Motor Fault Diagnosis Method Based on Adaptive Signal Decomposition and Hybrid Deep Learning

By employing an adaptive signal decomposition and hybrid deep learning approach, and utilizing the center collision optimization algorithm and the TCN-Transformer-BiLSTM model, the problems of modal aliasing and local optima in motor fault diagnosis are solved, enabling automatic and accurate motor fault diagnosis and improving the reliability and efficiency of diagnosis.

CN123115944APending Publication Date: 2026-11-17DONGFENG DANA AXLE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611303410.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing motor fault diagnosis methods suffer from modal aliasing and endpoint effects when dealing with non-stationary and multi-scale current and vibration signals. Furthermore, deep learning models struggle to simultaneously consider local temporal features, global dependencies, and bidirectional contextual information, leading to parameter combinations that easily fall into local optima and affect fault diagnosis accuracy.

Method used

An adaptive signal decomposition and hybrid deep learning approach is adopted. The variational mode decomposition (VMD) is adaptively solved by the center collision optimization algorithm, and combined with the TCN-Transformer-BiLSTM hybrid deep learning model to achieve automatic and accurate diagnosis of motor faults.

Benefits of technology

It improves the reliability and efficiency of motor fault diagnosis, can accurately extract fault-related impact features, reduce interference between signals, and enhance fault feature extraction capability and classification accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115944A_ABST
    Figure CN123115944A_ABST
Patent Text Reader

Abstract

This invention relates to the field of motor fault diagnosis technology, and provides a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning. The method includes: acquiring multi-channel signals of the motor under diagnosis during real-time operation and preprocessing them; performing dimensionality reduction and fusion on the preprocessed multi-channel signals to obtain a single-channel fused signal; using the optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO), performing VMD decomposition on the single-channel fused signal; inputting the decomposed intrinsic mode function components into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which has been optimized by the hybrid strategy frost-ice optimization algorithm (HSRIME), to output the probability distribution of various motor operating conditions; and using the operating condition category corresponding to the highest probability as the fault determination result. This invention can achieve automatic and accurate diagnosis of motor faults, improving the accuracy and robustness of diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, and more specifically, to a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning. Background Technology

[0002] Electric vehicles are widely used in urban mobility and logistics transportation. Permanent magnet synchronous motors (PMSMs), as the main drive motors in electric vehicles, offer advantages such as high efficiency, high power density, and high power, contributing to improved vehicle performance and driving range while reducing energy consumption. However, prolonged operation under harsh conditions such as high temperature, overload, and high vibration can lead to various faults, including insulation aging, inter-turn short circuits, permanent magnet demagnetization, and bearing damage. These faults can cause performance degradation, motor failure, and even safety accidents. Therefore, developing a multi-fault diagnosis solution for PMSMs to promptly detect fault types and severity is crucial for ensuring the safe and stable operation of electric vehicles.

[0003] Current motor fault diagnosis methods mainly include model-based methods and signal analysis-based methods. Model-based methods rely on accurate modeling of the motor's physical components, but actual operating conditions are complex and variable, making it difficult to obtain accurate model parameters. Signal analysis-based methods extract fault features through time-frequency analysis of signals such as current and vibration, and then combine this with a classifier for fault identification. However, existing methods have the following shortcomings: (1) The current and vibration signals of the motor during operation have strong non-stationarity and multi-scale characteristics, and the characteristic frequency band distributions corresponding to different fault types are different. Traditional signal decomposition methods (such as empirical mode decomposition (EMD), wavelet transform, etc.) have mode aliasing and endpoint effects, making it difficult to accurately separate the characteristic components of different faults. Although variational mode decomposition (VMD) overcomes the above problems to some extent, its decomposition effect is highly dependent on the selection of the number of modes K and the penalty factor α. Existing methods mostly adopt empirical selection or single optimization algorithms, which are prone to getting trapped in local optima.

[0004] (2) Existing deep learning fault diagnosis models (such as CNN, LSTM, TCN, etc.) each have their own focus in feature extraction, but a single model is difficult to simultaneously take into account local temporal features, global dependencies and bidirectional contextual information. In particular, for the PMSM multi-fault classification task, different fault types have significantly different manifestations in the signal, requiring more powerful multi-dimensional feature extraction and fusion capabilities.

[0005] (3) When optimizing VMD parameters and deep learning model hyperparameters at the same time, the existing optimization algorithm has insufficient global search capability, which can easily lead to the parameter combination getting stuck in local optima and affect the final fault diagnosis accuracy.

[0006] Therefore, it is necessary to develop a better motor fault diagnosis solution. Summary of the Invention

[0007] This invention addresses the technical problems existing in the prior art by providing a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning, which realizes automatic and accurate diagnosis of motor faults and improves the reliability and efficiency of diagnosis.

[0008] According to a first aspect of the present invention, a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning is provided, comprising: The multi-channel signals of the motor under diagnosis are collected and preprocessed in real time. The preprocessed multi-channel signals are then fused to obtain a single-channel fused signal. The optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are used to perform VMD decomposition on the single-channel fused signal to obtain the intrinsic mode function components. The intrinsic mode function components are input into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which is optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained. The model outputs the probability distribution of various operating conditions of the motor, and the operating condition category corresponding to the highest probability is used as the fault judgment result.

[0009] Based on the above technical solution, the present invention can also be improved as follows.

[0010] Optionally, the acquisition and preprocessing of multi-channel signals from the real-time operation of the motor under diagnosis includes: Collect multi-channel signals of the motor under diagnosis during its operation. The multi-channel signals include three-phase current signals and vibration signals. The operating state is normal state, inter-turn short circuit fault state, permanent magnet demagnetization fault state, or bearing damage fault state. The multi-channel signal is preprocessed by missing value filling and outlier removal, and the preprocessed multi-channel signal is divided into a multi-dimensional sample set to be diagnosed according to a preset scale. The signals of each channel in the multidimensional sample set to be diagnosed are subjected to min-max normalization to linearly map the feature data of each dimension to the [0,1] interval, thus obtaining the normalized multidimensional sample set.

