A method and system for fault diagnosis of a thyristor converter

Abstract

The application discloses a thyristor converter fault diagnosis method and system, and relates to the technical field of system fault diagnosis. Specifically, the following steps are included: obtaining normal operation parameters and abnormal operation parameters of a superconducting fusion device under a predetermined working condition to form an original data set; preprocessing the original data set to generate a training data set; constructing a superconducting fusion device diagnosis network model according to the training parameters and training the model; obtaining operation parameters of the superconducting fusion device in real time, and predicting whether the superconducting fusion device will fail through the superconducting fusion device diagnosis network model. The application improves the diagnosis efficiency and accuracy of the thyristor converter fault.

Description

Technical Field

[0001] This invention relates to the field of system fault diagnosis technology, and in particular to a method and system for diagnosing faults in thyristor converters. Background Technology

[0002] Thyristor converters are characterized by high output current and high voltage. The power supply system consists of converters, control modules, reactive power compensation, and quench protection subsystems, and is responsible for providing the high power voltage and current required for superconducting fusion device experiments.

[0003] High-power power supplies require long-pulse, high-power operation or are not maintained or repaired in a timely manner, which can lead to certain safety hazards during operation. Potential faults include thyristor open / short circuits, DC reactor short circuits, step-down transformer short circuits, and rectifier transformer short circuits. Severe incidents can result in magnetic energy leakage, equipment shutdown and damage, or spontaneous combustion, causing economic losses and even threatening lives.

[0004] To ensure the safe operation of the power supply and achieve real-time monitoring, diagnosis, and location of faults under various operating conditions, it is first necessary to monitor the operating status of the power supply and rationally arrange current and voltage monitoring points, including selecting the monitoring locations and numbers. Secondly, by processing, analyzing, and optimizing the collected data, the fault status can be effectively, comprehensively, and accurately reflected.

[0005] However, current fault diagnosis technologies are mainly based on models or signal processing. These methods diagnose faults based on changes in rectifier parameters exceeding set limits, resulting in low diagnostic efficiency, low accuracy, and inability to pinpoint the fault location. They also lack means to determine the location of the fault under different operating conditions, leading to increased costs, slow diagnostic speed, and unsatisfactory diagnostic results.

[0006] Therefore, improving the efficiency and accuracy of thyristor converter fault diagnosis has become an urgent technical challenge. Summary of the Invention

[0007] The main objective of this invention is to provide a method and system for diagnosing faults in thyristor converters, aiming to improve the efficiency and accuracy of diagnosing faults in thyristor converters.

[0008] To achieve the above objectives, this invention proposes a fault diagnosis method for thyristor converters, comprising the following steps:

[0009] The normal operating parameters and abnormal operating parameters of the superconducting fusion device under predetermined operating conditions are obtained to form the original dataset;

[0010] The original dataset is preprocessed to generate a training dataset;

[0011] Based on the training parameters, a diagnostic network model for a superconducting fusion device is constructed and trained.

[0012] The operating parameters of the superconducting fusion device are acquired in real time, and the superconducting fusion device diagnostic network model is used to predict whether the superconducting fusion device will malfunction.

[0013] In one embodiment of this application, the predetermined operating condition includes at least one of: rectification operating condition, steady-state operating condition, and inverter operating condition.

[0014] In one embodiment of this application, the original dataset is represented as:

[0015] Phase I i ={I i {j},i∈[A,B,C],j∈[M,N]};

[0016] Where A, B, and C represent one phase of the three-phase current, and current data of all monitoring points at M*N / f time points are acquired for each phase, where f is the sampling frequency of the current data of the monitoring point, and M and N are the number of rows and columns of the current data, respectively.

[0017] In one embodiment of this application, the process further includes:

[0018] The original dataset is packaged and encapsulated into a data package to form the training dataset;

[0019] Each data packet is a dataset for a single industrial frequency cycle. The number of data packets is f*T, and the total number of data packets is 3*M*N / f*T, where T is the time of a single industrial frequency cycle.

[0020] In one embodiment of this application, preprocessing the original dataset includes:

[0021] The data packets are labeled to generate sequence data packets with sequence numbers;

[0022] White noise data with the same latitude as the sequence data packets is generated and added to the sequence data packet set to form an expanded training dataset;

[0023] The expanded training set is normalized and mixed to generate a converged training dataset.

