Permanent magnet motor multi-fault diagnosis method and device based on MDFSDP and MobileViT

Abstract

The present application belongs to the technical field of permanent magnet motor fault diagnosis, more specifically, relates to a permanent magnet motor multi-fault diagnosis method and device based on MDFSDP and MobileViT, comprising: collecting three-phase induced electromotive force signals of the permanent magnet motor under four states of normal state, inter-turn short circuit fault, local demagnetization fault, inter-turn short circuit and local demagnetization composite fault under different rotating speeds; segmenting the collected three-phase induced electromotive force signals for processing, generating a multi-dimensional feature symmetric dot matrix; integrating the multi-dimensional feature symmetric dot matrix under each state to construct a feature image dataset; constructing a lightweight MobileViT fault diagnosis network model; inputting the feature image dataset into the model to output a fault diagnosis result. The present application realizes accurate diagnosis of permanent magnet motor multi-fault by fusing image multi-dimensional information to enhance the distinguishability of fault features, and utilizing the feature extraction and classification capabilities of the fault diagnosis network model.

Description

Technical Field

[0001] This invention belongs to the technical field of permanent magnet motor fault diagnosis, and more specifically, relates to a method and apparatus for multi-fault diagnosis of permanent magnet motors based on MDFSDP and MobileViT. Background Technology

[0002] Permanent magnet motors, due to their high efficiency and power density, are widely used in high-performance power systems such as new energy vehicles and aerospace. However, under harsh operating conditions, these motors are prone to two types of faults: local demagnetization and inter-turn short circuits, which seriously threaten the reliable operation of the system. Both types of faults can cause similar abnormal phenomena such as current harmonics and local overheating, making early characteristics mixed and difficult to identify. More importantly, faults often couple with each other, forming a vicious cycle and increasing the risk of compound faults. Existing diagnostic methods are mostly designed for single faults, and are prone to misdiagnosis or missed diagnosis under compound fault conditions. Therefore, researching diagnostic methods that can effectively distinguish between the above-mentioned faults and their compound faults is of great significance for achieving accurate maintenance and ensuring the reliable operation of the system.

[0003] Chinese patent document CN114254674A discloses a PMSM fault diagnosis method and system based on image color features, including acquiring the original vibration signals of the motor under test under different states and multiple operating conditions; performing empirical mode decomposition on the original vibration signals to solve for the intrinsic mode functions at different levels to obtain signal components at different levels; converting the original vibration signals into symmetrical dot matrix images of the three RGB channels according to the different level signal components; improving the shape distribution and color features of the DenseNet learning image by fusing the scSE module to construct a classification network model; inputting the symmetrical dot matrix images of the three RGB channels into the classification network model to output the fault diagnosis results.

[0004] Existing research on multi-fault diagnosis techniques for permanent magnet motors can be divided into two categories: model-driven methods and data-driven methods. Model-driven methods rely on establishing precise mathematical models of motor parameters; however, in systems with multi-parameter coupling and complex operating conditions, the modeling process is cumbersome and its adaptability is limited. With the development of artificial intelligence technology, data-driven methods have been widely applied in the field of motor fault diagnosis. Among these methods, converting motor operating signals into two-dimensional images and utilizing the powerful feature extraction and pattern recognition capabilities of deep learning models to achieve fault diagnosis has become a current research hotspot. However, such methods have the following limitations: First, two-dimensional images have a single dimension in feature representation, making it difficult to comprehensively integrate multi-dimensional information such as shape, color, and texture; second, the models used, such as traditional convolutional neural networks, have limited capabilities in global dependency modeling, while visual Transformers, although capable of modeling long-distance relationships, have a large computational complexity and number of parameters, making efficient deployment in mobile or embedded devices difficult. These factors result in existing diagnostic methods being insufficient in distinguishing subtle fault differences under complex operating conditions, thus limiting their practical application value. Summary of the Invention

[0005] This invention aims to overcome at least one of the shortcomings of the prior art. The existing technology lacks a diagnostic scheme that can simultaneously achieve multi-dimensional fault information fusion and lightweight deep model deployment. To this end, a multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT is provided. There is an urgent need to propose a diagnostic method that combines multi-dimensional feature expression capabilities with efficient computational structure to improve the accurate identification and engineering applicability of multi-faults in permanent magnet motors.

