Permanent magnet synchronous motor current sensor fault diagnosis method based on lightweight double-branch feature fusion neural network

By constructing an adaptive fusion gating mechanism using a lightweight bi-branch feature fusion neural network and a sandworm population-Bayesian optimization collaborative algorithm, we solved the problems of model complexity and manual parameter tuning dependence in the fault diagnosis of current sensors for permanent magnet synchronous motors, and achieved high-precision, real-time fault identification and diagnosis.

CN120993301APending Publication Date: 2025-11-21HEFEI UNIV OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511332893.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-09-18
Publication Date
2025-11-21

AI Technical Summary

Technical Problem

Existing fault diagnosis methods for permanent magnet synchronous motor current sensors suffer from high model complexity and strong reliance on manual parameter tuning, making it difficult to achieve efficient and real-time fault identification and diagnosis.

Method used

A lightweight dual-branch feature fusion neural network is adopted, which combines time-domain and frequency-domain features. The model parameters are automatically tuned using the sandworm population-Bayesian optimization collaborative algorithm (SSA-BOC), and an adaptive fusion gating mechanism is constructed to achieve fault diagnosis.

Benefits of technology

It achieves high-precision fault diagnosis of current sensors, reduces model complexity, and improves the accuracy and real-time performance of fault identification, making it suitable for embedded controllers and edge computing terminals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure QLYQS_1
    Figure QLYQS_1
Patent Text Reader

Abstract

The invention relates to the technical field of permanent magnet synchronous motor current sensor fault diagnosis, in particular to a permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network. The method comprises the steps of obtaining and preprocessing motor current data, constructing a current sensor fault diagnosis model, optimizing the current sensor fault diagnosis model, and performing fault diagnosis by using the current sensor fault diagnosis model. The method can realize high-accuracy fault identification in a small-sample and multi-working-condition environment, has the advantages of low model complexity, fast reasoning speed and strong edge deployment capability, is suitable for online fault diagnosis and predictive maintenance of the current sensor in the permanent magnet synchronous motor driving system, can monitor the health condition of the motor in real time, and has a good application prospect. And the service life of the motor can be prolonged.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of permanent magnet synchronous motor current sensor fault, in particular to a kind of light-weight double-branch feature fusion neural network's permanent magnet synchronous motor current sensor fault diagnosis method. BACKGROUND

[0002] Permanent magnet synchronous motor (PMSM) has been widely used in new energy vehicles, electric locomotives, industrial servo systems and aerospace equipment due to its high power density, high efficiency, low noise and other significant advantages. As a key sensor component in PMSM drive system, current sensor mainly undertakes real-time monitoring and feedback regulation of motor current, which is an important guarantee for high-precision closed-loop control and stable operation of the system.

[0003] However, due to the complex operating environment of PMSM and strong external interference, the current sensor is easily affected by factors such as temperature drift, electromagnetic interference, mechanical impact and line aging, resulting in faults such as sensor breakage, abnormal gain, offset drift and output saturation. Once the sensor fails, it will directly disrupt the balance of the current closed-loop control, leading to performance degradation of the drive system, increased vibration, and even equipment damage or safety accidents. Therefore, studying the fault diagnosis technology of current sensor is of great engineering significance and application value for ensuring the reliability and safety of PMSM system.

[0004] Currently, the detection methods for current sensor faults mainly include three categories: redundant hardware-based detection method, analytical model-based detection method and data-driven intelligent method. Although the first two methods can achieve a certain degree of fault identification, they have the problems of high cost, complex system or strong dependence on model. The third method can achieve good fault identification, but still has the problems of high model complexity and strong dependence on manual parameter adjustment. Therefore, designing a current sensor fault diagnosis method with light-weight model and adaptive parameter adjustment has become a topic worth exploring. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a permanent magnet synchronous motor current sensor fault diagnosis method based on light-weight double-branch feature fusion neural network. The method constructs a double-branch neural network structure and fuses time domain and frequency domain features, while adopting a model parameter automatic tuning strategy of structure-coupled sandworm colony-Bayesian optimization collaborative algorithm (SSA-BOC algorithm), which realizes higher accuracy diagnosis, reduces model complexity and avoids manual parameter adjustment.

[0006] To achieve the above-mentioned purposes, the technical solutions of the present application are as follows:

[0007] In a first aspect of the present application, a permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network is disclosed, which comprises the following steps:

[0008] A permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network, which comprises the following steps:

[0009] S1, data acquisition and preprocessing: acquiring three-phase current signals of a permanent magnet synchronous motor under different working conditions when the current sensor fails, and performing denoising and standardization processing on the signals to obtain processed current signal sequences.

[0010] S2, construction of fault diagnosis model: establishing a current sensor fault diagnosis model based on a lightweight double-branch feature fusion neural network; the current sensor fault diagnosis model comprises a time domain feature extraction branch, a frequency domain feature extraction branch and an adaptive fusion gating mechanism.

[0011] S3, optimization of fault diagnosis model: using a structure-coupled sandworm colony-Bayesian optimization collaborative algorithm (SSA-BOC algorithm) to jointly optimize the structure parameters and training parameters of the current sensor fault diagnosis model.

