Motor fault diagnosis method based on orthogonal sequence fusion and improved TABM

Abstract

The application belongs to the technical field of motor fault diagnosis, and provides a motor fault diagnosis method based on orthogonal sequence fusion and improved TABM. The application aims at the problems of difficult effective mutual mapping of cross-modal data, unbalanced feature representation and information redundancy in multi-modal data fault diagnosis of permanent magnet synchronous motor, proposes a multi-modal feature optimization mechanism based on orthogonal sequence fusion, aims at using multi-modal scientific sorting and orthogonal constraint to reduce redundant information, strengthen complementary and collaborative feature extraction, and provides high-quality feature representation for permanent magnet synchronous motor fault diagnosis. In view of the problems of slow training of existing permanent magnet synchronous motor fault diagnosis model in the face of massive data, low real-time reasoning efficiency and insufficient multi-modal adaptation ability, the application proposes an efficient fault reasoning architecture based on improved TABM algorithm, aims at relying on the advantages of high-speed derivation of large-scale data and simple deployment of the improved TABM model, and improves the effectiveness, reasoning efficiency and scene adaptability of permanent magnet synchronous motor fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of motor fault diagnosis technology, specifically relating to a motor fault diagnosis method based on orthogonal sequence fusion and improved TABM. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in key fields such as new energy vehicles, shipbuilding, and aerospace due to their excellent performance. However, they are prone to typical faults such as stator inter-turn short circuits, air gap eccentricity, and rotor inter-turn short circuits during long-term operation. Existing PMSM fault diagnosis methods fail to effectively explore the inherent correlations and mapping relationships between multi-dimensional modal data when extracting fault features. They lack comprehensive cross-modal fusion and representation mapping mechanisms, resulting in problems such as representation imbalance, information redundancy, and insufficient feature coupling in the extracted fault features. Furthermore, in terms of intelligent reasoning for PMSM fault models, existing PMSM fault diagnosis models generally suffer from complex structures and a large number of redundant parameters. When faced with massive amounts of data, the model training convergence speed is slow, and the inference computation efficiency is low, lacking efficient real-time fault reasoning capabilities and large-scale industrial deployment capabilities. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a motor fault diagnosis method based on orthogonal sequence fusion and improved TABM, thereby resolving the issues in the prior art. The technical solution adopted by this invention is as follows:

[0004] Motor fault diagnosis methods based on orthogonal sequence fusion and improved TABM include:

[0005] Step 1: Multimodal data acquisition and feature extraction;

[0006] Step 2: Multimodal data orthogonal sequence fusion to construct the data fusion layer of the permanent magnet synchronous motor fault diagnosis model;

[0007] Step 3: Based on the improved TABM model, construct the core prediction layer of the permanent magnet synchronous motor fault diagnosis model;

[0008] Step 4: Joint training and optimization of the fault diagnosis model for permanent magnet synchronous motors;

[0009] Step 5: Construct the fault classification layer of the permanent magnet synchronous motor fault diagnosis model to realize fault classification probability mapping and diagnosis result output.

[0010] Furthermore, step 1 includes:

[0011] Multimodal data of permanent magnet synchronous motors under normal and different fault conditions are collected; the multimodal data includes numerical data and image data.

[0012] Z-score normalization was performed separately for each modality of data;

[0013] The processed dataset is divided into a training set, a validation set, and a test set;

[0014] Each modal data is encoded, with numerical data processed using a dense layer and image data processed using the Inceptionv3 module.

[0015] The standardized features of each modality are mapped to a unified dimension d to generate the data embedding vector required for orthogonal sequence fusion.

[0016] Furthermore, numerical data includes three-phase current signals and vibration signals; image data includes infrared thermograms and stator winding images.

[0017] Furthermore, step 2 includes:

[0018] Each data embedding vector is input into the TABM model, and the fault diagnosis results output by the TABM model are tested sequentially. The data fusion order is determined based on the accuracy of the output fault diagnosis results.

[0019] Constructing a data fusion layer: Based on each data embedding vector, an embedding matrix E is generated according to the data fusion order; where, , is an embedding matrix for 4 modal data; each row Let be the data embedding vector for the t-th modality; during orthogonal sequence fusion of multimodal data, the mask is successively demasked and a single modality is incorporated, while all other modal data remain masked; the initialization form of the orthogonal sequence fusion representation is:

[0020] in: It is the initial orthogonal sequence fusion representation. This refers to the data embedding vector used in the initial data fusion process;

[0021] Represent the mask as a binary row vector Where 0 represents the masking state and 1 represents the demasking state; the data embedding vector used in the t-th step of orthogonal sequence fusion. , by binary row vector It is obtained by multiplying with the embedding matrix E; In orthogonal sequence fusion pairs After demasking, it is combined with the current fused representation. Integration; its expression is:

[0022]

[0023]

[0024]

[0025] Where: M t Let be the binary row vector at step t; Let be the embedding vector used in the orthogonal sequence fusion process at step t; orthogonal sequence fusion representation at step t; For data fusion functions;

[0026] After all modal data has been demasked and fused, the final orthogonal sequence fusion representation is as follows:

[0027]

[0028] in, For the final modal data fusion representation, This represents the third orthogonal sequence fusion characterization. This is the fourth data fusion function. This is the data embedding vector used for the 4th time.