[0024] In one embodiment of this application, generating white noise data with the same latitude as the sequence data packets and adding it to the sequence data packet set to form an expanded training dataset includes:

[0025] By iterating through each data packet in the sequence data packet set, Gaussian white noise with the same dimension as the data packet is added to the sequence data packet set, forming an expanded training dataset. The number of iterations is a multiple of the expanded dataset, and its mathematical expression is:

[0026] w(t)=Arandn(M,N)

[0027] SignalAug{i,1}=dataset+w(t),i∈{3*M*N / f*T}

[0028] Where w(t) is the Gaussian white noise signal, A is the amplitude of the signal, randn(M,N) is a function that generates an M*N dimensional random matrix that follows a standard normal distribution, which is used to generate a Gaussian white noise matrix of size M*N, dataset is the original dataset, and SignalAug is the expanded training dataset.

[0029] In one embodiment of this application, normalizing and mixing the expanded training set to generate a converged training dataset includes:

[0030] Normalize each element in SignalAug, mapping its value to the interval [0,1]. The function expression is as follows:

[0031] SignalAug'=(SignalAug{i,1}-min(SignalAug{i,1})) / (max(SignalAug{i,1})-min(SignalAug{i,1})); i∈{3*M·N / f·T};

[0032] Where SignalAug' is the normalized input data, min(SignalAug{i,1}) is the minimum parameter value for each data packet, and max(SignalAug{i,1}) is the maximum parameter value for each data packet.

[0033] In one embodiment of this application, the diagnostic network model of the superconducting fusion device is constructed using both the Inception and LSTM algorithms.

[0034] In one embodiment of this application, generating a diagnostic network model for a superconducting fusion device specifically includes:

[0035] The Inception module is defined with n parallel branches; the mathematical expression for the i-th branch can be represented as:

[0036] Conv i =Conv(X,W i );

[0037] Wherein, Conv(X,W) i This indicates that a convolution is performed on the input X, using a convolution kernel of W. i Connect the outputs of all branches to obtain the output of the Inception module:

[0038] Y=Concatenate([Pool1,Pool2,...,Pool n ]);

[0039] Here, Concatenate means connecting multiple feature maps together, and Y represents the output.

[0040] This application also discloses a fault diagnosis system for a thyristor converter, including: a thyristor converter;

[0041] The data acquisition unit is used to collect normal operating parameters and abnormal operating parameters of the thyristor converter; and

[0042] The control unit acquires normal operating parameters and abnormal operating parameters collected by the data acquisition unit, and predicts whether the superconducting fusion device will malfunction based on the thyristor converter fault diagnosis method described above.

[0043] By employing the above technical solution and constructing a diagnostic network model for superconducting fusion devices, it is possible to predict in real time whether the device will malfunction, and take corresponding maintenance and repair measures in advance, thereby reducing downtime and maintenance costs. This effectively improves the stability and efficiency of device operation. Simultaneously, the diagnostic network model constructed using machine learning and deep learning algorithms can automatically analyze a large amount of operating parameter data, achieving automated diagnosis of the superconducting fusion device, reducing the influence of human intervention and subjective judgment, and improving the accuracy and consistency of diagnosis. Attached Figure Description

[0044] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings, wherein:

[0045] Figure 1 This is a schematic diagram of the structure of the first embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram showing the accuracy results of the first embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram showing the recall rate results of the first embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following specific embodiments are only used to explain the invention and do not constitute a limitation thereof.

[0049] like Figure 1 As shown, in order to achieve the above objectives, this invention proposes a fault diagnosis method for thyristor converters, comprising the following steps:

[0050] The normal operating parameters and abnormal operating parameters of the superconducting fusion device under predetermined operating conditions are obtained to form the original dataset;

[0051] The original dataset is preprocessed to generate a training dataset;

[0052] Based on the training parameters, a diagnostic network model for a superconducting fusion device is constructed and trained.

[0053] The operating parameters of the superconducting fusion device are acquired in real time, and the superconducting fusion device diagnostic network model is used to predict whether the superconducting fusion device will malfunction.