[0006] The detailed technical solution of this invention is as follows:

[0007] A multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT, the method comprising:

[0008] S1. Collect the three-phase induced electromotive force signals of the permanent magnet motor under four states at different speeds: normal state, inter-turn short circuit fault, partial demagnetization fault, and combined inter-turn short circuit and partial demagnetization fault.

[0009] S2. The collected three-phase induced electromotive force signal is segmented and processed, and a multi-dimensional feature symmetrical dot matrix image that integrates shape, color and texture features is generated through an optimized symmetrical dot matrix image generation algorithm.

[0010] S3. Integrate the multi-dimensional feature symmetric dot matrix images under each state to construct a feature image dataset, and divide it into a training set, a validation set, and a test set according to a preset ratio;

[0011] S4. Construct a lightweight MobileViT fault diagnosis network model;

[0012] S5. Input the feature image dataset into the lightweight MobileViT fault diagnosis network model and output the fault diagnosis results.

[0013] Furthermore, S1 specifically includes:

[0014] First, the motor winding structure is adaptively adjusted: phase a and phase u are connected in series to form phase A, phase b and phase v are connected in series to form phase B, and phase c and phase w are connected in series to form phase C. Phases A, B, and C are then connected in a star configuration. Based on this, the following four states are simulated respectively:

[0015] (1) Normal state: Keep the motor windings and permanent magnets intact;

[0016] (2) Inter-turn short circuit fault: Short-circuit the u-phase part of the A-phase winding through a protective resistor;

[0017] (3) Local demagnetization fault: Remove permanent magnets PM1, PM7 and PM8 and replace them with non-magnetic stainless steel blocks;

[0018] (4) Combined fault of inter-turn short circuit and local demagnetization: The above two operations of inter-turn short circuit and local demagnetization are performed simultaneously;

[0019] Accordingly, fault labels are defined for the above four states to facilitate subsequent model training and diagnosis: normal state is labeled F1, inter-turn short circuit fault is labeled F2, local demagnetization fault is labeled F3, and combined inter-turn short circuit and local demagnetization fault is labeled F4. For the four states F1 to F4, three-phase induced electromotive force signals are collected at rated speed, 0.75 times the rated speed, and 0.5 times the rated speed, respectively.

[0020] Furthermore, the optimized symmetric dot matrix generation algorithm specifically includes:

[0021] S2.1: Expand the input to three channels, and map the three-phase timing data of the three-phase induced electromotive force to the three channels respectively, constructing a data matrix. S The rows of this matrix correspond to n Each sampling point corresponds to one column and three channels; elements S i,j Indicates the first i The sampling point at the th sampling point j The signal amplitude on each channel, of which i =1,2,…, n , j =1,2,3;

[0022] S2.2: Calculate the matrix S Each element S i,j polar radius in polar coordinate spaceR ( i , j ), counterclockwise deflection angle φ + ( i , j and clockwise deflection angle φ - ( i , j );

[0023] Each of the elements S i,j Mapped to two symmetrical points in polar coordinate space: P1( R ( i , j ), φ + ( i , j P2 and P2 R ( i , j ), φ - ( i , j ));

[0024] After the above mapping, the three-phase time series data form a more distinctive petal-shaped feature in polar coordinate space, specifically expressed as follows:

[0025] (1);

[0026] (2);

[0027] (3);

[0028] (4);

[0029] In formulas (1)-(4), S max Representation matrix S The maximum value among all positive elements in the set; S column This represents the total number of columns in the matrix. δ This is the rotation range parameter, with a value ranging from 0° to 60°; This is the time delay coefficient, with a value ranging from 1 to 10; φ j For the first j The initial rotation angle corresponding to the column;

[0030] S2.3: Define a new polar radius R' ( i,jAnd reuse the counterclockwise deflection angle described in S2.2. φ + ( i , j and clockwise deflection angle φ - ( i , j ), the matrix S Each element S i,j Remapped to two new symmetric points in polar coordinate space: P3( R' ( i,j ), φ + ( i , j )) and P4( R' ( i,j ), φ - ( i , j ));

[0031] Therefore, a new petal shape is generated concentrically within the existing petal shape, and the two overlap to form a layered texture feature. R' ( i,j This can be expressed as:

[0032] (5);

[0033] In formula (5), γ This is a scaling factor, ranging from 0.3 to 0.6, used to adjust the amplitude. λ This is an exponential factor, with a value ranging from 1 to 3, used to control nonlinear transformations;

[0034] S2.4: Apply the Jet color mapping method to two sets of symmetrical points respectively to form a visual pattern with a continuous color gradient;

[0035] Meanwhile, the image background is set to black, thus completing the construction of a multi-dimensional feature symmetrical bitmap that integrates shape, color, and texture features.

