Metalearning diagnosis method for sensing working condition of drive motor of fly-by-wire actuator

Abstract

The embodiment of the invention discloses a telex actuator driving motor working condition sensing meta learning diagnosis method, which comprises the following steps of: acquiring a feature embedding function value through a working condition encoder module, acquiring a class prototype of a fault class based on a prototype network, and training the prototype network based on prototype loss, establishing center loss and establishing a joint loss function to obtain a fault class model; and based on the trained prototype network, fault diagnosis in the actual working process is carried out. According to the method, high-robustness diagnosis on composite working condition changes under the condition of few samples is realized, the explicit modeling capability on operation conditions is remarkably improved, and the diagnosis precision is greatly improved. According to the method, the discrimination of the fault features in the embedding space is remarkably enhanced. The method has strong small sample rapid adaptive capability, so that when a new fault diagnosis task is encountered, rapid adaptation can be realized without retraining or only a small number of new samples, and the practicability and deployment flexibility of the method are greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation and high-reliability electric drive technology, and in particular to a meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor. Background Technology

[0002] In high-precision electro-hydraulic actuation systems such as aerospace and robotics, the reliability of brushless DC motors, as the core power source, directly determines the safety and performance of the entire actuation system. These critical motors in electro-hydraulic actuators often operate under extremely variable speeds, loads, and complex waveform commands, presenting more severe challenges to operational status monitoring and early fault diagnosis than in ordinary industrial scenarios. Due to their high power density, high efficiency, and excellent controllability, brushless DC motors have become core power components in modern electro-hydraulic actuation systems, industrial robots, and precision servo drives, making their operational reliability paramount. To ensure the high reliability of these systems, real-time and accurate fault diagnosis of the motors is crucial.

[0003] Currently, the technical solutions in this field can be mainly divided into three categories: model-based methods, signal processing-based methods, and data-driven methods. Model-based methods rely on accurate mathematical models of the motor and state observers, using residual analysis to detect faults. However, motors in electric actuators typically operate under complex and variable conditions, making it difficult to establish accurate physical models, and they are highly sensitive to model mismatch. Signal processing-based methods, such as Fast Fourier Transform and Wavelet Transform, diagnose faults by analyzing the time-frequency characteristics of signals such as vibration and current. However, these methods often require expert knowledge for manual feature extraction and parameter adjustment, and their performance is limited when handling non-stationary signals and strong noise interference.

[0004] In recent years, data-driven methods, represented by deep learning, have shown great potential in mechanical fault diagnosis, as they can automatically learn complex fault characteristics from data. However, the success of these methods heavily relies on large amounts of complete labeled data. In real-world industrial scenarios, especially for critical motors in electric actuators, obtaining a large number of samples of all potential fault modes under different operating conditions is costly and even impractical, leading to the challenge of small-sample learning for the models. Furthermore, when a trained model is deployed under operating conditions different from the training data (such as varying speeds and loads), model performance degrades significantly due to domain bias. Although strategies such as transfer learning have been used to alleviate this problem, aiming to learn domain-invariant features, their adversarial training mechanisms are prone to instability and have limited generalization ability in extreme cases where complex operating conditions with combined waveform and load variations are encountered, and the target domain samples are extremely limited, making it difficult to learn truly robust condition-invariant features.

[0005] In summary, due to the complex and variable operating conditions of motors, the distribution of monitoring data changes accordingly, leading to domain bias issues in traditional data-driven diagnostic models. Furthermore, acquiring large amounts of labeled data for all fault types in real-world industrial scenarios is costly, resulting in challenges related to few-shot learning. Existing few-shot learning and domain-adaptive methods lack explicit modeling capabilities for complex operating conditions involving combined waveform and load variations, significantly reducing diagnostic accuracy and robustness. Therefore, current technologies struggle to meet the high-reliability system requirements for diagnostic accuracy and robustness when facing the common and severe engineering challenges of key motors in electric actuators operating under limited samples and complex variable conditions. Summary of the Invention

