CN-122020280-A - Converter commutation failure diagnosis method and system based on Glow and neural network model

CN122020280ACN 122020280 ACN122020280 ACN 122020280ACN-122020280-A

Abstract

The invention provides a converter commutation failure diagnosis method based on a Glow and neural network model, which belongs to the field of phase control converters and converts output voltage into two-dimensional image data ; Will Inputting an improved Glow module, and obtaining an image after compression , Activating the normalization layer to obtain a sample , Through LU parameterization reversible 1X 1 convolution and reversible residual block transformation, output The incoming nerve spline coupling is split into , , Keep it unchanged, Reversible mapping by rational quadratic spline function, outputting Decompressing to obtain variable in potential space, and reversely passing data in potential space through improved Glow module to obtain enhanced image ; Use And Training a neural network to obtain a fault diagnosis model, inputting two-dimensional image data into the model to obtain prediction probabilities of various fault states, and providing a diagnosis system to solve the problem of low commutation failure diagnosis accuracy under the condition of limited samples or uneven distribution.

Inventors

CHEN XIAOJIAO
WANG ZEJING
XUE ZHENYU
HUANG LIANSHENG
ZHANG XIUQING
HE SHIYING
DOU SHENG
ZUO YING
SHEN YAN
LI LINGPENG

Assignees

中国科学院合肥物质科学研究院

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. The converter commutation failure diagnosis method based on the Glow and neural network model is characterized by comprising the following steps: converting the output voltage of converter operation into two-dimensional image data ; To two-dimensional image data Inputting an improved Glow module, wherein the improved Glow module comprises a compression layer and N Flow modules, each Flow module comprises an activation normalization layer, an LU parameterized reversible 1×1 convolution, a reversible residual block, nerve spline coupling and decompression, and two-dimensional image data, which are connected in sequence The image is obtained after compression by the compression layer Image of The sample is obtained after activating and normalizing the layer Samples of The LU parameterization reversible 1 multiplied by 1 convolution operation is carried out to obtain output Output of The output is obtained after the reversible residual block transformation Output of The incoming nerve spline coupling is split into , , The temperature of the liquid crystal is kept unchanged, Reversible mapping is carried out through a rational quadratic spline function to obtain output Output of The variable Z in the potential space Z is obtained after decompression processing, and the data sampled from the potential space Z reversely passes through the improved Glow module to obtain an enhanced image ; Using two-dimensional image data Enhancing images Training a PCNN-GRU-ECA neural network by using the composed data set to obtain a fault diagnosis model; The method comprises the steps of collecting output voltage of the operation of the converter, converting the output voltage into two-dimensional image data, inputting the two-dimensional image data into a fault diagnosis model, obtaining prediction probability of various fault states, and judging the fault type based on the prediction probability.
2. The method for diagnosing converter commutation failure based on Glow and neural network model as recited in claim 1, wherein the output voltage of converter operation is converted into two-dimensional image data The process of (1) comprises: normalizing the output voltage of the current transformer operation to obtain normalized data; Mapping the normalized data to an angle domain to obtain a one-dimensional angle sequence; Combining angles at different moments in a one-dimensional angle sequence in pairs, calculating angle difference, and encoding by adopting a sine function to obtain a two-dimensional GADF matrix reflecting the correlation between the moments; inputting a two-dimensional GADF matrix into an image processing program for visualization, and obtaining two-dimensional image data 。
3. The method for diagnosing converter commutation failure based on Glow and neural network model as recited in claim 1, wherein the method is characterized by comprising the following steps of The LU parameterization reversible 1×1 convolution operation comprises the steps of sampling Is defined by the spatial location pixel of (a) Corresponds to a channel vector For each spatial position pixel Performing linear transformation, and pixels after linear transformation The method comprises the following steps: Weight matrix Is decomposed into: Wherein P is a fixed permutation matrix, and the unit lower triangular matrix L and the unit upper triangular matrix U are respectively: ; is the actual scale factor; the corresponding logarithmic jacobian is: 。
4. The method for diagnosing converter commutation failure based on Glow and neural network model as recited in claim 1, wherein the reversible residual block pair outputs The processing procedure of (1) is that output As input features Provided with input features Divided into two parts in the channel dimension: the above transform is defined as: ; Wherein, the , For a nonlinear mapping function parameterized by a convolutional neural network, L is the number of convolutional layers, Is the first The convolution of the layers incorporates a nonlinear operator, Is a neural network parameter; Output of reversible residual block The method comprises the following steps: 。
5. the Glow and neural network model-based converter commutation failure diagnosis method of claim 1, wherein the method is characterized by: The temperature of the liquid crystal is kept unchanged, The process of reversible mapping by the rational quadratic spline function is as follows: Wherein, the Representing a rational quadratic spline transformation function, For the piecewise quadratic term coefficient, For the segmentation of the first order term coefficients, For constant term offset, spline functions are formed by The number of segments is constituted, and the parameter set is as follows: Wherein, the , 、、 Respectively the first Spline width, spline height, spline slope of each segment, and satisfies constraint conditions: final coupling mapping: The corresponding logarithmic jacobian is: Wherein, the For transformed features Is used for the number of channel dimensions.