[0054] Specifically, a fault diagnosis method for a thyristor converter includes the following steps:

[0055] The normal operating parameters and abnormal operating parameters of the superconducting fusion device under predetermined operating conditions are obtained to form the original dataset;

[0056] Under predetermined operating conditions, various operating parameters of the superconducting fusion device are collected in real time using sensors and monitoring equipment, recording relevant parameters under both normal and abnormal operating conditions. These parameters include three-phase current parameters, etc. The acquired normal and abnormal operating parameters are integrated to form the raw dataset. The predetermined operating conditions include at least one of the following: rectification condition, steady-state condition, and inverter condition.

[0057] The original dataset is preprocessed to generate a training dataset;

[0058] Preprocessing of the original dataset includes data labeling, serialization labeling, data augmentation, data standardization, and data shuffling to improve its suitability for training and prediction.

[0059] Based on the training parameters, a diagnostic network model for a superconducting fusion device is constructed and trained.

[0060] A diagnostic network model for superconducting fusion devices is constructed by combining Inception feature extraction with LSTM feature processing and classification. Inception refers to a deep learning model whose core idea is to use multiple convolutional kernels of different sizes to process input data in parallel and then concatenate their outputs. This parallel processing can capture features at different scales, giving the model a better receptive field and thus improving its performance in image recognition tasks.

[0061] The Inception model introduces "Inception modules," which consist of multiple parallel convolutional layers, including kernels of different sizes, 1x1 convolutional layers, and 1x1 convolutional layers followed by 3x3 convolutional layers. These parallel convolutional layers process the input data and extract features at different scales, and then concatenate their outputs to form a richer feature map. By using the stacking of Inception modules and global average pooling, the Inception model can maintain high model performance while reducing the number of parameters.

[0062] LSTM (Long Short-Term Memory) is a recurrent neural network implemented using long short-term memory. Its core structure includes the following parts:

[0063] Input Gate: Determines which information will be updated to the cell state. It controls the importance of the input using the sigmoid activation function. Forget Gate: Determines which older information will be forgotten. It controls the degree of forgetting in the cell state using the sigmoid activation function. Cell State: Serves as the memory unit of the LSTM network, responsible for storing and transmitting information. The cell state is updated at each time step, and information can be added or removed depending on the control of the input, forget, and output gates. Output Gate: Determines which information from the cell state will be output. It controls the importance and scope of the output using the sigmoid and tanh activation functions.

[0064] A diagnostic network model for superconducting fusion devices is constructed by combining Inception feature extraction with LSTM feature processing and classification. Since fusion devices typically involve signals and images at multiple scales, the parallel convolutional layers of the Inception model can capture features at different scales, providing more comprehensive information and improving the accuracy of data processing. The LSTM model can handle and model long-term dependencies in sequence data. In the diagnosis of superconducting fusion devices, signals and data often exhibit temporal correlations, and LSTM can effectively capture these temporal dependencies, improving the model's accuracy. The LSTM model has the ability to accept variable-length inputs, processing time-series data of varying lengths without requiring fixed-length inputs. This allows the model to adapt to superconducting fusion device data with different sampling frequencies or time steps. The LSTM model can remember and utilize previous contextual information to understand and predict the current input, thereby improving the accuracy of the diagnostic results.

[0065] The operating parameters of the superconducting fusion device are acquired in real time, and the superconducting fusion device diagnostic network model is used to predict whether the superconducting fusion device will malfunction.

[0066] During the actual operation of the superconducting fusion device, its operating parameters are acquired in real time, and a pre-trained diagnostic network model is used for fault prediction. The model analyzes the input operating parameters and generates fault prediction results, indicating whether a fault may occur in the superconducting fusion device.

[0067] By employing the above technical solution and constructing a diagnostic network model for superconducting fusion devices, it is possible to predict in real time whether the device will malfunction, and take corresponding maintenance and repair measures in advance, thereby reducing downtime and maintenance costs. This effectively improves the stability and efficiency of device operation. Simultaneously, the diagnostic network model constructed using machine learning and deep learning algorithms can automatically analyze a large amount of operating parameter data, achieving automated diagnosis of the superconducting fusion device, reducing the influence of human intervention and subjective judgment, and improving the accuracy and consistency of diagnosis.

[0068] In one embodiment of this application, the predetermined operating condition includes at least one of: rectification operating condition, steady-state operating condition, and inverter operating condition.

