An information fusion three-phase motor fault diagnosis method based on game mapping learning

Abstract

An information fusion three-phase motor intelligent fault diagnosis method based on game mapping learning, which automatically extracts fault features of multiple sensor sources through multiple self-learning feature mapping networks, constructs a sensor source discriminator, constructs a fault pattern identifier, and uses an Adam optimizer to update the diagnosis network parameters to achieve intelligent fault diagnosis of three-phase asynchronous motors; the present invention integrates fault information from different sensors, and the diagnosable fault types fully cover the electrical faults and mechanical faults of the three-phase motor. It has the advantages of higher diagnostic accuracy, wider fault type coverage, stronger robustness, and better stability, and can better meet the actual needs of engineering projects.

Description

Technical Field

[0001] The present invention relates to the field of mechanical engineering, and more specifically, to an information fusion three-phase motor fault diagnosis method based on game mapping learning. Background Art

[0002] Three-phase motors are widely used in various fields of industrial production and social life. Once a failure occurs, it may cause production work to stop, resulting in direct economic losses, or even cause safety accidents and casualties. Therefore, timely diagnosis of three-phase motor faults and taking corresponding maintenance measures are of great significance to ensure production safety, improve production efficiency and reduce maintenance costs. With the rapid development of the industrial Internet, various monitoring sensors such as vibration, current, and sound are widely used in three-phase motor status monitoring, resulting in an exponential increase in the scale of monitoring data for three-phase motors. Massive multi-sensor monitoring data provides a rich data foundation for intelligent fault diagnosis of three-phase motors based on machine learning algorithms, but also brings severe challenges. Therefore, it is necessary to invent a new intelligent fault diagnosis method for three-phase motors, which can automatically mine key fault features from massive multi-sensor monitoring data and timely diagnose fault types to ensure safe, reliable and efficient operation of three-phase motors.

[0003] With the powerful information mining and self-learning capabilities of deep neural networks, intelligent fault diagnosis methods based on deep learning can extract fault information from massive monitoring data, avoiding the tedious manual feature extraction process and achieving good diagnostic results. However, most of the existing intelligent fault diagnosis methods for three-phase motors based on deep learning use single sensor monitoring data as the input of the diagnostic network. Since three-phase motor faults include electrical faults and mechanical faults, and different types of faults have different physical representations, using only single sensor data will not be able to fully and effectively capture the fault information of the motor, which limits the range of diagnosable fault types and reduces the diagnostic accuracy. In addition, even if some intelligent fault diagnosis methods use multi-sensor monitoring data as input, they only integrate the multi-sensor information fusion process into the classification process, and do not consider the similarities and differences between different sensor data in the fault information mining and feature fusion process, which seriously affects the accuracy of multi-source information fusion fault diagnosis results. Summary of the invention

[0004] The present invention aims to provide an information fusion three-phase motor intelligent fault diagnosis method based on game mapping learning to overcome the shortcomings and limitations of the prior art.

[0005] The technical solution of the present invention is:

[0006] An information fusion three-phase motor fault diagnosis method based on game mapping learning includes the following steps:

[0007] 1) Obtain multi-sensor signals of three-phase motors under different working conditions Including vibration signal, current signal, speed signal, sound signal, among which, represents the i-th sample x i ={x i1 ,x i2 ,…,x iM}jth sensor data, Represents a matrix space with dimension H×1, where M is the number of sensors, N is the number of signal samples, and H is the number of data points contained in each sensor data;

[0008] 2) Construct M self-learning feature mapping networks, corresponding to the M sensor sources, and these M self-learning feature mapping networks have the same structure and do not share parameters with each other. It indicates that each self-learning feature mapping network is composed of three stacked convolutional layers, which is used to automatically extract multi-sensor fault features from different sensor data and map them into a common information metric space;

[0009] 3) Construct a sensor source discriminator, which consists of three fully connected layers, denoted by D(·), to identify the sensor data sources of various fault characteristics in the common information metric space;

[0010] 4) Construct a fault pattern recognizer, which is a three-layer fully connected network, denoted by C(·), to identify the fault category of the input sample;