[0006] This invention discloses a meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor, in order to overcome the above-mentioned technical problems.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A meta-learning diagnostic method for sensing the operating conditions of an electrically driven motor includes the following steps: S1: Acquire the three-phase current and triaxial vibration signals of the drive motor of the electric actuator under different operating conditions, and use the sliding window technique to acquire M initial sample data, and randomly select N initial sample data under different operating conditions to form sample data; where M is the total number of initial sample data; N is the total number of sample data, M>N; S2: Based on the sample data, a working condition encoder is used to obtain the feature embedding function value, i.e., the sample data features fused with the working condition; to obtain the first... Feature embedding function values ​​of sample data for each fault category; S3: According to the... Based on the prototype network, the feature embedding function values ​​of the sample data for the fault category are obtained. The class prototype for the fault category, i.e., the first fault category. The average value of the support set sample feature embedding function for each fault category; to obtain the first The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between the class prototypes of each fault category; and then to obtain the first... The query set of sample data for the fault category is predicted as follows: The probability of each fault category; S4: According to the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category is used to obtain the prototype loss; a joint loss function is then established to train the prototype network. S5: Acquire the three-phase current and three-axis vibration signals of the drive motor of the electric actuator during actual working time, so as to obtain the predicted fault category based on the trained prototype network and realize the diagnosis of the drive motor of the electric actuator.

[0008] Furthermore, the working condition encoder includes a feature encoding module and a layered fusion module; The feature encoding module is used to obtain the statistical features of a single channel and the cross-channel correlation features based on the sample data, so as to obtain a statistical feature vector, and then obtain a conditional encoding vector based on a multilayer perceptron. The hierarchical fusion module includes a learnable mapping network and a convolutional fusion module, used to obtain sample data features fused with the working conditions based on the conditional encoding vector and the sample data. The learnable mapping network is used to obtain the attention weight vector based on the conditional encoding vector; The convolutional fusion module is used to obtain the feature embedding function value based on the attention weight vector, that is, the sample data features fused with the working conditions.

[0009] Furthermore, the calculation formula for the convolutional fusion module is expressed as follows:

[0010]

[0011] In the formula: Indicates the first The output of a convolutional neural network layer. ; Indicates the convolution operation; For inputting sample data from the encoder under operating conditions; This represents the output vector after attention modulation. Indicates the relationship with the first Attention weight vectors corresponding to layers in a convolutional neural network; This indicates element-wise multiplication; This is the index of the convolutional neural network layer.

[0012] Furthermore, S3 includes: S31: Obtain the... The class prototypes for each fault category are represented by the following formula:

[0013] In the formula: For the first The class prototype for the fault category, i.e., the first fault category. The average value of the support set sample feature embedding function for each fault category; For the first The total number of support set samples for each fault category; For the first A set of support sets for sample data of each fault category; For the first Support set sample feature embedding function values ​​for sample data of each fault category; An index for fault categories; S32: Obtain the... The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between the class prototypes of each fault category; S33: Obtain the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category is calculated using the following formula:

[0014] In the formula: To make the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category; To make the first The predicted category of a query set of sample data for each fault category; Index of fault categories contained in a single meta-task 'The total number of fault categories contained in a single meta-task; For the first in a single meta-task The Euclidean distance between the feature embedding function value of each fault category and the class prototype; For the first The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between class prototypes of each fault category.

[0015] Furthermore, the original loss is obtained as follows:

[0016] In the formula: For prototype loss; A set of samples for the query set; The total number of samples in the query set; This is the index for the query set samples.

[0017] Furthermore, the joint loss function is established as follows:

[0018] In the formula: The coefficient of the center loss; Total loss; in,

[0019] In the formula: Indicates central loss; Index of fault categories contained in a single meta-task 'The total number of fault categories contained in a single meta-task; Represents the feature mapping function; Indicates the first in a single meta-task Samples of each fault category; Indicates the first in a single meta-task The class center for each fault category; This represents the L2 norm.