[0011] Optionally, the step of dimensionality reduction and fusion of the preprocessed multi-channel signals to obtain a single-channel fused signal includes: The normalized multidimensional samples of the normalized multidimensional sample set are used to form a multi-channel matrix, and the multi-channel matrix is ​​then decentered. Calculate the covariance matrix of the decentralized matrix, and perform eigenvalue decomposition on the covariance matrix to obtain each eigenvalue and its corresponding eigenvector. The projection direction, which is calculated and saved in advance using historical data, is called. The projection direction is the feature vector corresponding to the largest feature value among the feature values ​​calculated using historical training data. The multi-channel signals of the normalized multi-dimensional sample set are weighted and summed in the projection direction to obtain the first principal component signal as the single-channel fusion signal.

[0012] Optionally, the optimal parameters for adaptively solving the variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are obtained in advance during the offline training phase, including: Initialize the CCO algorithm parameters for the parameter combination to be optimized. The particle position is represented by the sum of the envelope entropies of all intrinsic mode functions after VMD decomposition, which serves as the fitness function. The particle position is updated through collision and bounce mechanisms, and the globally optimal parameter combination is obtained through iterative optimization. ; in, For modal number, As a penalty factor, To optimize the number of modes, Take the integer. This is for optimizing the penalty factor.

[0013] Optionally, the globally optimal parameter combination can be used. The step of performing VMD decomposition on the single-channel fused signal includes: Using the globally optimal parameter combination Optimized mode number Determine the number of modes to be decomposed, and use the optimized penalty factor. As the weighting coefficient of the secondary penalty term, a constrained variational model is established. The constrained variational model aims to minimize the difference between the single-channel fused signal and the sum of each modal component. Introducing Lagrange multipliers The constrained variational model is transformed into an unconstrained problem, and the alternating direction multiplier method (ADMM) is used to iteratively solve the unconstrained problem. In each iteration, each modal component and its center frequency are updated sequentially until the preset convergence condition is met. Output the single-channel fused signal corresponding to Each intrinsic mode function component, and the The intrinsic mode function components are organized as follows: The input tensor has 10 channels, where each channel corresponds to a time series of an intrinsic mode function component.

[0014] Optionally, the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning adopts an architecture of dual-channel parallel fusion based on the temporal convolutional network TCN channel and Transformer channel, followed by a bidirectional long short-term memory network BiLSTM. The step of processing the intrinsic mode function components obtained by performing VMD decomposition in the fault diagnosis model includes: The VMD decomposition results The input tensor, composed of eigenmode function components, is simultaneously fed into the parallel temporal convolutional network TCN channel and Transformer channel. The Temporal Convolutional Network (TCN) channel extracts local temporal features through stacked dilated causal convolutional residual blocks, while the Transformer channel extracts global dependency features through a multi-head self-attention mechanism. The local temporal features output by the TCN channel of the temporal convolutional network and the global dependency features output by the Transformer channel are concatenated along the feature dimension to obtain the fused features; The fused features are input into a bidirectional long short-term memory network (BiLSTM) to capture bidirectional temporal context. The output of the last time step is then passed through a fully connected layer and a softmax function to obtain the probability distribution of various operating conditions.

[0015] Optionally, the processing steps of the temporal convolutional network (TCN) channel include: The input tensor is received, and features are extracted through stacked dilated causal convolution residual blocks. Each residual block sequentially performs dilated causal convolution, weight normalization, ReLU activation, and Dropout operations. The input and output are added through residual connections to output local temporal features.

[0016] Optionally, the processing steps of the Transformer channel include: The input tensor receives the same input as the channel of the temporal convolutional network (TCN), maps the input to a specified dimension through linear projection, and passes through a multi-head self-attention layer and a feedforward network in sequence. Each sub-layer is followed by residual connections and layer normalization to output global dependency features.

[0017] Optionally, the step of training the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning includes: The hybrid strategy Frost Ice Optimization Algorithm (HSRIME) is used to jointly optimize several key hyperparameters of the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning. These key hyperparameters include the TCN convolution kernel size, the number of Transformer self-attention heads, the number of BiLSTM hidden units, and the model learning rate. After iteratively optimizing to obtain the optimal combination of key hyperparameters, the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning is configured using this combination of key hyperparameters. The model is trained on the training set using the Adam optimizer and cross-entropy loss function, and an early stopping mechanism is set to prevent overfitting.

[0018] Optionally, the hybrid strategy frost-ice optimization algorithm HSRIME is used to jointly optimize multiple key hyperparameters of the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, including: The population is initialized using a Circle chaotic map, which makes the initial individuals uniformly distributed in the search space, and each individual corresponds to a set of candidate values ​​for the multiple key hyperparameters; In each iteration, a Gaussian-Cauchy difference hybrid mutation operation is performed on a subset of individuals. This combines the difference strategy, Gaussian distribution, and Cauchy distribution to generate mutated individuals, enhancing their ability to escape local optima. The hybrid mutation formula is as follows:

[0019] in, , , Three distinct individuals are randomly selected from the population. This is the difference scaling factor. and Let be random variables that follow standard Gaussian and standard Cauchy distributions, respectively. and The mixing coefficient, It is a mutated individual; A greedy selection mechanism is adopted to compare the fitness of mutated individuals with that of the original individuals, and retain the individuals with better fitness to enter the next generation. The fitness function adopts a surrogate evaluation strategy, which is to calculate the misclassification rate on the validation set after training a small number of times on a subset of the training set, in order to measure the fault classification performance of the fault diagnosis model under the current combination of key hyperparameters. The optimal combination of key hyperparameters is obtained through iterative optimization.

[0020] According to a second aspect of the present invention, a motor fault diagnosis system based on adaptive signal decomposition and hybrid deep learning is provided, comprising: The acquisition and preprocessing module is used to acquire multi-channel signals of the motor under diagnosis in real time and preprocess them, and then perform dimensionality reduction and fusion on the preprocessed multi-channel signals to obtain a single-channel fused signal. The adaptive signal decomposition module is used to perform VMD decomposition on the single-channel fused signal using the optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) to obtain the intrinsic mode function components. The fault diagnosis module is used to input the intrinsic mode function components into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which has been optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained. The model outputs the probability distribution of various operating conditions of the motor and takes the operating condition category corresponding to the highest probability as the fault judgment result.

[0021] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the processor is configured to implement the steps of the above-described motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning when executing a computer management program stored in the memory.

