconstructed signal. Calculate the mean square error between the reconstructed signal and the original bearing signal and use it as the reconstruction loss. Optimize the domain encoder parameters and decoder through the reconstruction loss.

[0027] S223. Input the causal features into the classifier to obtain the fault classification probability, calculate the cross-entropy loss between the fault classification probability and the true fault label and use it as the classification loss, and optimize the classifier parameters and causal encoder parameters through the classification loss.

[0028] S224. Construct causal aggregation loss and redundancy reduction loss based on causal and non-causal features. Update the parameters of all components of the decoupled model by combining all losses into a multi-objective loss function until the set conditions are met and training stops.

[0029] As a further aspect of the present invention, the sub-steps of step S3 are as follows:

[0030] S31. Determine the number of data collectors participating in the causal strength embedding. H and obtain this H Causal characteristics of a data acquisition device;

[0031] S32. Utilize the original signals from multiple source domains to discover causality and construct a one-way causal strength matrix. B One-way causal strength matrix B It is an upper triangular matrix, where the matrix elements are... Indicates the first The causal characteristics of the first collector are related to the first... The causal strength of the causal characteristics of each collector;

[0032] S33. Calculate the weighting matrix W Weighted matrix W identity matrix E With the one-way causal strength matrix B The sum of, i.e. W = E + B ;

[0033] S34, Based on weighted matrix W right H The causal features of each data collector are weighted and fused, and the comprehensive causal features are calculated using a formula. Z :

[0034] ;

[0035] In the formula, Represents the natural constant; Indicates the first Causal characteristics of a data collector, subscript It serves as a marker of causal characteristics.

[0036] As a further aspect of the present invention, the reconstruction loss is expressed as follows:

[0037] ;

[0038] In the formula, Indicates the losses incurred during reconstruction; Indicates the total number of source domains; Indicates the first Number of samples in each source domain ; express The average number of samples in each source domain; Indicates the signal dimension of the input sample; Indicates the first In the source domain, the first The sample number, ; Indicates the first In the source domain, the first The original signal corresponding to each sample; Indicates decoder pair The reconstructed signal output after reconstruction; express L The square of the 2-norm.

[0039] As a further aspect of the present invention, the classification loss is expressed as follows:

[0040] ;

[0041] In the formula, Indicates classification loss; Indicates the first In the source domain, the first The true fault label of each sample; This represents the output of the classifier. The corresponding fault classification probability; It represents the natural logarithm.

[0042] As a further aspect of the present invention, the causal aggregation loss is expressed as follows:

[0043] ;

[0044] In the formula, Indicates loss due to causal aggregation; Indicates loss of causal characteristics; Indicates loss of non-causal characteristics;

[0045] ;

[0046] In the formula, Indicates the first One source domain, Indicates the first One source domain, ; Indicates the first Number of samples in each source domain Indicates the first The number of samples in each source domain; Indicates the first In the source domain, the first The sample number, ; Indicates the first In the source domain, the first The sample number, ; Indicates the first In the source domain, the first The true fault label of each sample Indicates the first In the source domain, the first The true fault label of each sample Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates from the first The first source domain The causal features extracted from each sample are in vector form; express The transpose of ; Indicates from the first The first source domain The causal features extracted from each sample are in vector form;

[0047] ;

[0048] In the formula, Indicates the first In the source domain, the first The sample number, Indicates the first In the source domain, the first The sample number, ; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form, with subscripts... It is a non-causal characteristic marker; for The transpose of ; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form; Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form.

[0049] As a further aspect of the present invention, the redundancy reduction loss is expressed as follows:

[0050] ;

[0051] In the formula, This indicates redundancy to reduce losses; Represents a matrix of all 1s; It represents the Hadamardi (or Hadama) stack; Represents the causal characteristic matrix; This represents the transpose of the causal characteristic matrix; Represents a non-causal feature matrix; Represents the transpose of the noncausal characteristic matrix; Denotes the square of the F-norm; Dimensions representing causal and non-causal characteristics.