[0011] 5) Train the sensor source discriminator D(·) to minimize the sensor source discrimination loss so that it can maximally distinguish the sensor sources of fault features in the common information metric space; and train M self-learning feature mapping networks at the same time. Maximize the sensor source identification loss so that the fault characteristics of different sensor data mapped to the common information metric space have the smallest difference and are difficult to distinguish; that is, the sensor source discriminator D(·) and the self-learning feature mapping network A "zero-sum game" is played between them; a correlation loss is constructed to train M self-learning feature mapping networks Minimizing the correlation loss vector further makes the multi-sensor fault features in the common information metric space more clustered within the class and more dispersed between classes; after iteration, the multi-sensor fault features obtained by different sensor data of the sample through different feature mapping networks have the smallest difference within the same type of fault and the largest difference between different types of faults; finally, the multi-sensor fault features in the common information metric space are fused as the input of the fault pattern recognizer for fault diagnosis.

[0012] Preferably, step 2) specifically includes the following steps:

[0013] (2.1) The input of the self-learning feature mapping network is sample x i ={x i1 ,x i2 ,…,x iM} a sensor data x ij , x ij The activation operation is performed through the Rectified linear unit (ReLU), and the convolution kernel k is used l Convolve with the activated data to get vector u ij , the calculation expression is as follows:

[0014] u ij =σ(k l *x ij +b l )

[0015] In the formula, k l is the convolution kernel of the convolution layer, b l is the bias term, * represents the convolution operation, and σ(·) represents the linear rectifier unit;

[0016] (2.2) for the vector u ij Perform average pooling to obtain the average value of the elements in the non-overlapping pooling area, and the pooled vector is y ij , the calculation expression is as follows:

[0017] y ij =pool(u ij ,p,s)

[0018] Where pool(·) represents the average pooling function, p is the pooling size, and s is the pooling moving step length;

[0019] (2.3) The vector y ij Execute steps (a) and (b) again, flatten the obtained vector into a one-dimensional vector, and finally obtain the fault feature X ij .

[0020] Preferably, step 3) comprises the following steps:

[0021] (3.1) The fault feature X output by the self-learning feature mapping network ij As the input of the sensor source discriminator; X ij Input the sensor source discriminator, extract the sensor source features and reduce the dimension to M dimensions (the number of sensor sources) to obtain the vector The calculation expression is as follows:

[0022]

[0023] in, represents the output of the l-1th layer, represents the output of the lth layer, W l is the weight matrix of the lth layer, b l is the bias term, σ(·) is the sigmoid nonlinear activation function;

[0024] (3.2) As the input of the softmax layer, the fault feature X is finally obtained. ij The sensor source prediction probability vector The expression of the softmax function is:

[0025]

[0026] In the formula, Representation vector The a-th value of .

[0027] Preferably, step 4) comprises the following steps:

[0028] (4.1) The M fault features {X i1 ,X i2 ,…,X iM}Add the corresponding elements to get the fused fault feature Z i , which is used as the input of the fault pattern identifier;

[0029] (4.2) Z i Input the fault pattern recognizer of the three-layer fully connected network, extract the deep fault features and reduce the dimension to the same dimension as the number of fault categories to obtain the fault category feature vector The calculation expression is as follows:

[0030]

[0031] in, represents the output of the l-1th layer, represents the output of the lth layer, W l is the weight matrix of the lth layer, b l is the bias term, σ(·) is the sigmoid nonlinear activation function;

[0032] (4.3) As the input of the softmax layer, we finally get the sample {x i1 ,x i2 ,…,x iM}fault category probability vector Y i .