[0022] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer management program is stored, wherein when executed by a processor, the computer management program implements the steps of the above-described motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning.

[0023] This invention provides a motor fault diagnosis method, system, electronic device, and storage medium based on adaptive signal decomposition and hybrid deep learning. It overcomes the shortcomings of traditional methods that rely on manual trial and error to determine the number of modes and penalty factors by adaptively solving the optimal parameters of variational mode decomposition through a center collision optimization algorithm, enabling signal decomposition to more accurately extract fault-related impact features. By dimensionality reduction and fusion of multi-channel signals to obtain a single-channel signal, it reduces redundancy in subsequent processing and avoids mutual interference between signals of different physical quantities. The invention also utilizes a hybrid strategy frost-ice optimization algorithm to optimize TCN. Transformer The BiLSTM hybrid deep learning model globally optimizes multiple key hyperparameters, avoiding the problem of getting trapped in local optima that is common with manual parameter tuning. Simultaneously, this hybrid deep learning model captures local temporal patterns through a Temporal Convolutional Network (TCN), captures global dependencies through a Transformer network, and fuses bidirectional temporal context through a Bidirectional Long Short-Term Memory (LSTM) network, effectively improving the ability to extract fault features and the accuracy of classification. This invention enables automatic and accurate diagnosis of motor faults without relying on human experience, improving the reliability and efficiency of diagnosis. Attached Figure Description

[0024] Figure 1 A flowchart of a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning provided by the present invention; Figure 2 A flowchart of a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning provided for one embodiment; Figure 3 A schematic diagram of a gating mechanism provided for one embodiment; Figure 4 A block diagram of a motor fault diagnosis system based on adaptive signal decomposition and hybrid deep learning provided by the present invention; Figure 5 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention; Figure 6 This is a schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0026] Figure 1 A flowchart of a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning provided by this invention is shown below. Figure 1 As shown, the method includes steps S1 to S3: S1: Collect multi-channel signals of the motor under diagnosis in real time and preprocess them. Then, perform dimension reduction and fusion on the preprocessed multi-channel signals to obtain a single-channel fused signal. S2, the optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are used to perform VMD decomposition on the single-channel fused signal to obtain the intrinsic mode function components; S3. Input the intrinsic mode function components into the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which has been optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained, to output the probability distribution of various operating conditions of the motor, and take the operating condition category corresponding to the highest probability as the fault judgment result.

[0027] Understandably, given the shortcomings in the background technology, this invention proposes a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning. This method adaptively solves for the optimal parameters of variational mode decomposition using a center collision optimization algorithm, overcoming the shortcomings of traditional methods that rely on manual trial and error to determine the number of modes and penalty factors. This allows signal decomposition to more accurately extract fault-related impact features. By dimensionality reduction and fusion of multi-channel signals to obtain a single-channel signal, redundancy in subsequent processing is reduced and mutual interference between signals of different physical quantities is avoided. The hybrid strategy frost-ice optimization algorithm is used to globally optimize several key hyperparameters of the TCN-Transformer-BiLSTM hybrid deep learning model, avoiding the problem of easily getting trapped in local optima with manual parameter tuning. Simultaneously, this hybrid deep learning model captures local temporal patterns through a temporal convolutional network (TCN), captures global dependencies through a Transformer network, and fuses bidirectional temporal context through a bidirectional long short-term memory network, effectively improving the extraction capability and classification accuracy of fault features.

[0028] This invention enables automatic and accurate diagnosis of motor faults without relying on human experience, thus improving the reliability and efficiency of diagnosis.

[0029] Based on the above technical solutions, the embodiments of the present invention can be further improved as follows.

[0030] refer to Figure 2 As shown in the figure, this embodiment provides a motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning, which is divided into an offline training stage and an online fault diagnosis stage. The offline stage is used to build and train the fault diagnosis model, and the online stage uses the trained fault diagnosis model to diagnose the real-time operating status of the motor.

[0031] I. Offline Training Phase The offline training phase mainly includes signal acquisition and preprocessing, dimensionality reduction and fusion, VMD parameter optimization and batch VMD decomposition based on the center collision optimization algorithm (CCO), construction of a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, optimization of model hyperparameters using the hybrid strategy frost ice optimization algorithm (HSRIME), and model training.

[0032] 1. Signal Acquisition and Preprocessing Multi-channel signals from a permanent magnet synchronous motor are acquired under various operating conditions, including normal operation, inter-turn short-circuit fault, permanent magnet demagnetization fault, and bearing damage fault. The multi-channel signals consist of three-phase current signals (U-phase, V-phase, and W-phase) and vibration signals, totaling four channels.

[0033] The acquired 4-channel raw signals are preprocessed, including missing value filling and outlier removal, to eliminate sensor noise and transient interference.

[0034] Then, a sliding window is used to segment the preprocessed continuous signal. The window length is set to L sampling points, the step size is set to L / 2, and each signal segment segmented by the window is a sample to be diagnosed. The shape of each sample is (4, L). Each sample to be diagnosed in the offline stage corresponds to a fault category label, forming a labeled time-series-based multidimensional sample set.

[0035] The multidimensional sample set was divided into a training set and a test set in a 7:3 ratio. Stratified sampling was used to ensure that the distribution ratio of various types of fault samples was consistent in the training set and the test set.

[0036] For each sample in both the training and test sets, min-max normalization was performed on all four channels to linearly map the signal to the [0,1] interval, thus eliminating the dimensional differences between the current and vibration signals. The normalization formula is as follows:

[0037] in, This represents the i-th sampling point in the original sequence. and These are the minimum and maximum values ​​of this feature dimension across all training samples in the training set, respectively. The values ​​are normalized dimensionless values, i = 1, 2, ..., L, where L is the sliding window length (corresponding to the number of sampling points in a single sample). The test set is obtained using statistics from the training set. and Normalization is performed, that is, the test set is transformed using the normalization parameters of the training set statistics to prevent information leakage.