[0012] S4, current sensor fault diagnosis: inputting the processed current signal sequence into the optimized current fault diagnosis model, using the time domain feature extraction branch and the frequency domain feature extraction branch to extract multi-domain information, and using the adaptive fusion gating mechanism to fuse the multi-domain information, and then diagnosing the current sensor fault, and outputting the fault category.

[0013] As a further improvement of the above technical solution, the step S2 specifically comprises the following steps:

[0014] S21, constructing a time domain feature branch: using an improved lightweight gating unit and a differentiable Fourier network module in series to construct a time domain feature branch;

[0015] S22, constructing a frequency domain feature branch: using a structure combining a shallow multilayer perceptron network and a spectrum gating mechanism, and constructing a frequency domain feature branch using the following formula:

[0016] ;

[0017] Where, F fg is the output result of the frequency domain feature branch, i.e. the frequency domain feature vector, which is obtained by element-wise multiplication of the spectrum gating coefficient G f and the output F MLP of the shallow multilayer perceptron network; the spectrum gating coefficient G f is an adaptive weight generated by the spectrum energy.

[0018] the output F of the shallow multilayer perceptron network MLP , is calculated by the following formula:

[0019] ;

[0020] wherein, F MLP is the output of the shallow multilayer perceptron network, W1, W2 are the weight matrices of the fully connected layer, b1, b2 are the bias terms of the fully connected layer, is the frequency domain input signal, represents a rectified linear unit activation function.

[0021] the spectrum gating coefficient G f , is calculated by the following formula:

[0022] ;

[0023] wherein, G f is the spectrum gating coefficient, is an adaptive weight generated by the spectrum energy; is a Sigmoid function, is a gating weight matrix, is the frequency domain input signal, is a gating bias term, represents element-wise multiplication.

[0024] S23, constructing an adaptive fusion gating mechanism: constructing a mutual information heat matrix to measure the information interaction strength between each channel in the time domain and the frequency domain; based on the mutual information heat matrix, extracting a channel dynamic weighting coefficient , and determining the fusion feature by the following formula:

[0025] ;

[0026] wherein, is the fused feature vector, is the time domain feature vector is the frequency domain feature vector, represents a channel dimension splicing operation, is a channel adaptive weight vector, represents element-wise multiplication.

[0027] As a further improvement of the above technical solution, the step S3 specifically comprises the following steps:

[0028] S31, individual search: the position of the individual in the sandworm colony algorithm is updated by the following formula:

[0029] ;

[0030] wherein, denotes the parameter vector of the i th individual in the t+1 th generation; denotes the parameter vector of the i th individual in the t th generation, denotes the global optimal solution in the current SSA search population; denotes the optimal point predicted by the Bayesian surrogate model; is a control guide and disturbance intensity adjustment factor; is a periodic disturbance term of the individual, is a sine function; N is the population size of SSA.

[0031] S32, Surrogate modeling: a Gaussian process regression model is constructed using the historical search trajectory of the sandworm colony algorithm SSA; based on the Gaussian process regression model, the predicted mean and uncertainty of the candidate hyperparameter point are calculated using the following formula, and the sampling point selection and update of the next generation candidate sample are performed:

[0032] ;

[0033] ;

[0034] wherein, , are the predicted mean and uncertainty of the candidate hyperparameter point respectively; GP is a Gaussian process regression model, is a historical search trajectory dataset of the sandworm colony algorithm SSA, is a sampling point of the next generation candidate sample, is an exploration coefficient; denotes the information density weight between the candidate point and the historical sampling point.

[0035] S33, Joint search: the sandworm colony-Bayesian optimization collaborative algorithm (SSA-BOC algorithm) is used to act on the parameter combination vector set , with the minimum validation set loss function as the optimization objective, and the parameter combination vector set is continuously updated in the training iteration process.

[0036] The minimum validation set loss function is as follows:

[0037] ;

[0038] wherein, denotes the parameter combination vector set; is the optimal parameter combination vector, and each dimension parameter denotes a structure parameter or a training hyperparameter to be optimized; denotes the neural network model output under the parameter combination vector set . is the original input signal; is the evaluation index of the validation set.

[0039] As a further improvement of the above technical solution, the step S4 specifically comprises the following steps:

[0040] S41, input the preprocessed test signal sequence to the optimized current sensor fault diagnosis model.

[0041] S42, forward propagation is performed through the current sensor fault diagnosis model to obtain a fault category prediction result y.

[0042] The fault category prediction result y realizes fault classification by using the following formula:

[0043] ;

[0044] Wherein, y is the fault category prediction result, is the fused feature vector; , are the weights and biases of the classification layer respectively; is the normalized output function; represent gain fault, zero offset fault, broken wire fault, stuck fault and intermittent fault respectively.

[0045] As a further improvement of the above technical solution, the state formula of the improved lightweight gating unit is:

[0046] ;

[0047] ;

[0048] ;

[0049] Wherein, represents the update gate; and represent the current input state and the previous hidden state respectively; and represent the candidate hidden state and the final hidden state respectively, , and , is the shared weight matrix, and and are the bias vectors of the update gate and the candidate hidden state respectively; the operator represents element-wise multiplication; is the Sigmoid function; is the hyperbolic tangent function.

