Converter commutation failure diagnosis method and system based on Glow and neural network model

Abstract

The invention provides a converter commutation failure diagnosis method based on Glow and a neural network model, and belongs to the field of phase-controlled converters, and the method comprises the steps: converting an output voltage into two-dimensional image data; the method comprises the following steps: inputting into an improved Glow module, compressing to obtain an image, activating a normalization layer, obtaining a sample, performing LU parameterized reversible 1 * 1 convolution and reversible residual block transformation, splitting output into neural spline coupling, keeping unchanged, performing reversible mapping through a rational quadratic spline function, and decompressing output to obtain variables in a potential space. Reversely passing the data in the potential space through an improved Glow module to obtain an enhanced image; training a neural network to obtain a fault diagnosis model; inputting the two-dimensional image data into the model to obtain prediction probabilities of various fault states; the invention further provides a diagnosis system. The problem that the commutation failure diagnosis accuracy is low under the condition of limited samples or uneven distribution is solved.

Description

Technical Field

[0001] This invention relates to the field of phase-controlled converter technology, and in particular to a method and system for diagnosing converter commutation failure based on Glow and neural network models. Background Technology

[0002] The poloidal field (PF) power system is a key component in tokamak devices for achieving magnetic confinement and plasma shape control. Its main function is to provide high-precision, high-current pulse power to the poloidal field coils to achieve tasks such as plasma current establishment, configuration stabilization, and longitudinal field adjustment. PF power supplies typically employ thyristor-based phase-controlled rectification technology, achieving precise voltage and current control by adjusting the firing angle. However, this method inherently suffers from commutation failure. Commutation failure can induce significant induced currents and forces, threatening the integrity of the vacuum chamber structure and causing overstress damage to converter components, the power system, and the superconducting magnet, leading to serious safety accidents.

[0003] Neural networks, with their powerful feature learning and highly nonlinear modeling capabilities, have become a key technology in the field of fault diagnosis. Compared to traditional methods that rely on manual feature extraction, neural networks can automatically mine deep features from raw signals through end-to-end learning and demonstrate superior advantages in the fusion of multi-dimensional spatiotemporal information. However, neural networks still have some significant limitations in fault diagnosis applications. Typically, training neural network models requires a large number of labeled samples, while in real-world engineering scenarios, fault data is often scarce and extremely costly to collect; furthermore, uneven sample distribution leads to low accuracy in diagnosing different faults.

[0004] In the existing technology, the Glow module used for sample expansion in the paper "A Small Sample Rolling Bearing Fault Diagnosis Method Based on Glow-ECNN Model" (Liu Xiaobo et al., Taiyuan University of Technology, Control Engineering, July 2025) adopts the following basic structural design: activation normalization layer, 1 1. Convolutional layer and affine transformation layer. This Glow module has the following shortcomings: the generated sample transformation forms are limited, resulting in a huge amount of computational data; affine transformation is essentially a linear scaling and translation, and the generated samples often exhibit linear perturbations of the original samples; it has limited ability to fit complex data distributions, making it difficult to characterize complex distribution structures near multimodal, low-probability regions or class boundaries; the enhanced samples provide insufficient effective information increment to the classification model, and contribute little to improving the model's generalization ability. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to solve the problem of low accuracy in commutation failure fault diagnosis of phase-controlled converters under conditions of limited or uneven sample distribution.

[0006] This invention solves the above-mentioned technical problems through the following technical solution: a converter commutation failure diagnosis method based on Glow and neural network models, comprising: Convert the output voltage of the converter into two-dimensional image data. ; Two-dimensional image data The input is an improved Glow module, which consists of a compression layer and N Flow modules. Each Flow module includes sequentially connected activation normalization, LU-parameterized invertible 1×1 convolution, invertible residual block, neural spline coupling, and decompression of the two-dimensional image data. The image is obtained after compression by the compression layer. ,image The sample was obtained after activation and normalization. ,sample The output is obtained after a LU-parameterized reversible 1×1 convolution operation. Output The output is obtained after reversible residual block transformation. Output The neural spline coupling is split into , , remain unchanged. The output is obtained by invertible mapping using rational quadratic spline functions. Output After decompression, the variable z in the latent space Z is obtained. The data sampled from the latent space Z is then passed inversely through the improved Glow module to obtain the enhanced image. ; Using two-dimensional image data and image enhancement The PCNN-GRU-ECA neural network was trained using the dataset to obtain a fault diagnosis model. The output voltage of the converter is collected and converted into two-dimensional image data. The two-dimensional image data is then input into the fault diagnosis model to obtain the predicted probability of various fault states. The fault type is then determined based on the predicted probability.

[0007] This invention constructs a diagnostic framework based on an improved Glow generative model and a convolutional neural network. The improved Glow module utilizes reversible flow transformation to accurately model the potential distribution of the original samples. Even under actual operating conditions with only a small number of fault samples, it can still generate high-quality amplified samples that are consistent with the characteristics of real data, alleviating the model underfitting problem caused by the scarcity of fault samples. Thus, it effectively solves the problem of difficulty in modeling small samples and limited diagnostic accuracy in the fault diagnosis of commutation failure in phase-controlled converters.

[0008] Preferably, the output voltage of the converter is converted into two-dimensional image data. The process includes: The output voltage of the converter is normalized to obtain normalized data; Mapping the normalized data to the angle domain yields a one-dimensional angle sequence; By pairing angles at different times in a one-dimensional angle sequence, calculating the angle difference, and encoding it using a sine function, a two-dimensional GADF matrix reflecting the correlation between different times is obtained. The two-dimensional GADF matrix is ​​input into an image processing program for visualization; two-dimensional image data. .

[0009] This invention employs the Gram angle difference field conversion method to transform the one-dimensional output voltage signal of a converter into two-dimensional image features. Through polar coordinate angle encoding and angle difference matrix construction, the inherent correlation and differences between different time points in the voltage sequence can be fully characterized, making the trend changes, phase characteristics, and abnormal fluctuations of the original time-series signal more intuitively presented in the two-dimensional image. This representation makes fault features clearer and more prominent in the image space, which is beneficial for the accurate identification of abnormal states such as commutation failure by subsequent deep learning models, improving the accuracy and noise resistance of fault diagnosis.

[0010] Preferred, sample The process after the LU-parameterized reversible 1×1 convolution operation is as follows: Sample Each spatial location pixel Corresponding to a channel vector For each spatial location pixel Perform a linear transformation, the resulting pixels for:

[0011] weight matrix Decomposed into:

[0012] Where P is a fixed permutation matrix, and the unit lower triangular matrix L and upper triangular matrix U are respectively: ;

[0013] The actual scale factor is given; the corresponding logarithmic Jacobian determinant is: .

[0014] The key requirement for improving Glow is to compute the logarithmic Jacobian determinant more efficiently; LU decomposition can reduce the complexity from O(c3 The training parameters are reduced to O(c), and the LU decomposition can directly calculate the determinant, which facilitates the calculation of the log-likelihood and improves the stability and efficiency of training.

