Variable working condition traction motor diagnosis and prediction method based on parameter identification and transfer learning driving

Abstract

This invention provides a diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning, comprising the following steps: S1. Synchronous acquisition and adaptive preprocessing of multi-source heterogeneous signals; S2. Variational Bayesian unscented Kalman filter parameter identification and operating condition discrimination; S3. Multi-scale time-frequency feature extraction and enhanced feature construction; S4. Contrastive learning-driven cross-modal graph attention transfer and feature alignment; S5. Parameter drift-driven virtual sample generation and small sample fault classification; S6. Fractional-order stochastic differential equation-driven Bayesian update for remaining service life prediction; S7. Full-process adaptive parameter update driven by diagnostic prediction results. This invention effectively solves the core technical problems of separation between diagnosis and prediction and poor adaptability to varying operating conditions in existing technologies, realizing high-precision intelligent fault detection, fault category identification, and remaining service life prediction of traction motors in rail transit under complex varying operating conditions.

Description

Technical Field

[0001] This invention relates to the fields of condition monitoring and health management of rail transit traction motors, fault prediction and fault diagnosis, and in particular to a diagnostic and prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning. Background Technology

[0002] As a core power component of rail transit vehicles, the traction motor's operating status directly affects the safe and reliable operation of trains. With the rapid development of the rail transit industry, higher demands are placed on the operation and maintenance of traction motors. Traditional periodic maintenance methods suffer from over-maintenance or under-maintenance, making it difficult to meet the maintenance needs of modern rail transit. Condition-based predictive maintenance technology can formulate reasonable maintenance plans based on the actual operating status of the motor, effectively reducing maintenance costs and improving system reliability, and has become a current research hotspot.

[0003] Existing traction motor fault diagnosis and remaining service life prediction technologies still have many shortcomings in practical applications. First, traction motors operate under complex and variable conditions; motor parameters and signal characteristics differ significantly under different conditions such as starting, braking, and steady-state operation. Traditional methods struggle to adapt to these changes, leading to decreased diagnostic and prediction accuracy. Second, motors experience parameter drift during long-term operation, and existing methods, mostly based on fixed-parameter models, cannot effectively compensate for the impact of parameter drift. Furthermore, model performance gradually degrades with increasing operating time. Third, fault diagnosis and remaining service life prediction are typically treated as two independent tasks. Diagnostic results cannot provide effective information for prediction, and prediction results cannot be fed back to optimize the diagnostic model; there is a lack of deep integration between the two. In addition, traction motor fault samples are scarce, especially rare fault samples, leading to insufficient generalization ability of data-driven methods and hindering their widespread application in practical engineering. Summary of the Invention

[0004] This invention provides a diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning. By deeply integrating and bidirectionally interacting three algorithms—variational Bayesian unscented Kalman filtering, contrastive learning-driven cross-modal graph attention transfer algorithm, and fractional-order stochastic differential equation Bayesian online update algorithm—a fully closed-loop adaptive diagnostic prediction joint system is constructed. This effectively solves the core technical problems in existing technologies, such as insufficient parameter drift compensation, separation of diagnosis and prediction, weak generalization ability, scarce fault samples, and poor adaptability to varying operating conditions. It achieves high-precision intelligent fault detection, fault category identification, and remaining service life prediction for rail transit traction motors under complex varying operating conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A diagnostic prediction method for traction motors under varying operating conditions, driven by parameter identification and transfer learning, includes the following steps: S1. Synchronously acquire three-phase stator current signal, three-phase stator voltage signal and speed signal, perform adaptive preprocessing to obtain dq-axis current component, dq-axis voltage component and preprocessed speed signal; S2. The variational Bayesian unscented Kalman filter algorithm is used to identify the motor parameters online for the dq axis current component, dq axis voltage component and preprocessed speed signal, so as to obtain the motor parameter identification value and parameter drift vector, and at the same time calculate the working condition discrimination factor. S3. Based on the working condition discrimination factor, adaptively extract multi-scale time-frequency features, and fuse the multi-scale time-frequency features with the parameter drift vector to construct an enhanced feature vector; S4. A contrastive learning-driven cross-modal graph attention transfer algorithm is adopted. The graph structure is constructed based on the current feature vector part in the enhanced feature vector. The graph features are extracted using the parameter drift vector as conditional information. Domain adaptive feature alignment is performed to obtain the aligned target domain feature vector. S5. Based on the motor parameter identification value and parameter drift vector, generate virtual fault samples, mix the virtual fault samples with real fault samples to construct a support set, and use the prototype calibration small sample classification algorithm combined with the target domain feature vector to perform fault diagnosis, and obtain the fault category and fault probability.

[0006] The diagnostic prediction method for traction motors under varying operating conditions, driven by parameter identification and transfer learning, also includes: S6. Calculate the comprehensive degradation index based on the fault category and parameter drift vector, and use the Bayesian update algorithm driven by fractional stochastic differential equations to predict the remaining useful life and obtain the predicted value of the remaining useful life. S7. Calculate the process noise adjustment coefficient, observation noise adjustment coefficient, domain alignment loss weight coefficient, and prototype calibration coefficient based on the remaining useful life prediction value, and feed each coefficient back to S2, S4, and S5 respectively to form a closed-loop adaptive optimization.

[0007] In this specification, the adaptive preprocessing in S1 specifically includes: using a third-order Butterworth low-pass filter to remove high-frequency noise, performing Park transform on the three-phase stator current signal and the three-phase stator voltage signal to obtain the dq-axis current component and the dq-axis voltage component respectively, and performing moving average filtering on the speed signal to remove speed fluctuation noise. The electrical angle required for the Park transform is obtained by integrating the speed signal.

[0008] In this specification, when the variational Bayesian unscented Kalman filter algorithm in S2 is used for online identification of motor parameters, it first generates sigma points through unscented transformation. Each sigma point corresponds to a set of motor parameter states. Then, an explicit nonlinear observation function is used to calculate the observation prediction value of each sigma point. The explicit nonlinear observation function depends only on the motor parameter state corresponding to the sigma point, the dq axis voltage components, and the preprocessed speed signal.

[0009] In this specification, the parameter drift vector in S2 is a vector composed of the normalized difference between the motor parameter identification value and the reference parameter. The operating condition discrimination factor is used to distinguish between dynamic operating conditions and steady-state operating conditions. When the operating condition discrimination factor is greater than the operating condition discrimination threshold, it is determined to be a dynamic operating condition. When the operating condition discrimination factor is less than or equal to the operating condition discrimination threshold, it is determined to be a steady-state operating condition.

[0010] In this specification, the adaptive extraction of multi-scale time-frequency features in S3 specifically includes: when the condition is determined to be dynamic, wavelet packet transform is used to extract time-frequency features; when the condition is determined to be steady-state, both time-domain and frequency-domain features are extracted simultaneously.

[0011] In this specification, after the contrastive learning-driven cross-modal graph attention transfer algorithm in S4 completes offline training, it extracts samples from the source domain vibration signal training dataset to form a fixed proxy source domain sample set, extracts and stores the source domain feature vectors, and directly uses the stored source domain feature vectors to perform domain adaptive feature alignment calculation in the online stage.