[0052] As a further aspect of the present invention, the multi-objective loss function is expressed as follows:

[0053] ;

[0054] In the formula, Indicates multi-target loss; Indicates the reconstruction loss coefficient; Represents the classification loss coefficient; Indicates the causal aggregation loss coefficient; This represents the redundancy reduction loss coefficient.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] By acquiring historical bearing fault data from multiple source domains, causal features (including fault-related information) and non-causal features (including domain-related information) from each data acquisition device are separated through causal characterization decoupling processing. Then, causal strength embedding processing is applied to all causal features to obtain comprehensive causal features, thereby achieving fault prediction. Its advantages are: by decoupling causal and non-causal features, interference from non-causal factors such as operating conditions and data acquisition device differences can be reduced; by fusing multi-source causal features through causal strength embedding, the effective information from each data acquisition device can be fully utilized; simultaneously, combining the decoder to reconstruct the signal ensures the integrity of the feature learning information, and the causal strength matrix can be updated with new data, enabling the method to maintain good diagnostic performance in cross-domain scenarios. This effectively improves the accuracy of fault diagnosis, the effect of multi-source data fusion, cross-domain adaptability, and the integrity of feature learning. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention.

[0058] Figure 2 This is a structural diagram of the decoupling model in this invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see Figure 1 In this embodiment of the invention, a fault diagnosis method for decoupling causal representation between data and embedding causal strength includes the following:

[0061] I. Obtaining historical bearing fault data from multiple source domains

[0062] This is used to collect raw bearing signals under different operating conditions and fault types, and to label them to provide a data foundation for subsequent model training.

[0063] 1. Determine the monitoring targets and operating conditions.

[0064] The monitoring object was selected as a 6205 deep groove ball bearing, and the corresponding industrial equipment was a 1.5kW three-phase asynchronous motor. The bearing's operating conditions are shown in Table 1.

[0065] Table 1 Bearing Operating Conditions

[0066] ;

[0067] By combining the different operating parameters in Table 1 in pairs, a total of 9 operating conditions can be obtained: 3 speeds × 3 loads, for a total of 9 operating conditions.

[0068] 2. Deployment of data acquisition devices

[0069] Three vibration sensors are placed at predetermined locations on the bearing as data acquisition units. The specific locations are as follows:

[0070] Sensor 1: Bearing radial direction, collects horizontally extending radial vibration signals;

[0071] Sensor 2: Axial bearing, collecting axial vibration signals extending horizontally;

[0072] Sensor 3: Motor end cover, for collecting comprehensive vibration signals;

[0073] All vibration sensors acquire signals synchronously via a data acquisition card to ensure consistent timestamps.

[0074] 3. Raw signal acquisition

[0075] For each operating condition, the bearings with the following four failure types were tested.

[0076] Fault type 1: No fault, i.e., normal state;

[0077] Fault type 2: Inner ring crack fault, with crack widths of 0.1mm, 0.3mm and 0.5mm respectively;

[0078] Fault type 3: Outer ring crack fault, with crack widths of 0.1mm, 0.3mm and 0.5mm respectively;

[0079] Fault type 4: Roller crack failure, with crack widths of 0.1mm, 0.3mm and 0.5mm respectively;

[0080] For each fault type, 200 samples of each crack width were collected under each operating condition. Each sample was a continuous time-domain signal, lasting 1 second, containing 10,000 data points. All samples under a single operating condition constituted a source domain, for a total of 9 source domains.

[0081] 4. Labeling

[0082] The collected raw signals are labeled to form historical bearing fault data.

[0083] Tags contain two types of information:

[0084] (1) Fault-related information: The fault types are no fault, inner ring crack, outer ring crack and roller crack, the crack sizes are 0.1mm, 0.3mm and 0.5mm respectively, and the crack locations are inner ring, outer ring and roller respectively.

[0085] (2) Domain-related information: operating conditions, speed + load, such as "1500r / min + 1.0kW"; vibration sensor number, 1, 2 and 3.

[0086] II. Decoupling of Causal Representations

[0087] By constructing a decoupling model, causal features related to failures and non-causal features related to the domain are decoupled from historical bearing failure data.

[0088] 1. Construct a decoupling model

[0089] like Figure 2 As shown, the decoupling model comprises four core components, as follows:

[0090] Domain encoder E d The system employs a Convolutional Neural Network (CNN) with three convolutional layers, each containing a 3×3 kernel. The kernel counts are 32, 64, and 128, respectively, with a stride of 1. Each convolutional layer is followed by a ReLU activation function and a 2×2 max-pooling layer. This is used to extract non-causal features containing domain-related information. Z d .

[0091] Causal encoder E c Structure and Domain Encoders E d Completely identical; used to extract causal features containing fault-related information. Z c .

[0092] Decoder D: Employs a deconvolutional neural network symmetrical to the encoder, containing 3 deconvolutional layers; Input Z d and Z cThe concatenated vector is used to output the reconstructed signal, which is used to recover the features of the original signal.

