A permanent magnet motor inter-turn short circuit and demagnetization fault diagnosis method based on CPO-WPD and BO-MSCNN

Abstract

The present application belongs to the technical field of permanent magnet motor fault diagnosis, and more particularly relates to a permanent magnet motor inter-turn short circuit and demagnetization fault diagnosis method based on CPO-WPD and BO-MSCNN. The method comprises: collecting the induced electromotive force signals of the permanent magnet motor under different fault states at different speeds; dividing into a training set, a validation set and a test set, and constructing a fault diagnosis data set; inputting the training set into the CPO-WPD module to iteratively optimize the double-target fitness function and obtain the optimal WPD parameters; inputting the training set processed by the optimal WPD into the BO-MSCNN module to adaptively search for the optimal hyperparameters; constructing a model architecture based on the optimal hyperparameters; inputting the data set processed by the optimal WPD into the model to output the diagnosis result. The present application solves the core bottleneck problems of the prior art, such as WPD parameter dependence on experience, insufficient adaptive capacity, inefficient hyperparameter optimization and difficulty in adapting to front-end features.

Description

Technical Field

[0001] This invention belongs to the technical field of permanent magnet motor fault diagnosis, and more specifically, relates to a method for diagnosing inter-turn short circuit and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN. Background Technology

[0002] Permanent magnet motors, with their advantages of high electromagnetic torque, high efficiency density, and high power factor, have been widely used in high-end equipment fields such as electric vehicles and aerospace. However, under complex working conditions and harsh operating environments, these motors are prone to two typical faults: inter-turn short circuits and local demagnetization. Both of these faults can cause current harmonic distortion and abnormal local temperature rise, exhibiting highly coupled fault characteristics. Existing research often focuses on designing diagnostic solutions for single faults, making it difficult to effectively distinguish between the two types of faults and provide early warnings. This highlights the urgent need for precise maintenance and ensuring reliable system operation.

[0003] Chinese patent document CN117192363A discloses a fault diagnosis method for permanent magnet synchronous traction motors based on transfer learning, including 1) multi-sensor signal acquisition; 2) signal preprocessing; 3) wavelet transform feature extraction; 4) convolutional neural network training; 5) transfer learning application; 6) fine-tuning network parameters; and 7) fault type diagnosis. This method aims to improve the fault detection performance of permanent magnet synchronous motors, is applicable to high-speed train traction systems, and can reduce the risk of accidents caused by permanent magnet motor faults, ensuring the smooth and safe operation of high-speed trains.

[0004] Data-driven intelligent fault diagnosis methods offer an effective approach to solving the aforementioned problems. Their core processes include signal acquisition, feature extraction, and state classification. Wavelet packet decomposition (WPD) can achieve multi-resolution analysis across the entire frequency band, effectively extracting weak fault features from complex electromagnetic signals. However, its key parameters, such as the number of decomposition layers and wavelet basis functions, are highly dependent on expert experience and lack adaptive optimization capabilities. This can easily lead to insufficient feature extraction or redundant feature dimensions, thus limiting the accuracy and efficiency of subsequent diagnosis.

[0005] In the state classification stage, the Multi-Scale Convolutional Neural Network (MSCNN) has demonstrated good performance in the field of motor fault diagnosis by capturing fault features of different ranges through parallel multi-scale convolutional kernels. However, the hyperparameters of MSCNN, such as kernel size, number of channels, and learning rate, have a significant impact on the model's generalization ability and diagnostic effect. Traditional methods such as network search and random search are inefficient and it is difficult to find the optimal combination of hyperparameters that matches the front-end WPD feature extraction, resulting in the model's performance not being fully realized.

[0006] Therefore, existing technologies suffer from two major bottlenecks: WPD parameters rely on experience and lack adaptive capabilities, while MSCNN hyperparameter optimization is inefficient and difficult to adapt to front-end features. These bottlenecks severely restrict the practical application value of accurate diagnosis of inter-turn short circuits and local demagnetization faults in permanent magnet motors. Summary of the Invention

[0007] This invention aims to overcome at least one of the defects of the prior art and provide a method for diagnosing inter-turn short circuits and demagnetization faults of permanent magnet motors based on CPO-WPD and BO-MSCNN, so as to solve the two major bottleneck problems of existing technologies: WPD parameters rely on experience and have insufficient adaptive ability, and MSCNN has low efficiency in hyperparameter optimization and difficulty in adapting to front-end features.

[0008] The detailed technical solution of this invention is as follows: A method for diagnosing inter-turn short circuits and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN, the method comprising: S1. Collect induced electromotive force signals of permanent magnet motors under different speeds in normal state, inter-turn short circuit fault, and local demagnetization fault to construct a fault diagnosis dataset, and divide it into training set, validation set and test set according to a preset ratio; S2. Set the CPO population size and maximum number of iterations, and predefine the parameter search space for the WPD wavelet basis function candidate set and the number of decomposition layers; take the training set data as input, use the wavelet packet decomposition parameters as the optimization space, and adaptively determine the optimal WPD decomposition parameters through the CPO algorithm; S3. Call the obtained optimal WPD decomposition parameters to perform WPD decomposition on the training set data to obtain sub-band signals that are suitable for the corresponding classification features; take the sub-band signal group as input, take the state classification performance as the objective function, and adaptively search for the optimal hyperparameter combination of MSCNN that is suitable for the sub-band signals through the BO algorithm. S4. Based on the optimal hyperparameter combination obtained by the BO algorithm, construct the MSCNN network model for fault diagnosis and set its training configuration to complete the model initialization for subsequent training and diagnosis of the dataset. S5. Using the training set data as input, after optimal WPD processing, it is input into the fault diagnosis MSCNN network model. The classification loss value is calculated through forward propagation, and the model weights are updated through backpropagation. After each round of training, the model is evaluated through the validation set. The above training and validation process is repeated iteratively until the model performance reaches the maximum number of training rounds. Finally, the fault diagnosis is completed through the test set.

[0009] Furthermore, S2 specifically includes: S2.1: Set the CPO population size and maximum number of iterations, and predefine the parameter search space for the WPD wavelet basis function candidate set and the number of decomposition layers, which is used to limit the range of selectable values ​​for wavelet basis functions and the number of decomposition layers in the subsequent optimization process and constrain the parameter optimization boundary. Then, a population initialization operation is performed, generating an initial position for each individual within a defined parameter space. , X g For the first g Each population position corresponds to a set of WPD decomposition parameters to be optimized, consisting of wavelet basis functions and the number of decomposition layers.

