CN-121784607-B - Inverter degradation fault diagnosis method

CN121784607BCN 121784607 BCN121784607 BCN 121784607BCN-121784607-B

Abstract

The invention belongs to the technical field of fault diagnosis and discloses a method for diagnosing degradation faults of an inverter. The diagnosis method comprises the steps of S1, taking a pyramid visual transducer as a main network, embedding a cross-scale feature fusion attention module and a KAN network layer module, constructing an inverter degradation fault diagnosis model based on the cascading pyramid visual transducer, S2, collecting time-frequency image data containing various inverter running states to obtain a time-frequency image data set, S3, training the inverter degradation fault diagnosis model by adopting the time-frequency image data set, S4, deploying the trained inverter degradation fault diagnosis model and the time-frequency image conversion module in an inverter, and diagnosing the inverter degradation fault in real time. The diagnosis method can accurately identify and decouple the complex parameter faults of the inverter through the association of the strong nonlinear mapping and the multi-scale characteristics, and improves the operation reliability and maintenance efficiency of the power electronic system.

Inventors

WANG LI
Shen Caoxin

Assignees

南通大学

Dates

Publication Date: 20260512
Application Date: 20260306

Claims (6)

1. A method for diagnosing an inverter degradation fault, the method comprising: s1, taking a pyramid visual transducer as a main network, fusing an attention module and a KAN network layer module by internal cross-scale characteristics, and constructing an inverter degradation fault diagnosis model based on the cascaded pyramid visual transducer; s2, acquiring time-frequency image data containing various inverter running states to obtain a time-frequency image data set; S3, training an inverter degradation fault diagnosis model by adopting the time-frequency image data set; s4, deploying the trained inverter degradation fault diagnosis model and the time-frequency image conversion module in the inverter, and diagnosing the inverter degradation fault in real time; The time-frequency image conversion module is used for converting output voltage signals acquired in real time into time-frequency images which can be input into the inverter degradation fault diagnosis model; the method for converting the output voltage signal acquired in real time into the time-frequency image which can be input into the inverter degradation fault diagnosis model comprises the following specific steps: mapping the output voltage signal to a time-frequency domain using a continuous wavelet transform, introducing a wavelet basis function Is the conjugate function of (2) At different scale factors And displacement factor Performing convolution operation to generate complex wavelet coefficient matrix : ; Wherein the method comprises the steps of Is a matrix of complex wavelet coefficients, Is the scale factor of the scale factor, Is a displacement factor which is a function of the displacement, Is a wavelet basis function Is a conjugate function of (2); calculating complex wavelet coefficient matrix with respect to displacement factor Obtaining the corresponding instantaneous frequency estimation value of each scale component ; ; Wherein the method comprises the steps of For the instantaneous frequency estimate, Is complex wavelet coefficient matrix to displacement factor Is a partial derivative of (2); estimating the instantaneous frequency Synchronous extrusion treatment is carried out to lead the energy scattered on the scale axis to be along the frequency axis to be centered at the frequency Synchronous extrusion rearrangement is carried out on the positions to obtain a synchronous extrusion coefficient matrix ; ; Wherein the method comprises the steps of In order to synchronize the extrusion coefficient matrix, For the width of the frequency interval, As a result of the center frequency, Is the scale factor Representing the scale sampling step size; adopts cascade structure pair Synchronous extrusion coefficient matrix with different conversion times Performing linear superposition or feature fusion to generate a comprehensive time-frequency characterization matrix ; ; Wherein J is the conversion times, T is the synchronous extrusion operator, Is the wavelet coefficient obtained in the jth transformation, Is the scale factor at the j-th transition, Is the translation factor at the jth transition; combining time-frequency characterization matrix Mapping the normalized pixel values into a two-dimensional pixel matrix to obtain a single-channel gray image matrix, wherein the pixel values of the single-channel gray image matrix correspond to the normalized energy intensity; converting the single-channel gray image matrix into color time-frequency image data by using a pseudo-color mapping technology, wherein the color codes energy information; the inverter degradation fault diagnosis model further comprises a classification module, the classification module performs dimension reduction processing on the feature matrix output by the KAN network layer module through a global average pooling layer, extracts a global feature vector, then sends the global feature vector into a full connection layer to perform linear transformation, outputs a classification matrix y, processes the classification matrix by using a Softmax normalization exponential function, calculates probability values of the inverter to be diagnosed belonging to each preset fault category, and takes the fault category with the largest probability value as a final inverter parameter fault diagnosis result, wherein the Softmax normalization exponential function is as follows: ; Wherein, the The inverter to be diagnosed is identified as the first The probability value of the class fault, K is the total number of preset inverter fault classes, Outputting feature vectors for full connection layer The first of (3) The value of each element, representing the corresponding value of the first element A fault-like feature score; the classification matrix y is obtained according to the following formula: ; Wherein, the Is a weight matrix of the full connection layer, Is a global feature vector that is used to determine the feature vector, Is a bias term.
2. The diagnostic method of claim 1 wherein the pyramid visual transducer dynamically adjusts the feature map resolution of each stage according to the energy distribution in the input time-frequency image using an adaptive downscaling mechanism; the adaptive reduction mechanism is shown as follows: ; Wherein R i is the actual space reduction ratio in the ith stage, beta i is the super parameter in the ith stage, the m value is a preset parameter, and m is more than or equal to 1 and less than or equal to 8.
3. The diagnostic method of claim 1, wherein the mechanism of the cross-scale feature fusion attention module is specifically: The time-frequency characteristic diagram X s−1 in the previous stage is processed by downsampling to obtain a characteristic diagram Then, through bilinear upsampling and 1X1 convolution, the time-frequency characteristic diagram X s of the current stage is aligned with the time-frequency characteristic diagram X fuse of the current stage in terms of resolution and channel number, and fusion calculation is carried out to obtain a fusion time-frequency characteristic diagram X fuse ; the calculation formula of bilinear upsampling and 1x1 convolution is shown as follows: ; In the formula, Is 1 The convolution of 1, Is bilinear upsampling; The formula of the fusion calculation is shown as follows: , ; Wherein X fuse is a fused time-frequency characteristic diagram, alpha is a weight, sigma represents a Sigmoid function, MLP is a multi-layer perceptron, as well as element-by-element multiplication, common characteristics are enhanced, and GAP represents global average pooling; Inputting the fused time-frequency characteristic diagram X fuse into an improved multi-head attention mechanism to perform cross-scale characteristic fused attention module calculation; Wherein the expression of the improved multi-head attention mechanism is: , , , , Wherein Q, K, V represent query, key and value matrices, respectively, W Q 、W K and W V are corresponding weight matrices, attention (Q, K, V) is an improved multi-headed Attention mechanism, d h is the dimension of the key vector, and h is the number of heads.
4. The diagnostic method of claim 1, wherein the output of the KAN network layer module is: ; Wherein, the The j-th nonlinear activation function of the ith node of the KAN network layer of the first layer is L, wherein L is the layer number of the KAN network layer, and the nonlinear activation function is a univariate continuous function and is obtained by linear combination of a base function and a spline function.
5. The diagnostic method of claim 1, wherein the plurality of inverter operating states are operating states with different degrees of component degradation.
6. The diagnostic method of claim 5, wherein the plurality of inverter operating states are operating states at different levels of input filter capacitance and output filter capacitance degradation.