[0069] Specifically, the rectification mode refers to the operating condition in a superconducting fusion device where the power system converts AC power into DC power. Under rectification mode, the power system uses electronic devices to convert AC power into DC power, supplying the DC power required by the superconducting fusion device.

[0070] Steady-state operating conditions refer to the operating conditions of a superconducting fusion device under normal operating conditions. Under steady-state operating conditions, all components and subsystems of the superconducting fusion device are in a stable working state, and the output plasma maintains a stable and continuous fusion reaction.

[0071] Inverter operation refers to the mode in which the power system of a superconducting fusion device converts DC power to AC power. In inverter mode, the power system uses an inverter to convert DC power to AC power to meet the needs of AC-driven equipment in the superconducting fusion device, such as the heating system and magnetic control system.

[0072] By adopting the above technical solution and separating the operating conditions, precise control can be performed for each condition to meet the requirements of specific working states, thereby improving the performance and efficiency of the device. Separating the operating conditions can help detect and diagnose potential faults or abnormalities under specific operating conditions, avoiding the masking or confusion of faults.

[0073] In one embodiment of this application, the original dataset is represented as:

[0074] Phase I i ={I i {j},i∈[A,B,C],j∈[M,N]};

[0075] Where A, B, and C represent one phase of the three-phase current, and current data of all monitoring points at M*N / f time points are acquired for each phase, where f is the sampling frequency of the current data of the monitoring point, and M and N are the number of rows and columns of the current data, respectively.

[0076] Specifically, the original dataset is represented as: Phase I i ={I i {j}, i∈[A,B,C],j∈[M,N]}; where I represents the current value, I i This represents the current value of a single phase in a three-phase current system. M rows and N columns of current data are acquired from each monitoring point. This current data is collected at certain time intervals with a sampling frequency of f. Therefore, within a given time period, a total of M multiplied by N divided by f current data points at different times will be acquired.

[0077] By employing the above technical solution, dividing the dataset into three phases of current values ​​and indexing them according to the location of monitoring points (M rows and N columns), the organizational structure of the current data can be clearly represented. This representation makes data storage and access more intuitive and convenient. Furthermore, placing current data from the same phase in the same set helps maintain the correlation between adjacent data points. This allows for better utilization of the temporal relationships between adjacent data points in subsequent analysis and processing tasks, such as time series analysis or feature extraction.

[0078] In one embodiment of this application, the process further includes:

[0079] The original dataset is packaged and encapsulated into a data package to form the training dataset;

[0080] Each data packet is a dataset for a single industrial frequency cycle. The number of data packets is f*T, and the total number of data packets is 3*M*N / f*T, where T is the time of a single industrial frequency cycle.

[0081] Specifically, before preprocessing the original dataset, the original data is packaged and encapsulated into data packets to form the training dataset. Each data packet contains data for one complete industrial frequency cycle, with the number of data points being f*T. A total of 3*M*N / f*T data packets will be generated, where 3*M*N represents the total amount of data in the original dataset.

[0082] Specifically, the original dataset is divided according to power frequency cycles, so that each data packet corresponds to a complete power frequency cycle. This division preserves the temporal relationships within the power frequency cycle, which helps the model learn and understand temporal features. At the same time, packaging the data from each power frequency cycle into a single data packet ensures that the data within each packet has consistent frequency and phase information. This helps improve the stability and accuracy of the model.

[0083] In one embodiment of this application, preprocessing the original dataset includes:

[0084] The data packets are labeled to generate sequence data packets with sequence numbers;

[0085] White noise data with the same latitude as the sequence data packets is generated and added to the sequence data packet set to form an expanded training dataset;

[0086] The expanded training set is normalized and mixed to generate a converged training dataset.

[0087] Specifically, preprocessing the original dataset includes:

[0088] For each data packet in the original dataset, a sequence number is labeled to generate a sequence of data packets with sequence numbers. The sequence number can be an index number or other identifier used to uniquely identify each data packet.

[0089] White noise data with the same dimension as the sequence data packets is generated and added to the sequence data packet set to expand the size of the training dataset, forming an expanded training dataset. The white noise data is a randomly generated Gaussian white noise signal with a uniform power spectral density across all frequencies, meaning its energy is uniformly distributed in the frequency domain. The energy of the noise signal is equal at any frequency, without any enhancement or attenuation at any specific frequency.

