Converter commutation failure fault diagnosis method and system

Abstract

The application provides a converter commutation failure fault diagnosis method, belongs to the field of phase-controlled converters, and comprises the following steps: converting output voltage of the converter into two-dimensional image data; inputting the output voltage of the converter and the two-dimensional image data into a pre-trained double-path convolutional neural network; the pre-trained double-path convolutional neural network comprises a one-dimensional data processing channel and a two-dimensional data processing channel; the output voltage of the converter enters the one-dimensional data processing channel to obtain a characteristic representation signal; the two-dimensional image data enters the two-dimensional data processing channel to obtain a characteristic representation signal; the characteristic representation signals are fused to obtain a fused characteristic representation signal; the fused characteristic representation signal is processed to obtain a prediction probability of various fault states; and the fault type is judged based on the prediction probability; the application also provides a converter commutation failure fault diagnosis system; two convolutional neural network paths establish and capture time sequence and spatial information of converter voltage data, and the algorithm accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of phase-controlled converters, and in particular to a converter commutation failure fault diagnosis method and system. Background Art

[0002] As the core subsystem of the tokamak, the fusion magnet power supply provides ultra-high currents to the superconducting coils, generating and maintaining a stable, powerful confining magnetic field. Thyristor phase-controlled rectification technology, with its advantages of high cost-effectiveness, high technical maturity, and high power capacity, is currently the primary technology for fusion magnet power supply. However, it suffers from the inherent flaw of commutation failure. Continuous commutation failures will trigger power failures, which not only affects the plasma discharge quality but can also damage internal device components due to the coupling of tens of thousands of amperes of circular current, threatening the safety of the fusion device.

[0003] Neural networks are widely used in the field of real-time diagnosis due to their good adaptability and nonlinear modeling characteristics. As a typical structure, convolutional neural networks have excellent feature extraction capabilities and parameter sharing mechanisms. They can effectively identify local and global features in complex signals. They are particularly effective in processing fault data with local temporal and spatial characteristics. They are especially suitable for tasks with high precision requirements and strong model generalization capabilities. However, neural network models are usually complex in structure, have a large number of parameters, and require a large amount of computation in the training process. Since the diagnosis of converter commutation failure faults requires high real-time and accuracy, when performing commutation failure fault diagnosis based on convolutional neural networks, the designed model must not only detect whether the device has a commutation failure in real time, but also have low hardware requirements and be able to distinguish the type of fault in real time so that corresponding safety protection actions can be taken.

[0004] In the prior art, the Chinese invention patent application with publication number CN113721162A, "An Intelligent Diagnosis Method for Fusion Magnet Power Supply Faults Based on Deep Learning," avoids the problem of differences in data characteristics in different fields through a dual-path heterogeneous network design. However, the method has the following problems: it only identifies two types of faults: inverter failure and pulse loss; the target domain network Ft structure is a shallow CNN (only two layers of convolution + full connection), which may have insufficient feature extraction capabilities when processing high-complexity fault signals, especially those with electromagnetic interference. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to improve the accuracy of converter commutation failure fault diagnosis.

[0006] The present invention solves the above technical problems through the following technical solutions: a fault diagnosis method for converter commutation failure, the method comprising: converting the output voltage of the converter into two-dimensional image data, inputting the output voltage of the converter and the two-dimensional image data into a pre-trained two-way convolutional neural network, the pre-trained two-way convolutional neural network comprising a one-dimensional data processing channel and a two-dimensional data processing channel, the output voltage of the converter enters the one-dimensional data processing channel, and obtains a feature representation signal , the two-dimensional image data enters the two-dimensional data processing channel and obtains the feature representation signal , the feature represents the signal and After fusion, the fused feature representation signal is obtained. After processing the fused feature representation signal, the predicted probability of each fault state is obtained, and the fault type is determined based on the predicted probability.

[0007] Beneficial effects: The present invention designs two convolutional neural network paths respectively, performs independent feature extraction for different feature sequences, fully establishes and captures the temporal and spatial information of the thyristor converter voltage data, completes the judgment of whether a commutation failure fault occurs and the type of fault, and improves the accuracy of the algorithm and the real-time performance of fault diagnosis.

[0008] Preferably, the process of converting the output voltage of the converter into two-dimensional image data includes:

[0009] Processing the output voltage of the converter to obtain one-dimensional signal data;

[0010] Mapping the one-dimensional signal data into a normalized difference sequence, mapping the normalized difference sequence to an angle on the unit circle, obtaining an angle sequence, and obtaining a first-order difference angle sequence;

[0011] A two-dimensional symmetric matrix is ​​constructed based on the first-order difference angle sequence as the two-dimensional image data.

[0012] Beneficial effects: The present invention adopts the FO-GADF conversion method to convert one-dimensional signal data into two-dimensional image data. The FO derivative reflects the rate of change of the sequence. By introducing first-order difference processing to enhance the expressiveness of prominent points, the mutation points can be captured more keenly. The features of the formed two-dimensional image at the prominent points are more obvious, which is more conducive to model learning.

[0013] Preferably, the normalized difference sequence for:

[0014]

[0015] in, , one-dimensional signal data , , , ;

[0016] An angle on the unit circle , ;Angle sequence ;

[0017] set up , , get the first-order difference angle sequence of length n-1 ,right and The difference between them is sine-encoded to construct a two-dimensional symmetric matrix FOGADF , for the two-dimensional symmetric matrix FOGADF After visualization, a two-dimensional image is obtained.