[0012] In this specification, the contrastive learning-driven cross-modal graph attention transfer algorithm in S4 constructs a graph using a sliding window approach before performing domain adaptive feature alignment. The nodes of the graph correspond to each dimension of the current feature vector, the parameter drift vector serves as the conditional information during graph attention calculation, the feature of each node is the time series of that dimension within the sliding window, and the edge weight of the graph is the Pearson correlation coefficient between the feature time series of two nodes.

[0013] In this specification, the generation of virtual fault samples in S5 specifically includes: using the motor parameter identification value as the reference parameter for generating virtual samples, adjusting the corresponding parameter values ​​according to the parameter change rules corresponding to different fault types, determining the range of fault severity values ​​by the parameter drift vector, and generating current signals under different fault types and different degradation degrees.

[0014] In this specification, when using the prototype calibration few-sample classification algorithm for fault diagnosis in S5, a linear mapping layer is used to map the parameter drift vector to a calibration vector. The calibration vector is then used to calibrate the prototype vector. The parameters of the linear mapping layer are jointly trained with the prototype network during the meta-training phase. In summary, the present invention has at least the following beneficial effects: This invention effectively compensates for parameter drift during long-term operation of traction motors, significantly improving parameter identification accuracy under varying operating conditions. It achieves deep integration of fault diagnosis and remaining service life prediction; the diagnostic results provide accurate degradation mode information for prediction, while the prediction results optimize the parameters of the diagnostic model. Both mutually promote each other, jointly improving system performance.

[0015] This invention effectively improves the model's cross-condition generalization ability by employing cross-modal graph attention transfer and hybrid domain adaptation techniques, enabling it to adapt to the complex and ever-changing operating environment of traction motors. Virtual fault samples generated based on real-time parameter identification results can reflect the personalized degradation characteristics of the current motor, effectively solving the problem of scarce fault samples and improving fault diagnosis accuracy under small sample conditions.

[0016] The fully closed-loop adaptive update mechanism of this invention enables the model to continuously adapt to changes in the motor's operating state, continuously optimizing diagnostic and predictive performance as operating time increases, thus avoiding model performance degradation. The overall solution of this invention is logically sound, highly implementable, and can provide reliable technical support for predictive maintenance of rail transit traction motors. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning, which is involved in this invention.

[0019] Figure 2 This is a schematic diagram of the online identification and drift calculation of variational Bayesian unscented Kalman filter parameters involved in this invention.

[0020] Figure 3 This is a schematic diagram of the contrastive learning cross-modal graph attention transfer and domain adaptive alignment process involved in this invention.

[0021] Figure 4 This is a schematic diagram of the process of parameter drift-driven virtual sample generation and prototype calibration small sample classification involved in this invention. Detailed Implementation

[0022] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0023] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this embodiment provides a diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning. It addresses the needs of intelligent fault detection, fault category identification, and remaining service life prediction for rail transit traction motors under multiple operating conditions, including starting, braking, and steady-state operation. A fully closed-loop adaptive diagnostic prediction joint system with deep integration of three algorithms is constructed. This method employs a variational Bayesian unscented Kalman filter parameter identification algorithm, a contrastive learning-driven cross-modal graph attention transfer algorithm, and a fractional-order stochastic differential equation-driven Bayesian update remaining service life prediction algorithm. The three algorithms achieve pairwise bidirectional interaction, and through deep integration, they address core issues in existing technologies such as insufficient parameter drift compensation, separation of diagnosis and prediction, weak generalization ability, and scarce fault samples.

[0026] The specific steps are as follows: S1. Synchronous Acquisition and Adaptive Preprocessing of Multi-Source Heterogeneous Signals; A synchronous acquisition card with a time synchronization accuracy of no less than 1 microsecond is used to synchronously acquire three types of signals, with the sampling frequency of all signals uniformly set to 20 kHz. The first type of signal is the three-phase stator current signal, denoted as follows: , , The first type of signal uses a Hall effect current sensor with an accuracy of 0.1%. The sensor is installed at the output terminal of the stator winding of the traction motor to ensure accurate acquisition of current changes during motor operation. The second type of signal is the three-phase stator voltage signal, denoted as follows: , , The voltage is acquired using a 0.2-accuracy voltage sensor, which is connected in parallel to the three-phase power input of the traction motor. The third type of signal is the motor speed signal, denoted as... The speed measurement is acquired using an incremental photoelectric encoder with a resolution of 1024 lines. The encoder is installed on the output shaft of the traction motor and is coaxially connected to the motor shaft to ensure the accuracy of the speed measurement.

[0027] The acquired signals undergo adaptive preprocessing. First, a third-order Butterworth low-pass filter is used to remove high-frequency noise, with a cutoff frequency set to 5 kHz. The third-order Butterworth filter features a flat passband and fast stopband attenuation, effectively removing high-frequency electromagnetic interference and mechanical vibration noise generated during motor operation, while preserving effective fault characteristics in the signal. Then, Park transform is performed on the three-phase current and three-phase voltage signals respectively, converting the AC signals in the three-phase stationary coordinate system into DC signals in the dq rotating coordinate system. The specific formula for the Park transform is: ; ; in, The electrical angle of the motor rotor is obtained by integrating the speed signal; the specific calculation formula is as follows: , The number of pole pairs of the motor is determined by the technical parameters provided by the motor manufacturer. The d-axis current component is denoted as... The q-axis current component is denoted as The d-axis voltage component is denoted as The q-axis voltage component is denoted as The Park transform can convert complex three-phase AC signals into simple DC signals, greatly simplifying the subsequent parameter identification and feature extraction process.

[0028] Finally, a moving average filter is applied to the speed signal with a window size of 10 sampling points to remove speed fluctuation noise, resulting in a preprocessed speed signal. The moving average filter effectively suppresses high-frequency noise in the speed signal while preserving the overall trend of speed variation, providing accurate input for subsequent operating condition determination.

[0029] S2. Variational Bayesian Unscented Kalman Filtering Parameter Identification and Operating Condition Determination: This step employs the variational Bayesian unscented Kalman filter algorithm for online motor parameter identification. This algorithm combines the probabilistic modeling capabilities of variational Bayesian inference with the nonlinear system estimation capabilities of unscented Kalman filtering. Traditional recursive least squares methods assume a constant noise covariance matrix, failing to adapt to changes in noise characteristics during traction motor operation under varying operating conditions, leading to decreased parameter identification accuracy. While traditional unscented Kalman filtering can handle nonlinear systems, it also assumes a known noise covariance matrix, making online noise characteristic estimation impossible. The variational Bayesian unscented Kalman filter algorithm can simultaneously estimate motor parameters and the noise covariance matrix online, significantly improving parameter identification accuracy under varying operating conditions.

[0030] Algorithm model construction process; The state-space model of the three-phase asynchronous motor is defined as follows: ; ; in, For the first The state vector at any given time contains five parameters of the motor's equivalent circuit, namely... , This represents the matrix transpose operation. For the first The stator resistance at time t, measured in ohms. For the first The rotor resistance at any given time, measured in ohms. For the first The magnetizing inductance at any given time, measured in Henry. For the first The stator leakage inductance at any given time, measured in Henrys. For the first The rotor leakage inductance at any given time, measured in Henry.