[0020] Beneficial Effects: This invention provides a meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor. Through an operating condition encoder module, it obtains feature embedding function values ​​and, based on a prototype network, acquires class prototypes of fault categories. This allows for the prediction of probability for fault sample data. Based on prototype loss, a central loss and a joint loss function are established to train the prototype network. Based on the trained prototype network, fault diagnosis is performed during actual operation. This achieves highly robust diagnosis of complex operating condition changes under limited sample conditions, significantly improving the explicit modeling capability of operating conditions and greatly enhancing diagnostic accuracy. This invention effectively addresses the simultaneous occurrence of speed waveform and load torque changes in motors during real-world operation. Even in extreme cases where there are only a few samples for each fault category, it maintains extremely high diagnostic accuracy, solving the problem of performance degradation due to domain offset in traditional data-driven models. It significantly enhances the discriminative power of fault features in the embedding space. The fault features learned by this invention have extremely high intra-class compactness and inter-class separation. Samples from different fault categories form clear and clustered clusters in the feature space, making distance-based limited-sample classification more accurate and reliable. It possesses a powerful ability to quickly adapt to small samples, enabling it to adapt rapidly to new fault diagnosis tasks (including new fault types or new combinations of operating conditions) without retraining or with only a very small number of new samples, greatly improving the practicality and deployment flexibility of the method. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the meta-learning diagnostic method for sensing the operating conditions of the electric actuator drive motor according to the present invention.

[0023] Figure 2 This is a schematic diagram of the working condition encoder architecture in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall architecture of OAD-PN in an embodiment of the present invention; Figure 4 This is a visualization result of T-sne in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] This embodiment introduces a meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor, including the following steps: Figure 1 As shown: S1: Acquire the three-phase current and three-axis vibration signals of the drive motor of the electric actuator under different working conditions, and use the sliding window technique to acquire M initial sample data, and randomly select N initial sample data under different working conditions to form sample data; and divide the sample data into a support set and a query set; Specifically, during operation, the electric actuator synchronously acquires the multi-channel three-phase current and three-axis vibration signals of the drive motor, and uses sliding window technology to obtain a large amount of initial sample data. From this, N initial sample data that can cover different working conditions are randomly selected to form the small sample data in this embodiment. The drive motor in this embodiment is a brushless DC motor.

[0026] In this embodiment, the diagnostic object is a 60BLDC 150W brushless DC motor. Three-phase current signals and triaxial vibration signals during motor operation are synchronously acquired using sensors, forming a 6-channel raw signal, denoted as... , The signal length is specified. Signal acquisition must cover a complete range of fault types and operating condition combinations: 1. Fault types: healthy, high-resistance fault, demagnetization fault, inter-turn short circuit, MOSFET degradation, a total of 5 types; 2. Operating conditions: speed waveforms include constant, sine, sawtooth, square wave, and random; load torque includes 0.2Nm and 0.6Nm, a total of 10 combinations; 3. Acquisition parameters: sampling frequency is set to 25.6kHz, and sufficient signal duration is acquired for each operating condition combination to ensure data is included during the stable operation phase of the motor.

[0027] In this embodiment, to eliminate noise interference and unify the data format, the original signal needs to be preprocessed to generate a standardized sample dataset. This includes: signal truncation: removing invalid data such as motor start-stop transition segments and abnormal fluctuation segments, retaining only the signal of the stable operation segment; sliding window segmentation: using a sliding window of length 5000 to segment the stable segment signal, with the window overlap rate set to 50%, generating a single sample matrix. Sufficient samples are generated for each working condition combination to meet the requirements of meta-learning task construction; normalization processing: Z-score normalization is performed on each sample matrix to eliminate dimensional differences.

[0028] In this embodiment, both the training and testing task sets adopt the "M-way K-shot" meta-learning setting to simulate small-sample diagnostic scenarios. The training task set is constructed by continuously sampling randomly from the source domain training dataset to build a "3-way 5-shot-5-query" task (3 types of faults, 5 support samples per type, and 5 query samples per type). A single task is denoted as... ,in For training the support set, The training query set is used. Twenty tasks are randomly sampled in each iteration for model parameter optimization. The test task set is constructed by sampling from the target domain test dataset to create a "3-way 5-shot-5-query" task of the same specifications as the training task set, denoted as... ,in To test the support set, This is the test query set. During testing, the prototype is computed using the test support set, and a task is tested using the best model through the test query set for model performance evaluation.