[0029] Furthermore, step 3 includes:

[0030] Step 3.1, Based on the improved TABM model, construct the core prediction layer:

[0031] First, the final data fusion representation F4 obtained by fusing the data embedding vectors generated in step 1 with the orthogonal sequence in step 2 is used as the original input of the improved TABM model. The original input is copied into k identical copies to form the input matrix. Batch integration is performed using the parameter sharing method in the BatchEnsemble mechanism, with the expression as follows:

[0032]

[0033] Where: ⊙ represents element-wise multiplication; S is the multiplication adapter matrix; W is the shared weight; R is the multiplication adapter matrix; B is the addition adapter bias matrix;

[0034] Finally, the core prediction layer is constructed and obtained, represented as follows:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] in: is the prediction head for each implicit sub-model; k is the total number of sub-models; i is the sub-model index; t is the network layer index; N is the total number of network layers; This is the output of layer 1; This is the multiplication adapter matrix for the first layer; This is the shared weight matrix for layer 1; X is the multiplication adapter matrix of layer 1; X is the data fusion representation vector; Here is the bias matrix of the additive adapter in layer 1; Dropout is a regularization technique; ReLU is the activation function. This is the output of the t-th layer; Let be the multiplication adapter matrix of the t-th layer; Let be the shared weight matrix of layer t; Let be the multiplication adapter matrix of the t-th layer; Let be the bias matrix of the additive adapter at layer t; This is the output of the first layer after ReLU activation and Dropout. This is the output of the t-th layer after ReLU activation and Dropout; The original confidence score vector is output for the i-th sub-model; Z is the average of the predictions of all k implicit sub-models, which is used as the final prediction score vector.

[0042] Step 3.2, introduce a dynamic pruning mechanism:

[0043] The dynamic pruning mechanism is executed periodically. An evaluation is triggered on the validation set every fixed training epoch E, pruning the sub-models without updating the model weights. The improved TABM model outputs an i-dimensional prediction value Z for each implicit sub-model. i The dynamic pruning mechanism determines the contribution of a sub-model by comparing the fault classification accuracy before and after removing the sub-model.

[0044] The fault classification accuracy after integrating all k sub-models is:

[0045]

[0046]

[0047] Where: y represents the actual fault category of the sample; Let be the model's final predicted category for sample x; P is the diagnostic accuracy indicator function for sample x. The total number of samples in the validation set; The fault diagnosis accuracy of the model on the validation set;

[0048] After removing the nth sub-model, the accuracy of the ensemble of the remaining k-1 sub-models is:

[0049]

[0050]

[0051] Where: y represents the actual fault category of the sample; This represents the final predicted category of the model for sample x after pruning the nth sub-model. This is the diagnostic accuracy indicator function for sample x after pruning the nth sub-model; The fault diagnosis accuracy on the validation set after pruning the nth sub-model;

[0052] The contribution of the nth sub-model is defined as If the sub-model If ≤0, then pruning is performed; if the sub-model's If the value is greater than 0, then it is retained; the expression is:

[0053]

[0054] Define a counter for each sub-model The initial value is 0; if the sub-model's If ≤0, then increment the count by 1; if the sub-model's If the count is greater than 0, the count value remains unchanged; when... If the value is greater than or equal to the set threshold L, the sub-model will be permanently pruned.

[0055] Furthermore, step 4 includes: the permanent magnet synchronous motor fault diagnosis model uses the minimization of cross-entropy of the classification task as the loss function, calculates the loss for multiple prediction results separately, and takes the average as the TABM classification loss; when training the model, the joint total loss function includes the orthogonal sequence fusion loss and the TABM classification loss, as expressed below:

[0056]

[0057]

[0058] Where: N is the total number of training samples; C is the total number of categories in the classification task; y i,c Let i be the true label of the i-th sample; Let J be the predicted probability of the TABM model that the i-th sample belongs to class c; J is the TABM classification loss. The loss is for orthogonal sequence fusion. The total loss is the combined loss.

[0059] After obtaining the joint total loss, gradient calculation is performed through backpropagation, and the gradient is passed to the parameters of each layer of the permanent magnet synchronous motor fault diagnosis model.

[0060] Based on the gradient information obtained from backpropagation, the ADAMW optimizer is used to update and optimize the parameters simultaneously.