[0018] Preferably, the output voltage of the converter is sampled and enters the one-dimensional data processing channel. First, the output voltage is convolved and pooled to obtain the first output feature. The first output feature enters three convolution kernels of different scales for feature extraction to obtain the first angle feature, the second angle feature and the third angle feature. The three angle features are sequentially passed through the SE-DRB module, the attention mechanism, and the FG-Bi-LSTM module to obtain the output features. 、 、 , output features 、 、 After fusion, the feature representation signal is obtained .

[0019] Beneficial effect: By processing the input features through the SE-DRB module, the feature extraction effect can be significantly improved while keeping the network computational complexity controllable.

[0020] Preferably, the SE-DRB module includes two 1×1 convolutional layers, a 3×1 convolutional layer, and an SE module, and each angle feature is used as an input feature. , enter the first 1×1 convolution layer to perform channel dimensionality reduction and obtain the output feature , C is the number of channels, L is the length of the time series, and the output feature Enter the 3×1 convolution layer to expand the receptive field and obtain the output features , output features After entering the second 1×1 convolutional layer to restore the original dimension, the output features are obtained , output features The SE module assigns different weights to the channels, broadcasts the weights to each time step, forms the attention weighted output, obtains the weighted features, and converts the input features into Add it element-wise to the weighted feature to get the second output feature.

[0021] Beneficial Effects: During the feature fusion phase, the system introduces a sequence position weighting mechanism, assigning dynamic weights based on the different channels of features in the time series. This effectively enhances the responsiveness to critical abnormal signals and improves the model's sensitivity to abnormal output voltage values ​​in phase-controlled converters. This optimizes the computational process, reduces the impact of redundant features on diagnostic results, and improves diagnostic accuracy and real-time performance.

[0022] Preferably, the SE module includes a global average pooling layer and a two-layer fully connected network, outputting features Enter the global average pooling layer and output the result z. The result z is passed through the two-layer fully connected network to obtain the channel weight for:

[0023] , ,result for: , Indicates the number of channels.

[0024] Preferably, the two-dimensional image data enters the two-dimensional data processing channel, and first performs two-dimensional convolution and pooling operations on the two-dimensional image data to perform preliminary feature extraction to obtain the third output feature, and the third output feature is sequentially passed through the convolution kernel, ReLU activation function, CBAM module and FG-Bi-LSTM module to obtain the feature representation signal .

[0025] Preferably, the output of the FG-Bi-LSTM module for:

[0026]

[0027]

[0028]

[0029] in, is the weight matrix of the fusion gate, is the bias vector of the fusion gate, represents the concatenation of the forward and backward hidden states, is the sigmoid activation function, Determine the weight of the forward hidden state, Determines the weight of the backward hidden state, ⊙ represents element-wise multiplication.

[0030] Beneficial effect: By adding the FG-Bi-LSTM module (fusion gate bidirectional LSTM structure) to the one-dimensional data processing channel and the two-dimensional data processing channel, the flexibility and response speed of the model in non-stationary fault signal analysis can be improved.

[0031] Preferably, the training process of the dual-path convolutional neural network includes:

[0032] Collect voltage signal data of various fault conditions during the commutation process and pre-process the voltage signal data to obtain a one-dimensional signal data set. The various fault conditions include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit and normal state;

[0033] The one-dimensional signal data set is divided into a one-dimensional training set, a one-dimensional validation set, and a one-dimensional test set. The one-dimensional training set and the one-dimensional validation set are respectively labeled with the corresponding fault status. The one-dimensional test set is used as the output voltage of the converter to be diagnosed.

[0034] Convert the divided one-dimensional signal data set into two-dimensional image data, and establish a two-dimensional training set, a two-dimensional validation set, and a two-dimensional test set that correspond one-to-one to the one-dimensional signal data;

[0035] The one-dimensional training set and the two-dimensional training set are input into the dual-path convolutional neural network together. The one-dimensional training set enters the one-dimensional data processing channel and outputs the one-dimensional feature training result. The two-dimensional training set enters the two-dimensional data processing channel and outputs the two-dimensional feature training result. The one-dimensional feature training result and the two-dimensional feature training result are fused to obtain the fault diagnosis result. The fault diagnosis result and the label of the fault state are compared, and the parameters of the two-path convolutional neural network are updated and retrained to obtain a trained two-path convolutional neural network.

[0036] The one-dimensional validation set and the two-dimensional validation set are input into the trained two-way convolutional neural network together, and the model parameters when the accuracy reaches the highest point and is stable are used as the parameters of the pre-trained two-way convolutional neural network.

[0037] The present invention also provides a converter commutation failure fault diagnosis system, the system comprising:

[0038] A data processing module, used for converting the output voltage of the converter into two-dimensional image data;

[0039] The feature extraction module is used to input the output voltage of the converter and the two-dimensional image data into a pre-trained two-way convolutional neural network. The pre-trained two-way convolutional neural network includes a one-dimensional data processing channel and a two-dimensional data processing channel. The output voltage of the converter enters the one-dimensional data processing channel to obtain a feature representation signal. , the two-dimensional image data enters the two-dimensional data processing channel and obtains the feature representation signal , the feature represents the signal and After fusion, the fusion feature representation signal is obtained;

[0040] The fault diagnosis module is used to process the fused feature representation signal to obtain the predicted probability of various fault states and determine the fault type based on the predicted probability.