Description

Inverter degradation fault diagnosis method Technical Field The invention belongs to the technical field of fault diagnosis, and particularly relates to a method for diagnosing degradation faults of an inverter. Background The power electronic system is continuously developed to the high frequency, integration and complex working condition directions, and the operation environment of the traditional inverter is increasingly complex. The traditional fault diagnosis multi-dependence Fourier transform, wavelet packet decomposition and the like frequency analysis method gradually show limitations. However, fourier transform cannot achieve both time domain localization and frequency domain resolution of transient faults, while wavelet transform can extract multi-scale features, but it is difficult to adaptively separate coupled harmonics and noise due to fixed basis function limitations. In addition, the artificial feature engineering relies on priori knowledge, has low sensitivity to recessive harmonic offset caused by capacitor parameter degradation, is easily interfered by high-frequency switch noise, causes insufficient feature robustness, and is difficult to adapt to multiple fault concurrency scenes under complex working conditions. The advantages of the Convolutional Neural Network (CNN), the transducer and other depth models in the aspect of automatic feature extraction enable the detection effect to be improved to a certain extent, but under the complex electromagnetic environment and multi-fault coupling background of the inverter, the existing depth model structure still has structural bottlenecks. Although the existing depth model (CNN, transformer) can automatically extract fault characteristics, the local convolution operation of CNN is difficult to model long Cheng Xiebo association, and the global self-attention calculation of a standard transducer is high in complexity and insufficient in capturing local high-frequency transient characteristics. In addition, the traditional multi-layer perceptron relies on a fixed activation function, has limited modeling capability on complex nonlinear mapping between capacitance parameters and harmonics, and is easy to cause model confusion due to characteristic coupling especially under concurrent faults. A diagnosis architecture that combines light weight, multi-scale sensing and nonlinear decoupling is needed to break through the performance bottleneck of the existing depth model in practical engineering application. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides a diagnosis method for the degradation faults of an inverter, which can accurately identify and decouple the complex parameter faults of the inverter through the association of strong nonlinear mapping and multi-scale characteristics and improve the operation reliability and maintenance efficiency of a power electronic system. In a first aspect of the present invention, there is provided a diagnostic method of an inverter degradation fault, the diagnostic method comprising: s1, taking a pyramid visual transducer as a main network, embedding a cross-scale feature fusion attention module and a KAN network layer module, and constructing an inverter degradation fault diagnosis model based on the cascading pyramid visual transducer; s2, acquiring time-frequency image data containing various inverter running states to obtain a time-frequency image data set; S3, training an inverter degradation fault diagnosis model by adopting the time-frequency image data set; s4, deploying the trained inverter degradation fault diagnosis model and the time-frequency image conversion module in the inverter, and diagnosing the inverter degradation fault in real time; the time-frequency image conversion module is used for converting output voltage signals acquired in real time into time-frequency images which can be input into the inverter degradation fault diagnosis model. In some embodiments of the present invention, the pyramid visual transformer dynamically adjusts the resolution of the feature map at each stage according to the energy distribution in the input time-frequency image by adopting an adaptive reduction mechanism; the adaptive reduction mechanism is shown as follows: Wherein R i is the actual space reduction ratio in the ith stage, beta i is the super parameter in the ith stage, the m value is a preset parameter, and m is more than or equal to 1 and less than or equal to 8. In some embodiments of the present invention, the mechanism of the cross-scale feature fusion attention module is specifically: The time-frequency characteristic diagram X s−1 in the previous stage is processed by downsampling to obtain a characteristic diagram Then, through bilinear upsampling and 1X1 convolution, the time-frequency characteristic diagram X s of the current stage is aligned with the time-frequency characteristic diagram X fuse of the current stage in terms of resolutio