[0031] For the first The state transition matrix at time t is given. In this scheme, it is assumed that the parameter changes are a random walk process. It is a 5th order identity matrix. For the first The process noise vector at time t follows a pattern with a mean of zero and a covariance matrix of . The Gaussian distribution. For the first The observation vector at time t contains the dq-axis current component, i.e. . It is an explicit nonlinear observation function, derived from the steady-state equivalent circuit model of the motor. For the first The dq-axis voltage vector at time t. For the first The motor speed at any given time. For the first The observation noise vector at time t follows a pattern with zero mean and a covariance matrix of . The Gaussian distribution.

[0032] The specific expression for the explicit nonlinear observation function is: ; ; The observation function depends only on the state variables. Input voltage and rotational speed This eliminates the need to observe the current. and The dependency on this solves the algebraic loop problem.

[0033] Algorithm training process; the algorithm training process is divided into two stages. The first stage is the offline initialization stage, which uses the parameter values ​​of the motor when it is a new product manufactured as the mean of the initial state vector, i.e. . This is the reference value for the stator resistance. This is the reference value for rotor resistance. This is the reference value for the magnetizing inductance. This is the reference value for stator leakage inductance. The rotor leakage inductance reference value. Initial state covariance matrix. Set as a diagonal matrix, with diagonal elements representing 1% of the baseline values ​​of each parameter. Initial process noise covariance matrix. and the initial observation noise covariance matrix The data was obtained through statistical analysis of offline experimental data. The offline experiments were conducted on a motor test bench, collecting operating data of the motor under different operating conditions. The initial noise covariance matrix was estimated using the maximum likelihood estimation method.

[0034] The second stage is the online training stage, which employs variational Bayesian inference to estimate the state vector, process noise covariance matrix, and observation noise covariance matrix in real time. Variational Bayesian inference achieves probabilistic estimation of unknown parameters by minimizing the Kullback-Leibler divergence between the true posterior distribution and the approximate posterior distribution. At each sampling time, the algorithm first performs a time update based on the state estimate and noise covariance matrix from the previous time step, then an observation update based on the current observation, and finally enters the variational Bayesian iterative update process.

[0035] Assume that the inverse of the predicted process noise covariance matrix follows a Wishart distribution, with its scaling matrix and degrees of freedom parameters being as follows: and The inverse of the observation noise covariance matrix follows a Wishart distribution, with its scale matrix and degrees of freedom parameters being as follows: and After each UKF filtering step, the following inner loop iteration is performed, with the number of iterations denoted as . (In this plan) ): (1) Calculate the expectation of the noise covariance matrix: , .

[0036] (2) Using the state estimate at the current time State covariance matrix and predicted values , The degrees of freedom of the noise covariance matrix during the update process: ; Degrees of freedom for updating the observation noise covariance matrix: .

[0037] (3) Update the scaling matrix of the process noise: .

[0038] (4) Update the scale matrix of the observation noise: .

[0039] (5) The updated noise covariance matrix is: , .

[0040] initial value , , , This was obtained through statistical analysis of offline experimental data.

[0041] The online parameter identification and drift calculation process for variational Bayesian unscented Kalman filters is as follows: Figure 2 As shown, the algorithm application process is as follows: 1. Initial Value Setting: Initial feedback values ​​are set upon system startup. Initial remaining service life prediction value. Hours, initial comprehensive degradation index Initial feature alignment These initial values ​​are used for the first parameter identification calculation when the system starts up, and will be replaced by the actual values ​​that are fed back later.

[0042] 2. Calculation of operating condition discrimination factor: Define the operating condition discrimination factor. This is used to characterize the current operating condition of the motor: ; in, For the first The motor speed at any given time, For the first The motor speed at any given time, For the first The sampling time interval at time . This refers to the rated speed of the motor. Operating condition discrimination factor. This reflects the rate of change of motor speed. When the rate of change of speed is large, the motor is in a dynamic operating condition of starting or braking; when the rate of change of speed is small, the motor is in a steady-state operating condition. When this occurs, it is determined to be a dynamic operating condition, including starting and braking conditions. When the condition is reached, it is determined to be a steady-state condition. To determine the threshold for operating conditions, this solution includes... This threshold was obtained through statistical analysis of a large amount of actual operating data from traction motors, and it can accurately distinguish between dynamic and steady-state operating conditions.

[0043] 3. Unscented Transform and Sigma-Point Sampling: According to the first... State estimate at time 1 and state covariance matrix Generate 2n+1 sigma points, where n is the dimension of the state vector; in this scheme, n=5. ; ; ; in, For the i-th sigma point, For the scale parameter, in this scheme ; The i-th column represents the square root of the matrix. The unscented transform generates sigma points through deterministic sampling, accurately capturing the mean and covariance of the state distribution and avoiding the linearization error present in the extended Kalman filter.

[0044] 4. Time Update: Perform a state transition for each sigma point to obtain the predicted sigma point: ; Calculate the predicted state mean and predicted state covariance matrix: ; ; in, Weighted by mean, The covariance weights have specific values: ; ; ; In this plan , ; Controlling the distribution range of sigma points Higher-order moments used to match state distributions.

[0045] 5. Observation Update: For each predicted sigma point, the observed predicted value is calculated using an explicit nonlinear observation function: ; in, The above is the explicit nonlinear observation function. Each sigma point corresponds to a different set of motor parameter values, so the observed prediction value is different for each sigma point, which solves the previous problem of the observation matrix depending on the state value.

[0046] Calculate the predicted observation mean, predicted observation covariance matrix, and cross covariance matrix: ; ; ; Calculate the Kalman gain matrix: ; Update the state estimates and state covariance matrix: ; .

[0047] 6. Adaptive Adjustment of Noise Covariance Matrix: Based on the operating condition discrimination results and the remaining service life prediction results fed back by S6, the process noise covariance matrix and the observed noise covariance matrix are adaptively adjusted. ; ; in, This is the process noise adjustment factor. This is the noise adjustment factor for observation. For the feedback The predicted remaining useful life at any given time. For the feedback The feature alignment at any given moment. For the feedback A comprehensive degradation index at any given moment.

[0048] The specific functional form of the process noise adjustment coefficient is determined by the adaptive update mechanism of the entire process parameters; this step directly uses the final adjustment coefficient value. That is, under dynamic operating conditions, the process noise adjustment coefficient increases, accelerating the update rate of parameter identification and ensuring timely tracking of rapid changes in motor parameters under dynamic operating conditions. When When the motor is nearing the end of its lifespan, the process noise adjustment coefficient is further increased to improve the sensitivity of parameter identification and promptly detect abnormal changes in motor parameters. That is, when the feature alignment is low, the observation noise adjustment coefficient Increase, thereby increasing the observation noise covariance matrix. This means reducing the weight of observation data in the filter update and reducing the impact of observation noise on the parameter identification results.