[0041] The advantages provided by the present invention are that the method for diagnosing commutation failure in a power converter can quickly and accurately identify the type of commutation failure, and exhibits stronger discrimination and anti-interference capabilities compared to single-path models. This method effectively addresses the high computational complexity of traditional neural network algorithms, significantly improving the efficiency and reliability of real-time diagnosis of commutation failure in phase-controlled power converters. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is the poloidal field power supply topology diagram;

[0043] Figure 2 Schematic diagram of FO-GADF conversion in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0044] FIG3 (a) is a schematic diagram of the conversion of a one-dimensional signal FO-GADF of a bridge arm short circuit in the converter commutation failure fault diagnosis method provided by Example 1 of the present invention;

[0045] FIG3( b ) is a schematic diagram of the conversion of a one-dimensional DC side short-circuit signal FO-GADF in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0046] FIG3( c ) is a schematic diagram of the conversion of a one-dimensional pulse loss signal FO-GADF in the converter commutation failure fault diagnosis method provided by Example 1 of the present invention;

[0047] FIG3( d ) is a schematic diagram of the conversion of the one-dimensional signal FO-GADF of the inverter failure in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0048] FIG3( e ) is a schematic diagram of the conversion of a one-dimensional signal FO-GADF in a normal state in the converter commutation failure fault diagnosis method provided by Example 1 of the present invention;

[0049] Figure 4 This is a framework diagram of a dual-path convolutional neural network model in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0050] Figure 5 This is a basic principle diagram of the SE-DRB module in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0051] Figure 6 Schematic diagram of the SE module in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0052] Figure 7Flowchart of the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0053] Figure 8 This is a flowchart of dual-path convolutional neural network training in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0054] Figure 9 A schematic diagram of the hierarchical cache and storage mechanism in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0055] Figure 10 A loss curve diagram of the dual-path convolutional neural network training process and verification process in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0056] Figure 11 A graph showing the training accuracy and verification accuracy of a dual-path convolutional neural network in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0057] Figure 12 This is a diagram showing the results of diagnosing an inverter failure (IF) fault using a dual-path convolutional neural network in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0058] Figure 13 This is a diagram showing the results of a pulse loss (PL) fault diagnosed by a dual-path convolutional neural network in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0059] Figure 14 This is a diagram showing the results of diagnosing a bridge arm short circuit (BS) fault using a dual-path convolutional neural network in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0060] Figure 15 This is a diagram showing the results of a DC side short circuit (DS) fault diagnosed by a dual-path convolutional neural network in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention;

[0061] Figure 16 This is a diagram showing the normal state (Normal) result of the dual-path convolutional neural network diagnosis in the converter commutation failure fault diagnosis method provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0062] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the following describes the technical solutions of the present invention clearly and completely with reference to specific embodiments and the accompanying drawings. It is obvious that the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] Example 1

[0064] The poloidal field power supply of the EAST device adopts a phase-controlled current conversion method. Each power supply consists of four three-phase fully controlled converter units. The topology diagram of the poloidal field power supply is as follows: Figure 1 As shown, a three-phase fully controlled converter unit contains six thyristors, connected in series (VT1 and VT4), then in parallel (VT3 and VT6, VT5 and VT2). The load is typically a large inductor. The three-phase power supplies Ua, Ub, and Uc are grounded at one end and connected to the center of each pair of thyristors via an inductor. Each thyristor is driven by an independent control signal to turn on and off, regulating the direction and magnitude of the current. The three-phase fully controlled converter unit consists of three bridge arms, each containing two complementary thyristors, forming a typical six-pulse rectifier circuit.

[0065] See also Figure 4 and Figure 7 This embodiment provides a method for diagnosing converter commutation failure faults, comprising the following steps:

[0066] Step 1: Convert the output voltage of the converter into two-dimensional image data. The specific process includes:

[0067] Step 1.1, collect the output voltage of the converter, process the output voltage of the converter, and obtain one-dimensional signal data. The output voltage of the converter is the real-time voltage data. In order to achieve efficient storage and subsequent analysis and processing of the collected signal, a hierarchical cache and storage mechanism is constructed. Figure 9 As shown in the figure, the collected data is stored separately for every 3000 sampling points. After each segment of raw signal data is sampled by the analog-to-digital converter (ADC), it is first sent to the cache FIFO (First-In First-Out) buffer for temporary buffering. When the data in the FIFO accumulates to 3000 points, the control logic starts DMA (Direct Memory Access) to write the data from the FIFO to the external SDRAM in blocks. The data can then be read and processed by reading the FIFO, with 500 data points read each time.

[0068] The output voltage of the converter obtained by acquisition is a one-dimensional discrete time series. By normalizing the acquired data, one-dimensional signal data is obtained. The normalization method is:

[0069]

[0070] in, is the voltage data after normalization, Represents the voltage value collected at the i-th sampling moment. If the total number of sampling points is n, then , μ and σ are the mean and standard deviation of the sampled voltage signal respectively.

[0071] Step 1.2: Convert one-dimensional signal data Mapping to normalized difference sequence , the normalized difference sequence Map to an angle on the unit circle and get the angle sequence , and obtain the first-order difference angle sequence ;

[0072] The difference sequence is defined as: , , normalized difference sequence for:

[0073]

[0074] in, , .

[0075] Normalize the difference series An angle mapped to the unit circle :

[0076]

[0077] in, .

[0078] Get angle sequence ,set up , , get the first-order difference angle sequence of length n-1 .

[0079] Step 1.3: Based on the first-order difference angle sequence Construct a two-dimensional symmetric matrix FOGADF , as two-dimensional image data.