[0015] Preferably, the reversible residual block outputs... The processing procedure is as follows: output as input features Let the input features be... It is divided into two parts in terms of channel dimension:

[0016] The transformation of the above expression is defined as: ;

[0017] in, , Let L be the nonlinear mapping function parameterized by the convolutional neural network, where L is the number of convolutional layers. For the first Layer convolution combined with nonlinear operators, These are the parameters of the neural network.

[0018] Output of reversible residual block for: .

[0019] The addition of reversible residual blocks improves gradient flow, alleviates the gradient vanishing problem during deep network training, and enhances model stability; without increasing likelihood computational complexity, it significantly improves the model's ability to fit complex local structures.

[0020] Preferred, remain unchanged. The process of invertible mapping using rational quadratic spline functions is as follows:

[0021]

[0022]

[0023] in, This represents the rational quadratic spline transformation function. The coefficients of the piecewise quadratic term are... The coefficient of the piecewise linear term, For the constant term offset, the spline function is derived from It consists of several segments, and its parameter set is as follows:

[0024] in, , , , The first The spline width, spline height, and spline slope of each segment must satisfy the following constraints:

[0025] Final coupling mapping:

[0026] The corresponding logarithmic Jacobian determinant is:

[0027] in Features to be transformed The number of channel dimensions.

[0028] Compared to traditional affine coupling layers, neural spline coupling can represent more complex nonlinear monotonic mappings, significantly improving the model's ability to model multimodal and non-Gaussian distributions, thereby enhancing the model's ability to fit complex distributions, improving expressive power, and ensuring the stability of the generation process.

[0029] Preferably, the PCNN-GRU-ECA neural network includes a feature extraction module, a feature fusion module, a GRU module, and an ECA module connected in sequence. The feature extraction module includes a first CNN branch, a second CNN branch, and a third CNN branch connected in parallel. The first CNN branch includes three sequentially connected... 3 convolutional layers, average pooling layer, 5 5 convolutional layers, average pooling layers, and the second CNN branch includes 3 sequentially connected layers. 3 convolutional layers, max pooling layer, 5 Five convolutional layers, a max pooling layer, and the third CNN branch consists of three sequentially connected layers. 3 convolutional layers, hybrid pooling layers, 5 The system consists of 5 convolutional layers and hybrid pooling layers. The first, second, and third CNN branches extract features from the input image. The features output from the three CNN branches are then fed into a feature fusion module for feature fusion. The fused features are then fed into a GRU module for sequence modeling. The output features of the GRU module are fed into an ECA module, which performs adaptive weight allocation on the channel features to obtain a weighted feature map. The weighted feature map is then fed into a classifier to obtain the predicted probabilities of various fault states. Dropout structures are added to the output of the feature fusion module and the input of the classifier.

[0030] The feature extraction module adopts a parallel convolutional structure, which can capture more feature details in the same stage and adapt to the different feature differences exhibited by different types of faults in the time series. With the help of the GRU module to model the dynamic evolution of the voltage sequence in the time dimension, the channel recalibration of the ECA module further enhances the ability to express key fault features, thereby significantly improving the accuracy and noise resistance of commutation failure fault diagnosis. Under the premise of ensuring the real-time performance of the model, it can accurately identify multiple fault states, thereby improving the safety and reliability of the fusion magnet power system during operation.

[0031] Preferably, the output characteristics of the GRU module are:

[0032] in, , The output characteristics of the GRU module at time t and time t-1 are respectively. To update the door, , To select a door, , To reset the door, , The features are fused at time t, and · represents the dot product operation. σ Here, Wr, Wz, and Wy are the coefficients and weights of each layer, representing the sigmoid function.

[0033] Preferably, the output features of the GRU module are input into the ECA module. The ECA module includes a globally average pooling layer, a one-dimensional convolutional layer, and channel weighting, connected sequentially. The globally average pooling layer performs global average pooling on the output features of the GRU module along the spatial dimension to obtain a global description vector for each channel. The one-dimensional convolutional layer convolves the global description vector to output the attention weight for each channel. Channel weighting multiplies the attention weights with the output features of the GRU module channel by channel to obtain a weighted feature map. Through the ECA mechanism, the network can automatically enhance the response to key channels while suppressing irrelevant or redundant channels, thereby improving feature representation capabilities. The ECA module can be effectively integrated into existing network architectures without significantly increasing model complexity.

[0034] Preferably, two-dimensional image data is used. and image enhancement The process of training the PCNN-GRU-ECA neural network using the dataset includes: Two-dimensional image data and image enhancement The dataset was divided into training and validation sets according to a preset ratio. Training strategies and network parameters were set, with the AdamW adaptive update method used as the optimizer and an initial learning rate of 0.0015. and The values ​​are 0.9 and 0.999 respectively, and the learning rate is gradually reduced during training using a cosine annealing strategy. Two-dimensional image data Or enhance the image Using the fault state as the real label as the input, the PCNN-GRU-ECA neural network is trained. The entire training process uses a batch size of 32 and a total of 100 epochs. The network parameters are continuously updated by calculating the loss between the predicted output and the real label. When the loss is minimized or the set number of training epochs is reached, the trained PCNN-GRU-ECA neural network is obtained. The validation set is input into the trained PCNN-GRU-ECA neural network for retraining. During the training process, the accuracy and stability of each iterative model on the validation set are continuously recorded, and the optimal iteration point is determined by comprehensive analysis. The model parameters corresponding to the optimal iteration point are fixed and saved to obtain the fault diagnosis model.

[0035] Preferably, the various fault states include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit, and normal state. The fault type is determined based on the predicted probability as follows: when the predicted probability of the normal state is the highest, it is determined whether the predicted probability of the normal state reaches the threshold. If it does, it is determined to be a normal state; if not, it is determined to be a potential anomaly. When the predicted probability of any of the fault states, such as bridge arm short circuit, inverter failure, pulse loss, and DC side short circuit, is the highest, it is directly determined to be the fault state corresponding to the highest predicted probability.

[0036] This invention also provides a converter commutation failure diagnosis system based on Glow and a neural network model, comprising: The data processing module is used to convert the output voltage of the converter into two-dimensional image data. ; Image enhancement module, used to enhance two-dimensional image data The input is an improved Glow module, which consists of a compression layer and N Flow modules. Each Flow module includes sequentially connected activation normalization, LU-parameterized invertible 1×1 convolution, invertible residual block, neural spline coupling, and decompression of the two-dimensional image data. The image is obtained after compression by the compression layer. ,image The sample was obtained after activation and normalization. ,sample The output is obtained after a LU-parameterized reversible 1×1 convolution operation. Output The output is obtained after reversible residual block transformation. Output The neural spline coupling is split into , , remain unchanged. The output is obtained by invertible mapping using rational quadratic spline functions. Output After decompression, the variable z in the latent space Z is obtained. The data sampled from the latent space Z is then passed inversely through the improved Glow module to obtain the enhanced image. ; Learning module for using two-dimensional image data and image enhancement The PCNN-GRU-ECA neural network was trained using the dataset to obtain a fault diagnosis model. The inference module is used to collect the output voltage of the converter and convert it into two-dimensional image data. The two-dimensional image data is then input into the fault diagnosis model to obtain the predicted probability of various fault states, and the fault type is determined based on the predicted probability.

