CN-122017561-A - Intelligent motor diagnosis method, system and medium

CN122017561ACN 122017561 ACN122017561 ACN 122017561ACN-122017561-A

Abstract

The application provides an intelligent diagnosis method, system and storage medium for a motor, wherein the method comprises the steps of obtaining multi-source heterogeneous information and real-time working condition parameters of the motor, transforming the multi-source heterogeneous information into a frequency domain to obtain a frequency spectrum tensor, mapping the real-time working condition parameters into a learnable spectrum attention mask, weighting the frequency spectrum tensor by the spectrum attention mask to obtain a mechanism weighted frequency spectrum, processing the multi-source heterogeneous information and the real-time working condition parameters by a physical simulation and generation model to obtain target enhancement data, constructing a hybrid diagnosis network, taking the mechanism weighted frequency spectrum and the target enhancement data as input, taking the spectrum attention mask as bias of a self-attention mechanism to extract global time sequence characteristics, inputting the fault type probability, the fault quantification parameters and a fault feature heat map into a multitask decoder based on the global time sequence characteristics, and obtaining a diagnosis result based on the fault type probability, the fault quantification parameters and the fault feature heat map. The application can realize motor diagnosis more accurately.

Inventors

XU YI
LIU PEIJUN
WANG JIASHUAI
CHENG LI
ZHONG BO

Assignees

杭州励贝电液科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. An intelligent motor diagnosis method, which is characterized by comprising the following steps: the method comprises the steps of obtaining multi-source heterogeneous information of a motor and real-time working condition parameters of the motor, converting the multi-source heterogeneous information into a frequency domain to obtain a frequency spectrum tensor, mapping the real-time working condition parameters into a learnable spectrum attention mask according to a physical mechanism of motor faults, and weighting the frequency spectrum tensor by the spectrum attention mask to obtain a mechanism weighted frequency spectrum; Processing the multi-source heterogeneous information and the real-time working condition parameters through a physical simulation and generation model to obtain physical information enhanced targeting enhancement data; Constructing a hybrid diagnostic network comprising a transducer, taking the mechanism weighted spectrum and the targeting enhancement data as inputs, and extracting global timing features in the transducer with the spectral attention mask as a bias for a self-attention mechanism; And inputting the global time sequence characteristics into a multi-task decoder to output the fault type probability, the fault quantification parameter and the fault characteristic heat map used for representing the diagnosis basis of the motor in parallel, and obtaining a diagnosis result based on the fault type probability, the fault quantification parameter and the fault characteristic heat map.
2. The method of claim 1, wherein transforming the multi-source heterogeneous information into the frequency domain to obtain a spectral tensor comprises: Respectively carrying out customized layering pretreatment adapted to the characteristics of each signal on each signal in the multi-source heterogeneous information to obtain a plurality of paths of pretreatment signals; Performing time-frequency conversion on the multi-channel preprocessing signals to obtain multi-channel frequency spectrums; channel screening is carried out on the multipath frequency spectrums to remove redundant frequency spectrums so as to obtain screened frequency spectrums; And aligning the screened frequency spectrums according to time stamps and splicing to obtain frequency spectrum tensors.
3. The method of claim 2, wherein the mapping the real-time operating condition parameters into a learnable spectral attention mask according to a physical mechanism of motor failure comprises: acquiring inherent parameters of a motor, and inputting the real-time working condition parameters and the inherent parameters into a fault frequency encoder for encoding a fault characteristic frequency formula to obtain a theoretical fault characteristic frequency set; Constructing a Gaussian spectrum attention basis mask comprising a leachable center frequency, a leachable bandwidth and a leachable attenuation coefficient by taking the theoretical fault characteristic frequency set as a center frequency basis; inputting the real-time working condition parameters into a lightweight adjustment network to predict working condition self-adaptive adjustment amounts corresponding to each leachable parameter; And superposing the working condition self-adaptive adjustment quantity and the corresponding leachable parameters in the Gaussian spectrum attention base mask to obtain a dynamic spectrum attention mask after nonlinear modulation of the working condition parameters.
4. The method of claim 3, wherein said weighting said spectral tensors with said spectral attention mask to obtain a mechanically weighted spectrum comprises: multiplying the spectral attention mask by the spectral tensor element by element to obtain a primary weighted spectrum; Inputting the primary weighted spectrum to a physical constraint convolution layer formed by fault physical rule codes for characteristic enhancement to obtain a mechanism enhancement characteristic diagram; and splicing the mechanism strengthening characteristic diagram and the primary weighted spectrum along the channel dimension, and performing attention fusion to obtain a fused spectrum characteristic serving as a mechanism weighted spectrum.
