CN-122020014-A - Asynchronous motor time-space diagram network diagnosis method and system integrating physical model

CN122020014ACN 122020014 ACN122020014 ACN 122020014ACN-122020014-A

Abstract

A method and a system for diagnosing a space-time diagram network of an asynchronous motor by fusing a physical model belong to the technical field of industrial automation and solve the problems that the existing method for diagnosing faults of the asynchronous motor ignores the space topological relation of a sensor, lacks the guidance of a physical mechanism and has poor generalization capability under variable working conditions. The method comprises the steps of collecting multi-source heterogeneous data of the asynchronous motor, preprocessing the multi-source heterogeneous data to obtain a signal snapshot sequence of a multi-sensor physical topological graph, constructing and training a PST-GATN model by adopting the preprocessed data, and carrying out forward propagation on the multi-source heterogeneous data of the asynchronous motor by adopting the trained PST-GATN model to obtain a diagnosis result. The invention is suitable for fault diagnosis scenes of the asynchronous motor.

Inventors

JIA GUANGXUE
DOU ZHIYONG
LIU XIAOJING
SHAO KEXUAN
DU SHANGMING

Assignees

国电电力发展股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251206

Claims (10)

1. The asynchronous motor time-space diagram network diagnosis method integrating the physical model is characterized by comprising the following steps of: S1, collecting multi-source heterogeneous data of an asynchronous motor, preprocessing the multi-source heterogeneous data, and obtaining a signal snapshot sequence of a multi-sensor physical topological graph; S2, constructing and training a PST-GATN model by adopting the preprocessed data, wherein the method comprises the following steps of: S21, constructing a multi-sensor physical topological graph of the asynchronous motor; s22, constructing a space-time feature extraction network based on a graph attention mechanism and a gating circulating unit, and processing the signal snapshot sequence to obtain a space-time fusion feature vector; s23, classifying the time fusion feature vectors by adopting a multi-layer perceptron to generate probability distribution vectors so as to obtain classification loss; S24, converting the space-time fusion feature vector into a space-time fusion feature vector through Park conversion The coordinate system is used for estimating the internal flux linkage and the derivative thereof by using a physical state decoder to obtain physical constraint loss; S25, weighting the classification loss and the physical constraint loss to obtain a composite loss function, and minimizing the composite loss function to obtain a trained PST-GATN model; and S3, forward propagation is carried out on multi-source heterogeneous data of the asynchronous motor by adopting a trained PST-GATN model, and a diagnosis result is obtained.
2. The method for diagnosing the space-time diagram network of the asynchronous motor fused with the physical model according to claim 1, wherein the multi-source heterogeneous data comprise three-phase current, three-phase voltage, multi-axis vibration signals and temperature.
3. The method for diagnosing a space-time diagram network of an asynchronous motor fused with a physical model according to claim 1, wherein the preprocessing comprises the steps of aligning, normalizing and time window slicing the acquired multi-source heterogeneous data.
4. The method for diagnosing a space-time diagram network of an asynchronous motor with a fusion physical model according to claim 1, wherein the step S21 is specifically implemented by abstracting each sensor into a diagram node based on a physical structure of the asynchronous motor, defining edges and connection weights between the nodes according to physical connectivity and functional coupling between the sensors, and constructing a weighted multi-sensor physical topological diagram expressed as: , Wherein, the Representing deployment on an electric machine A plurality of physical sensors are arranged on the surface of the substrate, Representing a set of edges that are to be joined, Representing a weighted adjacency matrix based on a priori knowledge, 。
5. The method for diagnosing a space-time diagram network of an asynchronous motor integrated with a physical model according to claim 4, wherein in S21, the sensor comprises at least a three-phase current sensor, a three-phase voltage sensor, a multi-axis vibration sensor and a temperature sensor.
6. The method for diagnosing a space-time diagram network of an asynchronous motor fused with a physical model according to claim 1, wherein the space-time characteristic extraction network is formed by stacking L layers of space-time diagram attention layers, and each layer comprises a diagram attention network and a gating circulation unit.
7. The method for diagnosing a space-time diagram network of an asynchronous motor with a fused physical model according to claim 1, wherein the step S24 is to construct a physical constraint loss function by using a d-q axis physical model equation of the asynchronous motor, thereby obtaining a physical constraint loss, and specifically comprises the steps of: estimating from the advanced feature vectors by a physical state decoder Shaft stator flux linkage and derivative thereof; Substituting the estimated value and the measured value of voltage and current together Calculating residual errors on two sides of the shaft stator voltage equation; The physical constraint penalty is the L2 norm of the residual.
8. An asynchronous motor space-time diagram network diagnosis system integrating a physical model, wherein the system is realized based on an asynchronous motor space-time diagram network diagnosis method integrating a physical model according to any one of claims 1 to 7, and the system comprises: The data processing module is used for acquiring multi-source heterogeneous data of the asynchronous motor, preprocessing the multi-source heterogeneous data and acquiring a signal snapshot sequence of a multi-sensor physical topological graph; The model training module is used for constructing and training a PST-GATN model by adopting the preprocessed data, and comprises the following steps: Constructing a multi-sensor physical topological graph of the asynchronous motor; Based on a graph attention mechanism and a gating circulating unit, constructing a space-time feature extraction network, and processing the signal snapshot sequence to obtain a space-time fusion feature vector; classifying the time-space fusion feature vectors by adopting a multi-layer perceptron to generate probability distribution vectors, thereby obtaining classification loss; converting the space-time fusion feature vector to by Park transformation The coordinate system is used for estimating the internal flux linkage and the derivative thereof by using a physical state decoder to obtain physical constraint loss; Weighting the classification loss and the physical constraint loss to obtain a composite loss function, and minimizing the composite loss function to obtain a trained PST-GATN model; and the fault diagnosis module is used for carrying out forward propagation on multi-source heterogeneous data of the asynchronous motor by adopting the trained PST-GATN model to obtain a diagnosis result.
9. A computer device, characterized in that it comprises a memory in which a computer program is stored and a processor which, when running the computer program stored in the memory, performs a method for asynchronous machine time space diagram network diagnosis according to any of claims 1-7, which incorporates a physical model.
10. A computer readable storage medium for storing a computer program for executing a physical model-fused asynchronous motor space-time diagram network diagnosis method according to any one of claims 1 to 7.