[0093] Classifier: Employs a multilayer perceptron (MLP) with two fully connected layers (128→64→4, where "4" corresponds to four fault types), and the output layer uses the SoftMax activation function; based on causal features. Zc Output the fault classification probability to determine the fault type.

[0094] 2. Feature extraction and splicing

[0095] Input the bearing's historical fault data into the domain encoder simultaneously. E d and causal encoder E c Non-causal features with dimension 64 were obtained respectively. Z d Causal features with dimension 64 Z c The two are then concatenated (cancatted) into a 128-dimensional concatenated vector.

[0096] 3. Calculate reconstruction losses

[0097] The concatenated vector is input into the decoder to obtain the reconstructed signal; the mean square error between the reconstructed signal and the original signal is calculated as the reconstruction loss, as shown in the following formula:

[0098] (1);

[0099] In the formula, Indicates the losses incurred during reconstruction;

[0100] This represents the total number of source domains, which is 9 in this embodiment.

[0101] Indicates the first The number of samples in each source domain is the same in this embodiment, which is 2400.

[0102] This represents the signal dimension of the input sample; in this embodiment, the value is 10000.

[0103] Indicates the first In the source domain, the first The sample number, ;

[0104] Indicates the first In the source domain, the first The original signal corresponding to each sample;

[0105] Indicates decoder pair The reconstructed signal output after reconstruction;

[0106] express L The square of the 2-norm.

[0107] 4. Calculate classification loss

[0108] causal characteristics Z c Input the data into the classifier to obtain the fault classification probability; calculate the cross-entropy between the classification probability and the true fault label as the classification loss, as shown in the following formula:

[0109] (2);

[0110] In the formula, Indicates classification loss; Indicates the first In the source domain, the first The true fault label of each sample; This represents the output of the classifier. The corresponding fault classification probability; It represents the natural logarithm.

[0111] 5. Calculate the causal aggregation loss

[0112] To enhance the consistency of causal features for the same fault type, while also increasing the differences in non-causal features across different domains, a causal aggregation loss is introduced, as shown in the following formula:

[0113] (3);

[0114] In the formula, Indicates loss due to causal aggregation; Indicates loss of causal characteristics; Indicates loss of non-causal characteristics;

[0115] (4);

[0116] In the formula, Indicates the first One source domain, Indicates the first One source domain, ; Indicates the first Number of samples in each source domain Indicates the first The number of samples in each source domain; Indicates the first In the source domain, the first The sample number, ; Indicates the first In the source domain, the first The sample number, ; Indicates the first In the source domain, the first The true fault label of each sample Indicates the first In the source domain, the first The true fault label of each sample Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates from the first The first source domain The causal features extracted from each sample are in vector form; express The transpose of ; Indicates from the first The first source domain The causal features extracted from each sample are in vector form.

[0117] (5);

[0118] In the formula, Indicates the first In the source domain, the first The sample number, Indicates the first In the source domain, the first The sample number, ; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form, with subscripts... It is a non-causal characteristic marker; for The transpose of ; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form; Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form.

[0119] 6. Calculate redundancy to reduce losses

[0120] To eliminate redundancy within causal features, non-causal features, and between the two, a redundancy reduction loss is introduced, as shown in the following formula:

[0121] (6);

[0122] In the formula, This indicates redundancy to reduce losses;

[0123] This represents a matrix of all ones with dimensions of 64×64, consistent with the feature dimension.

[0124] It represents the Hadamardi (or Hadama) stack;

[0125] This represents the identity matrix, with dimensions 64×64;

[0126] This represents the causal feature matrix, with the number of rows equal to the number of samples and 64 columns.

[0127] This represents the transpose of the causal characteristic matrix;

[0128] This represents a non-causal feature matrix with the number of rows equal to the number of samples and 64 columns.

[0129] Represents the transpose of the noncausal characteristic matrix;

[0130] Denotes the square of the F-norm;

[0131] The dimension representing causal and non-causal characteristics has a value of 64.

[0132] 7. Multi-objective loss function

[0133] The above losses are combined to form a multi-objective loss function, the formula of which is:

[0134] (7);

[0135] In the formula, Indicates multi-target loss; This represents the reconstruction loss coefficient, which is set to 1.0 in this embodiment. This represents the classification loss coefficient, which is 0.5 in this embodiment. This represents the causal aggregation loss coefficient, which is 0.3 in this embodiment. This represents the redundancy reduction loss coefficient, which is set to 0.2 in this embodiment.