[0080] Through and The difference between them is sine-encoded to construct a two-dimensional symmetric matrix FOGADF , two-dimensional symmetric matrix FOGADF Used as a two-dimensional image representation, the image processing program is used to process the two-dimensional symmetric matrix FOGADF After visualization, a two-dimensional image is obtained.

[0081] The present invention adopts the FO-GADF (First-Order Gramian Angular Difference Field) conversion method to convert one-dimensional signal data into two-dimensional image data. The FO derivative reflects the rate of change of the sequence. By introducing first-order difference processing to enhance the expressiveness of prominent points, the mutation points can be captured more sensitively. The resulting two-dimensional image has more obvious features at the prominent points, which is more conducive to model learning. Figure 2 Figure 3(a) through Figure 3(e) illustrate the FO-GADF conversion diagram for four fault states and one normal state. 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 (Normal).

[0082] Step 2: Input the output voltage of the converter and the two-dimensional image data into the pre-trained dual-channel convolutional neural network. The pre-trained dual-channel convolutional neural network includes a one-dimensional data processing channel and a two-dimensional data processing channel. The output voltage of the converter enters the one-dimensional data processing channel to obtain a feature representation signal. , the two-dimensional image data enters the two-dimensional data processing channel and obtains the feature representation signal , the feature represents the signal and After fusion, the fusion feature representation signal is obtained.

[0083] See also Figure 4 The output voltage of the converter is obtained through sampling. After normalization, the output voltage is processed to obtain one-dimensional signal data. The one-dimensional signal data enters the one-dimensional data processing channel and passes through the convolution layer and the maximum pooling layer in sequence. The one-dimensional signal data is convolved and pooled to perform preliminary feature extraction to obtain the first output feature. The first output feature enters three convolution kernels of different scales for feature extraction to extract multi-angle feature information. The three convolution kernels of different scales are 7 7 convolution kernels, 5 5 convolution kernels, 3 3 convolution kernels, the first output feature enters 7 After the 7 convolution kernel performs feature extraction, the first angle feature is obtained, and the first output feature enters 5 5 convolution kernels are used to extract features and obtain the second angle feature. The first output feature enters 3 The third angle feature is obtained after the 3 convolution kernel is used for feature extraction. The three angle features are sequentially passed through the SE-DRB module, the attention mechanism, and the FG-Bi-LSTM module. Among them, the first angle feature is sequentially passed through the first SE-DRB module, the attention mechanism, and the first FG-Bi-LSTM module for feature extraction to obtain the output feature. The second angle feature is extracted by the second SE-DRB module, the attention mechanism, and the second FG-Bi-LSTM module, and the output feature is obtained. The third angle feature is extracted by the third SE-DRB module, the attention mechanism, and the third FG-Bi-LSTM module, and the output feature is obtained. , output features 、 、 After fusion, the feature representation signal is obtained .

[0084] See also Figure 5 ,SE-DRB module includes two 1×1 convolution layers, 3×1 convolution layers, SE module, and each angle feature is used as input feature Enter the first 1×1 convolution layer to perform channel dimensionality reduction and convert the input features Map to low-dimensional space to obtain output features , C is the number of channels, and L is the length of the time series.

[0085] Output features for:

[0086] in, , BN represents the normalization operation, which is done by Channel dimensionality reduction can reduce the number of channels from C to C / 4.

[0087] Use a 3×1 convolution with a dilation rate r of 2 to output features Input 3×1 convolution layer to expand the receptive field, extract features within a larger time range, and obtain output features :

[0088]

[0089] Output features Enter the second 1×1 convolution layer to restore the number of channels to the original dimension, and then get the output features after restoring the original dimension. , output features It is the feature expression of the main branch that has not been weighted.

[0090]

[0091] in, .

[0092] Output features Assign different weights to channels through the SE module , broadcast the weights to each time step to form the attention weighted output and obtain the weighted features In order to retain the original input information and enhance the gradient propagation capability, the input features With weighted features Add element by element to get the second output feature .

[0093] Weighted Features for:

[0094] in, Represents element-wise multiplication.

[0095] Second output feature for:

[0096] By processing the input features through the SE-DRB module, the feature extraction effect can be significantly improved while keeping the network computational complexity controllable.

[0097] See also Figure 6 , the SE module includes a global average pooling layer and a two-layer fully connected network, outputting features Enter the global average pooling layer and output the result ,result The channel weights are obtained through a two-layer fully connected network for:

[0098] , ;result for: ,in, Indicates the number of channels (feature dimension).

[0099] The two-dimensional image data enters the two-dimensional data processing channel. First, two-dimensional convolution and pooling operations are performed on the two-dimensional image data to perform preliminary feature extraction to obtain the third output feature. The third output feature is sequentially passed through the convolution kernel, ReLU activation function, CBAM module and FG-Bi-LSTM module. The convolution kernel is used to extract local and global features, and the nonlinear expression ability of the network is enhanced by the ReLU activation function. Subsequently, the data is input into the CBAM, and feature optimization is performed in the spatial dimension and channel dimension respectively to further highlight important information related to commutation failure. After the model is processed by the FG-Bi-LSTM layer, it can automatically assign weights according to the importance of different input features, filter out features with higher diagnostic value, and finally obtain a feature representation signal. .