[0050] As a further improvement of the above technical solution, the differentiable Fourier network module adopts the following formula for periodic modeling:

[0051] ;

[0052] Wherein, is a time-domain periodic feature vector, is an amplitude parameter calculated based on frequency attention weight, k is the number of adjustable frequency bases, t is the input time index, is a cosine function, is a trainable frequency base function, is a phase.

[0053] As a further improvement of the above technical solution, the improved lightweight gate unit is connected with the output of the differentiable Fourier network module through a frequency domain residual connection mechanism, and the residual bridge is:

[0054] ;

[0055] Wherein, is a time-domain feature vector, is the output of the improved lightweight gate unit TinyGRU, is the output of the differentiable Fourier network module DFAN, is a learnable gating factor of the residual signal.

[0056] In the second aspect of the application, a permanent magnet synchronous motor current sensor fault diagnosis device based on a lightweight double-branch feature fusion neural network is disclosed, which comprises:

[0057] A data acquisition and preprocessing module, a fault diagnosis model construction module, a fault diagnosis model optimization module and a current sensor fault diagnosis module;

[0058] The data acquisition and preprocessing module is used to acquire three-phase current signals of the permanent magnet synchronous motor under different working conditions when the current sensor fails, and to denoise and standardize the signals to obtain a processed current signal sequence.

[0059] The fault diagnosis model construction module is used to establish a current sensor fault diagnosis model based on a lightweight double-branch feature fusion neural network; the current sensor fault diagnosis model comprises a time-domain feature extraction branch, a frequency-domain feature extraction branch and an adaptive fusion gate mechanism.

[0060] The fault diagnosis model optimization module is configured to adopt a structure-coupled sand-worm colony-Bayesian optimization collaborative algorithm to jointly optimize the structure parameters and training parameters of the current sensor fault diagnosis model.

[0061] The current sensor fault diagnosis module is configured to input the processed current signal sequence into the optimized current fault diagnosis model, extract multi-domain information by using a time-domain feature extraction branch and a frequency-domain feature extraction branch, perform feature fusion on the multi-domain information by using an adaptive fusion gate mechanism, and then perform fault diagnosis on the current sensor and output a fault category.

[0062] In a third aspect of the present application, an electronic device is disclosed, comprising: 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 permanent magnet synchronous motor current sensor fault diagnosis method based on the lightweight double-branch feature fusion neural network.

[0063] In a fourth aspect of the present application, a machine-readable storage medium is disclosed, characterized in that it stores executable instructions that, when executed, cause the machine to perform the permanent magnet synchronous motor current sensor fault diagnosis method based on the lightweight double-branch feature fusion neural network.

[0064] Compared with the prior art, the present application has the following advantages:

[0065] (1) The double-branch neural network model constructed by the present application adopts a heterogeneous design in structure, and the time-domain branch combines an improved TinyGRU and a differentiable Fourier attention network (DFAN), thereby realizing efficient modeling of non-stationary dynamics and periodic characteristics; the frequency-domain branch combines a shallow perceptron and a spectrum energy gate mechanism, thereby focusing on abnormal frequency band signals and improving the perception sensitivity to high-frequency disturbance faults.

[0066] (2) The present application introduces an adaptive fusion gate mechanism (AFGM), which realizes deep fusion and selective reinforcement of feature levels by constructing a time-frequency mutual information heat map and generating channel dynamic weights, thereby improving the comprehensive expression ability of the model to multi-source heterogeneous features.

[0067] (3) The present application realizes adaptive joint adjustment of model structure parameters and training hyperparameters by using a structure-coupled sand-worm colony-Bayesian optimization collaborative algorithm (SSA-BOC algorithm), thereby overcoming the problems of low efficiency and dependence on artificial experience in traditional parameter adjustment, and enhancing the generalization and adaptation ability of the model in different platforms and scenarios.

[0068] (4) Compared with existing fault diagnosis technologies, the neural network proposed in this invention has the characteristics of low model complexity and fast inference speed, and is suitable for embedded controllers and edge computing terminals, meeting the industrial needs of real-time fault monitoring and remote operation and maintenance management. Attached Figure Description

[0069] Figure 1 This is a flowchart of the method of the present invention;

[0070] Figure 2 This is a schematic diagram of the simulation results of the current sensor fault diagnosis model. Detailed Implementation

[0071] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0072] Example 1

[0073] like Figure 1 The method for fault diagnosis of permanent magnet synchronous motor current sensor based on lightweight dual-branch feature fusion neural network shown includes the following steps:

[0074] S1. Data Acquisition and Preprocessing: Acquire the three-phase current signals of the permanent magnet synchronous motor under different operating conditions when the current sensor fails, and perform noise reduction and standardization on the original current signals to obtain the processed current signal sequence.

[0075] In step S1, the three-phase current signals of the permanent magnet synchronous motor (PMSM) are acquired under different operating conditions such as rated load, variable load, and start / brake, when the current sensor experiences gain fault, zero bias fault, open circuit fault, jamming fault, and intermittent fault. In this embodiment, wavelet transform is used to denoise the original three-phase current signal to eliminate electromagnetic interference and measurement noise. Then, the denoised signal is normalized to a unified data scale using the Z-score normalization method, finally obtaining the processed current signal sequence x(t).