[0029] S2: Based on the sample data, a working condition encoder is used to obtain the feature embedding function value, i.e., the sample data features fused with the working condition; thus, the first... Feature embedding function values ​​of sample data for each fault category; Specifically, such as Figure 2As shown, the function of the operating condition encoder is to extract features related to the operating conditions from the original multi-channel signals (i.e., the three-phase current and three-axis vibration signals when the brushless DC motor is working) and generate multi-scale attention weights to guide the main feature extraction network and suppress condition-specific changes.

[0030] Preferably, the working condition encoder includes a feature encoding module and a layered fusion module; The feature encoding module is used to obtain the statistical features of a single channel and the cross-channel correlation features based on the sample data, so as to obtain a statistical feature vector, and then obtain a conditional encoding vector based on a multilayer perceptron. The hierarchical fusion module includes a learnable mapping network and a convolutional fusion module, used to obtain sample data features fused with the working conditions based on the conditional encoding vector and the sample data. Specifically, the feature encoding module takes 6 channels (three-phase current + triaxial vibration) of standardized sample data as input. First, it calculates the statistical characteristics (mean, standard deviation, root mean square, peak value, etc.) of each channel and the cross-channel correlation characteristics (correlation coefficient, energy ratio, etc.), forming a 55-dimensional statistical feature vector. Then, this statistical feature vector undergoes nonlinear transformation and compression using a two-layer multilayer perceptron (MLP), ultimately outputting a 64-dimensional conditional encoding vector. The feature encoding module and the multilayer perceptron are both applications of existing technologies, and their structures will not be described in detail here.

[0031] The learnable mapping network is used to obtain attention weight vectors based on the conditional encoding vectors. Specifically, obtaining four attention weight vectors based on the learnable mapping network and using the conditional encoding vectors is an application of existing technology for those skilled in the art, and the specific method for obtaining the attention weight vectors will not be described in detail here.

[0032] The convolutional fusion module is used to obtain the feature embedding function value based on the attention weight vector, that is, the sample data features fused with the working conditions; Preferably, the calculation formula for the convolutional fusion module is expressed as follows: (1) (2) In the formula: Indicates the first The output of a convolutional neural network layer. ; Indicates the convolution operation; For inputting sample data from the encoder under operating conditions; This represents the output vector after attention modulation. Indicates the relationship with the first Attention weight vectors corresponding to layers in a convolutional neural network; This indicates element-wise multiplication, with attention weights emphasizing conditionally relevant features while suppressing irrelevant variations; For the index of a convolutional neural network layer; Specifically, the attention-modulated output vector obtained after the fourth fusion stage is the final feature, which is then normalized to obtain the feature embedding function value, i.e., the sample data feature fused with the working conditions. Let the output of the last layer of the network be denoted as... .

[0033] Specifically, the layered fusion module will use the conditional encoding vector Projected onto four different depth stages of the main feature extraction network (convolutional neural network). In each stage corresponding to a convolutional neural network layer... The attention weight vector for this stage is generated through a learnable mapping network. Attention weights By modulating the convolutional feature map through element-wise multiplication, fault features most relevant to the current operating conditions are adaptively enhanced, while irrelevant interference is suppressed.

[0034] The core objective of this embodiment is to provide an end-to-end small-sample fault diagnosis scheme that can explicitly sense and decouple changes in operating conditions, thereby learning condition-invariant, highly discriminative fault feature representations. To this end, the embodiment proposes an operating-condition-aware discriminative prototypical network (OAD-PN) system framework as follows: Figure 3 As shown, it mainly includes the following three core modules and their collaborative relationships: Specifically, Prototype Networks (PN) are typical metric-based minority sample classification models in meta-learning. Their core idea is to compute a "prototype" vector for each class, defined as the average of all support samples for that class in the feature space. The fundamental principle of this method is to establish a metric space and classify by measuring the relative distance to the class prototype, rather than by parameter boundaries. This design has particular advantages in minority sample scenarios; by representing each class with a single prototype, the model is less prone to overfitting and can better generalize from a limited sample size. The averaging operation used in prototype computation provides a robust class representation that is relatively insensitive to outliers within the support set. The specific workflow is as follows: each meta-task samples the support set S and the query set Q. The embedding function maps the input samples to a low-dimensional feature space.