[0136] 8. Training termination conditions

[0137] The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and 200 iterations. Training was stopped when the validation set loss did not decrease for 10 consecutive iterations.

[0138] 9. Decoupling of causal and non-causal characteristics

[0139] Using the trained decoupling model, input the historical bearing fault data. E c and E d Output respectively:

[0140] Causal characteristics Z c Each causal feature has 64 dimensions, containing core information such as fault type and size.

[0141] Non-causal characteristics Z d Each non-causal feature has 64 dimensions, including domain information such as operating conditions and sensor number.

[0142] III. Causal Strength Embedding Process

[0143] By constructing causal strength relationships among multiple sensors, the causal features obtained from decoupling are weighted and fused to obtain comprehensive causal features.

[0144] 1. Determine the number of data acquisition devices and causal characteristics.

[0145] This embodiment uses three vibration sensors, namely H =3.

[0146] Obtain its corresponding causal features: Causal features of sensor 1 Z c1 Causal characteristics of sensor 2 Z c2 Causal characteristics of sensor 3 Z c3 .

[0147] 2. Construct a one-way causal strength matrix B

[0148] Based on historical bearing fault data from multiple sources, the classic PC algorithm for causal discovery is used to mine causal relationships, and an upper triangular matrix is ​​constructed. B 3x3, with elements Indicates the first The causal characteristics of the first collector are related to the first... The causal strength of the causal features of each collector.

[0149] ;

[0150] in: , representing the causal strength of the causal characteristics of sensor 1 on the causal characteristics of sensor 2. , representing the causal strength of the causal characteristics of sensor 1 on the causal characteristics of sensor 3. , representing the causal strength of the causal characteristics of sensor 2 on the causal characteristics of sensor 3.

[0151] 3. Calculate the weighting matrix W

[0152] Weighted matrix W identity matrix E With the one-way causal strength matrix B The sum of, i.e. W = E + B :

[0153] .

[0154] 4. Calculate the comprehensive causal characteristics

[0155] Based on weighted matrix W The causal features of the three vibration sensors are weighted and fused using the following formula:

[0156] (8);

[0157] In the formula, Represents the natural constant; Indicates the first Causal characteristics of a data collector The values ​​are 1, 2, and 3.

[0158] IV. Fault Prediction

[0159] The obtained comprehensive causal characteristics Z Input the trained classifier and output the fault classification probability. Take the category with the highest probability as the final bearing fault type.

[0160] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis decoupling inter-data causal representation from causal strength embedding, characterized in that, The method comprises the following steps: S1, acquiring bearing historical fault data of a multi-source domain collected by a plurality of data collectors; S2, performing causal representation decoupling processing on the bearing historical fault data, and decoupling causal features and acausal features of each data collector; The sub-steps of step S2 are as follows: S21, establishing a decoupling model comprising a domain encoder, a causal encoder, a decoder and a classifier, wherein the domain encoder is used to extract acausal features containing domain-related information, the causal encoder is used to extract causal features containing fault-related information, the decoder is used to reconstruct a signal, and the classifier is used to output a fault classification result based on the causal features; S22, jointly training the domain encoder, the causal encoder, the decoder and the classifier in the decoupling model using the bearing historical fault data, optimizing model parameters through a multi-objective loss function, and stopping training when a set condition is reached; the multi-objective loss function is represented as follows: ； wherein, denotes a multi-target loss; denotes a reconstruction loss coefficient, denotes a reconstruction loss; denotes a classification loss coefficient, denotes a classification loss; denotes a causal-aggregation loss coefficient, denotes a causal-aggregation loss; denotes a redundancy reduction loss coefficient, denotes a redundancy reduction loss; S23, using the trained decoupling model to process the bearing historical fault data through the causal encoder and the decoder respectively, and decoupling the causal features and the acausal features of each data collector; S3, performing causal strength embedding processing on the causal features and the acausal features to obtain comprehensive causal features; S4, performing fault prediction on the comprehensive causal features to predict a corresponding bearing fault type.