[0033] Preferably, step 5) comprises the following steps:

[0034] (5.1) For a small batch of samples with a size of B The different sensor data of the sample are input into the corresponding self-learning feature mapping network in turn to obtain The calculation expression is as follows:

[0035] X ij =F j (x ij ),j∈[1,M]

[0036] in, Represents the self-learning feature mapping network corresponding to M sensor sources;

[0037] (5.2) Calculate the sample The correlation loss L M , the specific steps are as follows:

[0038] (a) For a set of samples After the self-learning feature mapping network, the features are obtained The correlation loss between classes is calculated as follows:

[0039]

[0040] Among them, if x ik and x jm Belong to the same fault category, If x ik and x jm Belong to different fault categories, d(X ik ,X jm ) indicates X ik , X jm The Euclidean distance between two vectors is calculated as follows:

[0041]

[0042] Among them, X(a) represents the a-th data of the vector, and K represents the length of the vector;

[0043] (b) Calculate the correlation loss within the class. The calculation expression is as follows:

[0044]

[0045] (c) Calculate the total correlation loss: L M =L M1 +L M2

[0046] (5.3) For the multi-sensor fault signature of a set of samples obtained in (5.1) The fault feature X ijInput the sensor source discriminator in sequence to obtain the sensor source identification probability vectors of different fault characteristics. The calculation expression is as follows:

[0047]

[0048] Where D(·) represents the sensor source discriminator;

[0049] (5.4) Calculate the sample The sensor source discrimination loss L D , the sensor source identification loss is essentially the related entropy loss obtained by inputting the fault features of a set of training samples into the sensor source discriminator, and the calculation formula is as follows:

[0050]

[0051] In the formula, s j is the true label of the jth sensor, expressed in the form of a one-hot vector, s j (a) represents the a-th number of a vector;

[0052] (5.5) For the multi-sensor fault signature of a set of samples obtained in (5.1) The multi-sensor fault feature X of each sample i1 ,X i2 ,…,X iM Add the corresponding elements to get the fused fault features

[0053] (5.6) The fusion fault feature Z i Input the fault pattern identifier to obtain the fault category probability vector Y of the sample i ; The calculation expression is as follows:

[0054] Y i =C(Z i )

[0055] Where C(·) represents the fault mode identifier

[0056] (5.7) Calculate the sample The failure mode recognition loss L C , the calculation formula is as follows:

[0057]

[0058] where c i Indicates the true label of the fault category of the sample, in the form of a one-hot vector; F is the number of fault categories;

[0059] (5.8) Repeat steps (5.1)-(5.7), set the number of iterations E, perform game learning, and update the parameters of the self-learning feature mapping network, fault mode identifier, and sensor source identifier in turn. The specific steps are as follows:

[0060] (a) Based on the Adam optimization algorithm, to minimize L M +L C -L D To optimize the target, update the parameters of M self-learning feature mapping networks and fault mode identifiers;

[0061] (b) Based on the Adam optimization algorithm, to minimize L D To optimize the objective, the parameters of the sensor source discriminator are updated.

[0062] The beneficial effects of the present invention are:

[0063] 1) The present invention uses the multi-sensor monitoring signal of the three-phase motor as direct input, automatically extracts various types of multi-sensor fault features through a self-learning feature mapping network and maps them to a common information metric space. At the same time, in the process of fault information mining and feature fusion, the similarities and differences between different sensor data are fully considered. By constructing a game learning framework, the multi-sensor source fault features in the common information metric space have the smallest difference within the same type of faults and the largest difference between different types of faults, thereby effectively fusing the multi-sensor information and improving the accuracy of intelligent fault diagnosis of the three-phase motor.

[0064] 2) By integrating fault information from different sensors, the present invention can not only diagnose three-phase motor bearing faults, but also diagnose rotor faults, stator faults, etc. The diagnosable fault types comprehensively cover the electrical faults and mechanical faults of the three-phase motor. It has the advantages of higher diagnostic accuracy, wider fault type coverage, stronger robustness and better stability, and can better meet the actual needs of the project. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A flowchart of a method for diagnosing three-phase motor faults based on information fusion and game mapping learning provided by an embodiment of the present invention;

[0066] Figure 2 A characteristic t-SNE dimension reduction graph of an ablation experiment of an information fusion three-phase motor fault diagnosis method based on game mapping learning provided in an embodiment of the present invention;

[0067] Figure 3 A comparison chart of test accuracy of an ablation experiment of an information fusion three-phase motor fault diagnosis method based on game mapping learning provided by an embodiment of the present invention;

[0068] Figure 4A confusion matrix diagram of the test results of an information fusion three-phase motor fault diagnosis method based on game mapping learning provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0069] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. The implementation methods of the present invention are not limited thereto.