[0038] 2. Dimensional Reduction and Integration Principal component analysis (PCA) is performed on the normalized 4-channel signal for dimensionality reduction and fusion to obtain a single-channel fused signal. The specific steps include: First, the normalized 4-channel signals are combined into a 4×L multi-channel matrix, and the matrix is ​​then decentered to eliminate the DC offset of each channel. Then, the covariance matrix of the decentralized matrix is ​​calculated, and the covariance matrix is ​​decomposed into eigenvalues ​​to obtain each eigenvalue and its corresponding eigenvector. Next, the eigenvector corresponding to the largest eigenvalue is taken as the projection direction. The multi-channel signals are then weighted and summed along this projection direction, and the resulting first principal component signal is used as the single-channel fused signal. This projection direction is pre-calculated and saved from historical training data, and can be directly called up during the online phase.

[0039] 3. Adaptive optimization of VMD parameters and VMD decomposition based on the center collision optimization algorithm (CCO) The Center Collision Optimization (CCO) algorithm is used to adaptively optimize the parameters of Variational Mode Decomposition (VMD) to obtain the optimal combination of decomposition parameters. Then, the optimized combination of decomposition parameters is used to perform VMD decomposition on the PCA-fused single-channel signal to obtain multi-scale intrinsic mode components containing fault characteristics. The specific steps are as follows: (1) CCO algorithm initialization and parameter setting Set the population size and maximum number of iterations, for example, a population size of 30, a maximum number of iterations of 30, and a search space of 2. This will allow for the optimization of the parameter combinations. As the particle's position, where... For modal number, This is a penalty factor. The integer is set in this embodiment. The search range is [2,10]; α As a continuous value, this embodiment sets α The search range is [1000, 5000]. Each particle is encoded as a two-dimensional vector. , representing a set of candidate VMD parameter combinations. The positions of all particles are randomly initialized in the search space. After the update, round it down to the nearest integer.

[0040] (2) Define the fitness function The fitness function is used to evaluate each set of candidate parameters, that is, each combination of parameters to be optimized. The advantages and disadvantages. Specifically defined as: A representative sample is selected from the training set (this sample contains typical waveform segments of normal state and various fault types). VMD decomposition is then performed on the single-channel fused signal of this sample after PCA fusion to obtain... Each intrinsic mode function (IMF) component is represented. The envelope entropy of each IMF component is calculated, and all... The sum of the envelope entropies of each IMF component is used as the fitness value of the set of parameters to be optimized.

[0041] The calculation process for envelope entropy is as follows: First, for each eigenmode function component Perform a Hilbert transform to obtain the analytic signal. ,in, This represents the Hilbert transform, here. Represents the imaginary unit; Then, calculate the analytic signal. The magnitude of the signal is used to obtain the envelope signal. ; Next, the envelope signal Normalized to a probability distribution:

[0042] Where L is the signal length, corresponding to the number of sampling points in the sample. Refers to the envelope signal based on a continuous-time variable. Refers to the envelope signal of sampling point i in a discrete sequence. This refers to the envelope signal of sampling point j in a discrete sequence, where i and j are both sampling point indices. Finally, the information entropy of this probability distribution is calculated, which is the envelope entropy of this IMF component:

[0043] Calculate the envelope entropy of each IMF component, and then... The sum of the envelope entropies of each IMF component is used as the fitness value. The smaller the envelope entropy, the richer and more regular the impact fault features contained in the IMF component, the better the decomposition effect, and the more beneficial it is for subsequent fault diagnosis.

[0044] (3) Iterative optimization The CCO algorithm simulates the physical process of center collisions between objects. Particles update their positions in the search space through collision and bounce mechanisms: during the collision phase, particles move towards the current optimal solution, and during the bounce phase, particles diffuse in the opposite direction after the collision to explore new areas.

[0045] In each iteration, perform the following operations for each particle: 1) Fitness assessment: using the current combination of parameters represented by the particle. Perform VMD decomposition on the PCA fusion signal of representative samples and calculate the fitness value of the particle according to the defined fitness function.

[0046] 2) Update individual optimality: Compare the particle's current fitness value with its historical best fitness value. If the current fitness value is better, update the particle's individual optimal position. 3) Update the global optimum: Update the position of the individual with the best fitness among all particles to the global optimum position.

[0047] 4) Particle position update: Each particle is updated based on its current position, the global best position, and the population center position.

[0048] The particle position update formula is:

[0049] Here, For the number of iterations, It indicates that the i-th particle is in the... Position at the next iteration For the i-th particle in the... The position at the next iteration, i.e., the position of the particle after the update. This represents the current global optimal position of the particle. This represents the center position of all particles in the current population (i.e., the average position of all particles). and It is a random number in the interval [0,1].

[0050] It is worth noting that in the position update formula This represents the collision behavior of a particle moving towards the global optimal position, driving the particle to move towards the optimal solution; This represents the bounce behavior of particles moving away from the population center. The center-bounce mechanism enhances population diversity and avoids premature convergence. The particle position update formula, through the synergistic effect of collision and bounce mechanisms, ensures both the convergence of the algorithm and the maintenance of population diversity, preventing premature entrapment in local optima.

[0051] After the particle position is updated, the number of modes is... The components are rounded down to ensure... Always an integer.

[0052] Repeat the above iterative update steps until the maximum number of iterations is reached (e.g., the preset 30 times).

[0053] (4) Output the optimal parameters After the iteration is complete, the output is the globally optimal combination of parameters that minimizes the fitness. , To optimize the number of modes, Take the integer. This represents the optimized penalty factor. This optimal parameter combination represents the best number of modes and penalty factor that, under the current data conditions, allows VMD decomposition to obtain the richest fault characteristic information.

[0054] (5) Batch VMD decomposition Global optimal parameter combination obtained by CCO optimization VMD decomposition is performed on the PCA fused signals (i.e., the single-channel fused signals) of all samples in the training and test sets. The specific process of VMD decomposition is as follows: 1) Establish a constrained variational model Using the optimized modal number Determine the number of modes to be decomposed, and use the optimized penalty factor. As the weighting coefficients of the secondary penalty term, a constrained variational model is established. The constrained variational model aims to minimize the difference between the original signal and the sum of each modal component, with the constraint that the sum of each modal component equals the original signal.

[0055] The constrained variational model of VMD decomposition is expressed as:

[0056] The constraints are expressed as follows:

[0057] in, The original signal (i.e., the single-channel fused signal obtained by PCA fusion). For the k-th modal component, The center frequency of this component. Here, the Dirac function is used. The imaginary unit, This represents the convolution operation.