[0076] S2. Construction of Fault Diagnosis Model: Establish a current sensor fault diagnosis model based on a lightweight dual-branch feature fusion neural network; the current sensor fault diagnosis model includes a time-domain feature extraction branch, a frequency-domain feature extraction branch, and an adaptive fusion gating mechanism.

[0077] In the current sensor fault diagnosis model described in the application, the time domain feature extraction branch is constructed by connecting an improved TinyGRU (reduced gate recurrent unit) and a differentiable Fourier network (DFAN) module in series, which is used to capture the time sequence change characteristics and local fluctuation information of the current signal. The frequency domain feature extraction branch adopts a shallow multilayer perceptron network (MLP) combined with a spectrum energy gating strategy, which highlights the energy proportion of the fault-related frequency band and focuses on key frequency domain features. The adaptive fusion gating mechanism (AFGM) constructs a heat map by calculating the mutual information between time-frequency features, and combines a dynamic channel weight distribution algorithm to realize adaptive weighted fusion of time-frequency features, avoiding invalid feature interference.

[0078] S3, fault diagnosis model optimization: a structure-coupled sandworm population-Bayesian optimization collaborative algorithm (SSA-BOC algorithm) is used to jointly optimize the structure parameters and training parameters of the current sensor fault diagnosis model. The SSA-BOC algorithm is a population collaborative search method based on proxy modeling guidance.

[0079] In step S3, the structure parameters and training parameters of the current sensor fault diagnosis model are collaboratively optimized to achieve the highest model fault diagnosis accuracy and the smallest parameter quantity, thereby improving the model performance and running efficiency. The structure parameters include the number of hidden layer neurons of the improved TinyGRU, the number of Fourier bases of the DFAN, the number of layers and nodes of the MLP, etc. The training parameters include the initial learning rate, the batch size, the number of iterations, etc.

[0080] S4, current sensor fault diagnosis: the processed current signal sequence is input into the optimized current fault diagnosis model, which uses the time domain feature extraction branch and the frequency domain feature extraction branch to extract multi-domain information, and uses the adaptive fusion gating mechanism to fuse the multi-domain information, thereby diagnosing the fault of the current sensor and outputting the fault category.

[0081] In step S4, the preprocessed current signal sequence x(t) is input to the time domain feature extraction branch and the frequency domain feature extraction branch of the optimized fault diagnosis model at the same time. The time domain feature extraction branch is responsible for outputting the time domain feature vector representing the timing dynamics, and the frequency domain feature extraction branch outputs the frequency domain feature vector reflecting the frequency spectrum characteristics. The feature vectors of the two branches are intelligently fused through an adaptive fusion gating mechanism: first, based on the mutual information heat matrix, a channel weight is generated to weight and fuse the time domain and frequency domain features to generate a multi-domain fusion feature matrix with stronger discriminability. The fusion feature is then input into the full connection layer for dimension conversion and feature abstraction, which is compressed into a low-dimensional feature vector corresponding to the five types of faults. Finally, the Softmax normalization layer is processed: this layer converts the low-dimensional feature vector into a probability distribution to generate the probability distribution of gain fault, zero offset fault, broken line fault, stuck fault and intermittent fault, ensuring that the sum of all probabilities is 1. The system automatically selects the fault category corresponding to the maximum probability as the final diagnosis result, realizing accurate identification and quantitative output of fault types. The whole process can be completed within 2ms on the edge device, meeting the real-time online diagnosis requirements, while maintaining a high accuracy of 98.6%, effectively solving the problems of complex model and high diagnosis delay of traditional methods. This design cooperates through dynamic feature weighting and lightweight classification, which not only strengthens the fault sensitive features but also avoids redundant calculation, so that the system still maintains excellent generalization ability in multi-working-condition scenarios.

[0082] As a further improvement of the above technical solution, the step S2 specifically comprises the following steps:

[0083] S21, constructing a time domain feature branch: a modified lightweight gate unit (TinyGRU) and a differentiable Fourier network (DFAN) module are combined in series to construct the time domain feature branch.

[0084] Further, the state formula of the modified lightweight gate unit is: ;

[0085] ;

[0086] ;

[0087] wherein, represents the update gate; and represent the current input state and the previous hidden state, respectively; and represent the candidate hidden state and the final hidden state, respectively, , and , are shared weight matrices, and and are the bias vectors for updating gate and candidate hidden state respectively; operator denotes element-wise multiplication; is the Sigmoid function; is the hyperbolic tangent function.

[0088] TinyGRU is a lightweight gate unit after structure pruning. Compared with the standard gate unit GRU model, TinyGRU only retains a single update gate , removes the reset gate, and adopts channel reparameterization, so that the state update formula becomes the above formula. The improved lightweight gate unit TinyGRU reduces the parameter amount by about 41.8% compared with the standard GRU by fusing the candidate state calculation path and sharing the weight matrix.