[0070] like Figure 1 As shown, an information fusion three-phase motor intelligent fault diagnosis method based on game mapping learning includes the following steps:

[0071] 1) Obtain multi-sensor signals of three-phase motors under different working conditions Common signals include vibration signals, current signals, speed signals, sound signals, etc. represents the i-th sample x i ={x i1 ,x i2 ,…,x iM}jth sensor data, M is the number of sensors, N is the number of signal samples, and H is the number of data points contained in each sensor data;

[0072] 2) Construct M self-learning feature mapping networks, corresponding to the M sensor sources, and these M self-learning feature mapping networks have the same structure and do not share parameters with each other. Each self-learning feature mapping network is composed of three stacked convolutional layers, which is used to automatically extract multi-sensor fault features from different sensor data and map them into the common information metric space; the specific steps are as follows:

[0073] (2.1) The input of the self-learning feature mapping network is sample x i ={x i1 ,x i2 ,…,x iM} a sensor data x ij , x ij The activation operation is performed through the Rectified linear unit (ReLU), and the convolution kernel k is used l Convolve with the activated data to get vector u ij , the calculation expression is as follows:

[0074] u ij =σ(k l *x ij +b l )

[0075] In the formula, k l is the convolution kernel of the convolution layer, bl is the bias term, * represents the convolution operation, and σ(·) represents the linear rectifier unit;

[0076] (2.2) for the vector u ij Perform average pooling to obtain the average value of the elements in the non-overlapping pooling area, and the pooled vector is y ij ; The calculation expression is as follows:

[0077] y ij =pool(u ij ,p,s);

[0078] Where pool(·) represents the average pooling function, p is the pooling size, and s is the pooling moving step length;

[0079] (2.3) The vector y ij Execute steps (a) and (b) again, flatten the obtained vector into a one-dimensional vector, and finally obtain the fault feature X ij ;

[0080] 3) Construct a sensor source discriminator, which consists of three fully connected layers, denoted by D(·), to identify the sensor data sources of various fault characteristics in the common information metric space; the specific steps are as follows:

[0081] (3.1) The input of the sensor source discriminator is the fault feature X output by the self-learning feature mapping network ij ; X ij Input the sensor source discriminator, extract the sensor source features and reduce the dimension to M dimensions (the number of sensor sources) to obtain the vector The calculation expression is as follows:

[0082]

[0083] in, represents the output of the l-1th layer, represents the output of the lth layer, W l is the weight matrix of the lth layer, b l is the bias term, σ(·) is the sigmoid nonlinear activation function;

[0084] (3.2) As the input of the softmax layer, the fault feature X is finally obtained. ij The sensor source prediction probability vector The expression of the softmax function is:

[0085]

[0086] In the formula, Representation vector The a-th value of ;

[0087] 4) Construct a fault pattern recognizer, which is a three-layer fully connected network, represented by C(·), to identify the fault category of the input sample; the specific steps are as follows:

[0088] (4.1) The M fault features {X i1 ,X i2 ,…,X iM}Add the corresponding elements to get the fused fault feature Z i , which is used as the input of the fault pattern identifier;

[0089] (4.2) Z i Input the fault pattern recognizer of the three-layer fully connected network, extract the deep fault features and reduce the dimension to the same dimension as the number of fault categories to obtain the fault category feature vector The calculation expression is as follows:

[0090]

[0091] in, represents the output of the l-1th layer, represents the output of the lth layer, W l is the weight matrix of the lth layer, b l is the bias term, σ(·) is the sigmoid nonlinear activation function;

[0092] (4.3) As the input of the softmax layer, we finally get the sample {x i1 ,x i2 ,…,x iM}fault category probability vector Y i ;