[0100] By adding FG-Bi-LSTM modules (fusion gate bidirectional LSTM structure) to the one-dimensional data processing channel and the two-dimensional data processing channel, the flexibility and response speed of the model in non-stationary fault signal analysis can be improved. The output of the FG-Bi-LSTM module for:

[0101]

[0102]

[0103]

[0104] in, is the weight matrix of the fusion gate, is the bias vector of the fusion gate, represents the concatenation of the forward and backward hidden states, is the sigmoid activation function, Determine the weight of the forward hidden state, Determines the weight of the backward hidden state, ⊙ represents element-wise multiplication.

[0105] The two-way convolutional neural network is trained to obtain a pre-trained two-way convolutional neural network, see Figure 8 , the specific training process includes the following steps:

[0106] Step 2.1, use voltage data acquisition equipment to collect voltage signal data of various fault states during the commutation process, and obtain a one-dimensional signal data set after preprocessing the voltage signal data. Various fault states include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit and normal state. In the process of voltage data acquisition, in order to fully capture the dynamic changes of voltage during the commutation process, the sampling frequency is The frequency is set to 10kHz, and the collected original voltage signal is represented as a one-dimensional discrete time series ,in, = represents the voltage value collected at the i-th sampling moment, and n is the total number of sampling points. All collected voltage signal data are preprocessed using the normalization method in step 1.1. The normalized voltage data is used as a one-dimensional signal data set.

[0107] Step 2.2: Divide the one-dimensional signal dataset into a one-dimensional training set, a one-dimensional validation set, and a one-dimensional test set. Label the one-dimensional training set and the one-dimensional validation set with the corresponding fault state. The one-dimensional test set serves as the output voltage of the converter under diagnosis. The one-dimensional signal dataset can be divided into training, validation, and test sets in a ratio of 70%, 15%, and 15%. To partition the one-dimensional signal dataset, the signal is segmented into segments of equal length, with each segment corresponding to a sample. A sliding window is applied to the signal using a fixed window length (3000 points) and an overlapping step size to obtain multiple equal-length subsamples. These subsamples are randomly divided into training, validation, and test sets in a ratio of 70%:15%:15%, ensuring that each type of fault and normal state is represented in all three datasets.

[0108] Step 2.3: Convert the divided one-dimensional signal data set into two-dimensional image data, and establish a two-dimensional training set, a two-dimensional validation set, and a two-dimensional test set that correspond to the one-to-one signal data. Use the FO-GADF conversion method in steps 1.2 and 1.3 to convert the one-dimensional signal data into two-dimensional image data.

[0109] Step 2.4: Input the one-dimensional training set and the two-dimensional training set into the dual-path convolutional neural network together. The one-dimensional training set enters the one-dimensional data processing channel and outputs the one-dimensional feature training result. The two-dimensional training set enters the two-dimensional data processing channel and outputs the two-dimensional feature training result. The one-dimensional feature training result and the two-dimensional feature training result are integrated to obtain the fault diagnosis result. The fault diagnosis result and the label of the fault state are compared, the learning parameters are adjusted, and the parameters of the two-path convolutional neural network are updated and retrained to obtain a trained two-path convolutional neural network.

[0110] During model training, the Adam (Adaptive Moment Estimation) optimizer is used to update network parameters. It dynamically adjusts the learning rate based on the first- and second-order moment estimates of the gradient, thereby accelerating convergence and improving stability. The global learning rate is set to 0.002, and a Dropout layer is introduced before the fully connected layer of the network to prevent strong dependencies between neurons. The dropout layer's dropout rate is set to 0.3, and 30% of the neurons are randomly disabled during each training iteration. The ReLU activation function is uniformly used in the activation units of the convolutional and fully connected layers. Its function form is:

[0111]

[0112] The maximum pooling operation is used in the model to reduce the spatial dimension of the feature map, extract local features and enhance the translation invariance of the model.

[0113] Step 2.5: Input the one-dimensional validation set and the two-dimensional validation set together into the trained two-way convolutional neural network. Use the model parameters when the accuracy reaches the highest point and is relatively stable as the parameters of the pre-trained two-way convolutional neural network to form the final model for commutation failure detection.

[0114] Step 3: After processing the fusion feature representation signal, the predicted probability of each fault state is obtained, and the fault type is determined based on the predicted probability. and Connectivity fusion yields a comprehensive feature representation. After the fused feature representation signal is flattened through a fully connected layer, a softmax classifier is used to diagnose and classify commutation fault types. This yields predicted probabilities for various fault states, including bridge arm short circuit (BS), inverter failure (IF), pulse loss (PL), DC side short circuit (DS), and normal state (Normal). The maximum predicted probability for the normal state is then determined to be greater than a preset fault confidence threshold. For example, if the probability of commutation failure bridge arm short circuit (BS) is highest in the output, then a commutation failure fault is determined for that input segment. Alternatively, if the probability of the normal state is highest and does not exceed a preset value, a fault is also considered to have occurred. This often indicates that the converter may not be operating normally due to some factor. Only when the predicted probability of the normal state is highest and exceeds a preset value is a fault considered to have occurred.

[0115] In the fully connected layer and the Softmax layer, the model combines the final high-dimensional features through weighted combination and outputs the predicted probability of each type of fault. Assume that there is an input vector

[0116]

[0117] Where e is the base of the natural logarithm, i=1,2,…,K, W is the weight matrix, x is the input feature, and b is the bias.

[0118] The converter commutation failure fault diagnosis method of the present invention designs two convolutional neural network paths respectively, performs independent feature extraction for different feature sequences, fully establishes and captures the temporal and spatial information of the thyristor converter voltage data, completes the judgment of whether the commutation failure fault occurs and the fault type, and improves the algorithm accuracy and the real-time performance of fault diagnosis.