[0010] S2.2: Call the WPD decomposition parameters corresponding to the current population individuals, perform wavelet packet decomposition on the input training set induced electromotive force signal, decompose layer by layer to obtain each sub-band signal component, decompose to the preset decomposition layer, and finally obtain all sub-band signal components for subsequent steps to perform feature evaluation. S2.3: Based on the sub-band signal components of each state obtained from the decomposition, calculate the overall average envelope entropy and the ratio of inter-class and intra-class scatter, and construct and solve the bi-objective fitness function. F The bi-objective fitness values ​​corresponding to each WPD decomposition parameter combination are obtained, which are used to quantitatively evaluate the quality of the current WPD decomposition parameter combination in representing the fault features of the training set. Furthermore, S2.3 specifically includes: First, envelope analysis is performed on the sub-band signal components of each state type to calculate the envelope entropy. Then, the average of the samples in all states is calculated to obtain the overall average envelope entropy. : (1) In formula (1), C Total number of state types; N i For the first i The number of samples for each class state; l The length of the subband signal envelope sequence; p ij ( n ) is the first i The first class state j The sample in the subband signal envelope sequence is in the _ . n The probability distribution of points; the calculated probability distribution. It reflects the complexity of the signal after WPD decomposition. The lower the entropy value, the more concentrated the fault features and the better the distinguishability.

[0011] Subsequently, the sub-band signal components of each state type are used to construct eigenvectors, and the inter-class scatter matrix is ​​calculated. S b With the intra-class discreteness matrixS w The inter-class and intra-class dispersion ratios were obtained. J : (2) (3) (4) In formulas (2)-(4), det(·) represents the determinant of the matrix; S b This is the inter-class scatter matrix; representing the differences between feature vectors of different state types. S w The in-class discreteness matrix represents the degree of discreteness of eigenvectors within the same state type; m i For the first i The mean vector of samples in each state; m is the mean vector of samples in all states; X i For the first i The set of samples representing class states; x is the feature vector of the sample; J The larger the value, the better the feature discrimination between different states; Finally, the two indicators are weighted and fused together, and then substituted into the bi-objective fitness function to calculate the fitness value corresponding to each WPD decomposition parameter combination: (5) In formula (5), ω The weighting coefficient is used to balance the proportion of intra-class and inter-class dispersion ratios in the evaluation; fitness value. F The larger the value, the better the discrimination of the fault features extracted by the current WPD decomposition parameter combination, the lower the complexity, and the stronger the ability to represent the fault features of the training set.

[0012] S2.4: Determine the position of the current best individual based on the bi-objective fitness values ​​corresponding to each WPD decomposition parameter combination obtained; Then, the population position is updated according to the defense strategy of the CPO algorithm, and a new generation of WPD decomposition parameter combinations to be optimized is generated; steps S2.2-S2.4 are repeated iteratively until the preset maximum iteration termination condition is met, and finally the WPD decomposition parameter combinations adapted to the fault characteristics of the training set are output.

[0013] Furthermore, updating the population position according to the defense strategy of the CPO algorithm specifically includes: In each iteration, the population size is first dynamically adjusted, then the exploration or development phase is determined based on the triggering conditions. Next, a defense strategy for the corresponding phase is selected and executed based on random probability. This dynamically balances the exploration and development processes in the WPD parameter space, gradually increasing the fitness values ​​of individual populations until the algorithm converges. Cyclic population reduction technique: To maintain population diversity and avoid premature convergence to local optima during the optimization of WPD decomposition parameters, the population size is dynamically adjusted as follows: (6) In formula (6), N Indicates the current population size. q Indicates the current iteration number. Q Indicates the number of loops. N min For the minimum population size, Q max This represents the maximum number of loops. Exploration Phase: When the exploration condition is triggered, the algorithm randomly selects either the first defense strategy (Formula (7) or the second defense strategy (Formula (8)) according to a preset probability to update the individual position, and searches for potential high-quality parameter combinations in the WPD decomposition parameter space. The updates corresponding to the two strategies are as follows: (7) (8) In formulas (7)-(8), Indicates the first q In the +1st iteration, the... g The updated position of each individual corresponds to a new generation of WPD decomposition parameter combinations to be evaluated; Indicates the first q In the nth iteration g The updated position of each individual To indicate the first q The individual position corresponding to the optimal WPD decomposition parameter combination in the next iteration; U 1 represents the binary vector selected for the control strategy; y q g Indicates the first q The predator's position at the next iteration; τ 1. τ 2. τ 3 is a random number within the interval [0,1]. r 1. r 2 represents a random individual index; Development Phase: When development conditions are triggered, the algorithm randomly selects either the third defense strategy (odor) or the fourth defense strategy (physical attack) based on preset probabilities to update the individual's position. It then performs a refined search near the current high-quality WPD decomposition parameters to improve the parameters' fault characteristic representation capabilities. The updates corresponding to the two strategies are as follows: (9) (10) In formulas (9)-(10), X q r1 , X q r2 , X q r3 Indicates the first q In each iteration, three different individual positions are randomly selected from the current population, corresponding to three sets of randomly selected WPD decomposition parameter combinations. This is used to introduce diversity during the update process and avoid the algorithm getting stuck in local optima. r 1. r 2. r 3 represents distinct random individual indices; S q g It is an odor diffusion factor; δ These are the direction control parameters; α This is the convergence rate factor; τ 4. τ 5 is a random number within the interval [0,1]. r q As a defensive factor; F q g It is an inelastic collision factor.

[0014] Furthermore, S3 specifically includes: S3.1: Based on the optimal WPD decomposition parameters, the induced electromotive force signal of the training set is processed by WPD decomposition to obtain the sub-band signal; Then initialize the historical observation set of the BO algorithm. D 1:s : (11) In formula (11), λ s For the first s The MSCNN hyperparameter combination for each iteration includes kernel size, number of channels, batch size, initial learning rate, and weight decay coefficient; y s For the first sThe objective function value of the next iteration corresponds to the state classification performance index of MSCNN on the subband signal after WPD processing under the hyperparameter combination; D 1:s This indicates the number of iterations from the 1st iteration to the 2nd iteration. s All observation data from the next iteration.