6. The converter commutation failure diagnosis method based on a Glow and neural network model of claim 1, wherein the PCNN-GRU-ECA neural network comprises a feature extraction module, a feature fusion module, a GRU module and an ECA module which are sequentially connected, the feature extraction module comprises a first CNN branch, a second CNN branch and a third CNN branch which are connected in parallel, and the first CNN branch comprises 3 CNN branches which are sequentially connected 3 Convolutional layer, average pooling layer, 5 5 Convolution layers and an average pooling layer, wherein the second CNN branch comprises 3 connected in sequence 3 Convolutional layer, max pooling layer, 5 5 Convolution layers, a maximum pooling layer, and a third CNN branch comprising 3 connected in sequence 3 Convolutional layer, mix pooling layer, 5 And 5, a convolution layer and a mixed pooling layer, wherein the first CNN branch, the second CNN branch and the third CNN branch respectively extract characteristics of the input image, the characteristics output by the three CNN branches enter a characteristic fusion module together to perform characteristic fusion, the fused characteristics enter a GRU module to perform sequence modeling, the output characteristics of the GRU module are input to an ECA module, the ECA module performs self-adaptive weight distribution on the channel characteristics to obtain a weighted characteristic diagram, the weighted characteristic diagram enters a classifier to obtain the prediction probability of various fault states, and the output end of the characteristic fusion module and the input end of the classifier are respectively added with a Dropout structure.
7. The method for diagnosing a commutation failure of a converter based on a Glow and neural network model of claim 6, wherein the output characteristics of the GRU module are as follows: Wherein, the 、 The output characteristics of the GRU modules at the time t and the time t-1 are respectively, In order to update the door(s), , In order to select the gate(s), , In order to reset the gate, , For the feature fused at time t, the point multiplication operation is represented, σ is a sigmoid function, and Wr, wz and Wy are coefficient weights of each layer.
8. The converter commutation failure diagnosis method based on the Glow and neural network model of claim 6, wherein the output characteristics of the GRU module are input to the ECA module, the ECA module comprises global average pooling, one-dimensional convolution layer and channel weighting which are sequentially connected, the global average pooling is carried out on the output characteristics of the GRU module according to space dimension to obtain global description vectors of each channel, the one-dimensional convolution layer convolves the global description vectors to output attention weights of each channel, and the channel weighting multiplies the attention weights by the output characteristics of the GRU module channel by channel to obtain a weighted characteristic diagram.
9. The method for diagnosing a commutation failure of a converter based on a Glow and neural network model as recited in claim 1, wherein the two-dimensional image data is used for Enhancing images The process of training the PCNN-GRU-ECA neural network by the composed data set comprises the following steps: To two-dimensional image data Enhancing images The composed data set is divided into a training set and a verification set according to a preset proportion, a training strategy and network parameters are set, wherein, an optimizer adopts AdamW self-adaptive updating method, the initial learning rate is set to be 0.0015, And Respectively 0.9 and 0.999, and gradually reducing the learning rate in the training process by combining a cosine annealing strategy; In two-dimensional image data Or enhance images The method comprises the steps of taking a fault state as a real label, training a PCNN-GRU-ECA neural network in various fault states including bridge arm short circuit, inversion failure, pulse loss, direct current side short circuit and normal state, adopting Batch size of 32 and total 100 epochs in the whole training process, continuously updating network parameters through calculating the loss between predicted output and the real label, and obtaining the trained PCNN-GRU-ECA neural network when the loss is minimum or reaches a set training round; Inputting the verification set into the trained PCNN-GRU-ECA neural network for retraining, continuously recording the accuracy and stability indexes of each iteration model on the verification set in the training process, comprehensively judging the optimal iteration point, and fixing and storing model parameters corresponding to the optimal iteration point to obtain the fault diagnosis model.
10. The converter commutation failure diagnosis system based on the Glow and neural network model is characterized by comprising the following components: the data processing module is used for converting the output voltage of the converter operation into two-dimensional image data ; An image enhancement module for enhancing two-dimensional image data Inputting an improved Glow module, wherein the improved Glow module comprises a compression layer and N Flow modules, each Flow module comprises an activation normalization layer, an LU parameterized reversible 1×1 convolution, a reversible residual block, nerve spline coupling and decompression, and two-dimensional image data, which are connected in sequence The image is obtained after compression by the compression layer Image of The sample is obtained after activating and normalizing the layer Samples of The LU parameterization reversible 1 multiplied by 1 convolution operation is carried out to obtain output Output of The output is obtained after the reversible residual block transformation Output of The incoming nerve spline coupling is split into , , The temperature of the liquid crystal is kept unchanged, Reversible mapping is carried out through a rational quadratic spline function to obtain output Output of The variable Z in the potential space Z is obtained after decompression processing, and the data sampled from the potential space Z reversely passes through the improved Glow module to obtain an enhanced image ; A learning module for using two-dimensional image data Enhancing images Training a PCNN-GRU-ECA neural network by using the composed data set to obtain a fault diagnosis model; The reasoning module is used for collecting the output voltage of the current transformer operation, converting the output voltage into two-dimensional image data, inputting the two-dimensional image data into the fault diagnosis model, obtaining the prediction probability of various fault states, and judging the fault type based on the prediction probability.