[0061] The classification accuracy of the fault classification results diagnosed by the permanent magnet synchronous motor fault diagnosis model is used to evaluate the model performance; a dynamic pruning mechanism is triggered on the validation set once every fixed training epoch to evaluate the model, pruning low-contribution sub-models and dynamically updating the number of effective sub-models; an early stopping strategy is adopted during validation set training.

[0062] Repeat the training until the model training is complete to obtain a fault diagnosis model for permanent magnet synchronous motors optimized by dynamic pruning.

[0063] Furthermore, step 5 includes:

[0064] The multimodal data to be diagnosed is processed sequentially through data standardization and encoding in step 1 and orthogonal sequence fusion in step 2 to obtain a data fusion representation vector. The data fusion representation vector is then input into the permanent magnet synchronous motor fault diagnosis model optimized in step 4 to output the original confidence scores Z for various fault types.

[0065] A Softmax function is introduced as the fault classification layer; the original confidence scores Z of the i states are used as the input of the Softmax function, and the output is the fault label probability distribution of various permanent magnet synchronous motor faults;

[0066] The fault category corresponding to the maximum value in the fault label probability distribution is selected as the final diagnostic result.

[0067] The present invention has the following beneficial effects:

[0068] (1) The present invention adopts the improved TABM model framework as the core prediction layer for fault diagnosis of permanent magnet synchronous motor. Relying on the BatchEnsemble parameter sharing mechanism and the inherent characteristics of lightweight MLP network, it has the significant advantages of high-speed inference of massive data and low computational overhead, and can effectively solve the technical problems of slow training and insufficient inference efficiency of traditional deep learning models.

[0069] (2) This invention proposes an orthogonal sequence fusion mechanism as the model data fusion layer. Through modal performance ranking and orthogonal constraint mechanism, it accurately solves the problems of representation imbalance and information redundancy in feature extraction of existing methods, and efficiently extracts cross-modal complementary information and collaborative features, realizes effective mapping of multi-dimensional data, and greatly improves the identification accuracy and generalization performance of multiple types of faults of permanent magnet synchronous motor.

[0070] (3) The permanent magnet synchronous motor fault diagnosis model constructed by the present invention includes a data fusion layer, a core prediction layer and a fault classification layer. The structure is clear and simple, and it has both diagnostic reliability and convenient deployment. It solves the problems of complex structure and difficulty in large-scale deployment of existing fault diagnosis models, and greatly improves the credibility and practical application value of permanent magnet synchronous motor fault diagnosis. Attached Figure Description

[0071] Figure 1 This is the overall flowchart;

[0072] Figure 2 This is a diagram showing the overall architecture of a fault diagnosis model for permanent magnet synchronous motors. Detailed Implementation

[0073] The following will be described in conjunction with embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0074] This invention addresses the challenges of effective cross-modal data mapping, feature imbalance, and information redundancy in multimodal fault diagnosis of permanent magnet synchronous motors (PMSMs). It proposes a multimodal feature optimization mechanism based on orthogonal sequence fusion, aiming to reduce redundant information and enhance complementary and collaborative feature extraction through scientific multimodal sorting and orthogonal constraints, thereby providing high-quality feature representations for PMSM fault diagnosis. Furthermore, to address the issues of slow training, low real-time inference efficiency, and insufficient multimodal adaptability of existing PMSM fault diagnosis models with massive data, this invention proposes an efficient fault inference architecture based on an improved TABM algorithm. This architecture leverages the advantages of the improved TABM model—high-speed derivation on large-scale data and ease of deployment—to improve the effectiveness, inference efficiency, and scenario adaptability of PMSM fault diagnosis.

[0075] like Figure 1 A motor fault diagnosis method based on orthogonal sequence fusion and improved TABM includes:

[0076] Step 1, Multimodal data acquisition and feature extraction; Collect multimodal data of permanent magnet synchronous motor and divide it into corresponding datasets; Encode each single-modal data to generate the data embedding vector required for orthogonal sequence fusion, and extract the features of the single-modal data;

[0077] Step 2, orthogonal sequence fusion of multimodal data; sort each mode sequentially using the fault diagnosis results of single-modal data in the TABM model; construct the data fusion layer of the permanent magnet synchronous motor fault diagnosis model, form an embedding matrix by the embedding vectors according to the data fusion order, and form the data fusion representation sequentially using the orthogonal sequence fusion mechanism;

[0078] Step 3: Construct a fault diagnosis model for permanent magnet synchronous motors based on the improved TABM model; construct a TABM model framework based on the BatchEnsemble parameter sharing mechanism as the core prediction layer of the motor fault diagnosis model, and introduce a dynamic pruning mechanism to improve the model framework.

