CN-121997168-A - Meta learning diagnosis method for working condition sensing of drive motor of electric actuator

CN121997168ACN 121997168 ACN121997168 ACN 121997168ACN-121997168-A

Abstract

The embodiment discloses a meta-learning diagnosis method for working condition perception of an electric actuator driving motor, which comprises the steps of obtaining a characteristic embedding function value through a working condition encoder module, obtaining a class prototype of a fault class based on a prototype network, building a center loss based on prototype loss, building a joint loss function, training the prototype network, and carrying out fault diagnosis in an actual working process based on the trained prototype network. The method realizes high-robustness diagnosis on the change of the composite working condition under the condition of few samples, remarkably improves the explicit modeling capability on the operation condition, and greatly improves the diagnosis precision. The invention obviously enhances the discrimination of fault characteristics in the embedded space. The invention has strong small sample rapid self-adaptation capability, so that the method can be rapidly adapted without retraining or with a small amount of new samples when encountering a new fault diagnosis task, and the practicability and deployment flexibility of the method are greatly improved.

Inventors

MIAO QIANG
LUO QINGWEN
WANG JIANYU
CAO DONGYANG
ZHANG HENG
ZHANG YUJIE

Assignees

四川大学

Dates

Publication Date: 20260508
Application Date: 20260113

Claims (6)

1. A meta-learning diagnosis method for sensing working conditions of a drive motor of an electric actuator is characterized by comprising the following steps: S1, acquiring three-phase current and three-axis vibration signals of a driving motor of an electric actuator when the driving motor works under different working conditions, acquiring M initial sample data by adopting a sliding window technology, and randomly selecting N initial sample data under different working conditions to form the sample data, wherein M is the total number of the initial sample data, N is the total number of the sample data, and M is more than N; s2, acquiring characteristic embedded function values, namely sample data characteristics fused with working conditions, by adopting a working condition encoder according to the sample data to acquire a first step Characteristic embedding function values of sample data of each fault class; S3 according to the first Feature embedding function values of sample data of each fault category, and acquiring the first based on a prototype network Class prototypes of fault class, i.e. the first Average value of support set sample feature embedding function values of sample data of each fault class to obtain the first Query set sample feature embedding function value and th of sample data of each fault class The Euclidean distance between class prototypes of the fault class is further obtained Sample prediction of query set of sample data for each failure category as the first Probability of individual fault categories; s4, according to the following Sample prediction of query set of sample data for each failure category as the first Acquiring prototype loss according to the probability of each fault class, and training a prototype network by establishing a joint loss function; And S5, acquiring three-phase current and three-axis vibration signals of the drive motor of the telex actuator in actual working time, so as to acquire predicted fault types according to the trained prototype network and realize diagnosis of the drive motor of the telex actuator.
2. The meta-learning diagnosis method for sensing the working condition of the drive motor of the electric actuator according to claim 1, wherein the working condition encoder comprises a feature encoding module and a layered fusion module; the feature coding module is used for acquiring statistical features and cross-channel correlation features of a single channel according to the sample data so as to acquire statistical feature vectors, and further acquiring conditional coding vectors based on a multi-layer perceptron; The hierarchical fusion module comprises a learnable mapping network and a convolution fusion module and is used for acquiring sample data characteristics fused with working conditions according to the condition coding vector and the sample data; the learnable mapping network is used for acquiring an attention weight vector according to the condition coding vector; The convolution fusion module is used for acquiring characteristic embedding function values, namely sample data characteristics fused with working conditions, according to the attention weight vector.
3. The meta-learning diagnosis method for working condition sensing of an electric actuator driving motor according to claim 2, wherein the calculation formula of the convolution fusion module is as follows: Wherein: Represent the first The layer convolves the output of the neural network layer, ; Representing a convolution operation; Sample data of the working condition encoder is input; representing the vector output after the attention modulation; Representation and the first The attention weight vector corresponding to the layer convolution neural network layer; Representing multiplication by element; is an index of the convolutional neural network layer.
4. The meta-learning diagnosis method for sensing the working condition of the drive motor of the electric actuator according to claim 1, wherein the step S3 comprises: S31 acquisition of the first Class prototypes for each fault class are formulated as follows: Wherein: Is the first Class prototypes of fault class, i.e. the first The average value of the function value is embedded in the support set sample characteristics of the sample data of each fault class; Is the first A support set sample total number of sample data for each failure category; Is the first A set of support sets of sample data for each failure category; Is the first Embedding function values into the support set sample characteristics of the sample data of the fault categories; index for failure category; S32, acquiring the first Query set sample feature embedding function value and th of sample data of each fault class Euclidean distance between class prototypes of the individual fault classes; S33, acquisition of the first time Sample prediction of query set of sample data for each failure category as the first The probability of each fault class is expressed as follows: Wherein: To get the first Sample prediction of query set of sample data for each failure category as the first Probability of individual fault categories; To get the first A predictive category of a query set sample of sample data for each fault category; Indexing fault categories contained in a single unit task ' Is the total number of fault categories contained in a single unit task; is the first in a single unit task The feature embedding function value of each fault class and the Euclidean distance between the class prototypes; Is the first Query set sample feature embedding function value and th of sample data of each fault class Euclidean distance between class prototypes of individual fault classes.
5. The meta-learning diagnosis method for sensing the working condition of the drive motor of the electric actuator according to claim 1, wherein the original loss is obtained as follows: Wherein: Loss for prototype; A set of samples that is a query set; A total number of samples for the query set; Index for a sample of the query set.
6. The meta-learning diagnosis method for working condition sensing of an electric actuator driving motor according to claim 5, wherein the joint loss function is established as follows: Wherein: is the coefficient of center loss; Is the total loss; Wherein, the Wherein: Representing center loss; Indexing fault categories contained in a single unit task ' Is the total number of fault categories contained in a single unit task; representing a feature mapping function; Representing the first of the individual unit tasks Samples of individual fault categories; Representing the first of the individual unit tasks Class centers for the individual fault classes; Denote the L 2 norm.