[0058] In the objective function of the constrained variational model, each term represents the analytic signal obtained by performing a Hilbert transform on each modal component. This signal is then multiplied by an exponential term to shift its spectrum to baseband. Finally, the square of the L2 norm of its gradient is calculated to measure the bandwidth of that modal component. The constraint requires that the sum of all modal components equals the original signal, ensuring the completeness of the decomposition.

[0059] 2) Transform the constrained variational model into an unconstrained problem. Introducing a secondary penalty factor and Lagrange multipliers Among them, the optimized penalty factor is used. The weighting coefficients, serving as the second-order penalty term, transform the constrained variational model into an unconstrained augmented Lagrangian problem, as follows:

[0060] The quadratic penalty term is used to ensure the fidelity of the reconstructed signal in a Gaussian noise environment, while the Lagrange multiplier term is used to strictly enforce the constraints.

[0061] 3) Solve using ADMM iterative method The augmented Lagrange problem is solved iteratively using the alternating direction multiplier method (ADMM). For example, the following three update steps are executed sequentially in each iteration: A. Update modal components: Update each modal component in the frequency domain so that each modal component maintains a compact bandwidth around its center frequency.

[0062] B. Update center frequency: Reset the center frequency of each modal component to the centroid of its power spectrum.

[0063] C. Update the Lagrange multipliers.

[0064] After VMD decomposition is completed, the output of each sample is... Each intrinsic mode function component, and the The intrinsic mode function components are organized as follows: The input tensor has *n* channels, with each channel corresponding to a time series of an intrinsic mode function component. Each sample is ultimately represented as a tensor of shape *n*. The input tensor has rows and L columns, where L is the length of the sliding window.

[0065] 4. Construct a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning. A fault diagnosis model based on a hybrid deep learning approach (TCN-Transformer-BiLSTM) is constructed for multi-fault classification in motors. This model employs a dual-channel parallel architecture, consisting of a Temporal Convolutional Network (TCN) channel and a Transformer channel. Following this dual-channel parallel architecture is a Bidirectional Long Short-Term Memory (BiLSTM) network and a classification layer. Specifically, the TCN and Transformer channels process the same input tensor, outputting high-dimensional features which are then concatenated and fused before being input into the BiLSTM to capture the bidirectional temporal context. Finally, the model passes through a fully connected layer and a Softmax classifier to output the fault category probability distribution. The specific architecture is as follows: (1) TCN channel The TCN channel receives the input tensor and performs feature extraction through stacked dilated causal convolution residual blocks.

[0066] The TCN channel consists of stacked residual blocks. Each residual block contains a dilated causal convolutional layer, a weight normalization layer, a ReLU activation function, and a Dropout regularization layer. These operations are performed sequentially: dilated causal convolution, weight normalization, ReLU activation, and Dropout. The input and output are then summed via residual connections to output local temporal features. The mathematical expression for dilated causal convolution is:

[0067] In this formula, Given the input sequence, It is the expansion factor. Indicates the expansion factor as convolution, Here, k is the kernel size, and k is the kernel size. This is achieved by stacking residual blocks with exponentially growing inflation factors (such as...). =1,2,4), TCN can achieve a large receptive field with fewer layers, effectively capturing long-range dependencies in motor fault signals. The output of each residual block is added to the input through a residual connection:

[0068] in, This refers to the output of dilated causal convolutions and subsequent layers (WeightNorm, ReLU, Dropout, etc.) within the residual block. This is the input for the residual block; For activation functions; For 1×1 convolution, when the number of input and output channels is inconsistent, Used for dimension matching.

[0069] (2) Transformer channel The Transformer channel consists of a multi-head self-attention mechanism and a feedforward network. The Transformer channel receives the same input tensor as the TCN channel, maps the input to a specified dimension through linear projection, and then passes it through the multi-head self-attention layer and the feedforward network in sequence. Each sub-layer is followed by residual connections and layer normalization to output globally dependent features.

[0070] Specifically, the formula for calculating multi-head self-attention is:

[0071]

[0072]

[0073] In this formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. This is the transpose of the key matrix. The dimension of the key vector; The attention scores are transformed into a probability distribution by using a row-normalized function; For the number of self-attention heads, For the first The output of each attention head; For the splicing operation, splicing is performed along the feature dimension. Output of the head; , , and Let be a trainable projection matrix, where Indicates the first The query projection matrix of the head. Indicates the first The key projection matrix of the head, Indicates the first The projection matrix of the head value, This indicates the output projection matrix.

[0074] The multi-head self-attention mechanism enables the fault diagnosis model to capture global dependencies between different locations in the motor fault signal from different representation subspaces. Each self-attention sublayer is followed by a feedforward network: And residual connections and layer normalization are used. Here, For input vectors / matrices, This is the first layer weight matrix. This is the weight matrix for the second layer. This is the first layer bias vector. This is the second layer bias vector. (3) Dual-channel output feature fusion and BiLSTM The fused features are obtained by concatenating the local temporal features output from the TCN channel with the global dependency features output from the Transformer channel along the feature dimension. Specifically, the local temporal features output from the TCN channel are the feature vectors. (For example Figure 3 As shown The global dependency features output by the Transformer channels are feature vectors. (For example Figure 3 As shown ), to feature vector With feature vectors By concatenating the features along the feature dimension, a fused feature vector is obtained. (For example Figure 3 As shown ). (Merge feature vectors) Input BiLSTM, capture bidirectional timing context information, and take the output of the last time step.

[0075] BiLSTM consists of a forward LSTM and a backward LSTM. The output is obtained by concatenating the hidden states from the two directions. ,in, This indicates that the forward LSTM is at time [time]. The hidden state, Indicates the inverse LSTM at time... The hidden state.