[0089] Further, the differentiable Fourier network module adopts the following formula for periodic modeling:

[0090] ;

[0091] wherein, is a time-domain periodic feature vector, is an amplitude parameter calculated based on frequency attention weight, k is the number of adjustable frequency bases, and t is the input time index, is a cosine function, is a trainable frequency base function, is a phase.

[0092] The base function of the traditional Fourier expansion (non-parametric form) is fixed and untrainable; while the DFAN module introduces a trainable frequency base function and a phase During the model training process, the data characteristics are automatically adjusted and optimized, so that the Fourier expansion adapts to the data regularity of the specific task. The DFAN module replaces the structure of the traditional non-parametric Fourier expansion, introduces a trainable frequency base function and a phase . The DFAN module uses the superposition of multiple cosine functions to flexibly and adaptively model the periodic signal through the trainable frequency, phase parameters and attention weighted amplitude.

[0093] Further, the output between the improved lightweight gate unit and the differentiable Fourier network module is connected through a frequency domain residual connection mechanism, and the residual bridge of the two is:

[0094] ;

[0095] wherein, is a time-domain feature vector, The output of the improved lightweight gate unit TinyGRU, The output of the differentiable Fourier network module DFAN, The learnable gating factor of the residual signal.

[0096] S22, constructing a frequency domain feature branch: adopting a structure combining a shallow multilayer perceptron network MLP and a spectrum gating mechanism, constructing a frequency domain feature branch by using the following formula:

[0097] ;

[0098] Wherein, F fg is the output result of the frequency domain feature branch, i.e. the frequency domain feature vector, which is obtained by element-wise multiplication of the spectrum gating coefficient G f and the output F MLP of the shallow multilayer perceptron network; the spectrum gating coefficient G f is an adaptive weight generated by spectrum energy.

[0099] The output F MLP of the shallow multilayer perceptron network is calculated by using the following formula:

[0100] ;

[0101] Wherein, F MLP is the output of the shallow multilayer perceptron network, W1 and W2 are weight matrices of the fully connected layer, b1 and b2 are bias terms of the fully connected layer, is the frequency domain input signal, denotes a rectified linear unit activation function.

[0102] In order to suppress redundant frequency point learning, a frequency gating mechanism is introduced, and an adaptive weight, i.e. a spectrum gating coefficient G f , is generated according to spectrum energy. The spectrum gating coefficient G f is calculated by using the following formula:

[0103] ;

[0104] Wherein, G f is the spectrum gating coefficient, and is an adaptive weight generated by spectrum energy; is a Sigmoid function, is a gating weight matrix, is the frequency domain input signal, is a gating bias term, denotes element-wise multiplication.

[0105] S23, constructing an adaptive fusion gating mechanism: constructing a mutual information heat matrix The mutual information heat matrix is used to measure the information interaction intensity between each channel in the time domain and the frequency domain And the fusion feature is determined by using the following formula:

[0106] ;

[0107] Wherein, is the feature vector after fusion, is the feature vector in the time domain and the frequency domain respectively, represents a channel dimension splicing operation, is a channel adaptive weight vector, represents an element-wise multiplication.

[0108] As a further improvement of the above technical solution, the step S3 specifically comprises the following steps:

[0109] The adaptive fusion gating mechanism AFGM (Adaptive Fusion Gating Mechanism) is introduced in the feature fusion part, and the gating channel selection mechanism is introduced at the same time, the fusion bias is determined according to the channel interaction score, so as to improve the feature expression selectivity of the current sensor fault model in the multi-working-condition scene.

[0110] As a further improvement of the above technical solution, the step S3 specifically comprises the following steps:

[0111] S31, individual search: the position of the individual in the sandworm colony algorithm is updated by using the following formula:

[0112] ;

[0113] Wherein, represents the parameter vector of the i-th individual in the t+1 generation; represents the parameter vector of the i-th individual in the t generation, represents the global optimal solution in the current SSA search population; represents the optimal point predicted by the Bayesian proxy model; is a regulation factor of control guidance and disturbance intensity; is a periodic disturbance term of the individual, is a sine function; N is the population size of SSA.

[0114] In the individual search stage, the Bayesian proxy model prediction value is introduced to participate in the update process of the position of the individual in the sandworm colony algorithm (SSA), and a "guided search-diversified exploration" fusion mechanism is realized, and the update formula is shown in the above formula. The first half of the individual introduces the proxy guidance term to enhance the convergence, and the second half retains the nonlinear exploration ability of SSA itself.

[0115] S32, agent modeling: a Gaussian process regression model is constructed using the historical search trajectory of the sandworm colony algorithm (SSA); based on the Gaussian process regression model, the predicted mean and uncertainty of the candidate hyperparameter point are calculated using the following formula, and the sampling point selection and update of the next generation of candidate samples are performed:

[0116] ;

[0117] ;

[0118] wherein, , are the predicted mean and uncertainty of the candidate hyperparameter point, respectively; GP is the Gaussian process regression model, is the historical search trajectory data set of the SSA, is the sampling point of the next generation of candidate samples, is the exploration coefficient; represents the information density weight between the candidate point and the historical sampling point.