[0035] S3: According to the... Based on the prototype network, the feature embedding function values ​​of the sample data for the fault category are obtained. The class prototype for the fault category, i.e., the first fault category. The average value of the support set sample feature embedding function for each fault category; to obtain the first The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between the class prototypes of each fault category is calculated; then, the Softmax function is used to obtain the distance between the class prototypes of each fault category. The query set of sample data for the fault category is predicted as follows: The probability of each fault category; S31: Obtain the... Class prototypes for each fault category; Specifically, the prototype network calculates a class prototype for each fault category. The class prototype is defined as the average value of the feature embeddings of the support set samples for that category, as shown in the formula: (3) In the formula: For the first The class prototype for the fault category, i.e., the first fault category. The average value of the support set sample feature embedding function for each fault category; For the first The total number of support set samples for each fault category; For the first A set of support sets for sample data of each fault category; For the first Support set sample feature embedding function values ​​for sample data of each fault category; An index for fault categories; S32: Obtain the... The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between the class prototypes of each fault category; The formula used to obtain the Euclidean distance is as follows: (4) S33: Obtain the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category is calculated using the following formula: (5) In the formula: To make the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category; To make the first The predicted category of a query set of sample data for each fault category; An index of fault categories contained in a single meta-task; 'The total number of fault categories contained in a single meta-task; For the first in a single meta-task The Euclidean distance between the feature embedding function value of each fault category and the class prototype; S4: According to the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category is used to obtain the prototype loss; in order to establish the joint loss function. Preferably, the original loss is obtained as follows: In this embodiment, the prototype loss is the negative logarithmic probability loss on the query set: (6) In the formula: For prototype loss; A set of samples for the query set; The total number of samples in the query set; The index for the query set samples; Preferably, the joint loss function is established as follows: wherein the total loss of the joint loss function is a weighted average of the prototype loss and the center loss: (7) In the formula: The coefficient of the center loss; This represents the total loss.

[0036] Specifically, the prototype network is suitable for scenarios with few samples due to its simple structure and high computational efficiency. However, its reliance solely on Euclidean distance for classification can lead to fragmented intra-class features and insufficient inter-class differentiation when processing complex fault signals, potentially limiting diagnostic accuracy. Therefore, this embodiment introduces a condition encoder and a center loss on top of the prototype network. The prototype loss ensures that query samples are correctly classified to the nearest prototype, while the center loss enhances the clustering effect of features from within.

[0037] Specifically, this embodiment introduces a center loss as an auxiliary supervision signal based on the prototype network, jointly constraining the feature space. This makes the feature distribution of similar fault samples more compact and the feature separation of different fault classes more obvious. The center loss can maintain a learnable class center for each fault category. The center loss calculates the Euclidean distance between the feature embedding of each sample and its corresponding class center, and minimizes this distance, thereby forcing intra-class features to cluster, as shown below: (8) In the formula: Indicates central loss; Index of fault categories contained in a single meta-task 'The total number of fault categories contained in a single meta-task; Represents the feature mapping function; Indicates the first in a single meta-task Samples of each fault category; Indicates the first in a single meta-task The class center for each fault category; This represents the L2 norm.

[0038] S5: Acquire the three-phase current and three-axis vibration signals of the drive motor of the electric actuator during actual working time, so as to obtain the predicted fault category based on the trained prototype network and realize the diagnosis of the drive motor of the electric actuator.

[0039] Specifically, this embodiment proposes a unified end-to-end framework called Operational Condition-Aware Discriminative Prototype Network (OAD-PN), aiming to simultaneously address the challenges of few samples and cross-conditions. This framework is based on the standard prototype network PN, introduces a center loss to enhance feature discriminativity, and incorporates a condition encoder to enhance robustness to changes in operating conditions. The overall structure is as follows: Figure 3 As shown, the condition encoder module works in series with the main feature extraction network. The condition encoder acts as a "conditional sensor," and its output conditional codes are used to modulate the features of each layer of the main feature extraction network in real time, achieving condition-aware feature extraction. The extracted features are then fed into a prototype network module with enhanced central loss. This module jointly optimizes the loss function to complete discriminative feature learning and classification. The entire meta-learning framework trains the model by simulating few-sample tasks, enabling the model to quickly learn new tasks from a small number of samples. The process is as follows: a) Task Construction: During the meta-training phase, a large number of "M-way K-shot" tasks are continuously constructed from source domain data. Each task contains a support set and a query set.