[0119] During the feature fusion phase, the system introduces a sequence position weighting mechanism, assigning dynamic weights based on the different channels of features in the time series. This effectively enhances the responsiveness to key abnormal signals and improves the model's sensitivity to abnormal values ​​in phase-controlled converters. This optimizes the calculation process, reduces the impact of redundant features on diagnostic results, and improves diagnostic accuracy and real-time performance. Figure 10 The graph shows the changing loss function (loss) of the neural network model during training, including the trends in the training set loss and validation set loss. As can be seen from the figure, the loss value shows a clear and rapid downward trend from epochs 1 to 9, indicating that the model quickly learned the key features of the data in the early stages. Subsequently, between epochs 10 and 14, the rate of loss decline gradually slowed, indicating that the model's learning entered a relatively stable phase. From epochs 15 to 20, the training and validation losses stabilized, indicating that the model has essentially converged, overfitting is not significant, and the training effect is relatively satisfactory. Figure 11 The accuracy curves of the neural network model during training are shown, including the accuracy of the training and validation sets. It can be observed that the accuracy increases rapidly from the first to the ninth round of training, indicating that the model effectively learns the discriminative features of the samples in the early stages. Between the tenth and fourteenth rounds, the rate of increase in accuracy gradually slows, and model performance continues to improve but approaches saturation. From the fifteenth to the twentieth round, the training and validation accuracy remain largely stable, indicating that the model has achieved good generalization capabilities.

[0120] Figure 12 This figure shows the results of a dual-channel convolutional neural network diagnosing an inverter failure (IF) fault. The waveform formed by the first 2000 sampling points indicates a normal converter output voltage. After approximately the 2200th sampling point, the converter fails, and the fault is detected after the 2500th sampling point, indicating an inverter failure. Figure 13 The following figure shows the results of a dual-channel convolutional neural network diagnosing a pulse loss (PL) fault. As can be seen, the first 1000 sampling points correspond to the normal state of the converter output voltage; a fault occurs after approximately the 1300th sampling point, and is accurately diagnosed as a pulse loss fault around the 1500th sampling point. Figure 14 This figure shows the results of diagnosing a bridge arm short circuit (BS) fault using a dual-path convolutional neural network. The figure shows that the first 1000 sampling points are in a normal output state, and then the system fails after approximately the 1100th sampling point, and is diagnosed as a bridge arm short circuit fault at the 1500th sampling point. Figure 15 This figure shows the results of a dual-channel convolutional neural network diagnosing a DC short circuit (DS) fault. As shown in the figure, the system is normal for the first 1500 sampling points, but a fault occurs after approximately the 1600th sampling point. After the fault occurs, a DC short circuit fault is successfully diagnosed at the 2000th sampling point. Figure 16 This figure shows the results of the dual-channel convolutional neural network diagnosis under normal conditions. The figure shows no obvious faults from beginning to end. The model consistently correctly judges the system state and accurately outputs normal conditions, demonstrating the algorithm's stability and reliability under fault-free conditions.

[0121] Results show that the proposed method for diagnosing converter commutation failure can quickly and accurately identify the type of commutation failure, and exhibits stronger discrimination and anti-interference capabilities compared to single-path models. This method effectively addresses the computational complexity of traditional neural network algorithms and significantly improves the efficiency and reliability of real-time diagnosis of commutation failure in phase-controlled converters.

[0122] Example 2

[0123] This embodiment provides a converter commutation failure fault diagnosis system, including:

[0124] The data processing module is used to convert the output voltage of the converter into two-dimensional image data.

[0125] The process of converting the output voltage of the converter into two-dimensional image data includes:

[0126] Processing the output voltage of the converter to obtain one-dimensional signal data;

[0127] Mapping the one-dimensional signal data into a normalized difference sequence, mapping the normalized difference sequence to an angle on the unit circle, obtaining an angle sequence, and obtaining a first-order difference angle sequence;

[0128] A two-dimensional symmetric matrix is ​​constructed based on the first-order difference angle sequence as the two-dimensional image data.

[0129] Among them, the normalized difference sequence for:

[0130]

[0131] in, , one-dimensional signal data , , , ;

[0132] An angle on the unit circle , ;Angle sequence ;

[0133] set up , , get the first-order difference angle sequence of length n-1 ,right and The difference between them is sine-encoded to construct a two-dimensional symmetric matrix FOGADF , for the two-dimensional symmetric matrix FOGADF After visualization, a two-dimensional image is obtained.

[0134] The feature extraction module is used to input the output voltage of the converter and the two-dimensional image data into a pre-trained two-way convolutional neural network. The pre-trained two-way convolutional neural network includes a one-dimensional data processing channel and a two-dimensional data processing channel. The output voltage of the converter enters the one-dimensional data processing channel to obtain a feature representation signal. , the two-dimensional image data enters the two-dimensional data processing channel and obtains the feature representation signal , the feature represents the signal and After fusion, the fused feature representation signal is obtained.

[0135] The output voltage of the converter is sampled and enters the one-dimensional data processing channel. First, the output voltage is convolved and pooled to obtain the first output feature. The first output feature enters three convolution kernels of different scales for feature extraction to obtain the first angle feature, the second angle feature and the third angle feature. The three angle features are sequentially passed through the SE-DRB module, the attention mechanism, and the FG-Bi-LSTM module to obtain the output features. 、 、 , output features 、 、 After fusion, the feature representation signal is obtained .