[0015] S3.2: Based on historical observation sets D 1:s A Gaussian process surrogate model is constructed and updated, which is used to fit the mapping relationship between MSCNN hyperparameters and objective function values; Subsequently, an expected improvement acquisition function EI is constructed based on the Gaussian process surrogate model to guide the selection of the next set of hyperparameters: (12) In formula (12), μ ( λ )and σ ( λ ( ) represent Gaussian process models in hyperparameter combinations λ The predicted mean and standard deviation at a given point reflect the model's estimation and uncertainty regarding the performance of that hyperparameter. y best The optimal objective function value in the current historical observation set; Φ(·) and (·) represents the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0016] S3.3: By maximizing the acquisition function EI, the next combination of hyperparameters to be evaluated is obtained. λ s+1 ; S3.4: Based on hyperparameter combination λ s+1 Construct and train an MSCNN model, with the input being the sub-band signal processed by WPD and the output being the state classification result; Then, the state classification performance of the model on the validation set is calculated to obtain the corresponding objective function value. y s+1 ; S3.5: Combining new hyperparameters λ s+1 With the corresponding objective function value y s+1 Incorporate into the historical observation set and update the observation set to D 1:s+1 ; S3.6: If the current iteration count reaches the maximum iteration count, terminate the iteration and select the hyperparameter combination with the highest objective function value from the historical observation set as the global optimal hyperparameter of MSCNN; Otherwise, let s =s +1, return to step S3.2 and continue iterating.

[0017] Furthermore, the hyperparameter combination includes: kernel size, number of channels, batch size, initial learning rate, and weight decay coefficient.

[0018] Furthermore, a fault diagnosis MSCNN network model is constructed based on the optimal hyperparameter combination, specifically including: Multi-scale parallel convolutional branches: Three parallel convolutional branches are constructed, corresponding to feature extraction paths at different scales; the kernel sizes of branches 1, 2, and 3 are 5, 11, and 21, respectively, and the number of channels in each branch is uniformly set to 64; the stride of all branches is 1, the padding method is "same" padding adapted to the kernel size, the activation function is ReLU, and the weights are initialized using the Kaiming method; Internal structure of a single branch: Each convolutional branch consists of two cascaded CNN sub-blocks with the same structure. Each sub-block includes a convolutional layer and an average pooling layer. The average pooling layer has a pooling window size of 3×1 and a stride of 1, which is used to gradually reduce the feature dimension and enhance the translation invariance of the model. Multi-scale feature fusion layer: Multi-scale features extracted by three parallel branches are input to the feature fusion layer for concatenation, resulting in a one-dimensional fused feature vector. x fusion ; Fully connected layer and classification output layer: fusing feature vectors x fusion The inputs are sequentially fed into a fully connected layer and a classification output layer; the number of neurons in the fully connected layer is the same as the number of state types, and it is used to further map the fused features to the classification output layer; the classification output layer uses... Softmax The activation function maps the output to the probability distribution of each state type, thus obtaining the final fault diagnosis result. The specific steps for setting the model training configuration include: The optimizer used is Adam, and the loss function is the cross-entropy loss function. The initial learning rate, weight decay coefficient, and batch size are configured using the optimal hyperparameters output by S4, specifically: the initial learning rate is set to 5 × 10⁻⁶. -4 The weight decay coefficient is set to 1.5 × 10. -3 Set the batch size to 64.

[0019] In another aspect of the invention, an electronic device is also provided, comprising: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform a method for diagnosing inter-turn short circuits and demagnetization faults of a permanent magnet motor based on CPO-WPD and BO-MSCNN as described above.

[0020] In another aspect of the invention, a computer-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform a method for diagnosing inter-turn short circuits and demagnetization faults of a permanent magnet motor based on CPO-WPD and BO-MSCNN as described above.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention provides a method for diagnosing inter-turn short circuit and demagnetization faults of permanent magnet motors based on CPO-WPD and BO-MSCNN. The method takes minimizing the overall average envelope entropy and maximizing the inter-class and intra-class divergence ratio as the joint optimization objective. The WPD parameters are adaptively optimized by the CPO algorithm to achieve accurate multi-resolution decomposition of diagnostic signals, significantly improve the distinguishability of features between different fault categories, solve the problem of traditional WPD relying on expert experience and unstable feature extraction quality, and achieve high-quality adaptive feature extraction.

[0022] (2) The present invention provides a method for diagnosing inter-turn short circuit and demagnetization faults of permanent magnet motor based on CPO-WPD and BO-MSCNN. The method uses the BO algorithm to automatically optimize the key hyperparameters of MSCNN, which can quickly match the fault features extracted by the front end. This avoids the shortcomings of traditional grid search and other methods, such as low efficiency and difficulty in adaptation, and gives full play to the advantages of MSCNN in capturing multi-scale fault features.

[0023] (3) The present invention provides a method for diagnosing inter-turn short circuit and demagnetization faults of permanent magnet motor based on CPO-WPD and BO-MSCNN, and constructs a collaborative diagnosis framework of "adaptive feature extraction + hyperparameter optimization model" to achieve accurate identification and early warning of two types of typical faults, effectively improve the stability and reliability of fault diagnosis under complex working conditions, and achieve optimal matching of features and models. Attached Figure Description

[0024] Figure 1 This is a flowchart of the inter-turn short circuit and demagnetization fault diagnosis method for permanent magnet motors based on CPO-WPD and BO-MSCNN as described in this invention; Figure 2 This is a schematic diagram of the permanent magnet motor winding connection and permanent magnet numbering in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the feasible region of MSCNN hyperparameters in Embodiment 1 of the present invention; Figure 4is a schematic diagram of the fault diagnosis framework for permanent magnet motors based on optimized WPD and MSCNN in Embodiment 1 of the present invention; Figure 5 is a convergence curve diagram of accuracy and loss value during MSCNN training in Embodiment 1 of the present invention; Figure 6 is a confusion matrix diagram of fault diagnosis results in Embodiment 1 of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] It should be noted that all the following detailed descriptions are exemplary, and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as commonly understood by those of ordinary skill in the technical field to which the present invention pertains.