[0036] Furthermore, the application of the Jet color mapping method to the two sets of symmetrical points specifically includes:

[0037] Based on the size of the polar radius, a one-to-one correspondence is established between each set of symmetrical points and the continuously transitioning hue sequence—starting from the dark blue corresponding to the minimum polar radius value, as the polar radius value increases, it successively transitions to cyan, green, and yellow, and finally maps to red at the maximum polar radius value.

[0038] This invention generates a multi-dimensional symmetrical dot matrix image that integrates shape, color and texture features from three-phase induced electromotive force signals, breaking through the limitation of the single feature dimension of traditional two-dimensional images and significantly enhancing the distinguishability of fault features.

[0039] Furthermore, the construction of the lightweight MobileViT fault diagnosis network model specifically includes:

[0040] An initial convolutional layer, consisting of a 3×3 convolutional kernel and a stride of 2, is used to receive the input image, perform preliminary feature extraction and spatial downsampling, and output a feature map.

[0041] The initial convolutional layer is followed by a backbone feature extraction network consisting of an MV2 module and a MobileViT Block module. The MV2 module includes a non-downsampling MV2 module with a stride of 1 and a downsampling MV2 module with a stride of 2, namely the MV2↓2 module. The second MV2↓2 module and subsequent MV2↓2 modules are followed by a MobileViT Block module, which integrates a Transformer structure based on a self-attention mechanism to model global spatial dependencies, thereby overcoming the limitations of traditional convolutional neural networks in long-distance feature interactions.

[0042] The backbone feature extraction network is followed by a classifier, which includes a 1×1 convolutional layer for feature projection, a global average pooling layer for compressing spatial features into a 1×1 value, and a fully connected layer for outputting the probability of the fault category.

[0043] Furthermore, the input of the feature image dataset into the lightweight MobileViT fault diagnosis network model specifically includes: a customized training phase, a validation phase, and a testing phase;

[0044] Training Phase: The network is iteratively trained using the training set data. First, the parameters of the lightweight MobileViT fault diagnosis network model are initialized. The weights of the convolutional layers are initialized using the Kaiming normal distribution, the weights of the Transformer structure are initialized using the Xavier uniform distribution, and all bias terms are initialized to zero. Second, after forward propagation, the error between the prediction and the true label is calculated using the cross-entropy loss function. Subsequently, backpropagation is performed using the AdamW optimizer to update the network weights, and the learning rate is updated using a cosine annealing scheduling strategy, decaying from the base value to a minimum of 0.00001 according to the cosine function curve. After each training epoch, the model enters the validation phase.

[0045] Validation phase: Perform forward propagation using validation set data to calculate performance metrics such as accuracy and loss on the validation set; then determine whether the current validation result is better than the historical best performance. If so, save the current parameters as the best model; otherwise, retain the original best model; finally, terminate the training process after reaching the maximum number of training rounds.

[0046] Testing phase: After model training and validation are completed, the saved best model is loaded, and forward propagation is performed on the test set that did not participate in the training and hyperparameter optimization throughout the process to output the fault diagnosis results.

[0047] In another aspect of the invention, an apparatus is provided for implementing a method for diagnosing multiple faults in a permanent magnet motor based on a multidimensional feature symmetric dot matrix and MobileViT, the apparatus comprising:

[0048] At least one processor;

[0049] And a memory that stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT as described above.

[0050] In another aspect of the invention, a computer-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform the multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT as described above.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] The present invention provides a method and apparatus for multi-fault diagnosis of permanent magnet motors based on MDFSDP and MobileViT. It generates a multi-dimensional feature symmetrical dot matrix image that integrates shape, color and texture features from the three-phase induced electromotive force signal, breaking through the limitation of the single feature dimension of traditional two-dimensional images and significantly enhancing the distinguishability of fault features. At the same time, by using a lightweight MobileViT network model, it can accurately diagnose inter-turn short circuit faults, local demagnetization faults, and combined inter-turn short circuit and local demagnetization faults of permanent magnet motors, while effectively controlling the computational complexity, thus providing a feasible technical path for multi-fault diagnosis of permanent magnet motors. Attached Figure Description

[0053] Figure 1 This invention relates to a multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT.