[0037] This invention improves the Glow module by introducing three key improvements on the standard structure: LU-parameterized reversible 1×1 convolution, reversible residual blocks, and neural spline coupling layers. The 1×1 reversible convolution is parameterized using LU decomposition, ensuring that determinant calculation is only related to the diagonal elements, thus reducing computational complexity. Reduce to While ensuring sufficient channel mixing, the training stability of the overall Glow module is improved. Inverse residual blocks are introduced, incorporating deep nonlinear mapping within the channel partitioning structure. An additive structure ensures overall invertibility without increasing the computational burden of the Jacobian determinant. The inverse residual blocks ensure that features undergo a deep nonlinear reconstruction before entering the coupling layer, thereby increasing the degree of freedom in the feature space, expanding the range of variations in generated samples, and preventing generated samples from being merely slight perturbations of the original samples. Neural spline coupling achieves fine modulation of local probability density through learnable piecewise nonlinear functions, effectively covering the distribution tail and class boundary regions, resulting in more realistic and diverse sample generation. Compared to using activation normalization layers, 1 The improved Glow module of this invention, which includes convolutional and affine transformation layers, generates high-quality samples with low computational cost. Attached Figure Description

[0038] Figure 1 The topology diagram of the PF poloidal field power supply system; Figure 2 This is a schematic diagram of the principle of generating new samples in the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 3This is a schematic diagram of the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 4 This is a schematic diagram of the working principle of the improved Glow module in the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 5 The schematic diagram of PCNN in the PCNN-GRU-ECA neural network of the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention; Figure 6 The flowchart shows the training process of the PCNN-GRU-ECA neural network in the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 7 The accuracy curve of PCNN-GRU-ECA neural network training in the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention; Figure 8 The loss function curve of PCNN-GRU-ECA neural network training in the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention; Figure 9 The diagram shows the bridge arm short-circuit fault diagnosis results using the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 10 The figure shows the DC-side short-circuit fault diagnosis results of the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 11 The diagram shows the inverter failure diagnosis results using the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 12 The image shows the pulse loss fault diagnosis result of the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention. Figure 13 This is a normal state result diagram of the converter commutation failure diagnosis method based on Glow and neural network model provided in Embodiment 1 of the present invention; Figure 14 The flowchart shows the process of the converter commutation failure diagnosis system based on Glow and neural network model provided in Embodiment 2 of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0040] Example 1 The EAST device's poloidal field (PF) power system employs high-power thyristor phase-controlled converter technology to achieve precise bidirectional current regulation of the superconducting PF coils. Each PF power supply consists of four three-phase fully controlled bridge converter units connected together, with multi-bridge coordinated operation enhancing output current capability and regulation range. The poloidal field power supply topology is shown below. Figure 1 As shown, a single three-phase fully controlled converter unit consists of six high-power thyristors (VT1-VT6) forming a typical three-phase fully controlled bridge. The six thyristors are arranged in pairs according to a bridge structure: VT1 and VT4 form phase A, VT3 and VT6 form phase B, and VT5 and VT2 form phase C. The upper and lower thyristors in each bridge arm complement each other, realizing energy conversion between AC input and DC output. Three-phase AC power supplies Ua, Ub, and Uc are fed into the midpoint of each bridge arm, and the other end of the power supply is connected to ground. Since the PF coil is a strongly inductive load, a large inductor is connected in series at the power output terminal, making the DC current a continuous waveform. By adjusting the firing angle of the thyristors, the sign and amplitude of the DC output voltage can be controlled, thereby realizing the rise, maintenance, fall, and rapid reverse of the coil current.

[0041] like Figure 3 and Figure 4 As shown, this embodiment provides a converter commutation failure diagnosis method based on Glow and a neural network model, including the following steps: Step 1: Continuously record the converter's operating status during commutation using a voltage acquisition device. Acquire voltage change data under normal operating conditions and various typical fault conditions, including bridge arm short circuit, inverter failure, trigger pulse loss, and DC-side short circuit. To ensure clear capture of rapid dynamic details during commutation, the sampling frequency of the acquisition system is set to 15kHz to obtain sufficient density of time-series data. The acquired voltage signals are stored as a one-dimensional discrete sequence in chronological order, with each sampling point corresponding to the real-time voltage amplitude. This comprehensively reflects the transient evolution of the entire commutation process, converting the converter's operating output voltage into two-dimensional image data. .

[0042] The converter's output voltage is a real-time operating voltage signal. To achieve continuous storage and subsequent online analysis of the collected data, this embodiment employs a hierarchical storage strategy based on a "dual-port buffer + address counter + storage control state machine," the framework of which is as follows: Figure 14 As shown, firstly, the ADC samples the converter output voltage at a set sampling frequency, and the resulting one-dimensional discrete signal is directly written into a dual-port buffer established on the controller. This buffer consists of two dual-port RAMs of the same capacity, denoted as RAM_A and RAM_B, which take turns serving as the buffer during acquisition and the buffer to be written.

[0043] The acquisition buffer is driven by the ADC's dedicated write address counter. When the amount of data in RAM_X (X=A or B) reaches the preset segment length (5000 sampling points), the acquisition control logic immediately switches the buffer role: RAM_X is marked as "buffer to be written", and another RAM is cleared and used as a new acquisition buffer to continue receiving subsequent sampling points. The entire switching process is completed within one sampling cycle.

[0044] For the RAM marked as "to be written to buffer," the "external storage control state machine" within the controller initiates the write task: the state machine uses another port of the RAM to sequentially read 2048 sampling points, and generates row addresses, column addresses, and write enable signals according to the SDRAM address mapping rules, storing the data into the SDRAM in a sequential write manner. During the write process, the ADC acquisition is completely independent of this process because the acquisition channel has been switched to another RAM, achieving complete parallelism between acquisition and data writing.

[0045] This involves converting the output voltage of the converter into two-dimensional image data. The process includes: Step 1.1: Normalize the output voltage of the converter to obtain normalized data. The acquired converter output voltage is a one-dimensional discrete-time series. After normalizing the original sampled data, one-dimensional signal data is obtained. First, the series is normalized using min-max normalization:

[0046] in, This represents the voltage value at time i. For the normalized number, , The smallest number in the sequence. It is the largest number in the sequence.

[0047] All original voltage sequences are subjected to uniform normalization preprocessing to ensure that different data segments are consistent in magnitude and scale, thereby avoiding the impact of sampling device differences or external interference on subsequent feature extraction.

[0048] Step 1.2: Normalize the data Mapping to the angle domain yields a one-dimensional angle sequence. :

[0049] thereby, ,

[0050] Step 1.3: Combine the angles at different times in the one-dimensional angle sequence into pairs, calculate the angle difference, and encode it using a sine function to obtain a two-dimensional GADF (Gramian Angular Difference Field) matrix that reflects the correlation between each time point.

[0051] Define a two-dimensional GADF matrix as follows: ,in, , .Will After expansion, we get the algebraic form: This allows us to create a Gram matrix GADF(X) for the time series X.

[0052] Step 1.4: Input the two-dimensional GADF matrix into the image processing program for visualization to obtain the corresponding two-dimensional image representation, i.e., the two-dimensional image data. This is used for subsequent network feature extraction and fault diagnosis. Through normalization, angle mapping, and Gram operations, the time series is converted into a GADF matrix and further rendered as an image, so that the phase difference relationship of the time series can be input into the model for analysis in the form of a texture pattern.