[0027] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] Without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0029] Embodiment 1 Refer to Figure 1 , this embodiment provides an inter-turn short-circuit and demagnetization fault diagnosis method for permanent magnet motors based on CPO-WPD and BO-MSCNN, where CPO is the crested porcupine optimization algorithm, WPD is wavelet packet decomposition, BO is the Bayesian optimization algorithm, and MSCNN is multi-scale convolutional neural network. The method includes: S1, Constructing and dividing a fault diagnosis data set: collecting induced electromotive force signals of the permanent magnet motor under different rotating speeds and in a normal state, an inter-turn short-circuit fault state, and a local demagnetization fault state, to construct a fault diagnosis data set, and dividing the data set into a training set, a verification set and a test set according to a preset proportion.

[0030] In this embodiment, the adopted winding connection mode of the permanent magnet motor and the numbering of permanent magnets are as shown in Figure 2 , the pole pair number of the permanent magnets on the rotor side of the motor is 14, the pole pair number of the stator armature winding is 10, and the number of stator slots is 24. Each phase winding is formed by connecting 4 coils in series, and the total number of turns of each phase is 476.

[0031] In terms of fault settings, inter-turn short-circuit faults are implemented by introducing short-circuit turns in the A-phase winding. To simulate different fault severity levels, the number of short-circuit turns is set to 38 and 82, respectively, corresponding to short-circuit percentages of 8% and 17.2% of the total turns, i.e., 8% and 17.2% inter-turn short-circuit faults. For local demagnetization faults, three specific permanent magnets (PM1, PM7, PM8) on the rotor are selected as fault locations to simulate the randomness of permanent magnet faults in actual operation. To facilitate comparison and analysis with inter-turn short-circuit faults, demagnetization faults are set to residual magnetism at 20% and 50%, i.e., 20% and 50% demagnetization faults. Subsequently, induced electromotive force signals for each state type are collected at the motor's rated speed, 0.75 times the rated speed, 0.5 times the rated speed, and 0.25 times the rated speed. Preferably, in specific implementation, the collected state-induced electromotive force signals are divided into individual data samples with a duration of 0.28 seconds, and then divided into training set, validation set and test set in a ratio of 5:2:3.

[0032] S2. The optimal WPD decomposition parameters are adaptively determined using the CPO algorithm; The training set data (induced electromotive force signal) is input into the CPO-WPD module. The CPO population size and maximum number of iterations are set. CPO performs iterative optimization using a bi-objective fitness function as the optimization criterion. At the same time, the parameter search space of the WPD wavelet basis function candidate set and the number of decomposition layers is predefined. Using the parameters of wavelet packet decomposition as the optimization space, the optimal WPD decomposition parameters, i.e. the optimal wavelet basis and decomposition level parameters, are obtained by adaptive iterative selection and solution through the CPO algorithm.

[0033] CPO-WPD is not an independent "fixed structure module", but a collaborative combination of the Crowned Porcupine Optimization Algorithm (CPO) and Wavelet Packet Decomposition (WPD): it uses the parameters of wavelet packet decomposition (number of decomposition levels, basis functions) as the optimization space, and adaptively determines the optimal WPD decomposition parameters through the CPO algorithm.

[0034] The specific steps of S2 are as follows: S2.1: Set the CPO population size and maximum number of iterations; predefine the parameter search space for the WPD wavelet basis function candidate set and the number of decomposition layers to limit the range of selectable values ​​for wavelet basis functions and the number of decomposition layers in the subsequent optimization process and constrain the parameter optimization boundary. Then, a population initialization operation is performed, that is, an initial position is generated for each individual in the limited parameter space using the following formula. Each individual in the population corresponds one-to-one with a set of WPD decomposition parameter combinations to be optimized, consisting of wavelet basis functions and decomposition level.

[0035] The population location initialization formula is: (13) in, X g For the first g Population location, L , U These are the lower and upper limits of the parameter range, respectively. r A random number in the range [0,1].

[0036] Preferably, in specific implementation, the population size of the CPO algorithm is set to 80, the maximum number of iterations is 20; the candidate set of WPD wavelet basis functions is {sym2,sym4,sym5,sym8,coif3,coif5,db1,db2,db4,db8}, and the search space for the wavelet packet decomposition layer is 2 to 5 layers.

[0037] S2.2: Call the WPD decomposition parameters corresponding to the current population individuals, perform wavelet packet decomposition on the input training set induced electromotive force signal, and decompose layer by layer to obtain the signal components of each sub-band.

[0038] In practice, the wavelet packet decomposition process is implemented using the following recursive formula: (14) (15) in, u k+1,2v ( t ) is the first k Layer v After low-pass filtering, the node at the node is at the _th ... k The second generation generated at +1 level v Subband signal components of each node; u k+1,2v+1 ( t ) is the first k Layer v After passing through a high-pass filter, the node at the nth node... k The second generation generated at +1 level v +1 node sub-band signal components; h ( m ) represents the low-pass filter coefficients corresponding to the selected wavelet basis; g ( m These are the high-pass filter coefficients corresponding to the selected wavelet basis; t It is a time variable; m The shift factor is an integer. Z Represents the set of integers; In order to be in Time k Layer v Each sub-band corresponds to a sub-band signal component; uk,v (2 t - m ) indicates that the first k Layer v Each child carries a signal component u k,v ( t (time variable) t Transformed into 2 t - m This corresponds to the downsampling and integer delay operations in wavelet packet decomposition; The decomposition process is performed up to a preset number of decomposition layers, ultimately yielding all sub-band signal components for feature evaluation in subsequent steps.

[0039] S2.3: Based on the sub-band signal components of each state type obtained from the decomposition in step S3.2 above, calculate the overall average envelope entropy and the inter-class and intra-class dispersion ratio for all state types, and construct and solve the bi-objective fitness function. F We obtain the bi-objective fitness value corresponding to each WPD decomposition parameter combination, which is used to quantitatively evaluate the quality of the current WPD decomposition parameter combination in representing the fault features of the training set.

[0040] In practice, firstly, envelope analysis is performed on the sub-band signal components of each state type to calculate the envelope entropy. Then, the average of the samples in all states is calculated to obtain the overall average envelope entropy. : (1) In formula (1), C Total number of state types; N i For the first i The number of samples for each class state; l The length of the subband signal envelope sequence; p ij ( n ) is the first i The first class state j The sample in the subband signal envelope sequence is in the _ . n The probability distribution of points; the calculated probability distribution. It reflects the complexity of the signal after WPD decomposition. The lower the entropy value, the more concentrated the fault features and the better the distinguishability.