[0119] In step S32, a Gaussian process regression model is constructed using the historical search trajectory of the sandworm colony algorithm (SSA), and the predicted mean and uncertainty of the candidate hyperparameter point are calculated based on the model, combined with the exploration coefficient and the information density weight between the candidate point and the historical sampling point, to determine the next sampling point . This can achieve a dynamic balance between proxy prediction accuracy and diversity exploration, thereby more efficiently performing subsequent optimization and related work, improving the rationality and effectiveness of sampling point selection, and providing a better foundation for subsequent tasks based on these sampling points (such as model training, parameter tuning, etc.). The sampling strategy used in step S32 achieves a dynamic balance between proxy prediction accuracy and diversity exploration.

[0120] S33, joint search: the sandworm colony-Bayesian optimization collaborative algorithm (SSA-BOC algorithm) is used to act on the parameter combination vector set , with the minimum validation set loss function as the optimization objective, and the parameter combination vector set is continuously updated during the training iteration process.

[0121] The minimum validation set loss function is as follows:

[0122] ;

[0123] wherein, represents the parameter combination vector set, for the optimal parameter combination vector, each dimension parameter denotes a structure parameter to be optimized or a training hyperparameter, denotes a set of parameter combination vectors under the neural network model output; is the original input signal, is the evaluation index of the validation set.

[0124] The SSA-BOC collaborative optimization algorithm acts on the set of parameter combination vectors, and the optimization goal of the algorithm is to minimize the validation set loss function. In the training iteration process, the SSA-BOC collaborative process can continuously update and adjust , realizing joint search.

[0125] As a further improvement of the above technical solution, the step S4 specifically comprises the following steps:

[0126] S41, input the preprocessed test signal sequence x0(t) to the optimized current sensor fault diagnosis model.

[0127] S42, forward propagation is performed through the current sensor fault diagnosis model to obtain a fault category prediction result y;

[0128] The fault category prediction result y realizes fault classification by the following formula:

[0129] ;

[0130] wherein, is the fused feature vector; , are the weights and biases of the classification layer, respectively; is a normalized output function; denote gain fault, zero offset fault, broken wire fault, stuck fault, and intermittent fault, respectively.

[0131] As shown in Figure 2 , Raw Singal represents using original data features, Frea Domain represents using frequency domain branch features alone, Time Domain represents using time domain branch features alone, and Fusion represents using time-frequency domain branch fusion features. It can be seen from Figure 2 that the training result of the current sensor fault diagnosis model using time-frequency domain branch fusion features is optimal.

[0132] Table 1

[0133]

[0134] Table 1 shows a comparison of the number of model parameters and fault diagnosis accuracy between the current sensor fault model proposed in this invention and the current mainstream lightweight fault diagnosis models. As can be seen from Table 1, the current sensor fault diagnosis model built in this invention achieves the highest fault diagnosis accuracy while maintaining the minimum number of model parameters.

[0135] In summary, the fault diagnosis method of this invention first acquires and preprocesses the three-phase current signal of a permanent magnet motor to extract time-frequency domain composite features. Then, it constructs a heterogeneous dual-branch lightweight neural network model—a current sensor fault diagnosis model—and uses the sandworm swarm-Bayesian optimization collaborative algorithm (SSA-BOC algorithm) to fine-tune the model parameters. The model includes a time-domain branch based on joint modeling of an improved TinyGRU and a differentiable Fourier attention network (DFAN), and a frequency-domain branch that uses a shallow multilayer perceptron network (MLP) combined with a spectral energy gating strategy to focus on abnormal frequency band information. This invention employs an adaptive fusion gating mechanism (AFGM) to achieve adaptive weighted fusion between time-frequency features by constructing a mutual information heatmap and dynamic channel weights. Simultaneously, it uses a structurally coupled sandworm swarm-Bayesian optimization collaborative algorithm (SSA-BOC algorithm) to achieve the joint optimal configuration of model structure parameters and training parameters. This invention can achieve high-accuracy fault identification in small sample and multi-condition environments. It has the advantages of low model complexity, fast inference speed and strong edge deployment capability. It is suitable for online fault diagnosis and predictive maintenance of current sensors in permanent magnet synchronous motor drive systems, and can monitor the health status of the motor in real time, thus extending the service life of the motor.

[0136] Example 2

[0137] In a second aspect of the present invention, a fault diagnosis device for a current sensor of a permanent magnet synchronous motor based on a lightweight dual-branch feature fusion neural network is disclosed. The device includes: a data acquisition and processing module, a fault diagnosis model construction module, a fault diagnosis model optimization module, and a current sensor fault diagnosis module.

[0138] The data acquisition and preprocessing module is used to acquire the three-phase current signal of the permanent magnet synchronous motor under different operating conditions when the current sensor fails, and to perform noise reduction and standardization processing on the signal to obtain the processed current signal sequence.

[0139] The fault diagnosis model construction module is used to establish a current sensor fault diagnosis model based on a lightweight dual-branch feature fusion neural network; the current sensor fault diagnosis model includes a time-domain feature extraction branch, a frequency-domain feature extraction branch, and an adaptive fusion gating mechanism;

[0140] The fault diagnosis model optimization module is configured to adopt a structure-coupled sandworm colony-Bayesian optimization collaborative algorithm to jointly optimize the structure parameters and training parameters of the current sensor fault diagnosis model.