[0040] b) Forward Propagation and Loss Calculation: For each task, support set samples acquire features through a guided master feature extraction network to calculate various prototypes. Query set samples acquire features through the same network, and prototype losses are calculated based on distances to each prototype. Simultaneously, the central loss for both support set and query set samples is calculated.

[0041] c) Parameter update: The model parameters are updated by minimizing the total joint loss of all tasks.

[0042] Specifically, in the meta-testing phase, the same sampling method is used to divide the dataset into a support set and a query set. The support set is used to compute the prototype. The best model saved during the training phase is used to evaluate and classify the samples in the query set, and the accuracy is calculated.

[0043] To fully demonstrate the effectiveness of the present invention, it was implemented and verified on a specific BLDCM experimental platform.

[0044] 1. Experimental setup (1) Diagnostic object: 60BLDC150W brushless DC motor, covering five health states: healthy, high resistance fault, demagnetization fault, inter-turn fault, MOSFET degradation.

[0045] Operating conditions: a) Velocity waveform: constant, sine, sawtooth, square wave, random, a total of 5 types.

[0046] b) Load torque: 0.2 N·m and 0.6 N·m, two types in total.

[0047] (2) Data acquisition: Three-phase current and triaxial vibration signals were acquired synchronously at a sampling frequency of 25.6 kHz. Stable operating segments were extracted, and samples were generated using a sliding window of length 5000. Each sample was a 6×5000 matrix.

[0048] (3) Task Setup: To simulate a few-shot learning scenario, this embodiment trains 600 total samples per iteration, using a "3-way 5-shot-5-query" meta-learning setup. This means that each task randomly selects three fault types, providing five support samples and five query samples for each fault type. Twenty tasks are randomly sampled per iteration. The same sampling strategy is used for model selection and validation after each iteration. During the testing phase, a prototype is computed from the support set, and diagnostic performance is evaluated on the query set without model fine-tuning or parameter updates. Ultimately, the best model is tested in a few-shot learning scenario using a task containing 300 query samples.

[0049] 2. Comparative Experiment and Result Analysis To evaluate the performance of this embodiment, it was compared with several mainstream methods, including: Prototypical network (PN), Relation network (RN), Domain-adversarial neural network (DANN), and an improved prototype network (Meta-learning with intra-class and inter-class optimization, MLIIO).

[0050] (1) Cross-waveform diagnostic experiment In this embodiment, the training and test sets had the same load, but different velocity waveforms. A total of eight different cross-waveform diagnostic tasks were designed, as shown in Table 1. These include: constant waveform (con); sine waveform (sin); sawtooth waveform (saw); square wave (squ); and random waveform (ran). Table 1. Cross-waveform experiment setup

[0051] Table 2 Results of cross-waveform experiments

[0052] The accuracy rates of each method are shown in Table 2. The results demonstrate that the OAD-PN method proposed in this embodiment achieved the best diagnostic accuracy in all eight tasks, with an average accuracy of 98.5%, and achieved near 100% accuracy in several tasks. In contrast, the prototype network of the baseline method showed greater fluctuations, the relational network and MLIIO showed limited improvement, and although DANN performed well, it was still inferior to OAD-PN. This indicates that the present invention has excellent generalization ability and robustness when faced with unseen velocity waveforms, and its operating condition encoder effectively suppresses interference caused by waveform changes.

[0053] (2) Waveform and load combined diagnostic experiment This series of experiments simulates more demanding industrial scenarios. The training set and the test set are different in terms of speed waveform and load torque. A total of 8 composite diagnostic tasks are designed, as shown in Table 3.