[0041] Subsequently, the sub-band signal components of each state type are used to construct eigenvectors, and the inter-class scatter matrix is ​​calculated. S b With the intra-class discreteness matrix S w The inter-class and intra-class dispersion ratios were obtained. J : (2) (3) (4) In formulas (2)-(4), det(·) represents the determinant of the matrix; S b This is the inter-class scatter matrix, representing the differences between the feature vectors of different state types; S w The in-class discreteness matrix represents the degree of discreteness of eigenvectors within the same state type; m i For the first i The mean vector of samples in each state; m is the mean vector of samples in all states; X i For the first i The set of samples representing class states; x is the feature vector of the sample; J The larger the value, the better the feature discrimination between different states.

[0042] Finally, the two indicators mentioned above are weighted and fused, and then substituted into the bi-objective fitness function to calculate the fitness value corresponding to each WPD decomposition parameter combination. The calculation formula is as follows: (5) In formula (5), ω The weighting coefficient is used to balance the proportion of intra-class and inter-class dispersion ratios in the evaluation; fitness value. F The larger the value, the better the discrimination of the fault features extracted by the current WPD decomposition parameter combination, the lower the complexity, and the stronger the ability to represent the fault features of the training set.

[0043] S2.4: Determine the current optimal individual position based on the bi-objective fitness value corresponding to each WPD decomposition parameter combination calculated above; then update the population position according to the defense strategy of the CPO algorithm to generate a new generation of WPD decomposition parameter combinations to be optimized; repeat steps S2.2-S2.4 iteratively until the preset maximum iteration termination condition is met, and finally output the WPD decomposition parameter combination that adapts to the fault characteristics of the training set.

[0044] In practice, the CPO algorithm simulates four defensive behaviors of the crested porcupine and iteratively updates the positions of individuals representing the WPD decomposition parameter combinations in the population. Its core mechanism includes: (1) Cyclic population reduction technique: In order to maintain population diversity and avoid premature convergence to a local optimum during the WPD parameter optimization process, the population size is dynamically adjusted according to the following formula: (6) in, N Indicates the current population size. q Indicates the current iteration number.Q Indicates the number of loops. N min For the minimum population size, Q max This represents the maximum number of iterations.

[0045] (2) Exploration Phase (WPD Parameter Space Search): When the exploration condition is triggered, the algorithm randomly selects either the first defense strategy (visual) or the second defense strategy (sound) according to a preset probability to update the individual's position, and searches for potential high-quality parameter combinations in the WPD parameter space. The updates corresponding to the two strategies are as follows: (7) (8) in, Indicates the first q In the +1st iteration, the... g The updated position of each individual corresponds to a new generation of WPD decomposition parameter combinations to be evaluated; Indicates the first q In the nth iteration g The updated position of each individual To indicate the first q The individual position corresponding to the optimal WPD decomposition parameter combination in the next iteration; U 1 represents the binary vector selected for the control strategy; y q g Indicates the first q The predator's position at the next iteration; τ 1. τ 2. τ 3 is a random number within the interval [0,1]. r 1. r 2 represents a random individual index; (3) Development Phase (Local Optimization of WPD Decomposition Parameters): When the development condition is triggered, the algorithm randomly selects either the third defense strategy (odor) or the fourth defense strategy (physical attack) according to a preset probability to update the individual position, performing a refined search near the current high-quality WPD decomposition parameters to improve the fault characteristic representation capability of the parameters. The updates corresponding to the two strategies are as follows: (9) (10) In formulas (9)-(10), Indicates the first q In the +1st iteration, the... g The updated position of each individual corresponds to a new set of WPD decomposition parameters to be evaluated. X qr1 , X q r2 , X q r3 Indicates the first q In each iteration, three different individual positions are randomly selected from the current population, corresponding to three sets of randomly selected WPD decomposition parameter combinations. This is used to introduce diversity during the update process and avoid the algorithm getting stuck in local optima. r 1. r 2. r 3 represents distinct random individual indices; S q g It is an odor diffusion factor; δ These are the direction control parameters; α This is the convergence rate factor; τ 4. τ 5 is a random number within the interval [0,1]. r q As a defensive factor; F q g It is an inelastic collision factor.

[0046] In each iteration, one of the above defense strategies is selected for execution based on random probability. The exploration and development process of the WPD parameter space is dynamically balanced, and the fitness value of individuals in the population is gradually increased until the algorithm converges.

[0047] S3. Find the optimal hyperparameter combination of MSCNN that is adapted to the subband signal based on the BO algorithm; The optimal WPD decomposition parameters obtained by iterative solution in S2 are called to perform WPD decomposition on the induced electromotive force signal in the training set to obtain sub-band signals that are adapted to the fault characteristics. The sub-band signal group is used as input, and the optimal hyperparameter combination of MSCNN that is adapted to the sub-band signals is adaptively searched through the BO algorithm with state classification performance as the objective function.

[0048] In a preferred embodiment, the BO-MSCNN module adaptively searches for a hyperparameter combination of MSCNN (Multi-Scale Convolutional Neural Network) that matches the sub-band signal after optimal WPD decomposition using the BO algorithm. Specifically, this includes the following steps: S3.1: Based on the optimal WPD decomposition parameters output from S2, perform WPD decomposition of the induced electromotive force signal in the training set to obtain the sub-band signal; then initialize the historical observation set of the BO algorithm. D 1:s : (11) In formula (11), D1:s This indicates the number of iterations from the 1st iteration to the 2nd iteration. s All observation data from the next iteration λ s For the first s The MSCNN hyperparameter combination for each iteration includes kernel size, number of channels, batch size, initial learning rate, and weight decay coefficient; y s For the first s The objective function value of the next iteration corresponds to the state classification performance index of MSCNN on the subband signal after WPD processing under the hyperparameter combination.