[0141] The current sensor fault diagnosis module is configured to input the processed current signal sequence into the optimized current fault diagnosis model, extract multi-domain information by using a time-domain feature extraction branch and a frequency-domain feature extraction branch, perform feature fusion on the multi-domain information by using a self-adaptive fusion gate mechanism, and then perform fault diagnosis on the current sensor and output a fault category.

[0142] Embodiment three

[0143] An electronic 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 above-described permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network.

[0144] In this embodiment, the electronic device can include, but is not limited to, a personal computer, a server computer, a workstation, a desktop computer, a laptop computer, a notebook computer, a mobile computing device, a smart phone, a tablet computer, a cellular phone, a personal digital assistant (PDA), a handheld device, a messaging device, a wearable computing device, a consumer electronic device, and the like.

[0145] Embodiment four

[0146] A machine-readable storage medium stores executable instructions that, when executed, cause the machine to perform the above-described permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network.

[0147] In particular, a system or apparatus can be provided with a readable storage medium on which software program codes for implementing the functions of any of the above embodiments are stored, and a computer or processor of the system or apparatus is caused to read out and execute the instructions stored in the readable storage medium. In this case, the program codes read from the readable medium can themselves implement the functions of any of the above embodiments, and thus the machine readable code and the readable storage medium storing the machine readable code constitute a part of the present specification. Embodiments of the readable storage medium include floppy disks, hard disks, optical disks (e.g., CD-ROMs, CD-Rs, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD-RWs), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program codes can be downloaded from a server computer or a cloud over a communication network. Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, and the like) embodying computer readable program codes. The present application is described in reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each flow and / or block in the flowcharts and / or block diagrams, and a combination of flows and / or blocks in the flowcharts 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, a special purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowcharts and / or block diagrams flow or flows and / or block or blocks. Figure 1 Figure 1 Figure 1 Figure 1

[0148] Figure 1 ​​​​​one or more processes and / or blocks Figure 1 the steps of the functions specified in the one or more blocks.

[0149] The foregoing merely illustrates the principles of the application and application of its more particular aspects and embodiments. Those skilled in the art will appreciate that the application is not limited to these embodiments and examples and that various modifications and equivalents can be used without departing from the true spirit and scope of the application. The scope of the application is to be afforded the broadest interpretation so as to encompass all such modifications and equivalents. The specification and drawings are, accordingly to be regarded in an illustrative rather than a restrictive sense.

Claims

1. A method for permanent magnet synchronous motor current sensor fault diagnosis based on a lightweight double-branch feature fusion neural network, characterized in that, The method comprises the following steps: S1, data acquisition and preprocessing: acquiring three-phase current signals of a permanent magnet synchronous motor under different working conditions when a current sensor fails, and performing denoising and standardization processing on the signals to obtain a processed current signal sequence; S2, construction of a fault diagnosis model: establishing a current sensor fault diagnosis model based on a lightweight double-branch feature fusion neural network; the current sensor fault diagnosis model comprises a time domain feature extraction branch, a frequency domain feature extraction branch, and an adaptive fusion gating mechanism; S3, optimization of the fault diagnosis model: using a structure-coupled sandworm colony-Bayesian optimization collaborative algorithm to jointly optimize the structure parameters and training parameters of the current sensor fault diagnosis model; S4, current sensor fault diagnosis: inputting the processed current signal sequence into the optimized current fault diagnosis model, using the time domain feature extraction branch and the frequency domain feature extraction branch to extract multi-domain information, and using the adaptive fusion gating mechanism to fuse the multi-domain information, and then diagnosing the fault of the current sensor, and outputting the fault category.

2. The permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network according to claim 1, characterized in that: S2, specifically comprising the following steps: S21, constructing a time domain feature branch: using an improved lightweight gating unit and a differentiable Fourier network module in series to construct a time domain feature branch; S22, constructing a frequency domain feature branch: using a structure combining a shallow multilayer perceptron network and a spectrum gating mechanism, and constructing a frequency domain feature branch using the following formula: ; wherein F fg is the output result of the frequency domain feature branch, i.e., the frequency domain feature vector, generated by the spectrum gating coefficient G f and the output F MLP of the shallow multi-layer perceptron network; the spectrum gating coefficient G f is an adaptive weight generated by the spectrum energy; the output F of the shallow multilayer perceptron network MLP is calculated using the following equation: ; where F MLP is the output of the shallow multilayer perceptron network, W1, W2 are the weight matrices of the fully connected layers, b1, b2 are the bias terms of the fully connected layers, is the frequency domain input signal, denotes the rectified linear unit activation function; The spectral gating coefficient G f is calculated using the following equation: ; where G f is a spectral gating coefficient, is an adaptive weight generated by spectral energy; is a Sigmoid function; is a gating weight matrix; is a frequency domain input signal; is a gating bias term; denotes element-wise multiplication; S23, constructing an adaptive fusion gating mechanism: constructing a mutual information heat matrix , to measure the information interaction strength between each channel in the time domain and the frequency domain; based on the mutual information heat matrix, extracting a channel dynamic weighting coefficient , and determining the fusion feature using the following formula: ; wherein, is the fused feature vector, is the time domain feature vector is the frequency domain feature vector, denotes a channel dimension concatenation operation, is the channel adaptive weight vector, denotes an element-wise multiplication.