[0053] Based on the preprocessed one-dimensional voltage signal dataset, each one-dimensional sequence is further converted into two-dimensional image data that reflects its internal temporal correlation characteristics. Specifically, each equal-length one-dimensional voltage segment is first subjected to amplitude normalization and scale normalization to compress its waveform variations to a uniform range. Subsequently, the normalized sequence is mapped to the polar coordinate angular domain, and a two-dimensional difference relation matrix reflecting temporal dependence is constructed by establishing the angular difference relationship between sequence points. After GADF transformation, this matrix forms a two-dimensional representation image with clear structural texture. GADF images can present the correlation patterns, abrupt change locations, and dynamic characteristics of commutation behavior of one-dimensional signals in a two-dimensional plane, exhibiting stronger visualization and feature separability compared to the original voltage waveform. The converted two-dimensional image data is matched one-to-one with its corresponding fault category and used as training samples for subsequent two-dimensional channel neural networks, providing richer spatial feature information for fault identification.

[0054] This invention employs the GADF conversion method to map one-dimensional signal data into a two-dimensional image. Through this conversion process, the original time-varying one-dimensional waveform is expressed as a two-dimensional image with texture and structure distribution. This not only preserves the overall trend and temporal dependence of the signal but also enhances the pattern differences under different operating conditions in the image domain. This facilitates the extraction of more discriminative features by convolutional neural networks, thereby improving the accuracy and anti-interference capability of commutation failure fault identification. It should be noted that the fault types in this invention include bridge arm short circuit (BS), inverter failure (IF), pulse loss (PL), DC side short circuit (DS), and normal state.

[0055] Step 2: Transfer the two-dimensional image data The improved Glow module is used to perform data augmentation, generating a two-dimensional image with a feature distribution consistent with the original sample. See also... Figure 4 The improved Glow module mainly consists of a forward process and a backward process, with a reversible network at its core. The forward process maps the input image to the latent space, while the backward process reconstructs the image from the latent space. This invention constructs an improved Glow reversible generative model. The improved Glow module includes a compression layer and N Flow modules. Each Flow module includes, in sequence, an activation normalization layer, an LU-parameterized reversible 1×1 convolution, a reversible residual block, neural spline coupling, and decompression. The compression layer downsamples the input image, compressing the spatial dimension. The activation normalization layer performs learnable normalization on each channel. The LU-parameterized reversible 1×1 convolution performs a reversible linear transformation on the channels, achieving channel blending. The reversible residual block facilitates gradient flow and stable training. Neural spline coupling enables nonlinear transformations.

[0056] Two-dimensional image data The image is obtained after compression by the compression layer. ,image The sample was obtained after activation and normalization. ,sample The output is obtained after a LU-parameterized reversible 1×1 convolution operation. Output The output is obtained after reversible residual block transformation. Output The neural spline coupling is split into , , remain unchanged. The output is obtained by invertible mapping using rational quadratic spline functions. Output After decompression, the variable z in the latent space Z is obtained. The data sampled from the latent space Z is then passed inversely through the improved Glow module to obtain the enhanced image. .

[0057] Glow is a stream-based generative model designed to accurately capture the probability distribution of input samples. The core idea of ​​using an improved Glow module for small-sample augmentation is to map a complex image distribution to a simple latent space distribution through a series of invertible transformations, and then generate new images from the latent variable z through an inverse transformation.

[0058] ;

[0059] in, The input GADF image data is the two-dimensional image data of this invention. , It is an invertible transformation function. The image data generated by the reverse mapping, i.e., the enhanced image of the present invention. .

[0060] Assuming two-dimensional image data The probability density is The prior probability density of the latent variable z is Then the probability density and prior probability density satisfy:

[0061] Invertible transformation function Two-dimensional image data From probability density Mapping to prior probability density The formula is as follows:

[0062] in, yes In two-dimensional image data The Jacobian matrix at point , in log-likelihood form, is:

[0063] in, It is the Jacobian determinant. In normalized flow models, the Jacobian matrix explicitly characterizes how an invertible transformation locally alters the input space. It is crucial for accurate likelihood calculations, guarantees the invertibility of the transformation, and provides a quantitative measure of the volume change caused by the transformation.

[0064] After accurately capturing the probability density of the input image, an inverse transformation is performed. This allows the generation of a new image with the same features as the input image.

[0065] Within the generative flow modeling framework, this invention constructs an improved Glow reversible generative model, outputting... After sequentially passing through reversible residual blocks, neural spline coupling, and decompression, the variable z in the latent space Z is obtained. Through a series of reversible transformations with analytic Jacobian determinants, a bidirectional mapping from the image space to the latent space is achieved. Let the input image be... First, the input image undergoes compression to rearrange the spatial and channel dimensions. The compression operation changes the spatial and channel dimensions from H×W×C to... × ×4C. The mathematical expression is:

[0066] in, , Since this process only involves the permutation of tensor elements, its Jacobian matrix is ​​the permutation matrix, thus satisfying the determinant det(J). sq )=1, log|det(J) sq )|=0.

[0067] In the compressed feature space, an activation normalization layer is introduced to perform linear scaling correction on each channel. Let x sq Given the input features, their forward transform can be written as:

[0068] s is a learnable parameter, and b is a bias term. This represents channel-wise multiplication. The corresponding Jacobian matrix is ​​a diagonal block matrix with the following determinant:

[0069] Where h is w is , where c is the channel dimension, which is 4C in this case.

[0070] Therefore, the logarithmic determinant can be expressed as:

[0071] This structure achieves stable and reversible feature normalization without introducing batch dependencies.

[0072] Subsequently, the model is invertible through LU parameterization. Convolution achieves a global linear transformation along the channel dimension. (Sample) Each spatial location pixel Corresponding to a channel vector For each spatial location pixel Perform a linear transformation, the resulting pixels for:

[0073] Among them, the weight matrix determinant Not zero. To avoid directly calculating the determinant. The high cost of the weight matrix Decomposed into:

[0074] Where P is a fixed permutation matrix, L is a unit lower triangular matrix, and U is an upper triangular matrix. The unit lower triangular matrix L is:

[0075] The upper triangular matrix U is:

[0076] in, This is the actual scale factor.

[0077] The corresponding logarithmic Jacobian determinant is:

[0078] The key requirement for improving Glow is to be more efficient in computing the logarithmic Jacobian determinant, and LU decomposition can reduce the complexity from O(c³) to O(c).

[0079] Building upon this, the model incorporates invertible residual blocks to enhance its nonlinear modeling capabilities. The invertible residual blocks affect the output... The processing procedure is as follows: output as input features Let the input features be... It is divided into two parts in terms of channel dimension:

[0080] The transformation of the above expression is defined as: ;

[0081] in, , Let L be the nonlinear mapping function parameterized by the convolutional neural network. L represents the number of convolutional layers. For the first Layer convolution combined with nonlinear operators, For neural network parameters; Therefore, the output of the reversible residual block can be obtained. for: .

[0082] This residual structure significantly improves the model's ability to fit complex local structures without increasing the likelihood computation complexity.