[0049] S3.2: Based on historical observation sets D 1:s A Gaussian process surrogate model is constructed and updated to fit the mapping relationship between MSCNN hyperparameters and objective function values. Subsequently, an expected improvement acquisition function EI is constructed based on this surrogate model to guide the selection of the next set of hyperparameters. (12) In formula (12), μ ( λ )and σ ( λ ( ) represent Gaussian process models in hyperparameter combinations λ The predicted mean and standard deviation at a given point reflect the model's estimation and uncertainty regarding the performance of that hyperparameter. y best The optimal objective function value in the current historical observation set; Φ(·) and (·) represents the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0050] S3.3: By maximizing the acquisition function EI, the next combination of hyperparameters to be evaluated is obtained. λ s+1 : (16) In formula (16), Λ is the feasible region of the MSCNN hyperparameter combination, such as Figure 3 As shown, the parameters and their corresponding value ranges are: branch 1 convolution kernel size [3,5,7,9], branch 2 convolution kernel size [11,13,15,17], branch 3 convolution kernel size [19,21,23,25], number of channels [32,64,128], batch size [32,64,128], initial learning rate [2×10⁻⁶]. -4 8×10 -4 Weight decay coefficient [8×10] -4 3×10 -3 ]; Within this feasible region, the kernel size, number of channels, and batch size are discrete values, while the initial learning rate and weight decay coefficient are continuous values.

[0051] S3.4: Based on hyperparameter combination λ s+1 Construct and train an MSCNN model, with the input being the sub-band signal processed by WPD and the output being the state classification result; Then, the state classification performance of the model on the validation set is calculated to obtain the corresponding objective function value. y s+1 .

[0052] S3.5: Combining new hyperparameters λ s+1 With the corresponding objective function value y s+1 Incorporate into the historical observation set and update the observation set to D 1:s+1 : (17) S3.6: If the current iteration count reaches the maximum iteration count, terminate the iteration and select the hyperparameter combination with the highest objective function value from the historical observation set as the globally optimal hyperparameters for MSCNN; otherwise, let s = s +1, return to step S3.2 and continue iterating.

[0053] In a preferred embodiment, the globally optimal hyperparameter combination obtained after optimization is as follows: the kernel sizes corresponding to scale 1, scale 2, and scale 3 are 5, 11, and 21 respectively, the number of channels is 64, the batch size is 64, and the initial learning rate is 5×10. -4 The weight decay coefficient is 1.5 × 10⁻⁶. -3 .

[0054] S4. Construct the network architecture for the fault diagnosis network model; Based on the optimal hyperparameter combination obtained by the BO algorithm, a fault diagnosis MSCNN network model is constructed and its training configuration is set, thereby completing the model initialization (referring to the construction of the fault diagnosis MSCNN model and the setting of the model training configuration), which is used for subsequent training and diagnosis of the dataset.

[0055] In this preferred embodiment, the establishment and initialization of the fault diagnosis MSCNN model refers to the model used in step S3. The purpose of step S3 is to select the MSCNN parameter configuration that is compatible with the front-end WPD, and to use these parameters to build the fault diagnosis MSCNN model in step S4, laying the groundwork for subsequent training and diagnosis.

[0056] Using the obtained optimal hyperparameter combination as the core configuration basis, a fault diagnosis MSCNN model is established, such as... Figure 4 As shown, the specific implementation is as follows: Multi-scale parallel convolutional branches: Three parallel convolutional branches are constructed, corresponding to feature extraction paths at different scales. The kernel sizes for branches 1, 2, and 3 are 5, 11, and 21, respectively, with a uniform channel count of 64 for each branch. All branches have a stride of 1, use "same" padding adapted to the kernel size, employ ReLU activation, and initialize the weights using the Kaiming method.

[0057] Single-branch internal structure: Each convolutional branch consists of two cascaded CNN sub-blocks with identical structures. Each sub-block contains a convolutional layer and an average pooling layer in sequence; the average pooling layer has a pooling window size of 3×1 and a stride of 1, which is used to gradually reduce the feature dimension and enhance the translation invariance of the model.

[0058] Multi-scale feature fusion layer: Multi-scale features extracted by three parallel branches are input to the feature fusion layer for concatenation, resulting in a one-dimensional fused feature vector. x fusion .

[0059] Fully connected layer and classification output layer: The fused feature vector is sequentially input into the fully connected layer and the classification output layer. The number of neurons in the fully connected layer is the same as the number of state types, used to further map the fused features; the classification output layer uses... Softmax The activation function maps the output to the probability distribution of each state type, and outputs the final fault diagnosis result.

[0060] The specific steps for setting the model training configuration include: The optimizer used is Adam, and the loss function is the cross-entropy loss function. The initial learning rate, weight decay coefficient, and batch size are configured using the optimal hyperparameters output by S4, specifically: the initial learning rate is set to 5 × 10⁻⁶. -4 The weight decay coefficient is set to 1.5 × 10. -3 Set the batch size to 64.

[0061] S5. Iteratively train the MSCNN network model for fault diagnosis; The training set data is used as input, processed by the optimal WPD, and then input into the fault diagnosis MSCNN network model. The classification loss value is calculated through forward propagation, and the model weights are updated through backpropagation. After each training round, the model is evaluated through the validation set. The above training and validation process is repeated iteratively until the model performance reaches the maximum number of training rounds. Finally, the fault diagnosis is completed through the test set, and the fault diagnosis result is output.

[0062] In practice, Figure 4 The diagram shows a fault diagnosis framework for permanent magnet motors based on optimized WPD and MSCNN. First, the fault diagnosis dataset is processed by optimal wavelet packet decomposition, resulting in 16 sub-band signals for each sample. To effectively reduce data dimensionality while retaining the most discriminative key information, the top 8 sub-band signals in terms of energy of each sample are selected as the input features of the fault diagnosis network model.

[0063] Subsequently, the aforementioned features are input into the initialized fault diagnosis network model, and the model is trained by updating its weights using the training and validation sets. Using the training set data as input, the state classification loss is calculated through forward propagation, and the model weights are updated through backpropagation. After each training round, the model performance is evaluated on the validation set, and the training process is dynamically monitored based on the validation set accuracy to prevent overfitting. This training and validation process is iteratively repeated until the model performance reaches the maximum number of training rounds. The convergence curves of accuracy and loss values ​​during training are shown below. Figure 5 As shown, from the perspectives of accuracy and loss value, both curves converged rapidly and gradually stabilized within the first 20 iterations, with no obvious oscillations or overfitting during the training process.