[0079] Step 4: Joint training and optimization of the permanent magnet synchronous motor fault diagnosis model; The ADAMW optimizer, joint total loss function, dynamic pruning mechanism and early stopping strategy are used to jointly train and optimize the data fusion layer and the improved TABM model to obtain the optimized permanent magnet synchronous motor fault diagnosis model.

[0080] Step 5: Fault classification probability mapping and diagnostic result output; The optimized permanent magnet synchronous motor fault diagnosis model adds a Softmax function as a fault classification layer. The original fault type prediction scores obtained from the prediction layer are transformed into fault type probability distributions through the Softmax function, thereby achieving effective and accurate identification of various faults of permanent magnet synchronous motors.

[0081] Step 1 of this invention specifically includes:

[0082] The system collects three-phase current signals, vibration signals, infrared thermograms, and stator winding images under various conditions, including normal operation, stator inter-turn short circuit fault, air gap eccentricity fault, and rotor inter-turn short circuit fault.

[0083] In this invention, the multimodal data includes four categories: three-phase current signal, vibration signal, infrared thermogram, and stator winding image.

[0084] Z-score normalization is performed individually for each modality to eliminate differences in dimensions and orders of magnitude among multimodal data. Z-score normalization transforms raw data from different distributions into a standard normal distribution with a mean of 0 and a standard deviation of 1, making the data from different modalities comparable and fusionable. Its expression is as follows:

[0085]

[0086] Where: x i The input data is μ; the mean of the original data is μ; and the standard deviation of the original data is σ. These are the standardized values.

[0087] The processed dataset was then divided into training, validation, and test sets, with corresponding proportions of 70%, 20%, and 10%, respectively.

[0088] Each mode is encoded. Numerical data such as three-phase current signals and vibration signals are processed using dense layers, while image data such as infrared thermograms and stator winding images are processed using the Inceptionv3 module. Infrared thermograms are processed for temperature characteristics, and stator winding images for winding insulation damage characteristics. The standardized features of each mode are mapped to a unified dimension d, generating four different data embedding vectors required for orthogonal sequence fusion, thus completing feature extraction for each mode. The Inceptionv3 module refers to a classic convolutional neural network model for image feature extraction. This model, through optimization techniques such as decomposing large-size convolutional kernels and introducing auxiliary classifiers, can extract multi-scale image features while controlling the number of parameters, meeting the requirements for image data feature extraction and dimension mapping in this invention.

[0089] Step 2 of the present invention specifically includes:

[0090] First, the single-modal data extracted in step 1 is embedded into the TABM model as a vector. The fault diagnosis results output by the TABM model are tested sequentially. The data fusion order is determined based on the accuracy of the output fault diagnosis results.

[0091] Construct a data fusion layer: Based on the embedding vectors of four different types of data, generate an embedding matrix E according to the data fusion order.

[0092] set up An embedding matrix for four modalities, each row... Let be the embedding vector of the t-th modality. Orthogonal sequence fusion successively demasks and incorporates a single modality, while all other modalities remain masked; the initialization form of the corresponding data fusion representation is:

[0093]

[0094] in: It is the initial orthogonal sequence fusion representation. This refers to the embedding vector used during the initial data fusion process.

[0095] The mask is represented as a 4-dimensional binary row vector. Where 0 represents the masking state and 1 represents the demasking state. The embedding vector used in the t-th fusion step. , by binary row vector This is obtained by multiplying with the embedding matrix E. In orthogonal sequence fusion pairs... After demasking, it is combined with the current fused representation. Integration; its expression is:

[0096]

[0097]

[0098]

[0099] in: Let be the 4-dimensional binary row vector at step t; Let be the embedding vector used in the data fusion process at step t; Data at step t ; This is a data fusion function.

[0100] After all four modalities have been demasked and fused, the final data fusion representation is as follows:

[0101]

[0102] Regarding the data fusion loss function, L1-L2 regularization is applied to the fusion layers of N modalities and N-1 modalities. The aim is to ensure that the data fusion embeddings of each layer remain orthogonal, thereby helping to capture richer and more complementary information from each modality. This invention introduces hyperparameters. and These are used to control the constraint strength of the L1 and L2 regularization terms, respectively; four modalities and three modal fusion layers are set up; and optimization is performed using the standard backpropagation algorithm. The corresponding data fusion loss function expression is as follows:

[0103]

[0104] in: This represents the main loss function corresponding to the given task T; preds and outputs represent the model's predicted values ​​and true labels, respectively. This represents the embedding vector used in the data fusion process at step t; Let t represent the data fusion representation at step t.

[0105] The above steps construct the data fusion layer of the final permanent magnet synchronous motor fault diagnosis model. The data fusion result F4 in this step is a d-dimensional flat numerical vector, which requires no additional format conversion and only needs to be aligned with the input dimension of the TABM model before it can be directly input.