[0093] 5) Train the sensor source discriminator D(·) to minimize the sensor source discrimination loss so that it can maximally distinguish the sensor sources of fault features in the common information metric space; and train M self-learning feature mapping networks at the same time. Maximize the sensor source identification loss so that the fault characteristics of different sensor data mapped to the common information metric space have the smallest difference and are difficult to distinguish; that is, the sensor source discriminator D(·) and the self-learning feature mapping network Construct correlation loss and train M self-learning feature mapping networks Minimize the correlation loss vector, and further make the multi-sensor fault features in the public information metric space more concentrated within the class and more dispersed between classes. After iteration, the multi-sensor fault features obtained by different sensor data of the sample through different feature mapping networks have the smallest difference within the same type of fault and the largest difference between different types of faults. Finally, the multi-sensor fault features in the public information metric space are fused as the input of the fault pattern recognizer for fault diagnosis. The specific steps are as follows:

[0094] (5.1) For a small batch of samples with a size of B The different sensor data of the sample are input into the corresponding self-learning feature mapping network in turn to obtain The calculation expression is as follows:

[0095] X ij =F j (x ij ),j∈[1,M]

[0096] in, Represents the self-learning feature mapping network corresponding to M sensor sources;

[0097] (5.2) Calculate the sample The correlation loss L M , the specific steps are as follows:

[0098] (a) For a set of samples After the self-learning feature mapping network, the features are obtained The correlation loss between classes is calculated as follows:

[0099]

[0100] Among them, if x ik and x jm Belong to the same fault category, If x ik and x jm Belong to different fault categories, d(X ik ,X jm ) indicates X ik , X jm The Euclidean distance between two vectors is calculated as follows:

[0101]

[0102] Among them, X(a) represents the a-th data of the vector, and K represents the length of the vector;

[0103] (b) Calculate the correlation loss within the class. The calculation expression is as follows:

[0104]

[0105] (c) Calculate the total correlation loss: L M =L M1 +L M2

[0106] (5.3) For the multi-sensor fault signature of a set of samples obtained in (5.1) The fault feature X ij Input the sensor source discriminator in sequence to obtain the sensor source identification probability vectors of different fault characteristics. The calculation expression is as follows:

[0107]

[0108] Where D(·) represents the sensor source discriminator;

[0109] (5.4) Calculate the sample The sensor source discrimination loss L D , the sensor source identification loss is essentially the related entropy loss obtained by inputting the fault features of a set of training samples into the sensor source discriminator, and the calculation formula is as follows:

[0110]

[0111] In the formula, s j is the true label of the jth sensor, expressed in the form of a one-hot vector, s j (a) represents the a-th number of a vector;

[0112] (5.5) For the multi-sensor fault signature of a set of samples obtained in (5.1) The multi-sensor fault feature X of each sample i1 ,X i2 ,…,X iM Add the corresponding elements to get the fused fault features

[0113] (5.6) The fusion fault feature Z i Input the fault pattern identifier to obtain the fault category probability vector Y of the sample i ; The calculation expression is as follows:

[0114] Y i =C(Z i )

[0115] Where C(·) represents the fault mode identifier

[0116] (5.7) Calculate the sample The failure mode recognition loss L C , the calculation formula is as follows:

[0117]

[0118] where c i Indicates the true label of the fault category of the sample, in the form of a one-hot vector; F is the number of fault categories;

[0119] (5.8) Repeat steps (5.1)-(5.7), set the number of iterations E, perform game learning, and update the parameters of the self-learning feature mapping network, fault mode identifier, and sensor source identifier in turn. The specific steps are as follows:

[0120] (a) Based on the Adam optimization algorithm, to minimize L M +L C -L D To optimize the target, update the parameters of M self-learning feature mapping networks and fault mode identifiers;

[0121] (b) Based on the Adam optimization algorithm, to minimize L D To optimize the objective, update the parameters of the sensor source discriminator;

[0122] 6) The three-phase motor multi-sensor signal {x i1 ,x i2 ,…,x iM} is input into the self-learning feature mapping network after the corresponding training iteration to extract the corresponding multi-sensor fault features and obtain {X i1 ,X i2 ,…,X iM}, and fuse it with the fault features described in (5.5) to obtain the fused fault feature Z i , the fused fault features are input into the fault pattern recognizer after training iterations to realize intelligent fault diagnosis of mechanical and electrical faults of three-phase motors.