Description

Asynchronous motor time-space diagram network diagnosis method and system integrating physical model Technical Field The invention belongs to the technical field of industrial automation, and particularly relates to a time-space diagram network diagnosis method of an asynchronous motor. Background Asynchronous motors, also called induction motors, are the most core and most widely used power equipment in modern industrial systems, and the stability of the running state of the asynchronous motors is directly related to the safety, efficiency and economic benefits of the whole production system. Therefore, implementing accurate fault diagnosis and prospective state monitoring on the asynchronous motor so as to avoid unplanned shutdown caused by sudden faults and huge economic loss and potential safety hazards caused by the unplanned shutdown, is always a key challenge facing the fields of industrial automation and intelligent operation and maintenance. Currently, the mainstream motor fault diagnosis technology relies heavily on deep analysis of motor operation data. The data generally show multi-source heterogeneous characteristics, and the main data acquisition source comprises a monitoring and data acquisition system of the motor, and is used for acquiring macroscopic operation parameters such as current, voltage, rotating speed, power, temperature and the like, and comprises a high-frequency vibration sensor and the like additionally installed for precise diagnosis. Based on these data, the existing diagnostic methods have developed several main technical paths, the first category is methods based on classical signal processing, such as applying tools of fourier transform (FFT), wavelet transform, etc., by analyzing the frequency spectrum or time spectrum of the current or vibration signal, searching for characteristic frequencies and their sideband harmonics corresponding to specific faults (such as damage of the inner and outer rings of the bearing, rotor breakage or stator and rotor air gap eccentricity) in complex signals. The second type is based on traditional machine learning methods, such as Support Vector Machines (SVMs), decision trees, random forests, etc., which typically require a complex "feature engineering" pre-step of relying on a priori knowledge of a diagnostician to design and extract valid time-domain and frequency-domain statistical features from the original signal, and then inputting these features into a classifier for training and recognition. The third category is a deep learning-based method that has been rapidly developed in recent years, and is represented by Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). CNN is often used to process one-dimensional waveforms or two-dimensional time-frequency spectra of vibration signals after conversion by virtue of its strong local feature extraction capability, while RNN and its variants (e.g., LSTM and GRU) are used to process time-series dependencies of signals due to their advantages in capturing sequence data dynamics. While the above-described methods have achieved certain success under certain and desirable conditions, their inherent limitations and drawbacks have become particularly pronounced in the face of increasingly complex modern industrial environments. First, conventional methods based on signal processing have a strong "expert experience dependency". The accuracy of the diagnostics is highly dependent upon the deep understanding of the motor failure mechanism by the diagnostics engineer in order to manually set and search for specific failure feature frequencies. When the motor is in dynamic working conditions such as variable frequency, variable speed or variable load, the fault characteristic frequency can be remarkably shifted and modulated and even is submerged by strong background noise, so that the traditional frequency spectrum analysis method is extremely easy to fail or misjudge, and the front end of the traditional machine learning method still depends on manual characteristic extraction, so that the fundamental defect is inherited. Secondly, although the deep learning method realizes automatic feature extraction to a certain extent, the existing implementation mode has serious design defects generally. They mostly fall into a "black box" model, driven purely by data, whose learned features often lack clear physical interpretability, severely disjointing from the intrinsic physical mechanisms of the motor. This results in the model being extremely prone to becoming overfitted when the training data is insufficient or the operating conditions change, producing false diagnostic results that violate the basic motor laws, such as false identification of fault signatures that would only appear under high load conditions. More importantly, conventional deep learning models generally ignore the inherent "physical topological relationship" between sensors. Existing CNN or RNN models often take s