[0106] Step 3 of this invention specifically includes:

[0107] This invention introduces a dynamic pruning mechanism during training on top of the TABM model. This mechanism aims to remove sub-models that contribute no or negatively over a long period, thereby improving training efficiency, reducing redundant computation, and enhancing the model's generalization ability. A sub-model refers to multiple parallel prediction branches generated in the TABM model through the BatchEnsemble parameter sharing mechanism. Each branch consists of a corresponding set of private adapter parameters and an independent prediction head. All sub-models share all network weight parameters except for the adapter and prediction head, and each sub-model independently outputs a multi-dimensional original confidence score vector.

[0108] Step 3.1, Build the improved TABM model architecture:

[0109] First, the final data fusion representation F4, obtained by fusing the data embedding vector generated in step 1 with the orthogonal sequence in step 2, is used as the original input to the improved TABM model in this step. This input vector is first copied into k identical copies by the cloning module to form the input matrix. , where each row of X corresponds to the input representation of a sub-model.

[0110] The parameter-sharing mechanism in BatchEnsemble is adopted: all ensemble members share most of the network layer parameters, while retaining some private parameters. Using this parameter ensemble method, all ensemble members can be integrated into a single model framework. This framework can simultaneously receive k copies of the same input sample as input, convert the k identical copies into k different representations, and then execute the operations of the k implicit sub-models. The expression for batch ensemble of one or more linear layers is:

[0111]

[0112] in: This is element-wise multiplication; The matrix represents the result of stacking k copies of the fused representation output from step 2, with each implicit sub-model corresponding to a different representation; S is the multiplication adapter matrix, which is not shared and is used to scale the calculation results of the shared weights W using element-wise multiplication to adjust the output features of each implicit sub-model; R is the multiplication adapter matrix, which is not shared and is used to map k inputs to k different representations, which are then mixed with the weight matrix W in the tabular features; B is the additive adapter bias matrix, which is not shared and is used to provide additive bias for each sub-model to correct feature distribution shifts; Wd×d is the shared weight matrix.

[0113] An MLP is a model composed of N simple modules strung together, ending with a linear prediction head. A block is the basic building block in an MLP, including a linear layer, activation function, and Dropout.

[0114]

[0115]

[0116] Where Dropout is a regularization technique, ReLU is an activation function, and linear(x) is a linear layer.

[0117] In the TABM model framework, the W parameter is shared by all sub-networks. The key to integration lies in the three matrices R, S, and B (also known as adapters). Each MLP has N layers of blocks, and the entire model has a total of 3N adapters.

[0118] Except for the first one The other 2N-1 multiplication adaptation parameters are initialized to 1. These multiplication adaptation parameters, initialized to 1, are continuously learned during subsequent training, providing sub-model-specific adjustment capabilities for deep feature interactions. The predicted value of the TABM model framework is defined as the average of the prediction results of its k implicit sub-models, where the number of sub-models k is set to 32. The following is the overall framework of the core prediction layer of the model built in this invention:

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] in: is the prediction head for each implicit sub-model; k is the total number of sub-models; i is the sub-model index; t is the network layer index; N is the total number of network layers; This is the output of layer 1; This is the multiplication adapter matrix for the first layer; This is the shared weight matrix for layer 1; is the multiplication adapter matrix of layer 1, randomly initialized; X is the data fusion representation vector received by the input layer; Here is the bias matrix of the additive adapter in layer 1; Dropout is a regularization technique; ReLU is the activation function. This is the output of the t-th layer; Let be the multiplication adapter matrix of the t-th layer; Let be the shared weight matrix of layer t; Let be the multiplication adapter matrix of the t-th layer; Let be the bias matrix of the additive adapter at layer t; This is the output of the first layer after ReLU activation and Dropout, which is the input of the next layer; This is the output of the t-th layer after ReLU activation and Dropout; Z is the original confidence score vector of the i-th sub-model; Z is the average of the predictions of all k implicit sub-models, which is used as the final prediction score vector.

[0126] Step 3.2, introduce a dynamic pruning mechanism:

[0127] The dynamic pruning mechanism proposed in this invention periodically evaluates the contribution of each implicit sub-model during training and removes low-contribution sub-models based on the accumulated number of low-contribution instances.

[0128] Dynamic pruning is performed periodically, triggering an evaluation on the validation set at fixed intervals E training epochs. Only sub-model pruning is performed throughout the process; model weight parameters are not updated. In this invention, the initial number of implicit sub-models in the TABM model is set to k=32, and each implicit sub-model outputs an i-dimensional original score vector. The dynamic pruning mechanism determines the contribution of a sub-model by comparing its fault classification accuracy before and after removal.

[0129] The fault classification accuracy after integrating all k sub-models is:

[0130]

[0131]

[0132] Where: y represents the actual fault category of the sample; Let be the model's final predicted category for sample x; P is the diagnostic accuracy indicator function for sample x. The total number of samples in the validation set; The fault diagnosis accuracy of the model (predicted together by all sub-models) on the validation set.