[0123] Embodiment: Taking ABB's M2QA90L2A three-phase asynchronous motor as an example, the effectiveness of this method is verified based on the motor fault simulation experimental data. The fault simulation experimental data includes the motor data of the drive motor at rotation frequencies of 15Hz, 25Hz, 35Hz, and 45Hz, and two loads of 0N·m and 10N·m, including normal state and 11 fault states. The fault state includes electrical faults such as short circuit, motor broken bar, motor phase loss, and mechanical faults such as bearing fault, shaft bending, and rotor imbalance. Each fault state includes three fault degrees. The sensor sampling frequency is 25.6kHz, and 800 samples are obtained for each healthy state. Therefore, the motor fault simulation data set contains a total of 9600 samples, each of which contains 8 channels of data including the three-axis vibration acceleration signal of the drive end, the three-phase current signal of the motor, the speed signal, and the sound signal. Each channel has 2560 data points. 25% of the samples are randomly selected for model training, and 75% of the samples are used for model testing. The resulting data set is shown in Table 1.

[0124] Table 1

[0125]

[0126] like Figure 2 As shown, in order to verify the effectiveness of the game learning framework and correlation loss in the method of the present invention, the method of the present invention is subjected to an ablation experiment, and t-distributed Stochastic Neighbor Embedding (t-SNE) is used to reduce the dimension of the fault features extracted by the self-learning feature mapping network and visualize them.

[0127] The parameters of each component of the game mapping learning are set as follows: the number of self-learning feature mapping networks M is 8, corresponding to 8 sensor data channels, the convolution kernel size is 8, the pooling size p and the step size s are both set to 4; the number of small batch samples B is 30; the number of iterations E is 100. The method of the present invention is used to conduct fault diagnosis experiments on the motor fault simulation data set, and the average diagnostic accuracy on the test set reaches 98.25%.

[0128] Other methods for comparison are:

[0129] Method 2 is the method of the present invention, removing the game learning strategy and the sensor source identifier, retaining the self-learning feature mapping network to extract features, using the fault mode identifier to perform fault diagnosis, and using the correlation loss and the fault mode recognition loss back propagation to adjust the network parameters;

[0130] Method 3 is the method of the present invention without correlation loss, using a self-learning feature mapping network to extract features, training the feature mapping network under a game learning framework, and not adding correlation loss when updating the parameters of the feature mapping network.

[0131] from Figure 2 It can be clearly seen that the method of the present invention is effective in clustering similar features and dispersing different types of features. By comparing with the other two variant methods, it is shown that the game learning framework and correlation loss of the method of the present invention can effectively play a role, making the fault features in the public information metric space more dispersed between classes and more clustered within classes, which is very helpful for improving the diagnostic accuracy.

[0132] like Figure 3 As shown, ten experiments were conducted on various methods, and box plots of recognition accuracy were obtained. It can be seen that the addition of the game learning framework and correlation loss makes the diagnosis network more accurate and more stable.

[0133] like Figure 4 As shown in FIG, the diagnosis results of the method of the present invention on the test set are plotted into a confusion matrix. It can be seen that the fault types that can be diagnosed by the method of the present invention fully cover the electrical faults and mechanical faults of the three-phase motor, and the diagnosis of various faults has a very high accuracy rate; Poisson distribution noise and Gaussian noise with a variance of 0 and a standard deviation of 0.1 are added to the test data respectively. The experimental results are shown in Table 2. It can be seen that the method of the present invention has strong robustness and generalization.

[0134] Table 2

[0135]

[0136] In order to further verify the superiority of the method of the present invention, the method of the present invention is compared with convolutional neural networks (CNN) and stacked auto-encoders (SAE), and the results are shown in Table 3. It can be seen that the method of the present invention has higher accuracy and better stability in three-phase motor fault diagnosis.