[0083] Output The output is obtained after processing by the reversible residual block. Output The input is fed into a neural spline coupling layer, where the model uses this layer to refine the feature distribution modeling. The direct input to the neural coupling layer is one portion of the segmented features, while the other portion serves as conditional input. The neural spline coupling layer outputs... The operation process is as follows: Output as input vector It was split into:

[0084] in, Remain unchanged, while Invertible mapping via rational quadratic spline functions:

[0085]

[0086]

[0087] in, This represents the rational quadratic spline transformation function. The coefficients of the piecewise quadratic term are... The coefficient of the piecewise linear term, The constant term offset, all of which are trainable parameters, is derived from the spline function. It consists of several segments, and its parameter set is as follows:

[0088] in, , , , The first The spline width, spline height, and spline slope of each segment must satisfy the following constraints:

[0089] Final coupling mapping:

[0090] The corresponding logarithmic Jacobian determinant is:

[0091] in, Features to be transformed The number of channel dimensions.

[0092] Its analytical form can be directly calculated from the derivative of the spline function. Compared with traditional affine coupling layers, neural spline coupling can represent more complex nonlinear monotonic mappings, thereby significantly improving the model's ability to model multimodal and non-Gaussian distributions.

[0093] Glow's reverse generation process is to perform the reverse inversion of the forward transformation layer by layer. This strict reversibility ensures that Glow can accurately generate realistic samples from latent variables.

[0094] This invention uses the obtained 2D GADF image data as basic samples and inputs them into the Glow generative network based on an invertible flow model to expand the training data and enhance the feature distribution. The improved Glow module constructs the latent space representation of the original images through a series of invertible transformations, allowing each GADF image to be mapped into a structured latent variable vector, and perturbing and sampling the latent variables while maintaining the original feature distribution. Subsequently, using the inverse transformation of the Glow model, new 2D images are regenerated from the sampled latent variables, achieving diversified expansion of the original GADF samples. The enhanced images generated by Glow maintain consistency with the real data in terms of texture, intensity changes, and local structure, effectively improving the model's generalization ability and robustness under small sample conditions. Finally, the original GADF images and the enhanced images generated by Glow are combined to form an expanded 2D image training set, providing a sufficient number of continuously distributed 2D sample data for subsequent deep neural network training.

[0095] Step 3: Use two-dimensional image data and image enhancement The PCNN-GRU-ECA neural network was trained using the dataset to obtain a fault diagnosis model. See also... Figure 5 The PCNN-GRU-ECA neural network includes a feature extraction module, a feature fusion module, a GRU module, and an ECA module connected in sequence. The feature extraction module includes a first CNN branch, a second CNN branch, and a third CNN branch connected in parallel. The first CNN branch includes three sequentially connected... 3 convolutional layers, average pooling layer, 5 5 convolutional layers, average pooling layers, and the second CNN branch includes 3 sequentially connected layers. 3 convolutional layers, max pooling layer, 5 Five convolutional layers, a max pooling layer, and the third CNN branch consists of three sequentially connected layers. 3 convolutional layers, hybrid pooling layers, 5 5 convolutional layers and hybrid pooling layers.

[0096] Hybrid pooling layers are a weighted combination of max pooling and average pooling, and can be expressed by the following formula:

[0097] in, The maximum value within the window. This is the average value within the window. These are the weighting coefficients. In this embodiment Take 0.5.

[0098] In this embodiment, the hybrid pooling layer uses a hybrid pooling method with weighted coefficients of 0.6 for max pooling and 0.4 for average pooling to preserve different types of spatial information. The first, second, and third CNN branches extract features from the input image, and the features output from the three CNN branches are fed together into the feature fusion module for feature fusion. The fused features are then fed into the GRU module for sequence modeling. In this embodiment, the hidden layer dimension of the GRU module is set to 128. The output features of the GRU module are input into the ECA module, and in this embodiment, the local interaction range of the ECA module is k=5. The ECA module performs adaptive weight allocation on the channel features to obtain a weighted feature map. The weighted feature map is then fed into the classifier to obtain the predicted probabilities of various fault states. To suppress overfitting, Dropout structures are added to the output of the feature fusion module and the input of the classifier, with an inactivation ratio of 0.25, causing some neurons to be randomly turned off during training, thereby improving the model's generalization ability. The convolutional module and the feature mapping layer uniformly use the LeakyReLU activation function, with a negative slope of 0.02.

[0099] In this embodiment, a softmax classifier is used to achieve probabilistic identification of various commutation faults. The comprehensive feature vector obtained after multi-branch convolution, sequence modeling, and channel attention filtering is fed into a fully connected classification structure for mapping and compression, forming a compact representation for fault identification. The classification layer uses a softmax probability output mechanism to map various potential fault modes to corresponding predicted probabilities, including five categories: bridge arm short circuit, inverter failure, pulse loss, DC-side short circuit, and normal state. The model outputs the prediction result based on the principle of maximizing probability and performs a secondary judgment based on a preset normal state confidence threshold: when the normal state probability is significantly higher than the threshold, the system is considered to be in a healthy working state; if the normal state probability is the highest but does not reach the threshold, it is judged as having a potential anomaly to avoid latent faults being misjudged as normal; if any fault category has the highest probability, it is directly identified as the corresponding fault type.

[0100] The GRU primarily comprises two core structures: the update gate and the reset gate. The update gate determines the extent to which the current hidden state retains information from the previous time step, while the reset gate controls the degree of integration between new input information and historical states. The output characteristics of the GRU module are:

[0101] in, , The output characteristics of the GRU module at time t and time t-1 are respectively. To update the door, , To select a door, , To reset the door, , The features are fused at time t, and · represents the dot product operation. σ Here, Wr, Wz, and Wy are the coefficients and weights of each layer, representing the sigmoid function.

[0102] The output features of the GRU module are input into the ECA module for channel recalibration. The ECA module consists of sequentially connected global average pooling, one-dimensional convolutional layers, and channel weighting. Global average pooling is applied to the output features of the GRU module. Global average pooling is performed along the spatial dimension to obtain the global description vector for each channel. The relation is:

[0103] One-dimensional convolutional layers for global description vectors Perform convolution to output the attention weights for each channel. The relation is: , σ For the Sigmoid activation function, the attention weights .

[0104] Channel weighting adds attention weights Output characteristics of the GRU module Multiplying each channel sequentially yields the weighted feature map. The relation is: , where c = 1, 2, 3, ..., C.

[0105] See Figure 6 Using two-dimensional image data and image enhancement The process of training the PCNN-GRU-ECA neural network using the dataset includes: Step 3.1: Transfer the two-dimensional image data and image enhancement The dataset was divided into training and validation sets according to a preset ratio. Training strategies and network parameters were set, with the AdamW adaptive update method used as the optimizer and an initial learning rate of 0.0015. and The values ​​are 0.9 and 0.999 respectively, and the learning rate is gradually reduced during training using a cosine annealing strategy. Step 3.2: Using two-dimensional image data Or enhance the image Using the fault state as the real label as the input, the PCNN-GRU-ECA neural network is trained. The entire training process uses a batch size of 32 and a total of 100 epochs. The network parameters are continuously updated by calculating the loss between the predicted output and the real label. When the loss is minimized or the set number of training epochs is reached, the trained PCNN-GRU-ECA neural network is obtained. Step 3.3: Input the validation set into the trained PCNN-GRU-ECA neural network for retraining. During training, continuously record the accuracy and stability metrics of each iterative model on the validation set, and comprehensively determine the optimal iteration point. The optimal iteration point achieves the highest classification performance while exhibiting minimal fluctuation in the performance curve. Fix and save the model parameters corresponding to the optimal iteration point to obtain the fault diagnosis model, which can be used for subsequent long-term online or offline fault detection and deployment. To ensure generalization, a final test can be performed using an independent test set to verify the unbiasedness of the performance.