[0133] After temporarily removing the nth sub-model, the accuracy of the ensemble of the remaining k-1 sub-models is:

[0134]

[0135]

[0136] Where: y represents the actual fault category of the sample; This represents the final predicted category of the model for sample x after pruning the nth sub-model. This is the diagnostic accuracy indicator function for sample x after pruning the nth sub-model; This represents the fault diagnosis accuracy of the model (after pruning the nth sub-model) on the validation set.

[0137] The contribution of the nth sub-model is defined as If the sub-model If the value is ≤0, it means that the sub-model has no contribution or a negative contribution and should be pruned; if the sub-model's value is ≤0, it means that the sub-model has no contribution or a negative contribution and should be pruned. If the value is greater than 0, it indicates that the sub-model has a positive contribution and should be retained. Its expression is:

[0138]

[0139] Define a counter for each sub-model The initial value is 0. If the sub-model's If ≤0, then increment the count by 1; if the sub-model's If the count is greater than 0, the count value remains unchanged. When the value is ≥L (the set threshold), the sub-model is permanently pruned.

[0140] like Figure 2 As shown, based on the above orthogonal sequence fusion mechanism and improved TABM model architecture, a fault diagnosis model for permanent magnet synchronous motors is finally constructed. This model consists of three layers, from front to back: a data fusion layer, a core prediction layer, and a fault classification layer. Specifically, the data fusion layer uses the orthogonal sequence fusion mechanism described in step 2 to progressively fuse the multiple single-modal data embedding vectors generated in step 1 into a unified d-dimensional fusion representation vector. The core prediction layer uses the improved TABM model described in step 3, based on the BatchEnsemble parameter sharing mechanism and dynamic pruning mechanism, to perform parallel sub-model inference on the fusion representation vector output by the data fusion layer, outputting k sets of original fault confidence scores, and taking the average as the final fault score prediction value. The fault classification layer uses the Softmax function to transform the fault score prediction value output by the core prediction layer into a probability distribution of various faults, and selects the fault category corresponding to the highest probability as the diagnosis result. The following is the overall model flow: First, the multimodal raw data is standardized and encoded in step 1 to obtain a data embedding vector; then, the data embedding vector is input into the orthogonal sequence fusion module in step 2 to generate a fusion representation vector; next, the fusion representation vector is sent to the improved TABM core prediction layer in step 3 to output the original confidence score of the fault; finally, the original confidence score of the fault is transformed into a fault category probability distribution through the Softmax classification layer in step 5, and the final diagnosis result is output accordingly.

[0141] Step 4 of this invention specifically includes:

[0142] This step involves joint training and optimization of the model based on the data fusion layer in step 2 and the improved TABM model in step 3.

[0143] The training set undergoes feature fusion through a data fusion layer. The resulting fused feature vector is then input into the input layer of the improved TABM model framework, where it is copied k=32 times using a cloning module. Each input copy predicts 32 independent multidimensional raw fault confidence scores using a shared weight matrix, a private adapter matrix, and its own independent prediction head.

[0144] When training the improved TABM model, the loss function is the minimization of cross-entropy for the classification task. The loss is calculated for each of the 32 prediction results, and the average is taken as the TABM classification loss. During model training, the joint total loss function includes the orthogonal sequence fusion loss and the TABM classification loss. The corresponding expressions are as follows:

[0145]

[0146]

[0147] Where: N is the total number of training samples; C is the total number of categories in the classification task; Let i be the true label of the i-th sample. If sample i belongs to category c, then =1, otherwise 0; Let J be the predicted probability of the i-th sample belonging to class c by TABM; J is the classification loss of TABM. The loss is for orthogonal sequence fusion. The total loss is the combined loss.

[0148] After obtaining the joint total loss, gradients are calculated via backpropagation, passing the gradients to the parameters of each layer of the model. During training, a global gradient clipping threshold of 1.0 is used to prevent gradient explosion.

[0149] Based on the gradient information obtained from backpropagation, the ADAMW optimizer is used to simultaneously update the parameters of the data fusion layer, the shared matrix W, and the private parameter matrices S, R, and B of the effective sub-models, thereby achieving parameter optimization.

[0150] After training on each epoch of data, the model's performance is evaluated using a validation set. The classification accuracy of the fault classification results diagnosed by the model is used to evaluate model performance. A dynamic pruning mechanism is triggered on the validation set at fixed training epochs E intervals, pruning low-contribution sub-models and dynamically updating the number of effective sub-models k. Furthermore, an early stopping strategy is employed during validation set training, continuing to train the model until its performance does not significantly improve after several consecutive training epochs and reaches the required level (Patience). In this invention, Patience is set to 16.