[0049] 7. Parameter drift calculation: The parameter drift is defined as the normalized difference between the current identified parameter and the baseline parameter. ; in, For the first The parameter in the first... Normalized drift at time step. For the first The parameter in the first... The identification value at any given moment. For the first The baseline values ​​for each parameter (factory baseline parameters for new motors). Parameter drift reflects the degree of change in the motor's internal physical parameters relative to its new state, and is a key indicator for quantifying the inherent degradation of the motor. Constructing the parameter drift vector: ; The motor parameter identification values ​​output in this step The input to the parameter drift-driven virtual fault sample generation step is used to simulate current signals under different fault types and different degradation levels. Parameter drift vector. Simultaneously, the inputs are used for enhanced feature construction, virtual fault sample generation, and fractional-order stochastic differential equation degradation modeling, respectively, to construct enhanced feature vectors, generate virtual fault samples, and calculate the comprehensive degradation index. Operating condition discrimination results. The inputs are fed into the multi-scale time-frequency feature extraction module and the cross-modal graph attention transfer module, respectively, for adaptive selection of the feature extraction scale and adjustment of the domain alignment loss weights. (Process noise covariance matrix) and observation noise covariance matrix Parameter identification calculations for the next time step.

[0050] S3. Multi-scale time-frequency feature extraction and enhanced feature construction; based on the working condition discrimination results The feature extraction scale is adaptively selected. The feature distribution of motor current signals varies significantly under different operating conditions. Under dynamic operating conditions, the current signal contains rich transient features, which are suitable for feature extraction using time-frequency analysis methods; under steady-state operating conditions, the current signal is relatively stable, which are suitable for feature extraction using time-domain and frequency-domain analysis methods.

[0051] when In dynamic operating conditions, wavelet packet transform is used to extract time-frequency features with a 3-level decomposition, totaling 32 dimensions. The wavelet packet transform employs the db4 wavelet basis function, and the input signal is a combination of the d-axis current signal and the q-axis current signal. The specific decomposition process is as follows: the combined current signal is decomposed into three layers of wavelet packets to obtain wavelet packet coefficients in eight frequency bands. Then, the energy, mean, variance, and kurtosis of the wavelet packet coefficients in each frequency band are calculated to obtain a total of 32-dimensional time-frequency features.

[0052] when In steady-state operation, both time-domain and frequency-domain features are extracted simultaneously. Time-domain features include eight dimensions: mean, variance, peak value, peak factor, waveform factor, margin factor, impulse factor, and kurtosis factor. The mean reflects the average level of the current signal; the variance reflects the fluctuation of the current signal; the peak value reflects the maximum amplitude of the current signal; the peak factor, the ratio of the peak value to the root mean square (RMS) value, reflects the impulse characteristics of the current signal; the waveform factor, the ratio of the RMS value to the rectified average value, reflects the waveform shape of the current signal; the margin factor, the ratio of the peak value to the rectified average value, reflects the impulse margin of the current signal; the impulse factor, the ratio of the peak value to the rectified average value, reflects the impulse intensity of the current signal; and the kurtosis factor, the ratio of the fourth-order central moment to the square of the variance, reflects the kurtosis characteristics of the current signal. The frequency domain features include the amplitudes of the 1st, 2nd, 3rd, 4th, 5th, 6th, 7th, 8th, 9th, 10th, 11th, 12th, 13th, 14th, 15th, 16th, 17th, 18th, 19th, and 20th harmonics, as well as the centroid frequency, mean square frequency, frequency variance, and frequency skewness, totaling 24 dimensions. Harmonic amplitude reflects the content of each harmonic in the current signal; centroid frequency reflects the energy center of the current signal; mean square frequency reflects the energy distribution of the current signal; frequency variance reflects the frequency dispersion of the current signal; and frequency skewness, the ratio of the third central moment of the frequency distribution to the cube of the frequency standard deviation, reflects the asymmetry of the frequency distribution. A total of 32 dimensions.

[0053] Constructing current feature vectors The dimension is 32. The current feature vector is compared with the parameter drift vector output by S2. Concatenate the data to construct an enhanced feature vector: ; Enhanced feature vectors The dimension is 37. The introduction of parameter drift allows the enhanced feature vector to include not only the apparent features of the current signal, but also information on the changes in the internal parameters of the motor, thereby compensating for the impact of motor aging and changes in operating conditions on the diagnostic model, and significantly improving the accuracy and robustness of fault diagnosis.

[0054] S4. Contrastive Learning-Driven Cross-Modal Graph Attention Transfer and Feature Alignment; This step employs a contrastive learning-driven cross-modal graph attention transfer algorithm for cross-modal knowledge transfer and domain-adaptive feature alignment. This algorithm introduces graph attention networks into the field of cross-modal transfer learning, combining the feature discrimination capabilities of contrastive learning to effectively capture the complex relationships between different modes. This is a pioneering application in the field of traction motor fault diagnosis. Traditional cross-modal transfer methods typically use simple parameter initialization or feature alignment, which cannot effectively capture the nonlinear relationships between different modes, resulting in poor knowledge transfer performance. This algorithm captures the dependencies between features through graph attention networks, enhances the discriminative ability of features through contrastive learning, and adaptively aligns feature distributions under different operating conditions through hybrid domain adaptation, significantly improving the effectiveness of cross-modal knowledge transfer and the accuracy of fault diagnosis under varying operating conditions.

[0055] The algorithm model construction process involves constructing a cross-modal graph attention transfer network, which consists of a source domain graph encoder, a target domain graph encoder, a graph attention layer, a global pooling layer, a fully connected layer, and a domain alignment layer.

[0056] The source domain graph encoder is pre-trained on a vibration signal dataset to learn fault feature representations in vibration signals. The target domain graph encoder is fine-tuned on a current signal dataset to learn fault feature representations in current signals. Both encoders use the same graph convolutional network structure, containing two graph convolutional layers. Graph convolutional layers can effectively extract local features from graph-structured data. The embedding dimension of each node is set to 32, so after processing, the output node feature matrix of the 32-node graph has a dimension of 32×32.

[0057] The graph attention layer computes attention weights between different nodes, capturing dependencies between features. The global pooling layer transforms the node feature matrix into a global feature vector. The fully connected layer maps the global feature vector to a unified feature space. The domain alignment layer minimizes the feature distribution differences between the source and target domains.

[0058] The algorithm training process consists of three stages. The first stage is the source domain pre-training stage, which uses a publicly available traction motor vibration signal fault dataset to train the source domain graph encoder. Each vibration signal sample is represented as a graph, where the nodes of the graph represent features of different frequency bands, and the edges represent the correlations between features. The training objective is to minimize the classification cross-entropy loss. Through source domain pre-training, the source domain graph encoder can learn rich fault feature representations.

[0059] The second stage is the cross-modal transfer stage, where the parameters of the source domain graph encoder are transferred to the target domain graph encoder as its initial parameters. The target domain graph encoder is fine-tuned using a current signal dataset, with the training objective being to minimize the weighted sum of the cross-modal contrast loss and the domain alignment loss. Through cross-modal transfer, the target domain graph encoder can inherit the fault feature knowledge learned by the source domain graph encoder while adapting to the feature distribution of the current signal domain. The cross-modal contrast loss... k is defined as follows: for a batch of target domain samples The corresponding source domain features (through the sample features of the same fault category in the proxy source domain) are used as positive samples. Source domain features of other different fault categories are used as negative samples. The comparative loss is expressed in the form of InfoNCE: ; Where B is the batch size. ) is the cosine similarity function. Temperature coefficient (in this scheme) =0.1), The set of negative samples corresponding to the i-th target domain sample. The total loss function is: ,in .5 represents the weighting coefficient for the comparative loss.