Description

Meta learning diagnosis method for working condition sensing of drive motor of electric actuator Technical Field The invention relates to the technical field of industrial automation and high-reliability electric drive, in particular to a meta-learning diagnosis method for sensing working condition of an electric actuator driving motor. Background In high-precision electric actuating systems such as aerospace and robots, the reliability of a brushless direct current motor serving as a core power source directly determines the safety and performance of the whole actuating system. These critical electrical actuator motors often operate at extremely variable speeds, loads and complex waveform commands, and their operational status monitoring and early fault diagnosis present more serious challenges than in common industrial scenarios. Brushless DC motors have become a core power component in modern electric actuation systems, industrial robots and precision servo drives because of their high power density, high efficiency and excellent controllability, and their operational reliability is of paramount importance. To ensure high reliability operation of these systems, it is critical to perform real-time, accurate fault diagnosis of the motor. At present, the technical schemes in the field mainly can be divided into three types, namely a model-based method, a signal processing-based method and a data driving method. Model-based methods rely on accurate mathematical models of motors and state observers to detect faults through residual analysis, however, motors in teleactuators typically operate under complex and variable conditions, whose accurate physical models are difficult to build, and are very sensitive to model mismatch. Methods based on signal processing, such as fast fourier transform, wavelet transform, etc., diagnose faults by analyzing time-frequency characteristics of signals such as vibration, current, etc., but such methods often require expert knowledge to perform manual characteristic extraction and parameter adjustment, and have limited performance when dealing with non-stationary signals and strong noise interference. In recent years, data-driven methods typified by deep learning have exhibited great potential in mechanical failure diagnosis, and they are capable of automatically learning complex failure features from data. However, the success of these methods is severely dependent on large, complete labeling data. In practical industrial scenarios, especially for telex key motors, it is costly, or even impractical, to obtain a large number of samples of all potential failure modes under different conditions, resulting in models facing the challenge of small sample learning. In addition, when the trained model is deployed under different operating conditions (e.g., varying rotational speed, load) than the training data, the model performance may be significantly degraded due to domain shift issues. While strategies such as transfer learning are used to alleviate this problem, with the aim of learning domain invariant features, they are prone to destabilization of the challenge training mechanism and limited in the extreme case of dealing with complex conditions where waveforms and loads change in complex, and where target domain samples are few, and it is difficult to learn truly robust condition invariant features. In conclusion, the actual running conditions of the motor are complex and changeable, so that the distribution of the monitoring data changes along with the actual running conditions of the motor, and the traditional data driving diagnosis model faces the field deviation problem. Meanwhile, in an actual industrial scenario, it is costly to acquire a large number of labeling data for all fault types, resulting in a challenge with small sample learning. The existing small sample learning method and the field self-adaptive method lack explicit modeling capability for operation conditions when dealing with complex working conditions of waveform and load compound change, and the diagnosis precision and the robustness are obviously reduced. Therefore, in the prior art, when the key motor of the electric actuator is subjected to common and serious engineering challenge of few samples and complex variable working conditions, the requirements of a high-reliability system are difficult to meet in diagnosis precision and robustness. Disclosure of Invention The invention discloses a meta-learning diagnosis method for sensing working conditions of a drive motor of an electric actuator, which aims to overcome the technical problems. In order to achieve the above object, the technical scheme of the present invention is as follows: a meta-learning diagnosis method for sensing working conditions of a drive motor of an electric actuator comprises the following steps: S1, acquiring three-phase current and three-axis vibration signals of a driving motor of an electric actuator when the driving motor