[0136] The SE-DRB module consists of two 1×1 convolutional layers, a 3×1 convolutional layer, and a SE module. Each angle feature is used as an input feature. , enter the first 1×1 convolution layer to perform channel dimensionality reduction and obtain the output feature , C is the number of channels, L is the length of the time series, and the output feature Enter the 3×1 convolution layer to expand the receptive field and obtain the output features , output features After entering the second 1×1 convolutional layer to restore the original dimension, the output features are obtained , output features The SE module assigns different weights to the channels, broadcasts the weights to each time step, forms the attention weighted output, obtains the weighted features, and converts the input features into Add it element-wise to the weighted feature to get the second output feature.

[0137] Among them, the SE module includes a global average pooling layer and a two-layer fully connected network, and the output feature Enter the global average pooling layer and output the result ,result The channel weights are obtained through a two-layer fully connected network for:

[0138] , ,

[0139] result for: , Indicates the number of channels (feature dimension).

[0140] The two-dimensional image data enters the two-dimensional data processing channel. First, the two-dimensional convolution and pooling operations are performed on the two-dimensional image data to perform preliminary feature extraction to obtain the third output feature. The third output feature is sequentially passed through the convolution kernel, ReLU activation function, CBAM module and FG-Bi-LSTM module to obtain the feature representation signal. .

[0141] The output of the FG-Bi-LSTM module for:

[0142]

[0143]

[0144]

[0145] in, is the weight matrix of the fusion gate, is the bias vector of the fusion gate, represents the concatenation of the forward and backward hidden states, is the sigmoid activation function, Determine the weight of the forward hidden state, Determines the weight of the backward hidden state, ⊙ represents element-wise multiplication.

[0146] The training process of the two-way convolutional neural network includes:

[0147] Collect voltage signal data of various fault conditions during the commutation process and pre-process the voltage signal data to obtain a one-dimensional signal data set. The various fault conditions include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit and normal state;

[0148] The one-dimensional signal data set is divided into a one-dimensional training set, a one-dimensional validation set, and a one-dimensional test set. The one-dimensional training set and the one-dimensional validation set are respectively labeled with the corresponding fault status. The one-dimensional test set is used as the output voltage of the converter to be diagnosed.

[0149] Convert the divided one-dimensional signal data set into two-dimensional image data, and establish a two-dimensional training set, a two-dimensional validation set, and a two-dimensional test set that correspond one-to-one to the one-dimensional signal data;

[0150] The one-dimensional training set and the two-dimensional training set are input into the dual-path convolutional neural network together. The one-dimensional training set enters the one-dimensional data processing channel and outputs the one-dimensional feature training result. The two-dimensional training set enters the two-dimensional data processing channel and outputs the two-dimensional feature training result. The one-dimensional feature training result and the two-dimensional feature training result are fused to obtain the fault diagnosis result. The fault diagnosis result and the label of the fault state are compared, and the parameters of the two-path convolutional neural network are updated and retrained to obtain a trained two-path convolutional neural network.

[0151] The one-dimensional validation set and the two-dimensional validation set are input into the trained two-way convolutional neural network together, and the model parameters that reach the highest accuracy and are relatively stable are used as the parameters of the pre-trained two-way convolutional neural network.

[0152] The fault diagnosis module is used to process the fused feature representation signal to obtain the predicted probability of various fault states and determine the fault type based on the predicted probability.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for diagnosing converter commutation failure, characterized by: Methods include: The output voltage of the converter is converted into two-dimensional image data, and the output voltage of the converter and the two-dimensional image data are input into a pre-trained dual-path convolutional neural network. The pre-trained dual-path convolutional neural network includes a one-dimensional data processing channel and a two-dimensional data processing channel. The output voltage of the converter enters the one-dimensional data processing channel, and convolution and pooling operations are performed on the output voltage to obtain the first output feature. The first output feature enters three convolution kernels of different scales for feature extraction to obtain the first, second, and third angle features. The three angle features pass through the SE-DRB module, the attention mechanism, and the FG-Bi-LSTM module in sequence to obtain the output features respectively. 、 、 , output features 、 、 After fusion, the feature representation signal is obtained ,SE-DRB module includes two 1×1 convolution layers, 3×1 convolution layers, SE module, and the output of FG-Bi-LSTM module for: in, , ; ; is the weight matrix of the fusion gate, is the bias vector of the fusion gate, represents the concatenation of the forward and backward hidden states, is the sigmoid activation function, Determine the weight of the forward hidden state, Determine the weight of the backward hidden state, ⊙ represents element-wise multiplication; The two-dimensional image data enters the two-dimensional data processing channel to obtain the feature representation signal , the feature represents the signal and After fusion, the fused feature representation signal is obtained. After processing the fused feature representation signal, the predicted probability of each fault state is obtained, and the fault type is determined based on the predicted probability.

2. The converter commutation failure fault diagnosis method according to claim 1, characterized in that: The process of converting the output voltage of the converter into two-dimensional image data includes: Processing the output voltage of the converter to obtain one-dimensional signal data; Mapping the one-dimensional signal data into a normalized difference sequence, mapping the normalized difference sequence to an angle on the unit circle, obtaining an angle sequence, and obtaining a first-order difference angle sequence; A two-dimensional symmetric matrix is ​​constructed based on the first-order difference angle sequence as the two-dimensional image data.

3. The converter commutation failure fault diagnosis method according to claim 2, characterized in that: Normalized difference series for: in, , one-dimensional signal data , , , ; An angle on the unit circle , ;Angle sequence ; set up , , get the first-order difference angle sequence of length n-1 ,right and The difference between them is sine-encoded to construct a two-dimensional symmetric matrix FOGADF , for the two-dimensional symmetric matrix FOGADF After visualization, a two-dimensional image is obtained.