[0106] This invention combines Glow data augmentation with GADF-based two-dimensional image representation to effectively expand the spatial pattern of converter voltage signals and construct a more representative training sample set. It extracts feature information at different scales and dimensions through a three-branch parallel convolutional neural network, models the dynamic evolution of voltage sequences over time using a GRU module, and further enhances the expression of key fault features through channel recalibration of the ECA module. The synergistic effect of these structures significantly improves the accuracy and noise resistance of commutation failure fault diagnosis, and achieves accurate identification of multiple fault states while ensuring model real-time performance, thereby improving the safety and reliability of the fusion magnet power system during operation.

[0107] Step 4: Collect the output voltage of the converter and convert it into two-dimensional image data. Input the two-dimensional image data into the fault diagnosis model to obtain the predicted probability of various fault states. Based on the predicted probability, determine the fault type.

[0108] Various fault states include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit, and normal state. The fault type is determined based on the predicted probability as follows: when the predicted probability of the normal state is the highest, it is determined whether the predicted probability of the normal state reaches the threshold. If it does, it is judged as a normal state; if not, it is judged as a potential anomaly. When the predicted probability of any of the fault states, such as bridge arm short circuit, inverter failure, pulse loss, and DC side short circuit, is the highest, it is directly judged as the fault state corresponding to the highest predicted probability.

[0109] This invention utilizes the Glow model to augment the GADF two-dimensional image samples constructed from the converter's operating voltage with small samples, significantly enriching the fault sample set while maintaining the original data distribution characteristics. Subsequently, a PCNN-GRU-ECA diagnostic network, incorporating multi-branch convolutional structures, gated recurrent units, and channel attention mechanisms, is constructed to mine deep information about the voltage signals in terms of spatial texture, temporal correlation, and feature importance, enabling automatic identification of whether commutation failure has occurred and its specific fault type. By combining generative data augmentation with multi-module deep feature fusion, this invention achieves high diagnostic accuracy and anti-interference capability even with limited sample size, while also considering computational efficiency, thus improving the reliability and real-time performance of online diagnosis of commutation failure faults in phase-controlled converters.

[0110] In the feature fusion stage, the system introduces a sequence position weighting mechanism, assigning dynamic weights to features based on their different channels in the time series. This effectively enhances the system's responsiveness to key anomaly signals and improves the model's sensitivity to anomalies in phase-controlled converters. The computational process is optimized, reducing the impact of redundant features on diagnostic results and improving diagnostic accuracy and real-time performance.

[0111] Figure 7 The graph shows the accuracy change curve of the neural network model during training. As can be seen from the graph, in the first 1-12 epochs, the model accuracy increases rapidly, indicating that the model has fully learned the main discriminative information from the input features in the initial stage. Subsequently, in the 13-38 epoch range, the rate of accuracy improvement slows significantly; the model continues to optimize, but performance growth gradually approaches saturation. After epochs 39-100, the training accuracy and validation accuracy tend to stabilize and remain consistent, indicating that the model has basically reached its optimal generalization ability and no significant overfitting has occurred.

[0112] Figure 8 The curves showing the change of the loss function during training are presented. In the first 1-12 epochs, the loss value decreases rapidly, indicating that the model can quickly capture the main data structure; in the 13-38 epochs, the rate of loss decrease slows down, and the model's learning gradually enters the stable region; when training reaches 39-100 epochs, the training loss shows a smooth convergence trend with minimal change, indicating that the model training process is stable, converges sufficiently, and has good generalization performance.

[0113] Figure 9This paper demonstrates the diagnostic results of the converter commutation failure diagnosis method based on the Glow and neural network model in diagnosing bridge arm short-circuit (BS) faults. As shown in the figure, the output voltage waveform remains normal for the first approximately 1300 sampling points; then, an abnormal fluctuation appears at approximately the 1360th sampling point, indicating that the system has entered a fault state. The model accurately identifies this anomaly as a bridge arm short-circuit fault at approximately the 1500th sampling point, demonstrating the algorithm's sensitive ability to capture rapidly abrupt faults.

[0114] Figure 10 This diagram shows the diagnostic results of the converter commutation failure diagnosis method based on the Glow and neural network model for DC-side short-circuit (DS) faults. The signal in the diagram remains normal for the first approximately 1400 sampling points, but exhibits a significant anomaly after approximately 1540 sampling points. The model reacts rapidly to changes in fault modes and successfully identifies a DC-side short circuit at approximately 2000 sampling points, verifying the algorithm's accuracy in identifying high-current-impact faults.

[0115] Figure 11 The diagnostic results of the converter commutation failure diagnosis method based on the Glow and neural network model of this invention for inverter failure (IF) faults are presented. It can be observed that the waveform is stable in the first 2000 sampling points, reflecting normal system operation; from about the 2400th sampling point, abnormal signs appear in the waveform, and the model determines that it is an inverter failure fault at about the 2500th sampling point, proving that the model can effectively identify abnormal transmission characteristics in the commutation process.

[0116] Figure 12 This paper demonstrates the diagnostic performance of the converter commutation failure diagnosis method based on the Glow and neural network model in diagnosing pulse loss (PL) faults. The system maintains normal output for the first 2000 sampling points, but pulse triggering anomalies begin to appear after approximately 2100 sampling points. The model accurately determines pulse loss at approximately 2500 sampling points, indicating that the algorithm can promptly capture the voltage distortion changes caused by missing trigger pulses.

[0117] Figure 13 This paper presents the diagnostic results of the converter commutation failure diagnosis method based on Glow and a neural network model for the normal state. The signal remained stable throughout the entire sampling interval, with no fault characteristics appearing, and the model continuously outputting the normal category, demonstrating the stability and false alarm suppression capability of the algorithm under fault-free operating conditions.

[0118] The results show that the converter commutation failure diagnosis method based on Glow and neural network model proposed in this invention can fully explore the distribution of fault features when the number of samples is limited, and significantly alleviate the constraint of insufficient samples on the model training effect. On this basis, the deep fusion learning of two-dimensional image features greatly improves the accuracy of fault type discrimination and effectively improves the accuracy of commutation failure fault identification. It is suitable for converter commutation failure diagnosis under complex operating conditions of poloidal field power supply.

[0119] Example 2 See Figure 14 This embodiment provides a converter commutation failure diagnosis system based on Glow and a neural network model, including: The data processing module is used to convert the output voltage of the converter into two-dimensional image data. Specifically, it includes: The output voltage of the converter is normalized to obtain normalized data. This normalized data is then mapped to the angle domain to obtain a one-dimensional angle sequence. Angles at different times within the one-dimensional angle sequence are paired, and the angle differences are calculated and encoded using a sine function to obtain a two-dimensional GADF matrix reflecting the correlation between different times. This two-dimensional GADF matrix is ​​then input into an image processing program for visualization, resulting in two-dimensional image data. .