[0054] Figure 2 This is a schematic diagram of the winding connection and permanent magnet numbering of the 28-pole six-phase permanent magnet motor in Embodiment 1 of the present invention;

[0055] Figure 3 It is a schematic diagram of the induced electromotive force and the generated multi-dimensional characteristic symmetric dot matrix diagram of each operating state in Embodiment 1 of the present invention under the rated speed;

[0056] Figure 4 It is a schematic diagram of the induced electromotive force and the generated multi-dimensional characteristic symmetric dot matrix diagram of each operating state in Embodiment 1 of the present invention under 0.75 times the rated speed;

[0057] Figure 5 It is a schematic diagram of the induced electromotive force and the generated multi-dimensional characteristic symmetric dot matrix diagram of each operating state in Embodiment 1 of the present invention under 0.5 times the rated speed;

[0058] Figure 6 It is a schematic diagram of the MobileViT lightweight network model structure adopted in Embodiment 1 of the present invention;

[0059] Figure 7 It is a training and diagnosis flowchart of the lightweight MobileViT fault diagnosis network model in Embodiment 1 of the present invention;

[0060] Figure 8 It is a confusion matrix diagram of the diagnosis results of the feature image data set in Embodiment 1 of the present invention. Detailed implementation manners<了

[0061] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0062] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further descriptions of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0063] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0064] Without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0065] Embodiment 1

[0066] Refer[[ID=~43]] Figure 1This embodiment provides a multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT, wherein MDFSDP (Multi-dimensional Feature Symmetrized Dot Pattern) is a multi-dimensional feature symmetric dot pattern, and the method includes:

[0067] S1. Collect the three-phase induced electromotive force signals of the permanent magnet motor under four states at different speeds: normal state, inter-turn short circuit fault, partial demagnetization fault, and combined inter-turn short circuit and partial demagnetization fault.

[0068] Preferably, this embodiment takes a 28-pole six-phase permanent magnet motor as an example. First, the motor winding structure is adaptively adjusted, and its winding connections and permanent magnet numbering are as follows: Figure 2 As shown: In a six-phase winding, phases a and u are connected in series to form phase A, phases b and v are connected in series to form phase B, and phases c and w are connected in series to form phase C. Phases A, B, and C are then connected in a star configuration. Based on this, the following four states are simulated respectively:

[0069] (1) Normal state: Keep the motor windings and permanent magnets intact;

[0070] (2) Inter-turn short circuit fault: Short-circuit the u-phase part of the A-phase winding through a protective resistor;

[0071] (3) Local demagnetization fault: Remove permanent magnets PM1, PM7 and PM8 and replace them with non-magnetic stainless steel blocks;

[0072] (4) Combined fault of inter-turn short circuit and local demagnetization: The above two operations of inter-turn short circuit and local demagnetization are performed simultaneously.

[0073] Accordingly, fault labels are defined for the above four states to facilitate subsequent model training and diagnosis: normal state is labeled F1, inter-turn short circuit fault is labeled F2, local demagnetization fault is labeled F3, and combined inter-turn short circuit and local demagnetization fault is labeled F4. For the four states F1 to F4, three-phase induced electromotive force signals are collected at rated speed, 0.75 times the rated speed, and 0.5 times the rated speed, respectively.

[0074] S2. The collected three-phase induced electromotive force signal is segmented and processed, and a multi-dimensional feature symmetrical dot matrix image with fused shape, color and texture features is generated by an optimized symmetrical dot matrix image generation algorithm.

[0075] Preferably, the three-phase induced electromotive force signal collected by S1 is segmented into segments using a sliding window with a length of 2000 sampling points and a step size of 100 sampling points, and the processed signal segments are generated into a multi-dimensional feature symmetrical dot matrix map that integrates shape, color and texture features.

[0076] Figure 3 , Figure 4 and Figure 5 The induced electromotive force (EMF) signals and their generated multidimensional characteristic symmetric dot plots under various operating conditions at rated speed, 0.75 times the rated speed, and 0.5 times the rated speed are displayed respectively. It can be observed that the multidimensional characteristic symmetric dot plots generated by the induced EMF under different operating conditions have obvious distinguishability in terms of shape, color, and texture features, and the multidimensional characteristic symmetric dot plots generated by the induced EMF under the same operating condition have good stability across different speeds.