[0060] The third stage is the online fine-tuning stage, which uses only unlabeled real-time current signal data to fine-tune the target domain graph encoder online. The training objective is to minimize the domain alignment loss. Online fine-tuning enables the model to adapt to differences in different motors and operating environments, improving the model's generalization ability. The online stage does not calculate contrastive or classification losses, thus addressing the problem of not being able to obtain source domain samples and labels in the online stage.

[0061] The contrastive learning process for cross-modal graph attention transfer and domain adaptive alignment is as follows: Figure 3 As shown, the specific application process of the algorithm is as follows: 1. Proxy Source Domain Construction: After offline training, 1000 samples are randomly selected from the source domain vibration signal training dataset to form a fixed proxy source domain sample set. For each proxy source domain sample, features are extracted using a pre-trained source domain graph encoder to obtain a set of source domain feature vectors. The set of feature vectors is stored in system memory, and these fixed source domain features are used directly for domain alignment calculation during the online phase, without the need for real-time acquisition of vibration signals.

[0062] 2. Graph Construction: A sliding window method is used to construct the graph, with the window size set to 100 sampling points. For the first... At that moment, take the first To the Enhanced feature vector sequence at time step Extract the first 32 dimensions (i.e., the current feature vector portion) from each enhanced feature vector to construct a graph. .in, It is a set of nodes, containing 32 nodes, each node corresponding to a current feature vector. One dimension, where each node's feature is the time series of that dimension across 100 sampling points. Parameter drift vector. It is not used as a node in the graph, but rather as conditional information during graph attention calculation. For specific usage, please refer to the subsequent 'Graph Attention Calculation' section. Let be a set of edges, where the weight of each edge is the Pearson correlation coefficient between the feature time series of two nodes. The Pearson correlation coefficient reflects the degree of linear correlation between two feature time series; the larger the edge weight, the stronger the correlation between the two features. Representing the enhanced feature vector as a graph structure can effectively capture the dependencies between features and improve the ability to represent features.

[0063] 3. Target domain graph encoding: [This refers to encoding the graph in a specific domain.] The input is fed into the target domain graph encoder to obtain the node feature matrix: ; in, The initial node feature matrix has a dimension of 32×32. For target domain graph encoder. These are the parameters of the target domain graph encoder. The target domain graph encoder extracts local features of the graph through graph convolution operations and outputs the feature representation of each node.

[0064] 4. Graph Attention Calculation: The initial node feature matrix is ​​input into the graph attention layer to calculate the attention weights. The graph attention layer uses a fully connected graph structure, meaning that each node is connected to all other nodes, including the neighbor set. Including nodes All 31 nodes except: ; ; in, For nodes and nodes Attention coefficient between them. The attention mechanism function is implemented using a single-layer feedforward neural network. It is a learnable weight matrix with a dimension of 32×32. For nodes The initial feature representation has a dimension of 32. For nodes The initial feature representation has a dimension of 32. This represents the normalized attention weights. For nodes The set of neighboring nodes. Modifying the linear unit activation function to address leakage can solve the problem of gradient vanishing in the negative region for the ReLU function.

[0065] Based on the parameter drift vector output by S2 Conditional modulation is applied to graph attention. The parameter drift vector is passed through a linear mapping layer to obtain the conditional vector. Conditional weights Conditional bias Add the conditional vector to the feature representation of each node: .

[0066] Recalculate the attention weights based on the updated node features: ; .

[0067] Then calculate the updated node features: .

[0068] in, For nodes The updated feature representation has a dimension of 32. It is a sigmoid activation function that maps feature values ​​to the range of 0 to 1.

[0069] Construct the updated node feature matrix: ; in, This is the updated node feature matrix, with dimensions of 32×32.

[0070] 5. Global Pooling and Feature Mapping: Perform global average pooling on the updated node feature matrix to obtain the global feature vector. ; in, This is the feature vector after global average pooling, with a dimension of 32.

[0071] The global feature vector is input into the fully connected layer and mapped to a unified feature space: ; in, This is the final target domain feature vector, with a dimension of 128. This is the weight matrix of the fully connected layer, with dimensions of 128×32. This is the bias vector of the fully connected layer, with a dimension of 128.

[0072] 6. Hybrid Domain Adaptive Feature Alignment: Calculate the multi-kernel maximum mean difference loss to minimize the feature distribution difference between the source and target domains. In the online phase, the source domain features use a pre-stored proxy source domain feature set. The target domain features use the target domain features of the current batch: ; in, For the first Multi-core maximum mean difference loss at time step. This represents the number of samples from the proxy source domain. This represents the number of target domain samples in the current batch. This is a kernel function mapping. The method uses a linear combination of five Gaussian kernel functions as the kernel function, with bandwidths of 0.1, 0.5, 1.0, 5.0, and 10.0, respectively. The multi-kernel maximum mean difference effectively captures the differences between different distributions, improving the robustness of distribution alignment.

[0073] Based on the results of the working condition judgment Adaptively adjust the weights of the domain alignment loss: ; in, For the first Domain adaptive total loss at time step. The MMD loss weight coefficients, whose specific functional form is determined by the end-to-end parameter adaptive update mechanism. That is, under dynamic operating conditions, increasing the MMD loss weighting coefficient improves robustness to changes in operating conditions. When That is, under steady-state conditions, reduce the MMD loss weighting coefficient to reduce computational overhead.

[0074] 7. Feature Alignment Calculation: Calculate the feature alignment. This is used to feed back to S2 to adjust the observation noise covariance matrix. ; in, The maximum value of MMD loss was determined through offline experiments in this scheme. Feature alignment reflects the degree of alignment between the feature distributions of the source and target domains. The higher the feature alignment, the better the effect of cross-modal knowledge transfer.

[0075] The aligned target domain feature vector output in this step Input to the prototype calibration few-sample classification module. Feature alignment. Feedback is sent to S2 to adjust the observation noise covariance matrix. Domain adaptive total loss Parameters for online updating of the target domain graph encoder .

[0076] S5. Parameter drift-driven virtual sample generation and small sample fault classification; the parameter drift-driven virtual sample generation and prototype calibration small sample classification process is as follows: Figure 4 As shown. Based on motor parameter identification values. and parameter drift vector This study simulates current signals under different fault types and degradation levels. A T-type equivalent circuit model of a three-phase asynchronous motor is used for current signal simulation. The voltage equation of this model is: ; ; in, For differential operators, , .

[0077] Motor parameter identification values ​​output in real time by S2 These parameters serve as the baseline parameters for generating virtual samples. After adjusting the parameters according to the parameter variation patterns corresponding to different fault types, the adjusted parameters are substituted into the voltage equations described above. Model input voltage. Data from the same time series as the voltage signal actually acquired in S1 (or voltage curves under typical operating conditions) are used. The fourth-order Runge-Kutta method is used to numerically solve the above state equations, with a time step of [missing information]. To obtain the stator current , and rotor current , The numerical solution. Finally, the solved... , Perform the inverse Park transform and combine it with the actual electrical angle. The virtual three-phase current signal is obtained. , , Its sampling frequency remains consistent with that in S1 (20 kHz). The inverse Park transform formula is: ; The generated virtual three-phase current signal is subjected to the same feature extraction and enhancement processing as S3 to generate an enhanced feature vector of the virtual fault sample.