2. The method of claim 1, wherein, The sub-steps of step S1 are as follows: S11, selecting a bearing type to be monitored and an industrial equipment using the bearing, and specifying an operating condition of the bearing; S12, sequentially arranging a plurality of data collectors at each predetermined position of the bearing; S13, under each operating condition, performing operation tests on bearings of different fault types, and synchronously collecting original signals through the data collectors; a set number of samples are collected for each fault type of bearing under each operating condition, and the original signal of each sample is a continuous time-domain signal, and all original signals under a single operating condition constitute a source domain; S14, performing label annotation on the collected original signals to obtain bearing historical fault data; the label annotation includes fault-related information and domain-related information, the fault-related information includes a fault type, a fault size and a fault position, and the domain-related information includes an operating condition and a data collector number.

3. The method of claim 2, wherein, The training process of the decoupling model is as follows: S221, inputting the original signal into the domain encoder and the causal encoder to obtain acausal features and causal features respectively, and splicing the acausal features and the causal features to obtain a splicing vector; S222, inputting the splicing vector into the decoder to obtain a reconstructed signal, calculating a mean square error between the reconstructed signal and the original signal, and taking the mean square error as a reconstruction loss to optimize the domain encoder parameters and the decoder; S223, inputting the causal features into the classifier to obtain a fault classification probability, calculating a cross-entropy loss between the fault classification probability and a real fault label, and taking the cross-entropy loss as a classification loss to optimize the classifier parameters and the causal encoder parameters; S224, constructing a causal aggregation loss and a redundancy reduction loss based on the causal features and the acausal features, updating the parameters of all components of the decoupling model through a multi-objective loss function formed by combining all losses, and stopping training when a set condition is reached.

4. The method of claim 3, wherein, The sub-steps of step S3 are as follows: S31, determining the number of data collectors participating in the causal strength embedding is H and obtaining the causal features of the H data collectors; S32, using the original signals of the multi-source domain to make causal discovery and construct a one-way causal strength matrix B , the one-way causal strength matrix B is an upper triangular matrix, and the matrix elements in the matrix represent the causal strength of the causal feature of the i th data collector to the causal feature of the j th data collector; S33, compute the weighting matrix W , the weighting matrix W is the identity matrix E and the one-way causal strength matrix B , i.e. W = E + B ; S34, based on the weighting matrix W To H The causal characteristics of the data collectors are weighted and fused, and the comprehensive causal characteristics are calculated by a formula Z : ； wherein represents a natural constant; is a matrix element, representing the causal strength of the causal feature of the th data collector to the causal feature of the th data collector; represents the causal feature of the th data collector, the subscript is a causal feature index.

5. The method of claim 4, wherein, The reconstruction loss is represented as follows: ； In the formula, Indicates the total number of source domains; Indicates the first Number of samples in each source domain ; express The average number of samples in each source domain; Indicates the signal dimension of the input sample; Indicates the first In the source domain, the first The sample number, ; Indicates the first In the source domain, the first The original signal corresponding to each sample; Indicates decoder pair The reconstructed signal output after reconstruction; express L The square of the 2-norm.

6. The method of claim 5, wherein, The classification loss is represented as follows: ； In the formula, represents the true fault label of the i-th sample in the j-th source domain; represents the corresponding fault classification probability output by the classifier; represents the natural logarithm.​​​ 7. The method of claim 6, wherein, The causal aggregation loss is represented as follows: ； wherein represents a causal feature loss; represents a non-causal feature loss; ； In the formula, represents the first source domain, represents the first source domain, ; represents the number of samples in the first source domain, represents the number of samples in the first source domain; represents the number of samples in the first source domain, ; ; represents the number of samples in the first source domain, ; ; represents the true fault label of the first sample in the first source domain, represents the true fault label of the first sample in the first source domain, represents 1 when , otherwise 0; represents 1 when , otherwise 0; represents the causal feature extracted from the first sample in the first source domain, in the form of a vector; represents the transpose vector of ; represents the causal feature extracted from the first sample in the first source domain, in the form of a vector; ； In the formula, Indicates the first In the source domain, the first The sample number, Indicates the first In the source domain, the first The sample number, ; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form, with subscripts... It is a non-causal characteristic marker; for The transpose of ; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form; Indicates when When the time is right, the value is 1; otherwise, it is 0. Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form; Indicates from the first The first source domain The non-causal features extracted from each sample are in vector form.

8. The method of claim 7, wherein, The redundancy reduction loss is represented as follows: ； wherein denotes an all-ones matrix; denotes a Hadamard product; denotes a causal feature matrix; denotes a transpose of a causal feature matrix; denotes an acausal feature matrix; denotes a transpose of an acausal feature matrix; denotes a square of an F-norm; denotes a dimension of a causal feature and an acausal feature.

Images