[0076] refer to Figure 3 As shown, the gating mechanism inside the LSTM unit is as follows:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] In this formula, Input for the current time, for example Figure 3 As shown ; The hidden state of the previous time step represents the time step. The output carries short-term memory information; The current state is hidden, representing the time step. The output of is used as the input of the next layer / next time step, for example Figure 3 As shown ; Indicates will and Concatenate them along the feature dimensions to form a single vector; The cell state at the previous moment serves as the "long-term memory" carrier for LSTM; This represents the current cell state, which is then updated as long-term memory. The candidate cell state is represented by the hyperbolic tangent function. The generated candidate information values ​​will be written after being filtered by the input gate. These are the forget gate, input gate, and output gate; the forget gate The value (0,1) determines how much information to retain / discard from the old cell state; the input gate... The value (0,1) determines the candidate information. How much is written to the cell state; output gate The value (0,1) determines the cell state. Which parts of the output are ; , , , is the weight matrix, which corresponds to the linear transformation parameters of the forget gate, input gate, output gate, and candidate state, respectively; , , , These are bias vectors, each corresponding to a bias term for a gate / candidate state; σ is the Sigmoid activation function, with an output range of (0,1), used for gating (on / off); ⊙ is the Hadamard product, representing element-wise multiplication; Indicates will and Concatenate them along the feature dimension to form a vector.

[0083] Finally, the output vector of the last time step of the BiLSTM is input into the fully connected layer, and the probability distribution of various faults is obtained through the Softmax function:

[0084] In the probability distribution expression, Indicates input Belongs to the The posterior probability of a class. For the input sample, C represents the category label, and C is the total number of fault categories (including normal state). In this embodiment, C=4. This represents the value of the j-th output node in the fully connected layer. For the fully connected layer The value of each output node, Iterate through all C categories for normalization. Take the category with the highest probability as the fault diagnosis result.

[0085] 5. Hybrid Strategy Frost Ice Optimization Algorithm (HSRIME) for Optimizing Model Hyperparameters The Hybrid Frost Ice Optimization (HSRIME) algorithm is used to jointly optimize several key hyperparameters of the fault diagnosis model. The key hyperparameters to be optimized include the TCN convolution kernel size (search range [2,15]), the number of Transformer self-attention heads (search range [3,15]), the number of BiLSTM hidden units (search range [10,200]), and the model learning rate (search range [0.0001,0.2]).

[0086] The HSRIME algorithm, based on the standard Frost Ice Optimization (RIME) algorithm, introduces the following improvement strategies, and the specific optimization process is as follows: (1) Circle chaotic mapping to initialize the population A Circle chaotic mapping is used instead of random initialization to generate an initial population, ensuring that the initial individuals are uniformly distributed in the search space, with each individual corresponding to a set of candidate values ​​for the multiple key hyperparameters. This setting enhances the uniformity and ergodicity of the initial population in the search space. The expression for the Circle mapping is:

[0087] In this formula, For the first A chaotic variable, This is for modulo operations. The initial population generated by the Circle mapping is more evenly distributed in the search space, which helps improve the global search efficiency of the algorithm.

[0088] (2) Fitness assessment The fitness function employs a surrogate evaluation strategy: it trains a small number of times on a subset of the training set and then calculates the misclassification rate on the validation set to measure the fault classification performance of the model under the current hyperparameter combination.

[0089] During model optimization, the minimum misclassification rate on the training set is used as the fitness function.

[0090] In the fitness function expression, The number of samples in the training set. For the fault category predicted by the model, For actual fault categories, This is an indicator function that indicates when the predicted class is inconsistent with the true class. Select 1 if the value is 1, otherwise select 0.

[0091] (3) Gauss-Cauchy difference mixed variation strategy In each iteration, a hybrid mutation operation is performed on a subset of individuals, combining the difference strategy, Gaussian distribution, and Cauchy distribution to generate mutated individuals, thereby enhancing the algorithm's ability to escape local optima. The hybrid mutation formula is:

[0092] In this formula, , , Three distinct individuals are randomly selected from the population. This is the difference scaling factor. and Let be random variables that follow standard Gaussian and standard Cauchy distributions, respectively. and The mixing coefficient, Let be the position of the i-th mutated individual.

[0093] In the formula, the difference term Using differences among individuals in a population to guide the direction of variation, Gaussian terms A fine-grained search is performed near the mean, yielding Cauchy terms. By utilizing the thick-tailed characteristic of the Cauchy distribution, large-scale jumps can be performed, enhancing global exploration capabilities.

[0094] (4) Greedy selection mechanism This embodiment employs a greedy selection mechanism, comparing the fitness of the mutated individuals with that of the original individuals, and retaining the better one, i.e., the individual with higher fitness, to enter the next generation, as shown below:

[0095] In this formula, f(·) is the fitness function (i.e., the misclassification rate). When the fitness of the mutated individual is better than that of the original individual, it is replaced; otherwise, the original individual is retained.

[0096] The HSRIME algorithm in this embodiment outputs the optimal combination of key hyperparameters after a preset number of iterations (e.g., 30 times).

[0097] 6. Fault Diagnosis Model Training The optimal combination of key hyperparameters obtained through HSRIME optimization is used to configure a fault diagnosis model based on a TCN-Transformer-BiLSTM hybrid deep learning system. The configured fault diagnosis model is trained on the training set using the Adam optimizer and a cross-entropy loss function, where the cross-entropy loss function is:

[0098] In this formula, Let i be the true label of the i-th sample belonging to the j-th class (one-hot encoding). This represents the probability predicted by the model that the i-th sample belongs to the j-th class. In this embodiment, the batch size is set to 64, the training epochs are set to 200, and an early stopping mechanism is used to prevent overfitting.

[0099] After training is complete, save the model weights and all parameters determined in the offline phase, including normalization parameters, PCA projection direction, and VMD optimal parameters. The optimal combination of key hyperparameters for the fault diagnosis model is determined. Then, the test set data is input into the optimized model to perform fault classification and output fault results to verify the fault classification effect.

[0100] II. Online Fault Diagnosis Phase 1. Real-time signal acquisition and preprocessing Multi-channel signals from the real-time operation of the motor under diagnosis are acquired. These signals include three-phase current signals and vibration signals, totaling four channels. The acquired signals are then sequentially processed with missing value filling, outlier removal, and sliding window segmentation. The window length (L) and step size (L / 2) are the same as in the offline stage. Finally, the normalization parameters saved from the offline stage are used to perform min-max normalization on each channel signal.

[0101] 2. PCA Dimensionality Reduction and Fusion Using the PCA projection direction pre-calculated in the offline stage, the normalized 4-channel signals are weighted and fused into a single-channel fused signal.