[0077] In this embodiment, 400 images are generated for each state at each of the three rotation speeds, resulting in a total of 1200 images generated for each state. The multi-dimensional feature symmetrical dot matrix image is constructed by optimizing the shape features of a traditional symmetrical dot matrix image and introducing color and texture features, specifically including:

[0078] S2.1: Expand the input of the optimized symmetric dot matrix generation algorithm to three channels, and map the three-phase time-series data of the three-phase induced electromotive force to the three channels respectively to construct a data matrix. S The rows of this matrix correspond to n Each sampling point corresponds to one column and three channels; elements S i,j Indicates the first i The sampling point at the th sampling point j The signal amplitude on each channel, of which i =1,2,…, n , j =1,2,3, in matrix form as follows:

[0079] (6);

[0080] S2.2: Calculate the matrix S Each element S i,j polar radius in polar coordinate space R ( i , j ), counterclockwise deflection angle φ + ( i , j and clockwise deflection angle φ - ( i , j );

[0081] Each of the elements S i,j Mapped to two symmetrical points in polar coordinate space: P1( R ( i ,j ), φ + ( i , j P2 and P2 R ( i , j ), φ - ( i , j After the above mapping, the three-phase time series data form a more distinctive petal-shaped feature in polar coordinate space. Specifically, this is expressed as:

[0082] (1);

[0083] (2);

[0084] (3);

[0085] (4);

[0086] In formulas (1)-(4), S max Representation matrix S The maximum value among all positive elements in the set; S column This represents the total number of columns in the matrix. δ This is the rotation range parameter, with a value ranging from 0° to 60°; This is the time delay coefficient, with a value ranging from 1 to 10; φ j For the first j The initial rotation angle corresponding to the column.

[0087] In a preferred embodiment, the value of the rotation range parameter is... δ =30°, time delay coefficient is l When the value is 10, the generated petal shape features have the best distinguishability and clarity.

[0088] S2.3: Define a new polar radius R' ( i,j And reuse the counterclockwise deflection angle described in S2.2. φ + ( i , j and clockwise deflection angle φ - ( i , j ), the matrix S Each element S i,jRemapped to two new symmetric points in polar coordinate space: P3( R' ( i,j ), φ + ( i , j )) and P4( R' ( i,j ), φ - ( i , j Therefore, within the petal shape described in S2.2, a new petal shape is generated concentrically nested therein. The two overlap to form a layered texture feature, with a polar diameter... R' ( i,j This can be expressed as:

[0089] (5);

[0090] In formula (5), γ This is a scaling factor, ranging from 0.3 to 0.6, used to adjust the amplitude. λ It is an exponential factor, with a value ranging from 1 to 3, used to control nonlinear transformations.

[0091] In a preferred embodiment, scaling factor γ Set to 0.5 and exponential factor λ When set to 3, it can effectively generate texture features that are distinct in level and do not confuse with the original features.

[0092] S2.4: Apply the Jet color mapping method to two sets of symmetrical points respectively: using the polar radius of the two sets of symmetrical points as... R ( i , j )and R' ( i,j Based on the size, a one-to-one correspondence is established between each set of symmetrical points and the continuously transitioning hue sequence—starting from the dark blue corresponding to the minimum polar radius, as the polar radius value increases, it transitions to cyan, green, and yellow in sequence, and finally maps to red at the maximum polar radius value, thereby rendering a visual pattern with a continuous color gradient.

[0093] Meanwhile, the image background is set to black, thus completing the construction of a multi-dimensional feature symmetrical bitmap that integrates shape, color, and texture features.

[0094] S3. Integrate the multi-dimensional feature symmetric dot matrix images under each state to construct a feature image dataset, and divide it into a training set, a validation set, and a test set according to a preset ratio.

[0095] Preferably, in specific implementation, the multi-dimensional feature symmetric dot matrix diagrams of each state described in S2 are integrated as feature image datasets, and divided into training set, validation set and test set in a 5:2:3 ratio. Each state refers to four states: normal state, inter-turn short circuit fault, local demagnetization fault, and combined inter-turn short circuit and local demagnetization fault.

[0096] S4. Construct a lightweight MobileViT fault diagnosis network model.