[0120] Image enhancement module, used to enhance two-dimensional image data The input is an improved Glow module, which consists of a compression layer and N Flow modules. Each Flow module includes sequentially connected activation normalization, LU-parameterized invertible 1×1 convolution, invertible residual block, neural spline coupling, and decompression of the two-dimensional image data. The image is obtained after compression by the compression layer. ,image The sample was obtained after activation and normalization. ,sample The output is obtained after a LU-parameterized reversible 1×1 convolution operation. Output The output is obtained after reversible residual block transformation. Output The neural spline coupling is split into , , remain unchanged. The output is obtained by invertible mapping using rational quadratic spline functions. Output After decompression, the variable z in the latent space Z is obtained. The data sampled from the latent space Z is then passed inversely through the improved Glow module to obtain the enhanced image. .

[0121] sample The process after the LU-parameterized reversible 1×1 convolution operation is as follows: Sample Each spatial location pixel Corresponding to a channel vector For each spatial location pixel Perform a linear transformation, the resulting pixels for:

[0122] weight matrix Decomposed into:

[0123] Where P is a fixed permutation matrix, and the unit lower triangular matrix L and upper triangular matrix U are respectively: ;

[0124] The actual scale factor is given; the corresponding logarithmic Jacobian determinant is: .

[0125] Reversible residual block output The processing procedure is as follows: output as input features Let the input features be... It is divided into two parts in terms of channel dimension:

[0126] The transformation of the above expression is defined as: ;

[0127] in, , Let L be the nonlinear mapping function parameterized by the convolutional neural network, where L is the number of convolutional layers. For the first Layer convolution combined with nonlinear operators, These are the parameters of the neural network.

[0128] Output of reversible residual block for: .

[0129] The output of the neural spline coupling to the reversible residual block The operation process is as follows: Output as input vector It was split into:

[0130] in, remain unchanged. Invertible mapping via rational quadratic spline functions:

[0131]

[0132]

[0133] in, This represents the rational quadratic spline transformation function. The coefficients of the piecewise quadratic term are... The coefficient of the piecewise linear term, The constant term offset, all of which are trainable parameters, is derived from the spline function. It consists of several segments, and its parameter set is as follows:

[0134] in, , , , The first The spline width, spline height, and spline slope of each segment must satisfy the following constraints:

[0135] Final coupling mapping:

[0136] The corresponding logarithmic Jacobian determinant is:

[0137] in, Features to be transformed The number of channel dimensions.

[0138] Learning module for using two-dimensional image data and image enhancement The PCNN-GRU-ECA neural network was trained using the dataset to obtain a fault diagnosis model.

[0139] The pCNN-GRU-ECA neural network comprises a feature extraction module, a feature fusion module, a GRU module, and an ECA module connected in sequence. The feature extraction module includes a first CNN branch, a second CNN branch, and a third CNN branch connected in parallel. The first CNN branch includes three sequentially connected... 3 convolutional layers, average pooling layer, 5 5 convolutional layers, average pooling layers, and the second CNN branch includes 3 sequentially connected layers. 3 convolutional layers, max pooling layer, 5 Five convolutional layers, a max pooling layer, and the third CNN branch consists of three sequentially connected layers. 3 convolutional layers, hybrid pooling layers, 5 The system consists of 5 convolutional layers and hybrid pooling layers. The first, second, and third CNN branches extract features from the input image. The features output from the three CNN branches are then fed into a feature fusion module for feature fusion. The fused features are then fed into a GRU module for sequence modeling. The output features of the GRU module are fed into an ECA module, which performs adaptive weight allocation on the channel features to obtain a weighted feature map. The weighted feature map is then fed into a classifier to obtain the predicted probabilities of various fault states. Dropout structures are added to the output of the feature fusion module and the input of the classifier.

[0140] The output characteristics of the GRU module are:

[0141] in, , The output characteristics of the GRU module at time t and time t-1 are respectively. To update the door, , To select a door, , To reset the door, , The features are fused at time t, and · represents the dot product operation. σ Here, Wr, Wz, and Wy are the coefficients and weights of each layer, representing the sigmoid function.

[0142] The output features of the GRU module are input into the ECA module. The ECA module consists of a global average pooling layer, a one-dimensional convolutional layer, and channel weighting connected in sequence. The global average pooling layer performs global average pooling on the output features of the GRU module according to the spatial dimension to obtain the global description vector for each channel. The one-dimensional convolutional layer convolves the global description vector to output the attention weight for each channel. The channel weighting layer multiplies the attention weight with the output features of the GRU module channel by channel to obtain the weighted feature map.

[0143] Using two-dimensional image data and image enhancement The process of training the PCNN-GRU-ECA neural network using the dataset includes: processing the two-dimensional image data... and image enhancement The dataset was divided into training and validation sets according to a preset ratio. Training strategies and network parameters were set, with the AdamW adaptive update method used as the optimizer and an initial learning rate of 0.0015. and The values ​​are 0.9 and 0.999 respectively, and the learning rate is gradually reduced during training using a cosine annealing strategy. Two-dimensional image data Or enhance the image Using the fault state as the real label as the input, the PCNN-GRU-ECA neural network is trained. The entire training process uses a batch size of 32 and a total of 100 epochs. The network parameters are continuously updated by calculating the loss between the predicted output and the real label. When the loss is minimized or the set number of training epochs is reached, the trained PCNN-GRU-ECA neural network is obtained. The validation set is input into the trained PCNN-GRU-ECA neural network for retraining. During the training process, the accuracy and stability of each iterative model on the validation set are continuously recorded, and the optimal iteration point is determined by comprehensive analysis. The model parameters corresponding to the optimal iteration point are fixed and saved to obtain the fault diagnosis model.

[0144] The inference module is used to collect the output voltage of the converter and convert it into two-dimensional image data. The two-dimensional image data is then input into the fault diagnosis model to obtain the predicted probability of various fault states, and the fault type is determined based on the predicted probability.

[0145] Various fault states include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit, and normal state. The fault type is determined based on the predicted probability as follows: when the predicted probability of the normal state is the highest, it is determined whether the predicted probability of the normal state reaches the threshold. If it does, it is judged as a normal state; if not, it is judged as a potential anomaly. When the predicted probability of any of the fault states, such as bridge arm short circuit, inverter failure, pulse loss, and DC side short circuit, is the highest, it is directly judged as the fault state corresponding to the highest predicted probability.