[0102] 3. VMD decomposition The globally optimal parameter combination obtained by offline CCO optimization VMD decomposition is performed on the single-channel signal after PCA fusion to obtain Each intrinsic mode function component is organized into 1,000 eigenmode function components. The input tensor of each channel.

[0103] 4. Model Reasoning and Diagnosis The The input tensors of each channel are fed into a pre-trained fault diagnosis model based on a hybrid deep learning approach of TCN-Transformer-BiLSTM. The fault diagnosis model outputs the probability distribution of various operating conditions. The category with the highest probability is taken as the fault determination result for the current motor. When a fault condition is diagnosed, an alarm is triggered and the specific fault type is indicated, providing a basis for maintenance decisions.

[0104] This embodiment divides the method into two stages: offline training and online diagnosis. In the offline stage, CCO adaptively optimizes VMD parameters and HSRIME optimizes the fault diagnosis model hyperparameters, constructing and training a high-performance fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning. In the online stage, all parameters determined in the offline stage are reused, achieving rapid and accurate diagnosis of real-time motor operating signals. This method effectively solves problems such as the difficulty in adaptively determining signal decomposition parameters, the limited feature extraction capability of a single model, and the tendency to get trapped in local optima during multi-parameter joint optimization. It has the advantages of high diagnostic accuracy and strong robustness.

[0105] Figure 4 A structural diagram of a motor fault diagnosis system based on adaptive signal decomposition and hybrid deep learning provided in an embodiment of the present invention is shown below. Figure 4 As shown, a motor fault diagnosis system based on adaptive signal decomposition and hybrid deep learning includes an acquisition and preprocessing module, an adaptive signal decomposition module, and a fault diagnosis module, wherein: The acquisition and preprocessing module is used to acquire multi-channel signals of the motor under diagnosis in real time and preprocess them, and then perform dimensionality reduction and fusion on the preprocessed multi-channel signals to obtain a single-channel fused signal. The adaptive signal decomposition module is used to perform VMD decomposition on the single-channel fused signal using the optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) to obtain the intrinsic mode function components. The fault diagnosis module is used to input the intrinsic mode function components into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which has been optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained. The model outputs the probability distribution of various operating conditions of the motor and takes the operating condition category corresponding to the highest probability as the fault judgment result.

[0106] It is understood that the motor fault diagnosis system based on adaptive signal decomposition and hybrid deep learning provided by the present invention corresponds to the motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning provided in the foregoing embodiments. The relevant technical features of the motor fault diagnosis system based on adaptive signal decomposition and hybrid deep learning can be referred to the relevant technical features of the motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning, and will not be repeated here.

[0107] Please see Figure 5 , Figure 5 A schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 5 As shown, this embodiment of the invention provides an electronic device 500, including a memory 510, a processor 520, and a computer program 511 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 511, it performs the following steps: The multi-channel signals of the motor under diagnosis are collected and preprocessed in real time. The preprocessed multi-channel signals are then fused to obtain a single-channel fused signal. The optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are used to perform VMD decomposition on the single-channel fused signal to obtain the intrinsic mode function components. The intrinsic mode function components are input into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which is optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained. The model outputs the probability distribution of various operating conditions of the motor, and the operating condition category corresponding to the highest probability is used as the fault judgment result.

[0108] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 6 As shown, this embodiment provides a computer-readable storage medium 600, on which a computer program 511 is stored. When the computer program 511 is executed by a processor, it performs the following steps: The multi-channel signals of the motor under diagnosis are collected and preprocessed in real time. The preprocessed multi-channel signals are then fused to obtain a single-channel fused signal. The optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are used to perform VMD decomposition on the single-channel fused signal to obtain the intrinsic mode function components. The intrinsic mode function components are input into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which is optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained. The model outputs the probability distribution of various operating conditions of the motor, and the operating condition category corresponding to the highest probability is used as the fault judgment result.

[0109] This invention provides a motor fault diagnosis method, system, electronic device, and storage medium based on adaptive signal decomposition and hybrid deep learning. The method adaptively determines the number of modes and penalty factor in variational mode decomposition (VMD) using the center collision optimization algorithm (CCO), enabling signal decomposition to automatically match the optimal decomposition scale based on motor operating data, thereby more fully separating the intrinsic mode components containing fault impact characteristics. Based on this, a hybrid deep learning model is constructed, in which a temporal convolutional network (TCN) and a Transformer extract local temporal features and global dependency features in parallel, and a bidirectional long short-term memory (BiLSTM) network fuses bidirectional temporal information. Furthermore, a hybrid strategy, the BiLSTM optimization algorithm, is used to globally optimize key hyperparameters such as kernel size, number of attention heads, number of hidden units, and learning rate, avoiding the blindness of manual parameter tuning and the insufficient feature extraction capabilities of a single model. This invention combines adaptive signal decomposition with a hybrid deep learning model of hyperparameter optimization, enabling online fault diagnosis of motors to no longer rely on human experience. It can more accurately distinguish between various operating conditions such as normal state, inter-turn short circuit, permanent magnet demagnetization, and bearing damage, thereby improving the accuracy and robustness of diagnosis.

[0110] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0111] 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.

[0112] 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 computer, 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. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] 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.

[0114] 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.

[0115] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0116] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning, characterized in that, include: The multi-channel signals of the motor under diagnosis are collected and preprocessed in real time. The preprocessed multi-channel signals are then fused to obtain a single-channel fused signal. The optimal parameters obtained by adaptively solving variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are used to perform VMD decomposition on the single-channel fused signal to obtain the intrinsic mode function components. The intrinsic mode function components are input into a fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, which is optimized by the hybrid strategy frost-ice optimization algorithm HSRIME and trained. The model outputs the probability distribution of various operating conditions of the motor, and the operating condition category corresponding to the highest probability is used as the fault judgment result.

2. The method of claim 1, wherein, The process of acquiring and preprocessing multi-channel signals from the real-time operation of the motor under diagnosis includes: Collect multi-channel signals of the motor under diagnosis during its operation. The multi-channel signals include three-phase current signals and vibration signals. The operating state is normal state, inter-turn short circuit fault state, permanent magnet demagnetization fault state, or bearing damage fault state. The multi-channel signal is preprocessed by missing value filling and outlier removal, and the preprocessed multi-channel signal is divided into a multi-dimensional sample set to be diagnosed according to a preset scale. The signals of each channel in the multidimensional sample set to be diagnosed are subjected to min-max normalization to linearly map the feature data of each dimension to the [0,1] interval, thus obtaining the normalized multidimensional sample set.