[0097] Preferably, in specific implementation, the MobileViT lightweight network model structure adopted in this invention is illustrated as follows: Figure 6 As shown, the input level of this model is an initial convolutional layer, which consists of a 3×3 convolutional kernel and a stride of 2. It is used to receive an image input of size 224×224, perform preliminary feature extraction and spatial downsampling, output a feature map of size 112×112, and increase the number of channels to 16.

[0098] The initial convolutional layer is followed by a backbone feature extraction network consisting of the MobileNetV2 (MV2) module and the MobileViT Block module;

[0099] The MV2 module includes a non-downsampling MV2 module with a step size of 1 and a downsampling MV2 module with a step size of 2, namely the MV2↓2 module. The non-downsampling MV2 module performs local feature extraction and nonlinear transformation through residual connections. The residual connections are linear superpositions of the features extracted by the non-downsampling MV2 module and the input before feature extraction, which are nonlinearly transformed to enhance the expression of local detailed features while maintaining the feature map resolution. The MV2↓2 module achieves gradual compression of the feature image size through spatial downsampling, while simultaneously increasing the number of feature channels, thus completing the dimensionality reduction of the feature space size and the dimensionality increase of the channel dimension.

[0100] The second MV2↓2 module and subsequent MV2↓2 modules are connected to a MobileViT Block module, which integrates a Transformer structure based on a self-attention mechanism to model global spatial dependencies, thereby overcoming the limitations of traditional convolutional neural networks in long-distance feature interactions.

[0101] The backbone feature extraction network is followed by a classifier, which includes a 1×1 convolutional layer for feature projection, a global average pooling layer for compressing spatial features into a 1×1 value, and a fully connected layer for outputting the probability of the fault category.

[0102] S5. Input the feature image dataset into the lightweight MobileViT fault diagnosis network model and output the fault diagnosis results.

[0103] In specific implementation, such as Figure 7 As shown, the feature image dataset is input into the lightweight MobileViT fault diagnosis network model, and the following stages are executed sequentially to complete model training and output diagnostic results:

[0104] Training phase: The network is iteratively trained using the training set data; First, the parameters of the lightweight MobileViT fault diagnosis network model are initialized: the weights of the convolutional layers are initialized using the Kaiming normal distribution, the weights of the Transformer structure are initialized using the Xavier uniform distribution, and all bias terms are initialized to zero.

[0105] During training, the batch size is set to 32. The batch data is input into the model, and after forward propagation, the error between the prediction and the true label is calculated using the cross-entropy loss function.

[0106] Subsequently, backpropagation was performed using the AdamW optimizer to update the network weights. The optimizer's base learning rate was set to 0.002, momentum parameters were β1=0.9 and β2=0.999, and the weight decay coefficient was 0.01. A cosine annealing scheduling strategy was used to update the learning rate, decaying it from the base value to a minimum of 0.00001 according to a cosine function curve. After each training epoch, the model entered the validation phase.

[0107] Validation Phase: As part of the training cycle, after each training cycle, forward propagation is performed using the validation set data to calculate the performance metrics of accuracy and loss on the validation set. Then, it is determined whether the validation result is better than the historical best performance. If so, the current parameters are saved as the best model; otherwise, the original best model is retained. Finally, this phase checks whether the preset maximum number of training epochs has been reached. If the maximum number of epochs has been reached, the training process terminates; otherwise, the next training cycle continues, returning to the forward propagation step of the training phase.

[0108] Testing phase: After model training and validation are completed, the saved best model is loaded, and forward propagation is performed on the test set that did not participate in the training and hyperparameter optimization throughout the process to output the fault diagnosis results.

[0109] The confusion matrix diagram of the diagnostic results of the feature image dataset in this embodiment is shown below. Figure 8 As shown, the diagonal elements account for a very high proportion, indicating that the classification accuracy of the four states F1, F2, F3, and F4 is extremely high, at 100%, 99.44%, 99.72%, and 100% respectively, demonstrating a high degree of consistency between the model's predictions and the true labels. Calculations show that the model achieves a final comprehensive diagnostic accuracy of 99.79% on the test set, effectively verifying the high precision and reliability of the diagnostic method of this invention.

[0110] The method provided in this embodiment constructs a complete solution from multi-dimensional fault feature representation to lightweight model diagnosis. Its core advantage lies in the fact that the proposed multi-dimensional feature symmetric dot matrix overcomes the limitation of the single feature dimension of traditional two-dimensional images, and effectively controls the computational complexity with the help of a lightweight model. Thus, it provides a feasible technical path for the accurate diagnosis of multiple fault states of permanent magnet motors and their deployment in resource-constrained scenarios.