[0054] Table 3. Waveform + Load Experiment Setup

[0055] Table 4 Results of cross-waveform + load experiments

[0056] The experimental results are shown in Table 4. The results demonstrate that under this more challenging setting, the operational condition-aware discriminative prototype network OAD-PN in this embodiment exhibits a more significant advantage. Its average diagnostic accuracy remains at a high level of 95.2%, while the performance of other comparative methods shows a significant decline. In particular, DANN's accuracy plummeted to 31.97% in tasks transferred from a low-load training set to a high-load test set, indicating that its adversarial training mechanism fails under extremely limited sample sizes and composite domain shifts. The accuracy of the prototype network and MLIIO also dropped to the 60%-70% range. This fully demonstrates the necessity of the design of this invention, which can simultaneously cope with concurrent changes in multiple operational conditions and exhibits strong composite domain generalization ability.

[0057] 3. Ablation test To further verify the contributions of each core component in this embodiment, a systematic ablation study was conducted, comparing the following model variants: a) PN: Basic Prototype Network.

[0058] b) OPN: A prototype network with only an operating condition encoder added.

[0059] c) PCN: A prototype network with only the central loss added.

[0060] d) OAD-PN: Complete model.

[0061] Table 5 Ablation Experiment Results

[0062] The results are shown in Table 5. On the composite diagnostic task, both OPN and PCN significantly outperformed the basic PN, demonstrating the effectiveness of both operational condition awareness and feature discriminative enhancement. The complete OAD-PN model, combining both, achieved the best performance, with an accuracy exceeding the improvement brought by either individual component. This indicates a synergistic effect between the operational condition encoder and the central loss; their combined effect is key to solving the problem of small-sample diagnostics in composite domains.

[0063] 4. Feature Visualization Analysis The t-SNE technique is used to reduce high-dimensional features to two dimensions for visualization, such as... Figure 4 As shown in the figure, the results indicate that the feature distribution of the basic method is loose, with significant inter-class overlap. The feature clustering compactness of OPN and PCN is improved. The feature visualization generated by the complete OAD-PN model is the most ideal: fault samples of the same type are tightly clustered together, forming clear clusters, while the separation between clusters of different types is the highest. This provides intuitive evidence for the superior performance of this invention.

[0064] This embodiment introduces a working condition encoder module to obtain feature embedding function values ​​and, based on a prototype network, obtains class prototypes of fault categories. It then predicts the probability of fault sample data and establishes a joint loss function based on prototype loss and music center loss to train the prototype network. Based on the trained prototype network, fault diagnosis is performed during actual operation. This embodiment explicitly extracts operating condition features from the original multi-channel signal and generates hierarchical attention weights to modulate the main feature extraction network, thereby actively suppressing feature changes related to the working condition and guiding the network to learn fault feature representations with invariant conditions. This fundamentally improves the model's generalization ability to unseen complex working conditions. It achieves highly robust diagnosis of complex working condition changes under limited sample conditions. This embodiment can effectively handle the simultaneous speed waveform and load torque changes that occur in motors during real operation. Even in extreme cases where there are only a few samples for each fault category, it maintains extremely high diagnostic accuracy, solving the problem of performance degradation due to domain bias in traditional data-driven models. The center loss and prototype network loss are jointly optimized. The center loss acts as an auxiliary supervision signal, forcing features of similar faults to converge towards a learnable class center during training. This complements the prototype loss, which aims to increase inter-class distance, and together constrains the model to learn highly discriminative feature embeddings. This significantly enhances the discriminativeness of fault features in the embedding space. The fault features learned in this embodiment exhibit extremely high intra-class compactness and inter-class separation. Samples from different fault categories form clear and clustered groups in the feature space, making distance-based few-sample classification more accurate and reliable. This embodiment provides an end-to-end general meta-learning framework by constructing an Operation Condition Aware Discriminative Prototype Network (OAD-PN) and organically integrates the operating condition encoder and center loss into the meta-learning training paradigm. By training on a large number of tasks covering different combinations of faults and operating conditions, the model learns how to quickly extract key information from a small number of support samples and construct discrimination boundaries. It has a strong ability to quickly adapt to small samples. The "learning framework" trained on a large number of meta-tasks enables it to quickly adapt to new fault diagnosis tasks (including new fault types or new combinations of operating conditions) without retraining or with only a very small number of new samples, which greatly improves the practicality and deployment flexibility of the method.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor, characterized in that, Includes the following steps: S1: Acquire the three-phase current and triaxial vibration signals of the drive motor of the electric actuator under different operating conditions, and use the sliding window technique to acquire M initial sample data, and randomly select N initial sample data under different operating conditions to form sample data; where M is the total number of initial sample data; N is the total number of sample data, M>N; S2: Based on the sample data, a working condition encoder is used to obtain the feature embedding function value, i.e., the sample data features fused with the working condition; to obtain the first... Feature embedding function values ​​of sample data for each fault category; S3: According to the... Based on the prototype network, the feature embedding function values ​​of the sample data for the fault category are obtained. The class prototype for the fault category, i.e., the first fault category. The average value of the support set sample feature embedding function for each fault category; to obtain the first The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between the class prototypes of each fault category; and then to obtain the first... The query set of sample data for the fault category is predicted as follows: The probability of each fault category; S4: According to the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category is used to obtain the prototype loss; a joint loss function is then established to train the prototype network. S5: Acquire the three-phase current and three-axis vibration signals of the drive motor of the electric actuator during actual working time, so as to obtain the predicted fault category based on the trained prototype network and realize the diagnosis of the drive motor of the electric actuator.