[0078] The parameter adjustment rules for different fault types are as follows: Normal state: All parameters remain at the S2 identification value. No further adjustments will be made in the vicinity. Rotor bar breakage fault: Rotor resistance identified by S2 Based on the baseline, increase the ratio by ,in The value represents the severity of the fault and ranges from 0 to 10. Bearing rolling element failure: Stator leakage inductance identified by S2 and rotor leakage inductance Based on the baseline, increase simultaneously, with the increase ratio being [missing information].

[0079] Bearing inner ring fault: stator leakage inductance identified by S2 Based on the baseline, increase the ratio by The rotor leakage inductance identified by S2 Based on the baseline, increase the ratio by

[0080] Bearing outer ring fault: stator leakage inductance identified by S2 Based on the baseline, increase the ratio by The rotor leakage inductance identified by S2 Based on the baseline, increase the ratio by

[0081] Stator winding inter-turn short circuit fault: stator resistance identified by S2 Based on the baseline, the reduction ratio is

[0082] Air gap eccentricity fault: excitation inductance identified by S2 Based on the baseline, the reduction ratio is

[0083] Severity of the fault The range of values ​​is determined by the parameter drift vector output by S2. Dynamically determined. Calculate the L2 norm of the parameter drift vector. This maps it to the range of values ​​for fault severity, i.e. When generating virtual samples, the severity of the fault... From 0 The values ​​are taken evenly between the ranges to ensure that the generated virtual samples can cover all fault severity that may occur under the current actual degradation state of the motor.

[0084] Using the above method, the motor parameter identification value output by S2 The parameter drift vector serves as the parameter benchmark for virtual samples. Both the range of values ​​used to determine the severity of the fault and the virtual sample itself are truly involved in the generation process. The generated virtual samples can reflect the actual degradation state and individual differences of the current motor, rather than a general fault mode, significantly improving the realism and effectiveness of the virtual samples.

[0085] The simulated current signal undergoes the same feature extraction and enhancement processing as in S3 to generate enhanced feature vectors for virtual fault samples. These virtual fault samples are then mixed with real fault samples to construct a support set for small-sample classification. The generation of virtual fault samples effectively addresses the problem of scarce traction motor fault samples, improving the accuracy of small-sample classification.

[0086] A prototype calibration network was used as a few-shot learning classifier, and a meta-learning mechanism was employed to quickly adjust the classification boundary using a small number of real fault samples and a large number of virtual fault samples. Meta-training was configured as a 7-way 5-shot process, with each episode randomly sampling 5 support set samples and 15 query set samples from each of the 7 classes. The meta-learning mechanism enabled the model to quickly adapt to new tasks and achieve good classification results with limited samples. After meta-training, the model parameters were transferred to the online inference stage. The support set for the online inference stage consisted of offline-generated virtual samples and a small number of online-collected real samples, and the support set was updated every 24 hours.

[0087] The training process of the prototype calibration mapping layer; the parameters of the prototype calibration mapping layer. and The network is jointly trained with the prototype network during the meta-training phase. The training objective is to minimize the weighted sum of the classification loss and the prototype calibration loss. The dataset used in the meta-training phase is generated through a simulation model, where each sample contains not only a fault category label but also a simulated parameter drift vector. Simulated Random sampling is performed within its typical value range [-0.3, 0.3] to cover different degrees of parameter drift. During training for each episode, both the support set and query set samples are accompanied by a set of parameters matching their degradation state. This enables the prototype calibration mapping layer to learn from The mapping relationship to the prototype offset. The specific training process is as follows: 1. In each training episode, the initial prototype vector is first calculated based on the support set samples. .

[0088] 2. For each support set sample, calculate its corresponding parameter shift vector. Calculate the calibration vector .

[0089] 3. Calibrate the initial prototype vector to obtain the calibrated prototype vector. . This is the initial calibration coefficient (set to 0.1 in this scheme).

[0090] 4. Calculate the Mahalanobis distance between the query sample and the calibrated prototype vector to obtain the classification probability distribution.

[0091] 5. Calculate the classification cross-entropy loss. and prototype calibration loss The prototype calibration loss is defined as the product of the difference between the prototype vectors before and after calibration and the severity of the fault. ; in, To query the number of samples in the set, For the first The severity of the fault in each sample.

[0092] 6. The total loss function is: ; in, To calibrate the loss weighting coefficient, which is used to balance the classification loss and the prototype calibration loss, this scheme sets it to 0.1 based on cross-validation.

[0093] 7. Simultaneously optimize prototype network parameters and prototype calibration mapping layer parameters using stochastic gradient descent algorithm. and .

[0094] Through joint training, the prototype calibration mapping layer can learn the mapping relationship between parameter drift and prototype vector offset, thereby achieving effective prototype calibration.

[0095] Algorithm application process; the prototype vector for each fault category is the mean of the feature vectors of the support set samples for that category: ; in, For the first Time-based fault categories The prototype vector has a dimension of 128. The number of support set samples for each category in this method . For the first Time Category The The support set contains feature vectors of 128 dimensions. The prototype vector represents the feature center of the fault category. The closer the query sample is to the prototype vector, the greater the probability that the query sample belongs to the fault category.

[0096] Based on the parameter drift vector output by S2 The prototype vector is calibrated to compensate for the impact of motor aging. First, a linear mapping layer is used to map the 5-dimensional parameter drift vector into a 128-dimensional calibration vector. ; in, This is the calibration vector, with a dimension of 128. It is a learnable weight matrix with dimensions of 128×5. The bias vector is a learnable vector with a dimension of 128. Then the prototype vector is calibrated: ;in, This is the calibrated prototype vector. The calibration coefficient is the overall degradation index fed back by S6. Sure: ; This is the initial calibration coefficient, and its specific value is determined by the full-process parameter adaptive update mechanism. For the feedback A comprehensive degradation index at any given time. Motor aging causes all parameters to drift, resulting in a shift in the prototype vector. Calibrating the prototype vector using the parameter drift vector can effectively compensate for the effects of motor aging and improve the accuracy of fault diagnosis.

[0097] The probability distribution of fault categories for the query samples is obtained by calculating the softmax of the Mahalanobis distance between their feature vectors and each calibrated prototype vector: ; in, To query the category to which the sample belongs The probability of. To query the feature vector of the sample. The Mahalanobis distance function is calculated using the following formula: ; in, The covariance matrix of the feature distribution is estimated from the feature vectors of the offline training set and remains unchanged during the online inference stage. Mahalanobis distance takes into account the correlation between features and can more accurately measure the distance between two feature vectors, achieving better classification results compared to Euclidean distance.

[0098] The category with the highest probability is selected as the final fault diagnosis result: arg ; The corresponding failure probability is: .