[0151] Repeat the above training, validation, pruning, and early stopping process until the model training is complete. Finally, use the test set to perform final performance validation on the trained model, and obtain the optimal fault diagnosis model for permanent magnet synchronous motors optimized by dynamic pruning.

[0152] Step 5 of this invention specifically includes:

[0153] The multimodal data of the permanent magnet synchronous motor to be diagnosed are processed sequentially through data standardization and feature encoding in step 1 and orthogonal sequence fusion in step 2 to obtain a d-dimensional data fusion representation vector. This data fusion representation vector is then input into the optimal model for fault diagnosis of the permanent magnet synchronous motor, and the model outputs the original confidence scores Z for various fault types.

[0154] After training the optimal fault diagnosis model for permanent magnet synchronous motors, a Softmax function is added as a fault classification layer. The original confidence scores of the i states of Z are used as input to the Softmax function, and the output is the label probability distribution of various permanent magnet synchronous motor faults. The expression of the Softmax function is as follows:

[0155] n

[0156] Where: Softmax(Z) i ) represents the output probability value for the i-th fault category; represents the original predicted value for the i-th fault category; n is the total number of fault categories.

[0157] The fault category corresponding to the maximum value in the fault label probability distribution is selected as the final diagnostic result of the model, so as to achieve effective and accurate differentiation and diagnostic output of various fault types such as stator inter-turn short circuit, air gap eccentricity, and rotor inter-turn short circuit of permanent magnet synchronous motor.

[0158] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, alterations, or substitutions made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A motor fault diagnosis method based on orthogonal sequence fusion and improved TABM, characterized in that, include: Step 1: Multimodal data acquisition and feature extraction; Step 2: Multimodal data orthogonal sequence fusion to construct the data fusion layer of the permanent magnet synchronous motor fault diagnosis model; Step 3: Based on the improved TABM model, construct the core prediction layer of the permanent magnet synchronous motor fault diagnosis model; Step 4: Joint training and optimization of the fault diagnosis model for permanent magnet synchronous motors; Step 5: Construct the fault classification layer of the permanent magnet synchronous motor fault diagnosis model to realize fault classification probability mapping and diagnosis result output; Step 1 includes: Multimodal data of permanent magnet synchronous motors under normal and different fault conditions are collected; the multimodal data includes numerical data and image data. Z-score normalization was performed separately for each modality of data; The processed dataset is divided into a training set, a validation set, and a test set; Each modal data is encoded, with numerical data processed using a dense layer and image data processed using the Inceptionv3 module. The standardized features of each modality are mapped to a unified dimension d to generate the data embedding vector required for orthogonal sequence fusion; Step 2 includes: Each data embedding vector is input into the TABM model, and the fault diagnosis results output by the TABM model are tested sequentially. The data fusion order is determined based on the accuracy of the output fault diagnosis results. Constructing a data fusion layer: Based on each data embedding vector, an embedding matrix E is generated according to the data fusion order; where, , is an embedding matrix for 4 modal data; each row Let be the data embedding vector for the t-th modality; during orthogonal sequence fusion of multimodal data, the mask is successively demasked and a single modality is incorporated, while all other modal data remain masked; the initialization form of the orthogonal sequence fusion representation is: in: It is the initial orthogonal sequence fusion representation. This refers to the data embedding vector used in the initial data fusion process; Represent the mask as a binary row vector Where 0 represents the masking state and 1 represents the demasking state; the data embedding vector used in the t-th step of orthogonal sequence fusion. , by binary row vector It is obtained by multiplying with the embedding matrix E; In orthogonal sequence fusion pairs After demasking, it is combined with the current fused representation. Integration; its expression is: Where: M t Let be the binary row vector at step t; Let be the embedding vector used in the orthogonal sequence fusion process at step t; orthogonal sequence fusion representation at step t; For data fusion functions; After all modal data has been demasked and fused, the final orthogonal sequence fusion representation is as follows: in, For the final modal data fusion representation, This represents the third orthogonal sequence fusion characterization. This is the fourth data fusion function. This is the data embedding vector used for the 4th time.

2. The motor fault diagnosis method based on orthogonal sequence fusion and improved TABM according to claim 1, characterized in that, Numerical data includes three-phase current signals and vibration signals; image data includes infrared thermograms and stator winding images.