[0137] Table 3

[0138]

[0139] From the above results of motor fault diagnosis and ablation experiments as well as comparison with existing methods, it can be found that the method of the present invention can extract fault features from multi-sensor monitoring data, so that the features in the common information metric space are clustered within the class and dispersed between classes, and then the multi-sensor data is effectively fused to improve the accuracy of three-phase motor fault diagnosis. In addition, by fusing fault information from different sensors, the present invention can not only diagnose three-phase motor bearing faults, but also diagnose rotor faults, stator faults, etc., and its diagnosable fault types fully cover the electrical faults and mechanical faults of three-phase motors. In summary, the method of the present invention has the advantages of higher diagnostic accuracy, wider fault type coverage, stronger robustness, and better stability, and can better meet the actual needs of engineering projects.

[0140] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the processes in the accompanying drawings are not necessarily required to implement the present invention.

Claims

1. An information fusion three-phase motor fault diagnosis method based on game mapping learning, characterized in that: The following steps are involved: 1) Obtain multi-sensor signals of three-phase motors under different working conditions Including vibration signal, current signal, speed signal, sound signal, among which, represents the i-th sample x i ={x i1 ,x i2 ,…,x iM }jth sensor data, Represents a matrix space with dimension H×1, where M is the number of sensors, N is the number of signal samples, and H is the number of data points contained in each sensor data; 2) Construct M self-learning feature mapping networks, corresponding to the M sensor sources, and these M self-learning feature mapping networks have the same structure and do not share parameters with each other. It indicates that each self-learning feature mapping network is composed of three stacked convolutional layers, which is used to automatically extract multi-sensor fault features from different sensor data and map them into a common information metric space; 3) Construct a sensor source discriminator, which consists of three fully connected layers, denoted by D(·), to identify the sensor data sources of various fault characteristics in the common information metric space; 4) Construct a fault pattern recognizer, which is a three-layer fully connected network, denoted by C(·), to identify the fault category of the input sample; 5) Train the sensor source discriminator D(·) to minimize the sensor source discrimination loss so that it can maximally distinguish the sensor sources of fault features in the common information metric space; and train M self-learning feature mapping networks at the same time. Maximize the sensor source identification loss so that the fault characteristics of different sensor data mapped to the common information metric space have the smallest difference and are difficult to distinguish; that is, the sensor source discriminator D(·) and the self-learning feature mapping network "Zero-sum game" between them; construct correlation loss and train M self-learning feature mapping networks Minimize the correlation loss, further make the multi-sensor fault features in the common information metric space more clustered within the class and more dispersed between classes; after iteration, the multi-sensor fault features obtained by different sensor data of the sample through different feature mapping networks have the smallest difference within the same type of fault and the largest difference between different types of faults; finally, the multi-sensor fault features in the common information metric space are fused as the input of the fault pattern recognizer for fault diagnosis; Wherein, step 2) specifically comprises the following steps: (2.1) The input of the self-learning feature mapping network is sample x i ={x i1 ,x i2 ,…,x iM } a sensor data x ij , x ij The activation operation is performed through the Rectified Linear Unit (ReLU), and the convolution kernel k is used l Convolve with the activated data to get vector u ij , the calculation expression is as follows: you ij =σ(k l *x ij +b l ) In the formula, k l is the convolution kernel of the convolution layer, b l is the bias term, * represents the convolution operation, and σ(·) represents the linear rectifier unit; (2.2) for the vector u ij Perform average pooling to obtain the average value of the elements in the non-overlapping pooling area, and the pooled vector is y ij , the calculation expression is as follows: y ij =pool(u ij ,p,s) Where pool(·) represents the average pooling function, p is the pooling size, and s is the pooling moving step length; (2.3) The vector y ij Execute steps (a) and (b) again, flatten the obtained vector into a one-dimensional vector, and finally obtain the fault feature X ij .