[0111] Example 2

[0112] This embodiment provides an apparatus for implementing a multi-fault diagnosis method for permanent magnet motors based on multi-dimensional feature symmetric dot matrix diagrams and MobileViT. The apparatus includes:

[0113] At least one processor; and

[0114] The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT as described above.

[0115] In this embodiment, the electronic devices involved in the device include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile computing devices, wearable computing devices, consumer electronic devices, etc.

[0116] Example 3

[0117] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed, cause the machine to perform the multi-fault diagnosis method for permanent magnet motors based on MDFSDP and MobileViT as described above.

[0118] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0119] In this case, the program code itself, which can be read from the readable medium, can perform the functions of any of the above embodiments, and therefore the computer-readable code and the readable storage medium storing the computer-readable code constitute a part of this specification.

[0120] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0121] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0122] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0123] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0124] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0125] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A permanent magnet motor multi-fault diagnosis method based on MDFSDP and MobileViT, characterized in that, The method comprises: S1, collecting three-phase induction electromotive force signals under four states of permanent magnet motor, including normal state, inter-turn short circuit fault, local demagnetization fault, and inter-turn short circuit and local demagnetization composite fault under different rotating speeds; S2, segmenting the collected three-phase induction electromotive force signals, and generating a multi-dimensional feature symmetry dot matrix with fused shape, color and texture features through an optimized symmetry dot matrix generation algorithm; S3, integrating the multi-dimensional feature symmetry dot matrix under each state, constructing a feature image dataset, and dividing it into a training set, a validation set and a test set according to a preset proportion; S4, constructing a lightweight MobileViT fault diagnosis network model; S5, inputting the feature image dataset into the lightweight MobileViT fault diagnosis network model, and outputting a fault diagnosis result; The optimized symmetry dot matrix generation algorithm specifically comprises: S2.1: extend the input to three channels, and correspond the three-phase time-series data of the three-phase induction electromotive force to the three channels respectively, to construct a data matrix S ; the rows of the matrix correspond to n sampling points, the columns correspond to 3 channels, and the elements S i,j represent the signal amplitude of the i sampling point on the j channel, wherein i =1, 2, …, n , j =1, 2, 3; S2.2: Calculate the matrix S Each element S i,j Polar radius in polar coordinate space R i , j Counter-clockwise deflection angle φ + i , j Clockwise deflection angle φ - i , j , is calculated as follows:​​​ （1）； （2）； （3）； （4）； In formulas (1)-(4), S max denotes a matrix S the maximum value among all positive elements in the matrix; S column denotes the total number of columns of the matrix; δ denotes a rotation range parameter; denotes a time delay coefficient; φ j denotes the initial rotation angle corresponding to the j column Each of the elements S i,j Mapped to two symmetrical points in polar coordinate space: P1( R ( i , j ), φ + ( i , j P2 and P2 R ( i , j ), φ - ( i , j After the above mapping, the three-phase time series data form more distinctive petal-shaped features in polar coordinate space; S2.3: define a new polar radius R' i,j ), and reuse the counter-clockwise deflection angle φ + i j ) and the clockwise deflection angle φ - i j ), map the matrix S each element S i,j again to two new symmetric points in polar coordinate space: P3( R' i, j φ + i j ) and P4( R' i,j φ - i j ); inside the petal shape, generate a new petal shape that is concentrically nested with it, and the two superimposed together form a texture feature with a sense of hierarchy, the polar radius R' i,j expression:​​​​​​​​​​​​​​ （5）； In equation (5), γ is a scaling factor for adjusting the amplitude; λ is an exponential factor for controlling the non-linear transformation; S2.4: Apply the Jet color mapping table to the two groups of symmetry dots respectively to generate a color gradient corresponding to the polar radius size; At the same time, set the image background to black, and finally complete the construction of the multi-dimensional feature symmetry dot matrix with fused shape, color and texture features.