5. The method of claim 1, wherein the processing the multi-source heterogeneous information and the real-time operating condition parameters to obtain physical information enhanced targeting enhancement data through a physical simulation and generation model comprises: Acquiring inherent parameters and a preset fault type of a motor, and inputting the inherent parameters and the fault type into a dynamic simulation model constructed based on digital twinning to generate a simulation signal containing fault characteristics; Inputting the multi-source heterogeneous information into a pre-trained diagnosis model to identify fault categories with classification confidence below a threshold as difficulty samples; inputting the simulation signal and the difficulty sample to a diffusion model that generates conditions with the difficulty sample to generate initial enhancement data for the difficulty sample; and inputting the initial enhancement data and the real fault samples in the multi-source heterogeneous information into a loop generation countermeasure network to perform style migration so as to align the characteristic distribution of the initial enhancement data with the characteristic distribution of the real fault samples and output targeted enhancement data.
6. The method of claim 1, wherein the constructing a hybrid diagnostic network comprising a transducer comprises: Decomposing the mechanism weighted spectrum into time-step embedding vectors to form a token sequence, applying masking constraints to a self-attention score matrix of the token sequence with the spectral attention mask in an encoder layer of a fransformer to generate a physically enhanced hidden layer representation; extracting a fault feature map layer by cascading a plurality of encoder layers, and introducing a differentiable wavelet transform layer between the encoder layers to perform multi-scale time-frequency decomposition on the hidden layer representation to extract residual features; the residual signature is skip connected to the fault signature to build a hybrid diagnostic network that includes a physically enhanced transducer.
7. The method of claim 6, wherein the extracting global timing features in the fransformer with the spectral attention mask as a bias for self-attention mechanisms comprises: Linearly projecting the hidden layer representations to three different feature spaces respectively to obtain a query matrix, a key matrix and a value matrix; Superimposing the spectral attention mask as a bias term onto the dot product result of the query matrix and the key matrix to generate a physical bias attention score; Normalizing the physical bias attention score to obtain an attention weight matrix; and carrying out weighted aggregation operation on the attention weight matrix and the value matrix to obtain a global time sequence characteristic guided by fault mechanism priori.
8. The method of claim 1, wherein the obtaining a diagnostic result based on the fault type probability, fault quantification parameter, and fault signature heat map comprises: Performing confidence calibration on the fault type probability to obtain calibrated confidence; constructing damage quantification indexes according to the fault quantification parameters; the fault characteristic heat map is superimposed on an original signal waveform corresponding to the multi-source heterogeneous information to generate a visual diagnosis map; fusing the calibrated confidence coefficient, the damage quantification index and the visual diagnostic map, and processing and outputting a diagnostic result containing a diagnostic basis description through an interpretability report generator.
9. The intelligent motor diagnosis system is characterized by comprising an embedding module, an enhancing module, a network module and a diagnosis module, wherein, The embedding module is used for acquiring multi-source heterogeneous information of the motor and real-time working condition parameters thereof, converting the multi-source heterogeneous information into a frequency domain to obtain a frequency spectrum tensor, mapping the real-time working condition parameters into a learnable spectrum attention mask according to a physical mechanism of motor faults, and weighting the frequency spectrum tensor by the spectrum attention mask to obtain a mechanism weighted frequency spectrum; The enhancement module is used for processing the multi-source heterogeneous information and the real-time working condition parameters through a physical simulation and generation model to obtain physical information enhanced targeting enhancement data; the network module is used for constructing a hybrid diagnosis network comprising a transducer, taking the mechanism weighted spectrum and the targeting enhancement data as inputs, and extracting global time sequence characteristics in the transducer by taking the spectrum attention mask as the bias of a self-attention mechanism; The diagnosis module is used for inputting the global time sequence characteristics into the multi-task decoder to output the fault type probability, the fault quantification parameter and the fault characteristic heat map used for representing the diagnosis basis of the motor in parallel, and obtaining a diagnosis result based on the fault type probability, the fault quantification parameter and the fault characteristic heat map.
10. A computer readable storage medium having stored thereon a computer program executable on a processor, characterized in that the computer program, when executed by the processor, implements a motor intelligent diagnostic method according to any of claims 1 to 8.