[0099] S6. Remaining Service Prediction Driven by Fractional Stochastic Differential Equations (FSDEs); This step employs a Bayesian update algorithm driven by fractional stochastic differential equations (FSDEs) for remaining service life prediction. This algorithm leverages the ability of fractional calculus to accurately describe the memory- and hereditary nature of degradation processes. Combined with a Bayesian online update mechanism, it significantly improves the accuracy of remaining service life prediction. This method has not yet been applied in the field of traction motor remaining service life prediction. Traditional remaining service life prediction methods typically use integer-order stochastic differential equations for modeling, assuming the degradation process is memoryless, which fails to accurately describe the memory and hereditary nature of traction motor degradation. Fractional stochastic differential equations introduce fractional derivatives, effectively capturing the historical dependencies of the degradation process and improving prediction accuracy.

[0100] Algorithm model construction process: Define the comprehensive degradation index as the weighted norm of the parameter drift vector, with the weights determined by the parameter sensitivity corresponding to the fault category: ; in, For the first A comprehensive degradation index at any given moment. For the first Each parameter in the fault category The degradation weights are assigned accordingly. Different fault types have different sensitivities to different parameters. For example, rotor bar breakage faults are most sensitive to rotor resistance, while stator winding inter-turn short circuit faults are most sensitive to stator resistance. By assigning different weights to different parameters, the degree of motor degradation can be reflected more accurately.

[0101] The correspondence between the fault category index c and the fault type is defined as follows: c=1 corresponds to normal state, c=2 corresponds to rotor bar breakage fault, c=3 corresponds to bearing rolling element fault, c=4 corresponds to bearing inner ring fault, c=5 corresponds to bearing outer ring fault, c=6 corresponds to stator winding inter-turn short circuit fault, and c=7 corresponds to air gap eccentricity fault. Then, the degradation weight of the i-th parameter under fault category c is... The degradation weights for each fault type are as follows: The degradation weight corresponding to the normal state is , , , , .

[0102] The degradation weight corresponding to rotor bar breakage fault is , , , , .

[0103] The degradation weight corresponding to bearing rolling element failure is , , , , .

[0104] The degradation weight corresponding to the bearing inner ring failure is , , , , .

[0105] The degradation weight corresponding to bearing outer ring failure is , , , , .

[0106] The degradation weight corresponding to the stator winding inter-turn short circuit fault is , , , , .

[0107] The degradation weight corresponding to the air gap eccentricity fault is , , , , .

[0108] The evolution of the comprehensive degradation index is modeled using fractional stochastic differential equations: ;in, for Caputo fractional derivative operator. The fractional order is initially set to 0.8 in this scheme. for A comprehensive degradation index at any given moment. Fault Category The corresponding degradation rate. The standard deviation of the degradation noise. It is standard Gaussian white noise. Fractional order. It reflects the memory of the degradation process. The smaller the size, the stronger the memory.

[0109] Algorithm training process; the algorithm training process is divided into two stages. The first stage is the offline parameter estimation stage, which uses historical degradation data to estimate the fractional order. Initial degradation rate and the standard deviation of initial degradation noise The fractional order is estimated using the least squares method. The optimal fractional order is obtained by minimizing the error between the model's predicted and actual values. The initial degradation rate and the standard deviation of the initial degradation noise are estimated using the maximum likelihood estimation method.

[0110] The second stage is the online training stage, which uses a Bayesian online update mechanism to estimate the degradation rate in real time. and degradation noise standard deviation The prior distribution adopts a conjugate normal-inverse Gamma distribution, and the likelihood function is given by the transition density of the fractional-order stochastic differential equation. The Bayesian online update mechanism can continuously update the model parameters using new observation data, thereby improving prediction accuracy.

[0111] The specific application process of the algorithm is as follows: 1. Discretization of fractional-order stochastic differential equations: Discretizing continuous-time fractional-order stochastic differential equations into discrete-time form. The time step for discretization... The sampling interval is consistent with that of S1, that is To improve the computational stability of long-term forecasts, a prediction-correction method is used for discretization. ; in, The fractional discretization coefficients are calculated using the following formula: ; The coefficients are binomial coefficients. , This is the Gamma function. Standard Gaussian white noise. Fractional discretization coefficients. It reflects the impact of degradation indicators at historical moments on degradation indicators at the current moment, demonstrating the memory nature of the degradation process.

[0112] 2. Bayesian Online Update: A Bayesian online update mechanism is used to estimate the degradation rate in real time. and degradation noise standard deviation Assuming the degradation rate Follows a normal distribution Degradation noise accuracy Follows Gamma distribution .

[0113] Let the observation increment be: ; The expected value of the observed increment is then: The variance is .

[0114] The update formula for the prior distribution parameters is: ; ; ; ; ; ; in, The equivalent sample size, initial value Using the comprehensive degradation index observed in actual data. Calculate observation increment and use Update the mean of the degradation rate.

[0115] 3. Remaining useful life point estimation: Let the failure threshold be... In this method The failure threshold is determined by the technical specifications provided by the motor manufacturer. When the overall degradation index reaches the failure threshold, the motor has reached the end of its service life and requires repair or replacement. The point estimate of the remaining service life is the sum of the overall degradation index values ​​from the current value. Evolved to the fault threshold Time required: ; in, For the first The estimated remaining useful life at a given time.

[0116] 4. Estimation of the remaining useful life probability distribution: The conditional probability density function of the remaining useful life is: ; in, Given the current degradation index The remaining useful life conditional probability density function. The remaining useful life probability distribution can provide richer predictive information, helping maintenance personnel to formulate reasonable maintenance plans.

[0117] The comprehensive degradation index output in this step Feedback is sent to S5 to adjust the prototype calibration coefficients. Predicted remaining useful life Feedback is sent to S2 to adjust the process noise covariance matrix. .

[0118] S7. Full-process parameter adaptive update driven by diagnostic prediction results; based on S5 fault categories. and failure probability and comprehensive degradation indicators and remaining useful life prediction The parameters are adaptively updated throughout the entire process. This adaptive updating of parameters enables the model to continuously adapt to changes in the motor's operating state, thereby continuously optimizing diagnostic and predictive performance.

[0119] Update the process noise adjustment coefficients and observation noise adjustment coefficients of the parameter identification module. The final formula for the process noise adjustment coefficient is: ; The final formula for the observation noise adjustment factor is: ; When the comprehensive degradation index At this time, the motor is in a moderately degraded state, and the rate of parameter change accelerates. Therefore, the process noise adjustment coefficient is increased to speed up the update rate of parameter identification. When the remaining service life is predicted... When the motor is nearing the end of its lifespan, its parameters may change abruptly. Therefore, it is necessary to further increase the process noise adjustment coefficient to improve the sensitivity of parameter identification. When the feature alignment... At this time, the cross-modal knowledge transfer effect is not good, and the reliability of the observation data is reduced. Therefore, the observation noise adjustment coefficient is increased and the weight of the observation data is reduced.

[0120] Update the domain alignment loss weights for the cross-modal graph attention transfer module. The final formula for the domain alignment loss weights is: ; in, This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. This is a set of rare fault types, defined as fault types that occur less than 5% of the historical fault data. When the fault probability... When the confidence level of the fault diagnosis result is low, it indicates that the discriminative ability of the target domain features is insufficient. Therefore, the domain alignment loss weight coefficient is increased to optimize the feature representation ability of the target domain encoder. For rare fault types, due to the small number of samples, feature distribution alignment is more difficult, so the domain alignment loss weight coefficient is increased by an additional 0.2.