4. The method for diagnosing converter commutation failure according to claim 1, wherein: Each angle feature is used as an input feature , enter the first 1×1 convolution layer to perform channel dimensionality reduction and obtain the output feature , C is the number of channels, L is the length of the time series, and the output feature Enter the 3×1 convolution layer to expand the receptive field and obtain the output features , output features After entering the second 1×1 convolutional layer to restore the original dimension, the output features are obtained , output features The SE module assigns different weights to the channels, broadcasts the weights to each time step, forms the attention weighted output, obtains the weighted features, and converts the input features into Add it element-wise to the weighted feature to get the second output feature.

5. The converter commutation failure fault diagnosis method according to claim 4, characterized in that: The SE module includes a global average pooling layer and a two-layer fully connected network, outputting features Enter the global average pooling layer and output the result , the result z gets the channel weight through the two-layer fully connected network for: , ,result for: , Indicates the number of channels.

6. The converter commutation failure fault diagnosis method according to claim 1, characterized in that: The two-dimensional image data enters the two-dimensional data processing channel. First, the two-dimensional convolution and pooling operations are performed on the two-dimensional image data to perform preliminary feature extraction to obtain the third output feature. The third output feature is sequentially passed through the convolution kernel, ReLU activation function, CBAM module and FG-Bi-LSTM module to obtain the feature representation signal. .

7. The converter commutation failure fault diagnosis method according to claim 1, characterized in that: The training process of the two-way convolutional neural network includes: Collect voltage signal data of various fault conditions during the commutation process and pre-process the voltage signal data to obtain a one-dimensional signal data set. The various fault conditions include bridge arm short circuit, inverter failure, pulse loss, DC side short circuit and normal state; The one-dimensional signal data set is divided into a one-dimensional training set, a one-dimensional validation set, and a one-dimensional test set. The one-dimensional training set and the one-dimensional validation set are respectively labeled with the corresponding fault status. The one-dimensional test set is used as the output voltage of the converter to be diagnosed. Convert the divided one-dimensional signal data set into two-dimensional image data, and establish a two-dimensional training set, a two-dimensional validation set, and a two-dimensional test set that correspond one-to-one to the one-dimensional signal data; The one-dimensional training set and the two-dimensional training set are input into the dual-path convolutional neural network together. The one-dimensional training set enters the one-dimensional data processing channel and outputs the one-dimensional feature training result. The two-dimensional training set enters the two-dimensional data processing channel and outputs the two-dimensional feature training result. The one-dimensional feature training result and the two-dimensional feature training result are fused to obtain the fault diagnosis result. The fault diagnosis result and the label of the fault state are compared, and the parameters of the two-path convolutional neural network are updated and retrained to obtain a trained two-path convolutional neural network. The one-dimensional validation set and the two-dimensional validation set are input into the trained two-way convolutional neural network together, and the model parameters when the accuracy reaches the highest point and is stable are used as the parameters of the pre-trained two-way convolutional neural network.

8. The converter commutation failure fault diagnosis method according to claim 1, characterized in that: The three convolution kernels of different scales are 7 7 convolution kernels, 5 5 convolution kernels, 3 3 convolution kernels, the first output feature enters 7 After the 7 convolution kernel performs feature extraction, the first angle feature is obtained, and the first output feature enters 5 5 convolution kernels are used to extract features and obtain the second angle feature. The first output feature enters 3 3. After feature extraction using convolution kernel, the third angle feature is obtained.

9. The converter commutation failure fault diagnosis method according to claim 1, characterized in that: The process of processing the fused feature representation signal to obtain the predicted probability of each fault state includes: after the fused feature representation signal is flattened through a fully connected layer, a Softmax classifier is used to complete the diagnosis and classification of the commutation fault type to obtain the predicted probability of each fault state.

10. The converter commutation failure fault diagnosis system is characterized by: The system includes: A data processing module, used for converting the output voltage of the converter into two-dimensional image data; The feature extraction module is used to input the output voltage of the converter and the two-dimensional image data into a pre-trained two-way convolutional neural network. The pre-trained two-way convolutional neural network includes a one-dimensional data processing channel and a two-dimensional data processing channel. The output voltage of the converter enters the one-dimensional data processing channel, and the output voltage is convolved and pooled to obtain the first output feature. The first output feature enters three convolution kernels of different scales for feature extraction to obtain the first, second, and third angle features. The three angle features pass through the SE-DRB module, the attention mechanism, and the FG-Bi-LSTM module in sequence to obtain the output features. 、 、 , output features 、 、 After fusion, the feature representation signal is obtained The SE-DRB module includes two 1×1 convolutional layers, a 3×1 convolutional layer, a SE module, and the output of the FG-Bi-LSTM module. for: in, , ; ; is the weight matrix of the fusion gate, is the bias vector of the fusion gate, represents the concatenation of the forward and backward hidden states, is the sigmoid activation function, Determine the weight of the forward hidden state, Determine the weight of the backward hidden state, ⊙ represents element-wise multiplication; The two-dimensional image data enters the two-dimensional data processing channel to obtain the feature representation signal , the feature represents the signal and After fusion, the fusion feature representation signal is obtained; The fault diagnosis module is used to process the fused feature representation signal to obtain the predicted probability of various fault states and determine the fault type based on the predicted probability.

Images