Description

Intelligent motor diagnosis method, system and medium Technical Field The application relates to the technical field of motors, in particular to an intelligent motor diagnosis method, system and medium. Background The motor is used as core power equipment in industrial production, and the real-time monitoring and accurate diagnosis of the running state of the motor are important for guaranteeing production safety and reducing maintenance cost. Currently, the motor fault diagnosis technology is mainly divided into two main categories, namely a method based on a physical mechanism model, wherein the method relies on expert experience, and faults are judged by analyzing specific frequency components (such as bearing fault characteristic frequencies) of vibration signals. The method has strong interpretability, but the diagnosis effect is highly dependent on the accuracy of expert knowledge, and early weak faults and complex compound faults are difficult to identify. The other is a purely data-driven method, which uses machine learning (such as a support vector machine, a convolutional neural network and the like) to train massive historical data and automatically learn fault modes. The method can effectively process complex nonlinear relations, but is essentially a black box model, the diagnosis result lacks of physical basis and has poor interpretability, and in the actual industrial scene of scarce fault data samples, the generalization capability and the diagnosis accuracy of the model can be drastically reduced. In order to combine the advantages of the two, the prior art tries to fuse the mechanism and the data method, but a shallow fusion mode of feature extraction and model series connection is generally adopted, namely, a plurality of mechanism features are manually extracted from signals by means of expert experience, and then the features are provided as input to a data classification model for fault identification. The fusion mode is not only free from dependence on expert experience, and has low automation degree, but also the mechanism knowledge and the data model are mutually independent in the training process, and the deep coordination of the algorithm level cannot be realized, so that the accuracy and the adaptability of the diagnosis of the model are still limited when the model faces to the individual difference of equipment, the working condition change and the small sample scene. Disclosure of Invention In order to be able to more accurately realize the diagnosis of the motor. The embodiment of the application provides a motor intelligent diagnosis method, a motor intelligent diagnosis system and a motor intelligent diagnosis medium. In a first aspect, a motor intelligent diagnosis method is provided, the method comprising: the method comprises the steps of obtaining multi-source heterogeneous information of a motor and real-time working condition parameters of the motor, converting the multi-source heterogeneous information into a frequency domain to obtain a frequency spectrum tensor, mapping the real-time working condition parameters into a learnable spectrum attention mask according to a physical mechanism of motor faults, and weighting the frequency spectrum tensor by the spectrum attention mask to obtain a mechanism weighted frequency spectrum; Processing the multi-source heterogeneous information and the real-time working condition parameters through a physical simulation and generation model to obtain physical information enhanced targeting enhancement data; Constructing a hybrid diagnostic network comprising a transducer, taking the mechanism weighted spectrum and the targeting enhancement data as inputs, and extracting global timing features in the transducer with the spectral attention mask as a bias for a self-attention mechanism; And inputting the global time sequence characteristics into a multi-task decoder to output the fault type probability, the fault quantification parameter and the fault characteristic heat map used for representing the diagnosis basis of the motor in parallel, and obtaining a diagnosis result based on the fault type probability, the fault quantification parameter and the fault characteristic heat map. In some of these embodiments, said transforming the multi-source heterogeneous information into the frequency domain to obtain a spectral tensor comprises: Respectively carrying out customized layering pretreatment adapted to the characteristics of each signal on each signal in the multi-source heterogeneous information to obtain a plurality of paths of pretreatment signals; Performing time-frequency conversion on the multi-channel preprocessing signals to obtain multi-channel frequency spectrums; channel screening is carried out on the multipath frequency spectrums to remove redundant frequency spectrums so as to obtain screened frequency spectrums; And aligning the screened frequency spectrums according to time stamps and splicing to obtain frequency sp