[0121] Update the prototype calibration coefficients of the few-shot classification module. The final formula for the prototype calibration coefficients is: ; When the remaining useful life is predicted At the beginning, the motor was in a severely degraded state, with large parameter drift and significant prototype vector offset. Therefore, the initial calibration coefficient was increased. This increases the accuracy of prototype calibration. If the actual fault category after repair is available (e.g., through a repair report), the actual fault category is confirmed based on the repair report after the motor is actually repaired. If it differs from the diagnostic results, the prototype vector for that fault category is recalculated as the mean of the feature vectors of all historical samples for that category. If there is no actual feedback, this step is skipped.

[0122] Update the prior distribution parameters of the fractional-order stochastic differential equation degradation modeling module. Update the prior distribution parameters of degradation rate and degradation noise accuracy based on the prediction error of the comprehensive degradation index; the larger the prediction error, the larger the variance of the prior distribution, to improve the robustness of the model. Update the fractional-order number based on the prediction error of the remaining useful life. By minimizing the remaining useful life prediction error, the optimal fractional order is obtained, and the update formula is: ; in, This represents the actual remaining service life and will be updated when actual maintenance data becomes available. For a sign function, when When the sign function value is 1, the fractional order is decreased; when When the sign function value is -1, the fractional order is increased. This approximation method avoids the problem of calculating the derivative of the absolute value function and can adjust the fractional order according to the prediction error.

[0123] The updated process noise adjustment coefficient is as described above. and observation noise adjustment factor Parameter identification calculation for the next sampling time S2; update the domain alignment loss weight coefficients. and prototype calibration coefficients These are used respectively for subsequent calculations of S4 and S5; the updated fractional order will be used. Used for predicting the remaining useful life in the subsequent S6 step. Through continuous iterative updates of the above parameters, a closed-loop adaptive optimization process is formed.

[0124] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0125] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0126] 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 the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0127] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0128] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0129] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

Claims

1. A variable operating condition traction motor diagnosis and prediction method based on parameter identification and transfer learning driving, characterized in that, Includes the following steps: S1. Synchronously acquire three-phase stator current signal, three-phase stator voltage signal and speed signal, perform adaptive preprocessing to obtain dq-axis current component, dq-axis voltage component and preprocessed speed signal; S2. The variational Bayesian unscented Kalman filter algorithm is used to identify the motor parameters online for the dq axis current component, dq axis voltage component and preprocessed speed signal, so as to obtain the motor parameter identification value and parameter drift vector, and at the same time calculate the working condition discrimination factor. S3. Based on the working condition discrimination factor, adaptively extract multi-scale time-frequency features, and fuse the multi-scale time-frequency features with the parameter drift vector to construct an enhanced feature vector; S4. A contrastive learning-driven cross-modal graph attention transfer algorithm is adopted. The graph structure is constructed based on the current feature vector part in the enhanced feature vector. The graph features are extracted using the parameter drift vector as conditional information. Domain adaptive feature alignment is performed to obtain the aligned target domain feature vector. S5. Based on the motor parameter identification value and parameter drift vector, generate virtual fault samples, mix the virtual fault samples with real fault samples to construct a support set, and use the prototype calibration small sample classification algorithm combined with the target domain feature vector to perform fault diagnosis, and obtain the fault category and fault probability.

2. The variable operating condition traction motor diagnostic prediction method based on parameter identification and transfer learning driving according to claim 1, characterized in that, Also includes: S6. Calculate the comprehensive degradation index based on the fault category and parameter drift vector, and use the Bayesian update algorithm driven by fractional stochastic differential equations to predict the remaining useful life and obtain the predicted value of the remaining useful life. S7. Calculate the process noise adjustment coefficient, observation noise adjustment coefficient, domain alignment loss weight coefficient, and prototype calibration coefficient based on the remaining useful life prediction value, and feed each coefficient back to S2, S4, and S5 respectively to form a closed-loop adaptive optimization.

3. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, The adaptive preprocessing in S1 specifically includes: using a third-order Butterworth low-pass filter to remove high-frequency noise; performing Park transform on the three-phase stator current signal and the three-phase stator voltage signal to obtain the dq-axis current component and the dq-axis voltage component, respectively; and performing moving average filtering on the speed signal to remove speed fluctuation noise. The electrical angle required for the Park transform is obtained by integrating the speed signal.

4. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, When performing online identification of motor parameters using the variational Bayesian unscented Kalman filter algorithm in S2, sigma points are first generated through unscented transformation. Each sigma point corresponds to a set of motor parameter states. Then, an explicit nonlinear observation function is used to calculate the observation prediction value of each sigma point. The explicit nonlinear observation function depends only on the motor parameter states corresponding to the sigma point, the dq axis voltage components, and the preprocessed speed signal.

5. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, In S2, the parameter drift vector is a vector composed of the normalized difference between the motor parameter identification value and the reference parameter. The operating condition discrimination factor is used to distinguish between dynamic operating conditions and steady-state operating conditions. When the operating condition discrimination factor is greater than the operating condition discrimination threshold, it is determined to be a dynamic operating condition. When the operating condition discrimination factor is less than or equal to the operating condition discrimination threshold, it is determined to be a steady-state operating condition.

6. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 5, characterized in that, The adaptive extraction of multi-scale time-frequency features in S3 specifically includes: when the condition is determined to be dynamic, wavelet packet transform is used to extract time-frequency features; when the condition is determined to be steady-state, both time-domain and frequency-domain features are extracted simultaneously.

7. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, After offline training, the contrastive learning-driven cross-modal graph attention transfer algorithm in S4 extracts samples from the source domain vibration signal training dataset to form a fixed proxy source domain sample set, extracts and stores the source domain feature vectors, and directly uses the stored source domain feature vectors to perform domain adaptive feature alignment calculation in the online stage.

8. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, Before performing domain-adaptive feature alignment, the contrastive learning-driven cross-modal graph attention transfer algorithm in S4 constructs a graph using a sliding window approach. The nodes of the graph correspond to the various dimensions of the current feature vector, and the parameter drift vector serves as the conditional information for graph attention calculation. The feature of each node is the time series of that dimension within the sliding window, and the edge weights of the graph are the Pearson correlation coefficients between the feature time series of two nodes.

9. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, The generation of virtual fault samples in S5 specifically includes: using the motor parameter identification value as the reference parameter for generating virtual samples, adjusting the corresponding parameter values ​​according to the parameter change rules corresponding to different fault types, determining the range of fault severity values ​​by the parameter drift vector, and generating current signals under different fault types and different degradation degrees.

10. The diagnostic prediction method for traction motors under varying operating conditions based on parameter identification and transfer learning as described in claim 1, characterized in that, When using the prototype calibration few-sample classification algorithm for fault diagnosis in S5, the parameter drift vector is mapped to the calibration vector through the linear mapping layer. The calibration vector is then used to calibrate the prototype vector. The parameters of the linear mapping layer are jointly trained with the prototype network during the meta-training stage.

Images