3. The permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network according to claim 1, characterized in that: S3, specifically comprising the following steps: S31, individual search: updating the position of an individual in the sandworm colony algorithm using the following formula: ; wherein, represents the parameter vector of the i-th individual in the t+1-th generation; represents the parameter vector of the i-th individual in the t-th generation, represents the global optimal solution in the current SSA search population; represents the optimal point predicted by the Bayesian surrogate model; is a control guide and a disturbance intensity adjustment factor; is a periodic disturbance term of the individual, is a sine function; N is the population size of the SSA; S32, proxy modeling: using the historical search trajectory of the sandworm colony algorithm SSA to construct a Gaussian process regression model; based on the Gaussian process regression model, calculating the predicted mean and uncertainty of the candidate hyperparameter point using the following formula, and selecting and updating the sampling points of the next generation of candidate samples: ; ; wherein, , are the predicted mean and uncertainty of the candidate hyperparameter point, respectively; GP is a Gaussian process regression model; is the historical search trajectory dataset of the sandworm colony algorithm SSA; is the sampling point of the next generation candidate sample, is the exploration coefficient; denotes the information density weight between the candidate point and the historical sampling point. S33, joint search: using the earthworm colony-bayesian optimization collaborative algorithm to act on the parameter combination vector set continuously updating the parameter combination vector set in the training iteration process to minimize the validation set loss function as the optimization objective The minimum validation set loss function is as follows: ; wherein, denotes a set of parameter combination vectors; is an optimal parameter combination vector, each dimension parameter denotes a structure parameter to be optimized or a training hyperparameter; denotes a set of parameter combination vectors under the neural network model output; is an original input signal; is a validation set evaluation indicator.

4. The permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network according to claim 1, characterized in that: S4, specifically comprising the following steps: S41. The preprocessed test signal sequence Input into the optimized current sensor fault diagnosis model; S42, performing forward propagation through the current sensor fault diagnosis model to obtain a fault category prediction result y; The failure category prediction result Failure classification is achieved using the following formula: ; wherein y is the failure category prediction result, is the fused feature vector; , are the weight and bias of the classification layer, respectively; is the normalized output function; respectively represent the gain fault, zero offset fault, open circuit fault, stuck fault, and intermittent fault five types of labels.

5. The permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network according to claim 2, characterized in that: The state formula of the improved lightweight gating unit is: ; ; ; wherein, denotes an update gate; and denote a current input state and a previous hidden state, respectively; and denote a candidate hidden state and a final hidden state, respectively; , and , is a shared weight matrix, while and are bias vectors for the update gate and the candidate hidden state, respectively; the operator denotes element-wise multiplication; is a Sigmoid function; is a hyperbolic tangent function.

6. The permanent magnet synchronous motor current sensor fault diagnosis method based on a lightweight double-branch feature fusion neural network according to claim 2, characterized in that: The differentiable Fourier network module uses the following formula for periodic modeling: ; wherein, is a time-domain periodicity feature vector, is an amplitude parameter based on frequency attention weight computation, k is the number of adjustable frequency bases, t is the input time index, is a cosine function, is a trainable frequency base function, is a phase.

7. The method of claim 2, wherein the improved lightweight gating unit is connected to the output of the differentiable Fourier network module through a frequency domain residual connection mechanism, and the residual bridge is: The device comprises a data acquisition and preprocessing module, a fault diagnosis model construction module, a fault diagnosis model optimization module, and a current sensor fault diagnosis module. ; wherein, is a time domain feature vector, is an output of the improved lightweight gate unit TinyGRU, is an output of the differentiable Fourier network module DFAN, is a learnable gating factor of the residual signal.

8. A permanent magnet synchronous motor current sensor fault diagnosis device based on a lightweight double-branch feature fusion neural network, characterized in that, The data acquisition and preprocessing module is configured to acquire three-phase current signals of the PMSM under different working conditions when the current sensor fails, and to perform denoising and standardization processing on the signals to obtain a processed current signal sequence. The fault diagnosis model construction module is configured to establish a current sensor fault diagnosis model based on a lightweight double-branch feature fusion neural network. The fault diagnosis model optimization module is configured to use a structure-coupled sandworm colony-Bayesian optimization collaborative algorithm to jointly optimize the structure parameters and training parameters of the current sensor fault diagnosis model. The current sensor fault diagnosis module is configured to input the processed current signal sequence into the optimized current fault diagnosis model, extract multi-domain information using the time domain feature extraction branch and the frequency domain feature extraction branch, and perform feature fusion on the multi-domain information through the adaptive fusion gating mechanism, thereby performing fault diagnosis on the current sensor and outputting a fault category. comprises:

9. An electronic device, comprising: 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 method of any one of claims 1 to 7. The machine has executable instructions stored therein, which, when executed, cause the machine to perform the method of any one of claims 1 to 7.

10. A machine-readable storage medium, characterized in that, ​

Citation Information

Cited By

  • Battery fault diagnosis method, device and system based on artificial intelligence

    CN121454355A