[0146] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A converter commutation failure diagnosis method based on Glow and neural network models, characterized in that: include: Convert the output voltage of the converter into two-dimensional image data. ; Two-dimensional image data The input is an improved Glow module, which consists of a compression layer and N Flow modules. Each Flow module includes sequentially connected activation normalization, LU-parameterized invertible 1×1 convolution, invertible residual block, neural spline coupling, and decompression of the two-dimensional image data. The image is obtained after compression by the compression layer. ,image The sample was obtained after activation and normalization. ,sample The output is obtained after a LU-parameterized reversible 1×1 convolution operation. Output The output is obtained after reversible residual block transformation. Output The neural spline coupling is split into , , remain unchanged. The output is obtained by invertible mapping using rational quadratic spline functions. Output After decompression, the variable z in the latent space Z is obtained. The data sampled from the latent space Z is then passed inversely through the improved Glow module to obtain the enhanced image. ; Using two-dimensional image data and image enhancement The PCNN-GRU-ECA neural network was trained using the dataset to obtain a fault diagnosis model. The output voltage of the converter is collected and converted into two-dimensional image data. The two-dimensional image data is then input into the fault diagnosis model to obtain the predicted probability of various fault states. The fault type is then determined based on the predicted probability.

2. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 1, characterized in that: Convert the output voltage of the converter into two-dimensional image data. The process includes: The output voltage of the converter is normalized to obtain normalized data; Mapping the normalized data to the angle domain yields a one-dimensional angle sequence; By pairing angles at different times in a one-dimensional angle sequence, calculating the angle difference, and encoding it using a sine function, a two-dimensional GADF matrix reflecting the correlation between different times is obtained. The two-dimensional GADF matrix is ​​input into an image processing program for visualization; two-dimensional image data. .

3. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 1, characterized in that: sample The process after the LU-parameterized reversible 1×1 convolution operation is as follows: Sample Each spatial location pixel Corresponding to a channel vector For each spatial location pixel Perform a linear transformation, the resulting pixels for: weight matrix Decomposed into: Where P is a fixed permutation matrix, and the unit lower triangular matrix L and upper triangular matrix U are respectively: ； The actual scale factor is given; the corresponding logarithmic Jacobian determinant is: 。 4. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 1, characterized in that: Reversible residual block output The processing procedure is as follows: output as input features Let the input features be... It is divided into two parts in terms of channel dimension: The transformation of the above expression is defined as: ; in, , Let L be the nonlinear mapping function parameterized by the convolutional neural network, where L is the number of convolutional layers. For the first Layer convolution combined with nonlinear operators, For neural network parameters; Output of reversible residual block for: .

5. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 1, characterized in that: remain unchanged. The process of invertible mapping using rational quadratic spline functions is as follows: in, This represents the rational quadratic spline transformation function. The coefficients of the piecewise quadratic term are... The coefficient of the piecewise linear term, For the constant term offset, the spline function is derived from It consists of several segments, and its parameter set is as follows: in, , , , The first The spline width, spline height, and spline slope of each segment must satisfy the following constraints: Final coupling mapping: The corresponding logarithmic Jacobian determinant is: in, Features to be transformed The number of channel dimensions.

6. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 1, characterized in that: The pCNN-GRU-ECA neural network comprises a feature extraction module, a feature fusion module, a GRU module, and an ECA module connected in sequence. The feature extraction module includes a first CNN branch, a second CNN branch, and a third CNN branch connected in parallel. The first CNN branch includes three sequentially connected... 3 convolutional layers, average pooling layer, 5 5 convolutional layers, average pooling layers, and the second CNN branch includes 3 sequentially connected layers. 3 convolutional layers, max pooling layer, 5 Five convolutional layers, a max pooling layer, and the third CNN branch consists of three sequentially connected layers. 3 convolutional layers, hybrid pooling layers, 5 The system consists of 5 convolutional layers and hybrid pooling layers. The first, second, and third CNN branches extract features from the input image. The features output from the three CNN branches are then fed into a feature fusion module for feature fusion. The fused features are then fed into a GRU module for sequence modeling. The output features of the GRU module are fed into an ECA module, which performs adaptive weight allocation on the channel features to obtain a weighted feature map. The weighted feature map is then fed into a classifier to obtain the predicted probabilities of various fault states. Dropout structures are added to the output of the feature fusion module and the input of the classifier.

7. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 6, characterized in that: The output characteristics of the GRU module are: in, , The output characteristics of the GRU module at time t and time t-1 are respectively. To update the door, , To select a door, , To reset the door, , The features are fused at time t, and · represents the dot product operation. σ Here, Wr, Wz, and Wy are the coefficients and weights of each layer, representing the sigmoid function.

8. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 6, characterized in that: The output features of the GRU module are input into the ECA module. The ECA module consists of a global average pooling layer, a one-dimensional convolutional layer, and channel weighting connected in sequence. The global average pooling layer performs global average pooling on the output features of the GRU module according to the spatial dimension to obtain the global description vector for each channel. The one-dimensional convolutional layer convolves the global description vector to output the attention weight for each channel. The channel weighting layer multiplies the attention weight with the output features of the GRU module channel by channel to obtain the weighted feature map.

9. The converter commutation failure diagnosis method based on Glow and neural network model according to claim 1, characterized in that: Using two-dimensional image data and image enhancement The process of training the PCNN-GRU-ECA neural network using the dataset includes: Two-dimensional image data and image enhancement The dataset was divided into training and validation sets according to a preset ratio. Training strategies and network parameters were set, with the AdamW adaptive update method used as the optimizer and an initial learning rate of 0.0015. and The values ​​are 0.9 and 0.999 respectively, and the learning rate is gradually reduced during training using a cosine annealing strategy. Two-dimensional image data Or enhance the image Using fault states as input and fault states as true labels, including bridge arm short circuit, inverter failure, pulse loss, DC side short circuit and normal state, the PCNN-GRU-ECA neural network is trained. The entire training process uses a batch size of 32 and a total of 100 epochs. The network parameters are continuously updated by calculating the loss between the predicted output and the true label. When the loss is minimized or the set number of training epochs is reached, the trained PCNN-GRU-ECA neural network is obtained. The validation set is input into the trained PCNN-GRU-ECA neural network for retraining. During the training process, the accuracy and stability of each iterative model on the validation set are continuously recorded, and the optimal iteration point is determined by comprehensive analysis. The model parameters corresponding to the optimal iteration point are fixed and saved to obtain the fault diagnosis model.

10. A converter commutation failure diagnosis system based on Glow and neural network models, characterized in that: include: The data processing module is used to convert the output voltage of the converter into two-dimensional image data. ; Image enhancement module, used to enhance two-dimensional image data The input is an improved Glow module, which consists of a compression layer and N Flow modules. Each Flow module includes sequentially connected activation normalization, LU-parameterized invertible 1×1 convolution, invertible residual block, neural spline coupling, and decompression of the two-dimensional image data. The image is obtained after compression by the compression layer. ,image The sample was obtained after activation and normalization. ,sample The output is obtained after a LU-parameterized reversible 1×1 convolution operation. Output The output is obtained after reversible residual block transformation. Output The neural spline coupling is split into , , remain unchanged. The output is obtained by invertible mapping using rational quadratic spline functions. Output After decompression, the variable z in the latent space Z is obtained. The data sampled from the latent space Z is then passed inversely through the improved Glow module to obtain the enhanced image. ; Learning module for using two-dimensional image data and image enhancement The PCNN-GRU-ECA neural network was trained using the dataset to obtain a fault diagnosis model. The inference module is used to collect the output voltage of the converter and convert it into two-dimensional image data. The two-dimensional image data is then input into the fault diagnosis model to obtain the predicted probability of various fault states, and the fault type is determined based on the predicted probability.

Images