3. The method of claim 2, wherein the method further comprises: The step of dimensionality reduction and fusion of the preprocessed multi-channel signals to obtain a single-channel fused signal includes: The normalized multidimensional samples of the normalized multidimensional sample set are used to form a multi-channel matrix, and the multi-channel matrix is ​​then decentered. Calculate the covariance matrix of the decentralized matrix, and perform eigenvalue decomposition on the covariance matrix to obtain each eigenvalue and its corresponding eigenvector. The projection direction, which is calculated and saved in advance using historical data, is called. The projection direction is the feature vector corresponding to the largest feature value among the feature values ​​calculated using historical training data. The multi-channel signals of the normalized multi-dimensional sample set are weighted and summed in the projection direction to obtain the first principal component signal as the single-channel fusion signal.

4. The method of claim 1, wherein, The optimal parameters for adaptively solving the variational mode decomposition (VMD) based on the center collision optimization algorithm (CCO) are obtained in advance during the offline training phase, including: Initialization of CCO algorithm parameters, to be optimized parameter combination As the particle position, the sum of the envelope entropy of all the intrinsic mode functions after VMD decomposition is taken as the fitness function; the particle position is updated through the collision and rebound mechanism, and the global optimal parameter combination is obtained through iterative optimization ; wherein, is the modal number, is the penalty factor, is the optimized modal number, is the integer, is the optimized penalty factor.

5. The motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning according to claim 4, characterized in that, adopting the globally optimal parameter combination The step of performing VMD decomposition on the single-channel fusion signal comprises: Utilize the global optimal parameter combination Optimized modal number in Determine the number of modes to be decomposed, and utilize the optimized penalty factor As the weight coefficient of the secondary penalty term, establish a constrained variational model, which aims to minimize the difference between the single-channel fusion signal and the sum of each modal component Introducing Lagrange multipliers The constraint variational model is converted into an unconstrained problem, and the unconstrained problem is solved iteratively by using an alternating direction method of multipliers (ADMM), and each time the modal components and the center frequency are updated in turn until a preset convergence condition is met. Output the single-channel fused signal corresponding to Each intrinsic mode function component, and the The intrinsic mode function components are organized as follows: The input tensor has 10 channels, where each channel corresponds to a time series of an intrinsic mode function component.

6. The method of claim 1 or 5, wherein, The fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning adopts an architecture of dual-channel parallel fusion based on the temporal convolutional network TCN channel and Transformer channel, followed by a bidirectional long short-term memory network BiLSTM. The step of processing the intrinsic mode function components obtained by performing VMD decomposition in the fault diagnosis model includes: The VMD decomposition results The input tensor, composed of eigenmode function components, is simultaneously fed into the parallel temporal convolutional network TCN channel and Transformer channel. The Temporal Convolutional Network (TCN) channel extracts local temporal features through stacked dilated causal convolutional residual blocks, while the Transformer channel extracts global dependency features through a multi-head self-attention mechanism. The local temporal features output by the TCN channel of the temporal convolutional network and the global dependency features output by the Transformer channel are concatenated along the feature dimension to obtain the fused features; The fused features are input into a bidirectional long short-term memory network (BiLSTM) to capture bidirectional temporal context. The output of the last time step is then passed through a fully connected layer and a softmax function to obtain the probability distribution of various operating conditions.

7. The motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning according to claim 6, characterized in that, The processing steps for the TCN channel of the temporal convolutional network include: The input tensor is received, and features are extracted through stacked dilated causal convolution residual blocks. Each residual block sequentially performs dilated causal convolution, weight normalization, ReLU activation, and Dropout operations. The input and output are added through residual connections to output local temporal features.

8. The motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning according to claim 6, characterized in that, The processing steps of the Transformer channel include: The input tensor receives the same input as the channel of the temporal convolutional network (TCN), maps the input to a specified dimension through linear projection, and passes through a multi-head self-attention layer and a feedforward network in sequence. Each sub-layer is followed by residual connections and layer normalization to output global dependency features.

9. The motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning according to claim 6, characterized in that, The steps for training the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning include: The hybrid strategy Frost Ice Optimization Algorithm (HSRIME) is used to jointly optimize several key hyperparameters of the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning. The key hyperparameters include TCN convolution kernel size, number of Transformer self-attention heads, number of BiLSTM hidden units, and model learning rate. After iteratively optimizing to obtain the optimal combination of key hyperparameters, the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning is configured using this combination of key hyperparameters. The model is trained on the training set using the Adam optimizer and cross-entropy loss function, and an early stopping mechanism is set to prevent overfitting.

10. The motor fault diagnosis method based on adaptive signal decomposition and hybrid deep learning according to claim 9, characterized in that, The hybrid strategy frost-ice optimization algorithm HSRIME is used to jointly optimize multiple key hyperparameters of the fault diagnosis model based on TCN-Transformer-BiLSTM hybrid deep learning, including: The population is initialized using a Circle chaotic map, which makes the initial individuals uniformly distributed in the search space, and each individual corresponds to a set of candidate values ​​for the multiple key hyperparameters; In each iteration, a Gaussian-Cauchy difference hybrid mutation operation is performed on a subset of individuals. This combines the difference strategy, Gaussian distribution, and Cauchy distribution to generate mutated individuals, enhancing their ability to escape local optima. The hybrid mutation formula is as follows: in, , , Three distinct individuals are randomly selected from the population. This is the difference scaling factor. and Let be random variables that follow standard Gaussian and standard Cauchy distributions, respectively. and The mixing coefficient, It is a mutated individual; A greedy selection mechanism is adopted to compare the fitness of mutated individuals with that of the original individuals, and retain the individuals with better fitness to enter the next generation. The fitness function adopts a surrogate evaluation strategy, which is to calculate the misclassification rate on the validation set after training a small number of times on a subset of the training set, in order to measure the fault classification performance of the fault diagnosis model under the current combination of key hyperparameters. The optimal combination of key hyperparameters is obtained through iterative optimization.