[0064] Finally, fault diagnosis was completed using the test set, and the overall diagnostic accuracy of the model reached 99.46%, with the corresponding confusion matrix as shown below. Figure 6 As shown, the 17.2% inter-turn short circuit (ITSC-17.2%) and 50% partial demagnetization (LD-50%) states achieved 100% error-free identification; only a small number of misclassifications occurred between the normal state and the 8% inter-turn short circuit (ITSC-8%) state, and between the normal state and the 20% demagnetization fault (LD-20%) state, with misclassification rates of 0.83% and 0.94% respectively, and no serious misclassifications across fault types occurred.

[0065] The above results demonstrate that the proposed method has excellent identification capabilities for inter-turn short circuits and local demagnetization faults of varying severity, and the misjudgment rate for minor faults is extremely low, fully verifying the robustness and diagnostic reliability of the proposed scheme under complex operating conditions.

[0066] Example 2 This embodiment also provides an electronic device, including: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the wire foreign object detection method based on the improved YOLOv11 as described above.

[0067] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable computing device, consumer electronic device, etc.

[0068] Example 3 This embodiment also provides a machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the wire foreign object detection method based on the improved YOLOv11 as described above.

[0069] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0070] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0071] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0072] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing inter-turn short circuit and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN, characterized in that, The method includes: S1. Collect induced electromotive force signals of permanent magnet motors under different speeds in normal state, inter-turn short circuit fault, and local demagnetization fault to construct a fault diagnosis dataset, and divide it into training set, validation set and test set according to a preset ratio; S2. Set the CPO population size and maximum number of iterations, and predefine the parameter search space for the WPD wavelet basis function candidate set and the number of decomposition layers; take the training set data as input, use the wavelet packet decomposition parameters as the optimization space, and adaptively determine the optimal WPD decomposition parameters through the CPO algorithm; S2 specifically includes: S2.1: Set the CPO population size and maximum number of iterations, and predefine the parameter search space for the WPD wavelet basis function candidate set and the number of decomposition layers; Then, a population initialization operation is performed, generating an initial position for each individual within a defined parameter space. , X g For the first g Each population position corresponds to a set of WPD decomposition parameters to be optimized, consisting of wavelet basis functions and decomposition level. S2.2: Call the WPD decomposition parameters corresponding to the current population individual, perform wavelet packet decomposition on the input training set induced electromotive force signal, decompose layer by layer to obtain each sub-band signal component, decompose to the preset decomposition layer, and finally obtain all sub-band signal components. S2.3: Based on the sub-band signal components of each state obtained from the decomposition, calculate the overall average envelope entropy and the ratio of inter-class and intra-class scatter, and construct and solve the bi-objective fitness function. F The bi-objective fitness values ​​corresponding to each WPD decomposition parameter combination are obtained. S2.4: Determine the position of the current best individual based on the bi-objective fitness values ​​corresponding to each WPD decomposition parameter combination obtained; Subsequently, the population position is updated according to the defense strategy of the CPO algorithm, and a new generation of WPD decomposition parameter combinations to be optimized is generated; Repeat steps S2.2-S2.4 iteratively until the preset maximum iteration termination condition is met, and finally output the WPD decomposition parameter combination that adapts to the fault characteristics of the training set. S3. Call the obtained optimal WPD decomposition parameters to perform WPD decomposition on the training set data to obtain sub-band signals that are suitable for the corresponding classification features; take the sub-band signal group as input, take the state classification performance as the objective function, and adaptively search for the optimal hyperparameter combination of MSCNN that is suitable for the sub-band signals through the BO algorithm. S3 specifically includes: S3.1: Based on the optimal WPD decomposition parameters, the induced electromotive force signal of the training set is processed by WPD decomposition to obtain the sub-band signal; Then initialize the historical observation set of the BO algorithm. D 1:s : (11) In formula (11), λ s For the first s The MSCNN hyperparameter combination for each iteration includes kernel size, number of channels, batch size, initial learning rate, and weight decay coefficient; y s For the first s The objective function value of the next iteration corresponds to the state classification performance index of MSCNN on the subband signal after WPD processing under the hyperparameter combination; D 1:s This indicates the number of iterations from the 1st iteration to the 2nd iteration. s All observation data from the next iteration; S3.2: Based on historical observation sets D 1:s A Gaussian process surrogate model is constructed and updated, which is used to fit the mapping relationship between MSCNN hyperparameters and objective function values; Subsequently, an expected improvement acquisition function EI is constructed based on the Gaussian process surrogate model to guide the selection of the next set of hyperparameters: (12) In formula (12), μ ( λ )and σ ( λ ( ) represent Gaussian process models in hyperparameter combinations λ The predicted mean and standard deviation at a given point reflect the model's estimation and uncertainty regarding the performance of that hyperparameter. y best The optimal objective function value in the current historical observation set; Φ(·) and (·) represent the cumulative distribution function and probability density function of the standard normal distribution, respectively; S3.3: By maximizing the acquisition function EI, the next combination of hyperparameters to be evaluated is obtained. λ s+1 ; S3.4: Based on hyperparameter combination λ s+1 Construct and train an MSCNN model, with the input being the sub-band signal processed by WPD and the output being the state classification result; Then, the state classification performance of the model on the validation set is calculated to obtain the corresponding objective function value. y s+1 ; S3.5: Combining new hyperparameters λ s+1 With the corresponding objective function value y s+1 Incorporate into the historical observation set and update the observation set to D 1:s+1 ; S3.6: If the current iteration count reaches the maximum iteration count, terminate the iteration and select the hyperparameter combination with the highest objective function value from the historical observation set as the global optimal hyperparameter of MSCNN; Otherwise, let s = s +1, return to step S3.2 and continue iterating; S4. Based on the optimal hyperparameter combination obtained by the BO algorithm, construct the MSCNN network model for fault diagnosis and set its training configuration to complete the model initialization. S5. Using the training set data as input, after optimal WPD processing, it is input into the fault diagnosis MSCNN network model. The classification loss value is calculated through forward propagation, and the model weights are updated through back propagation. After each training round, the model is evaluated using a validation set. The training and validation process is repeated iteratively until the model performance reaches the maximum number of training rounds. Finally, fault diagnosis is completed using a test set.