2. The information fusion three-phase motor fault diagnosis method based on game mapping learning according to claim 1 is characterized in that: Step 3) comprises the following steps: (3.1) The fault feature X output by the self-learning feature mapping network ij As the input of the sensor source discriminator; X ij Input the sensor source discriminator, extract the sensor source features and reduce the dimension to M dimensions (the number of sensor sources) to obtain the vector The calculation expression is as follows: in, represents the output of the l-1th layer, represents the output of the lth layer, W l is the weight matrix of the lth layer, b l is the bias term, σ(·) is the sigmoid nonlinear activation function; (3.2) As the input of the softmax layer, the fault feature X is finally obtained. ij The sensor source prediction probability vector The expression of the softmax function is: In the formula, Representation vector The a-th value of .

3. The information fusion three-phase motor fault diagnosis method based on game mapping learning according to claim 1 is characterized in that: Step 4) comprises the following steps: (4.1) The M fault features {X i1 ,X i2 ,…,X iM }Add the corresponding elements to get the fused fault feature Z i , which is used as the input of the fault pattern identifier; (4.2) Z i Input the fault pattern recognizer of the three-layer fully connected network, extract the deep fault features and reduce the dimension to the same dimension as the number of fault categories to obtain the fault category feature vector The calculation expression is as follows: in, represents the output of the l-1th layer, represents the output of the lth layer, W l is the weight matrix of the lth layer, b l is the bias term, σ(·) is the sigmoid nonlinear activation function; (4.3) As the input of the softmax layer, we finally get the sample {x i1 ,x i2 ,…,x iM }fault category probability vector Y i .

4. The information fusion three-phase motor fault diagnosis method based on game mapping learning according to claim 1 is characterized in that: Step 5) comprises the following steps: (5.1) For a small batch of samples with a size of B The different sensor data of the sample are input into the corresponding self-learning feature mapping network in turn to obtain The calculation expression is as follows: X ij =F j (x ij ),j∈[1,M] in, Represents the self-learning feature mapping network corresponding to M sensor sources; (5.2) Calculate the sample The correlation loss L M , the specific steps are as follows: (a) For a set of samples After the self-learning feature mapping network, the features are obtained The correlation loss between classes is calculated as follows: Among them, if x ik and x jm Belong to the same fault category, If x ik and x jm Belong to different fault categories, d(X ik ,X jm ) indicates X ik , X jm The Euclidean distance between two vectors is calculated as follows: Among them, X(a) represents the a-th data of the vector, and K represents the length of the vector; (b) Calculate the correlation loss within the class. The calculation expression is as follows: (c) Calculate the total correlation loss: L M =L M1 +L M2 (5.3) For the multi-sensor fault signature of a set of samples obtained in (5.1) The fault feature X ij Input the sensor source discriminator in sequence to obtain the sensor source identification probability vectors of different fault characteristics. The calculation expression is as follows: Where D(·) represents the sensor source discriminator; (5.4) Calculate the sample The sensor source discrimination loss L D , the sensor source identification loss is essentially the related entropy loss obtained by inputting the fault features of a set of training samples into the sensor source discriminator, and the calculation formula is as follows: In the formula, s j is the true label of the jth sensor, expressed in the form of a one-hot vector, s j (a) represents the a-th number of a vector; (5.5) For the multi-sensor fault signature of a set of samples obtained in (5.1) The multi-sensor fault feature X of each sample i1 ,X i2 ,…,X iM Add the corresponding elements to get the fused fault features (5.6) The fusion fault feature Z i Input the fault pattern identifier to obtain the fault category probability vector Y of the sample i ; The calculation expression is as follows: Y i =C(Z i ) Where C(·) represents the fault mode identifier (5.7) Calculate the sample The failure mode recognition loss L C , the calculation formula is as follows: where c i Indicates the true label of the fault category of the sample, in the form of a one-hot vector; F is the number of fault categories; (5.8) Repeat steps (5.1)-(5.7), set the number of iterations E, perform game learning, and update the parameters of the self-learning feature mapping network, fault mode identifier, and sensor source identifier in turn. The specific steps are as follows: (a) Based on the Adam optimization algorithm, to minimize L M +L C -L D To optimize the target, update the parameters of M self-learning feature mapping networks and fault mode identifiers; (b) Based on the Adam optimization algorithm, to minimize L D To optimize the objective, the parameters of the sensor source discriminator are updated.

Images