2. The meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor according to claim 1, characterized in that, The working condition encoder includes a feature encoding module and a hierarchical fusion module; The feature encoding module is used to obtain the statistical features of a single channel and the cross-channel correlation features based on the sample data, so as to obtain a statistical feature vector, and then obtain a conditional encoding vector based on a multilayer perceptron. The hierarchical fusion module includes a learnable mapping network and a convolutional fusion module, used to obtain sample data features fused with the working conditions based on the conditional encoding vector and the sample data. The learnable mapping network is used to obtain the attention weight vector based on the conditional encoding vector; The convolutional fusion module is used to obtain the feature embedding function value based on the attention weight vector, that is, the sample data features fused with the working conditions.

3. The meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor according to claim 2, characterized in that, The calculation formula for the convolutional fusion module is expressed as follows: In the formula: Indicates the first The output of a convolutional neural network layer. ; Indicates the convolution operation; For inputting sample data from the encoder under operating conditions; This represents the output vector after attention modulation. Indicates the relationship with the first Attention weight vectors corresponding to layers in a convolutional neural network; This indicates element-wise multiplication; This is the index of the convolutional neural network layer.

4. The meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor according to claim 1, characterized in that, S3 includes: S31: Obtain the... The class prototypes for each fault category are represented by the following formula: In the formula: For the first The class prototype for the fault category, i.e., the first fault category. The average value of the support set sample feature embedding function for each fault category; For the first The total number of support set samples for each fault category; For the first A set of support sets for sample data of each fault category; For the first Support set sample feature embedding function values ​​for sample data of each fault category; An index for fault categories; S32: Obtain the... The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between the class prototypes of each fault category; S33: Obtain the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category is calculated using the following formula: In the formula: To make the first The query set of sample data for the fault category is predicted as follows: The probability of each fault category; To make the first The predicted category of a query set of sample data for each fault category; Index of fault categories contained in a single meta-task 'The total number of fault categories contained in a single meta-task; For the first in a single meta-task The Euclidean distance between the feature embedding function value of each fault category and the class prototype; For the first The query set of sample feature embedding function values ​​for the sample data of the fault category and the first fault category The Euclidean distance between class prototypes of each fault category.

5. The meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor according to claim 1, characterized in that, The original loss is obtained as follows: In the formula: For prototype loss; A set of samples for the query set; The total number of samples in the query set; This is the index for the query set samples.

6. The meta-learning diagnostic method for sensing the operating conditions of an electric actuator drive motor according to claim 5, characterized in that, The joint loss function is established as follows: In the formula: The coefficient of the center loss; Total loss; in, In the formula: Indicates central loss; Index of fault categories contained in a single meta-task 'The total number of fault categories contained in a single meta-task; Represents the feature mapping function; Indicates the first in a single meta-task Samples of each fault category; Indicates the first in a single meta-task The class center for each fault category; This represents the L2 norm.

Images