2. The MDFSDP and MobileViT-based permanent magnet motor multi-fault diagnosis method according to claim 1, characterized in that, The application of the Jet color mapping table to the two groups of symmetry dots specifically comprises: The polar radius of two groups of symmetry points is R ( i , j ) and R' ( i,j ) size is based on, for each group of symmetry points and the corresponding relationship of the continuous transition of hue sequence, that is, from the minimum polar radius corresponding to the dark blue, with the polar radius value increasing, in turn transition for cyan, green, yellow, and finally at the maximum polar radius is mapped to red, rendering a visual pattern with continuous color gradient.

3. The MDFSDP and MobileViT-based permanent magnet motor multi-fault diagnosis method according to claim 2, characterized in that, The construction of the lightweight MobileViT fault diagnosis network model specifically comprises: The input level of the model is an initial convolutional layer, which is used to receive an input image, perform preliminary feature extraction and spatial down-sampling, and output a feature map; The initial convolutional layer is connected to a backbone feature extraction network composed of MV2 modules and MobileViT Block modules; wherein the MV2 module includes a non-down-sampling MV2 module with a step size of 1 and a down-sampling MV2 module with a step size of 2, i.e. MV2↓2 module; a second MV2↓2 module and the MV2↓2 module thereafter are connected to a MobileViT Block module, which integrates a Transformer structure based on a self-attention mechanism inside, for modeling global spatial dependency; The backbone feature extraction network is connected to a classifier, which includes a 1×1 convolutional layer for feature projection, a global average pooling layer for compressing spatial features to 1×1, and a fully connected layer for outputting fault category probabilities.

4. The MDFSDP and MobileViT-based permanent magnet motor multi-fault diagnosis method according to claim 3, characterized in that, The S5 specifically comprises a self-defined training phase, a validation phase and a test phase: The training phase: the network is iteratively trained using the training set data; first, the parameters of the lightweight MobileViT fault diagnosis network model are initialized, wherein the weights of the convolutional layer are initialized with Kaiming normal distribution, the weights of the Transformer structure are initialized with Xavier uniform distribution, and all bias items are initialized to zero; second, the error between the prediction and the true label is calculated using the cross-entropy loss function after forward propagation; then, the network weights are updated using the AdamW optimizer to perform back propagation, and the learning rate is updated using the cosine annealing scheduling strategy, which is decayed to the minimum 0.00001 according to the cosine function curve from the base value, after each training round, the model enters the verification phase; The verification phase: the forward propagation is performed using the validation set data, and the accuracy and loss value performance indicators on the validation set are calculated; then, it is judged whether the current verification result is better than the historical best performance, if yes, the current parameters are saved as the best model, if not, the original best model is retained; finally, the training process is terminated when the maximum number of training rounds is reached. The test phase: after the model training and verification are completed, the saved best model is loaded, and the test set which has not participated in the training and hyperparameter optimization is forward propagated, and the fault diagnosis result is output after processing.

5. The MDFSDP and MobileViT-based permanent magnet motor multi-fault diagnosis method according to claim 1, characterized in that, The S1 specifically includes: First, the motor winding structure is adaptively adjusted: the a phase and the u phase in the six-phase winding are connected in series to form an A phase, the b phase and the v phase are connected in series to form a B phase, the c phase and the w phase are connected in series to form a C phase, and the A, B and C phases are connected in star connection; on this basis, the following four states are simulated respectively: (1) Normal state: keep the motor winding and the permanent magnet intact; (2) Inter-turn short circuit fault: short-circuit the u phase part of the A phase winding through a protection resistor; (3) Local demagnetization fault: remove the permanent magnets PM1, PM7 and PM8 and replace them with non-magnetic stainless steel blocks; (4) Inter-turn short circuit and local demagnetization combined fault: implement both the inter-turn short circuit and the local demagnetization operations; Accordingly, the fault labels are defined for the above four states for subsequent model training and diagnosis: the normal state is marked as F1, the inter-turn short circuit fault is marked as F2, the local demagnetization fault is marked as F3, and the inter-turn short circuit and local demagnetization combined fault is marked as F4; For the F1 to F4 four states, three-phase induced electromotive force signals are collected at rated speed, 0.75 times rated speed and 0.5 times rated speed.

6. A device for implementing a permanent magnet motor multi-fault diagnosis method based on MDFSDP and MobileViT, characterized in that, The device includes: At least one processor; And a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the steps of the MDFSDP and MobileViT based permanent magnet motor multi-fault diagnosis method as claimed in any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program, wherein the computer program comprises the following steps of: The computer program, when executed by a processor, implements the steps of the method as claimed in any one of claims 1 to 5.

Images