Description

Converter commutation failure diagnosis method and system based on Glow and neural network model Technical Field The invention relates to the technical field of phase control converters, in particular to a converter commutation failure diagnosis method and system based on a Glow and neural network model. Background The polar field (Poloidal Field, PF) power supply system is a key device for realizing magnetic confinement and plasma shape control of a Tokamak device, and has the main functions of providing a high-precision large-current pulse power supply for a polar field coil so as to realize the tasks of plasma current establishment, configuration stabilization, longitudinal field adjustment and the like. The PF power supply generally adopts a thyristor-based phase control rectification technology, and realizes accurate control of voltage and current by adjusting a trigger angle, but has the inherent defect of commutation failure. Failure of commutation can cause a large amount of induced current and force to act, threaten the integrity of the vacuum chamber structure, and cause overstress damage to the converter assembly, the power supply system and the superconducting magnet, resulting in serious safety accidents. Neural networks have become one of the key technologies in the field of fault diagnosis due to their strong feature learning ability and highly nonlinear modeling ability. Compared with the traditional method relying on artificial feature extraction, the neural network can automatically mine deep features in the original signals through end-to-end learning and has excellent advantages in fusion of multidimensional space-time information. However, neural networks still have some non-negligible limitations in fault diagnosis applications. In general, a large number of labeling samples are needed for training the neural network model, but in an actual engineering scene, fault data are often scarce, the acquisition cost is extremely high, and meanwhile, the model has low accuracy rate for diagnosing different faults due to uneven sample distribution. In the prior art, a Glow module for sample expansion in paper 'Glow-ECNN model-based small sample rolling bearing fault diagnosis method' (Liu Xiaobo et al, tai Ji university, control engineering, month 7 of 2025) adopts the following basic structural design of activating a normalization layer, 11 Convolution layer, affine transformation layer. The Glow module has the following defects that a generated sample is single in change form and extremely large in calculated data quantity, affine transformation is essentially linear scale and translation, the generated sample is mostly represented as linear disturbance of an original sample, the fitting capacity of complex data distribution is limited, complex distribution structures near multi-mode and small-probability areas or category boundaries are difficult to be described, the effective information increment of a classification model is insufficient due to the fact that the enhanced sample has limited contribution to improving the generalization capacity of the model. Disclosure of Invention The invention aims to solve the technical problem of low accuracy of phase-change failure fault diagnosis of the phase-control converter under the condition of limited samples or uneven distribution. The invention solves the technical problems through the following technical scheme that the converter commutation failure diagnosis method based on the Glow and neural network model comprises the following steps: converting the output voltage of converter operation into two-dimensional image data ; To two-dimensional image dataInputting an improved Glow module, wherein the improved Glow module comprises a compression layer and N Flow modules, each Flow module comprises an activation normalization layer, an LU parameterized reversible 1×1 convolution, a reversible residual block, nerve spline coupling and decompression, and two-dimensional image data, which are connected in sequenceThe image is obtained after compression by the compression layerImage ofThe sample is obtained after activating and normalizing the layerSamples ofThe LU parameterization reversible 1 multiplied by 1 convolution operation is carried out to obtain outputOutput ofThe output is obtained after the reversible residual block transformationOutput ofThe incoming nerve spline coupling is split into,,The temperature of the liquid crystal is kept unchanged,Reversible mapping is carried out through a rational quadratic spline function to obtain outputOutput ofThe variable Z in the potential space Z is obtained after decompression processing, and the data sampled from the potential space Z reversely passes through the improved Glow module to obtain an enhanced image; Using two-dimensional image dataEnhancing imagesTraining a PCNN-GRU-ECA neural network by using the composed data set to obtain a fault diagnosis model; The method comprises the steps of coll