3. The motor fault diagnosis method based on orthogonal sequence fusion and improved TABM according to claim 1, characterized in that, Step 3 includes: Step 3.1, Based on the improved TABM model, construct the core prediction layer: First, the final data fusion representation F4 obtained by fusing the data embedding vectors generated in step 1 with the orthogonal sequence in step 2 is used as the original input of the improved TABM model. The original input is copied into k identical copies to form the input matrix. Batch integration is performed using the parameter sharing method in the BatchEnsemble mechanism, with the expression as follows: Where: ⊙ represents element-wise multiplication; S is the multiplication adapter matrix; W is the shared weight; R is the multiplication adapter matrix; B is the addition adapter bias matrix; Finally, the core prediction layer is constructed and obtained, represented as follows: in: is the prediction head for each implicit sub-model; k is the total number of sub-models; i is the sub-model index; t is the network layer index; N is the total number of network layers; This is the output of layer 1; This is the multiplication adapter matrix for the first layer; This is the shared weight matrix for layer 1; X is the multiplication adapter matrix of layer 1; X is the data fusion representation vector; Here is the bias matrix of the additive adapter in layer 1; Dropout is a regularization technique; ReLU is the activation function. This is the output of the t-th layer; Let be the multiplication adapter matrix of the t-th layer; Let be the shared weight matrix of layer t; Let be the multiplication adapter matrix of the t-th layer; Let be the bias matrix of the additive adapter at layer t; This is the output of the first layer after ReLU activation and Dropout. This is the output of the t-th layer after ReLU activation and Dropout; The original confidence score vector is output for the i-th sub-model; Z is the average of the predictions of all k implicit sub-models, which is used as the final prediction score vector. Step 3.2, introduce a dynamic pruning mechanism: The dynamic pruning mechanism is executed periodically. An evaluation is triggered on the validation set every fixed training epoch E, pruning the sub-models without updating the model weights. The improved TABM model outputs an i-dimensional prediction value Z for each implicit sub-model. i The dynamic pruning mechanism determines the contribution of a sub-model by comparing the fault classification accuracy before and after removing the sub-model. The fault classification accuracy after integrating all k sub-models is: Where: y represents the actual fault category of the sample; Let be the model's final predicted category for sample x; P is the diagnostic accuracy indicator function for sample x. The total number of samples in the validation set; The fault diagnosis accuracy of the model on the validation set; After removing the nth sub-model, the accuracy of the ensemble of the remaining k-1 sub-models is: Where: y represents the actual fault category of the sample; This represents the final predicted category of the model for sample x after pruning the nth sub-model. This is the diagnostic accuracy indicator function for sample x after pruning the nth sub-model; The fault diagnosis accuracy on the validation set after pruning the nth sub-model; The contribution of the nth sub-model is defined as If the sub-model If ≤0, then pruning is performed; if the sub-model's If the value is greater than 0, then it is retained; the expression is: Define a counter for each sub-model The initial value is 0; if the sub-model's If ≤0, then increment the count by 1; if the sub-model's If the count is greater than 0, the count value remains unchanged; when... If the value is greater than or equal to the set threshold L, the sub-model will be permanently pruned.

4. The motor fault diagnosis method based on orthogonal sequence fusion and improved TABM according to claim 3, characterized in that, Step 4 includes: The permanent magnet synchronous motor fault diagnosis model uses the minimum cross-entropy of the classification task as the loss function, calculates the loss for multiple prediction results separately, and takes the average as the TABM classification loss; when training the model, the joint total loss function includes the orthogonal sequence fusion loss and the TABM classification loss, as expressed below: Where: N is the total number of training samples; C is the total number of categories in the classification task; Let i be the true label of the i-th sample; Let J be the predicted probability of the TABM model that the i-th sample belongs to class c; J is the TABM classification loss. The loss is for orthogonal sequence fusion. The total loss is the combined loss. After obtaining the joint total loss, gradient calculation is performed through backpropagation, and the gradient is passed to the parameters of each layer of the permanent magnet synchronous motor fault diagnosis model. Based on the gradient information obtained from backpropagation, the ADAMW optimizer is used to update and optimize the parameters simultaneously. The classification accuracy of the fault classification results diagnosed by the permanent magnet synchronous motor fault diagnosis model is used to evaluate the model performance; a dynamic pruning mechanism is triggered on the validation set once every fixed training epoch to evaluate the model, pruning low-contribution sub-models and dynamically updating the number of effective sub-models; an early stopping strategy is adopted during validation set training. Repeat the training until the model training is complete to obtain a fault diagnosis model for permanent magnet synchronous motors optimized by dynamic pruning.

5. The motor fault diagnosis method based on orthogonal sequence fusion and improved TABM according to claim 1, characterized in that, Step 5 includes: The multimodal data to be diagnosed is processed sequentially through data standardization and encoding in step 1 and orthogonal sequence fusion in step 2 to obtain a data fusion representation vector. The data fusion representation vector is then input into the permanent magnet synchronous motor fault diagnosis model optimized in step 4 to output the original confidence scores Z for various fault types. A Softmax function is introduced as the fault classification layer; the original confidence scores Z of the i states are used as the input of the Softmax function, and the output is the fault label probability distribution of various permanent magnet synchronous motor faults; The fault category corresponding to the maximum value in the fault label probability distribution is selected as the final diagnostic result.

Images