2. The method for diagnosing inter-turn short circuit and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN according to claim 1, characterized in that, Specifically, S2.3 includes: First, envelope analysis is performed on the sub-band signal components of each state type to calculate the envelope entropy. Then, the average of the samples in all states is calculated to obtain the overall average envelope entropy. : (1) In formula (1), C Total number of state types; N i For the first i The number of samples for each class state; l The length of the subband signal envelope sequence; p ij ( n ) is the first i The first class state j The sample in the subband signal envelope sequence is in the _ . n The probability distribution of points; Subsequently, the sub-band signal components of each state type are used to construct eigenvectors, and the inter-class scatter matrix is ​​calculated. S b With the intra-class discreteness matrix S w The inter-class and intra-class dispersion ratios were obtained. J : (2) (3) (4) In formulas (2)-(4), det(·) represents the determinant of the matrix; S b This is the inter-class scatter matrix, representing the differences between the feature vectors of different state types; S w The in-class discreteness matrix represents the degree of discreteness of eigenvectors within the same state type; m i For the first i The mean vector of samples in each state; m is the mean vector of samples in all states; X i For the first i The set of samples representing class states; x is the feature vector of the sample; J The larger the value, the better the feature discrimination between different states; Finally, the inter-class and intra-class dispersion ratios of the two indicators mentioned above are... J and the overall average envelope entropy Weighted fusion, substituting into the bi-objective fitness function, yields the fitness values ​​corresponding to each WPD decomposition parameter combination: (5) In formula (5), ω This is a weighting coefficient used to balance the proportion of intra-class and inter-class dispersion ratios and the overall average envelope entropy in the evaluation.

3. The method for diagnosing inter-turn short circuits and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN according to claim 2, characterized in that, The specific steps of updating the population position according to the defense strategy based on the CPO algorithm include: In each iteration, the population size is first dynamically adjusted; Cyclic population reduction technique: dynamically adjusting the population size, as follows: (6) In formula (6), N Indicates the current population size. q Indicates the current iteration number. Q Indicates the number of loops. N min For the minimum population size, Q max This represents the maximum number of loops. Then, based on the triggering conditions, it is determined whether to enter the exploration or development phase. Next, a defense strategy for the corresponding phase is selected and executed based on random probability. This dynamically balances the exploration and development processes in the WPD parameter space, gradually increasing the fitness values ​​of individuals in the population until the algorithm converges. Exploration Phase: When the exploration condition is triggered, the algorithm randomly selects either the first defense strategy (Formula (7) or the second defense strategy (Formula (8)) according to a preset probability to update the individual position, and searches for potential high-quality parameter combinations in the WPD decomposition parameter space. The updates corresponding to the two strategies are as follows: (7) (8) In formulas (7)-(8), Indicates the first q In the +1st iteration, the... g The updated position of each individual corresponds to a new generation of WPD decomposition parameter combinations to be evaluated; Indicates the first q In the nth iteration g The updated position of each individual To indicate the first q The individual position corresponding to the optimal WPD decomposition parameter combination in the next iteration; U 1 represents the binary vector selected for the control strategy; y q g Indicates the first q The predator's position at the next iteration; τ 1. τ 2. τ 3 is a random number within the interval [0,1]. r 1. r 2 represents a random individual index; Development phase: When the development condition is triggered, the algorithm randomly selects either the third defense strategy (formula (9) or the fourth defense strategy (formula (10)) according to the preset probability to update the individual position. The updates corresponding to the two strategies are as follows: (9) (10) In formulas (9)-(10), X q r1 , X q r2 , X q r3 Indicates the first q In each iteration, three different individual positions are randomly selected from the current population, corresponding to three sets of randomly selected WPD decomposition parameter combinations. This is used to introduce diversity during the update process and avoid the algorithm getting stuck in local optima. r 1. r 2. r 3 represents distinct random individual indices; S q g It is an odor diffusion factor; δ These are the direction control parameters; α This is the convergence rate factor; τ 4. τ 5 is a random number within the interval [0,1]. r q As a defensive factor; F q g It is an inelastic collision factor.

4. The method for diagnosing inter-turn short circuits and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN according to claim 1, characterized in that, The hyperparameter combination includes: kernel size, number of channels, batch size, initial learning rate, and weight decay coefficient.

5. The method for diagnosing inter-turn short circuits and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN according to claim 1, characterized in that, A fault diagnosis MSCNN network model is constructed based on the optimal hyperparameter combination, specifically including: Multi-scale parallel convolutional branches: Three parallel convolutional branches are constructed, corresponding to feature extraction paths at different scales; the kernel sizes of branches 1, 2, and 3 are 5, 11, and 21, respectively, and the number of channels in each branch is uniformly set to 64; the stride of all branches is 1, the padding method is "same" padding adapted to the kernel size, the activation function is ReLU, and the weights are initialized using the Kaiming method; Internal structure of a single branch: Each convolutional branch consists of two cascaded CNN sub-blocks with the same structure. Each sub-block includes a convolutional layer and an average pooling layer. The average pooling layer has a pooling window size of 3×1 and a stride of 1, which is used to gradually reduce the feature dimension and enhance the translation invariance of the model. Multi-scale feature fusion layer: Multi-scale features extracted by three parallel branches are input to the feature fusion layer for concatenation, resulting in a one-dimensional fused feature vector. x fusion ; Fully connected layer and classification output layer: fusing feature vectors x fusion The inputs are sequentially fed into a fully connected layer and a classification output layer; the number of neurons in the fully connected layer is the same as the number of state types, and it is used to further map the fused features to the classification output layer; the classification output layer uses... Softmax The activation function maps the output to the probability distribution of each state type, thus obtaining the final fault diagnosis result.

6. The method for diagnosing inter-turn short circuit and demagnetization faults in permanent magnet motors based on CPO-WPD and BO-MSCNN according to claim 1, characterized in that, The training configuration specifically includes: The optimizer uses Adam, and the loss function is the cross-entropy loss function; the initial learning rate, weight decay coefficient, and batch size are configured using the optimal hyperparameters from the output.

7. An electronic device, characterized in that, The electronic device includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements the steps of a method for diagnosing inter-turn short circuits and demagnetization faults of a permanent magnet motor based on CPO-WPD and BO-MSCNN as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images