[0090] Normalization is performed on the expanded training set to ensure that the numerical ranges of different features are consistent. Common normalization methods include min-max scaling and standardization. Min-max scaling scales the data to a specified range (e.g., 0 to 1), while standardization makes the mean of the data 0 and the variance 1.

[0091] The normalized, expanded training set is shuffled, meaning the order of the data samples is randomly shuffled. Shuffling reduces order correlation in the training dataset and improves the model's generalization ability.

[0092] After normalization and mixing, the final converged training dataset is formed. This dataset contains the original data packets with sequence numbers and data packets with added white noise, as well as samples after normalization and mixing.

[0093] Its manifestations are as follows:

[0094]

[0095] Where k = 1, 2, 3 represent the three operating conditions of rectification, steady state, and inversion, respectively, and T ij ,i∈[A,B,C],j∈[a,b,c] represent the open circuit of the thyristors on the positive and negative terminals of the thyristor rectifier bridge arm.

[0096] By employing the above technical solutions, adding sequence numbers to sequential data packets facilitates packet tracking and management, ensuring data integrity and consistency. Adding white noise data increases the diversity of the training dataset, improving the model's robustness to noise and its generalization ability. Normalization ensures consistent numerical ranges for different features, preventing certain features from dominating model training and improving model stability and convergence speed. By rearranging the order of data samples, order correlations in the dataset are reduced, allowing the model to better learn the relationships between features and improving its generalization ability. After preprocessing, the generated convergent training dataset contains more diverse data samples with better data distribution and balance, contributing to improved model generalization ability and performance.

[0097] In one embodiment of this application, generating white noise data with the same latitude as the sequence data packets and adding it to the sequence data packet set to form an expanded training dataset includes:

[0098] By iterating through each data packet in the sequence data packet set, Gaussian white noise with the same dimension as the data packet is added to the sequence data packet set, forming an expanded training dataset. The number of iterations is a multiple of the expanded dataset, and its mathematical expression is:

[0099] w(t)=Arandn(M,N)

[0100] SignalAug{m,1}=dataset+w(t),m∈{3*M*N / f*T}

[0101] Where w(t) is the Gaussian white noise signal, A is the amplitude of the signal, randn(M,N) is a function that generates an M*N dimensional random matrix that follows a standard normal distribution, which is used to generate a Gaussian white noise matrix of size M*N, dataset is the original dataset, and SignalAug is the expanded training dataset.

[0102] Specifically, by iterating through each data packet in the sequence data packet set, and generating a Gaussian white noise signal with the same dimension as the data packet according to a given multiple, the signal is added to the sequence data packet set to form an expanded training dataset.

[0103] For each packet in the sequence of packets, perform the following steps.

[0104] The function `randn(M,N)` generates a random matrix of size M×N, whose elements follow a standard normal distribution, representing a Gaussian white noise signal. The dimensions of this random matrix are consistent with the dimensions of the data packets to ensure element-wise addition with the data packets.

[0105] The generated Gaussian white noise signal is multiplied by an amplitude factor A to control the intensity of the noise signal. This amplitude factor can be set according to specific needs to adapt to different noise levels and application scenarios.

[0106] The generated Gaussian white noise signal is added element-wise to the original data packet to obtain the expanded data packet. This addition is performed element-wise, meaning data at corresponding positions is added to ensure dimensionality consistency.

[0107] The number of iterations required is determined based on a specified multiple. Each iteration generates a new Gaussian white noise signal, which is added to the original data packet to form an expanded data packet.

[0108] After a specified number of iterations, the resulting set of expanded data packets becomes the expanded training dataset.

[0109] By employing the above technical solution, the addition of Gaussian white noise expands the training dataset, increasing the diversity of data samples. This helps the model better learn and adapt to data under different noise environments. Introducing Gaussian white noise increases the noise level of the dataset, thereby improving the model's robustness. The model is exposed to more noisy samples during training, which helps improve its ability to handle noise.

[0110] In one embodiment of this application, normalizing and mixing the expanded training set to generate a converged training dataset includes:

[0111] Normalize each element in SignalAug, mapping its value to the interval [0,1]. The function expression is as follows:

[0112] SignalAug'=(SignalAug{i,1}-min(SignalAug{i,1})) / (max(SignalAug{i,1})-min(SignalAug{i,1})); m∈{3*M·N / f·T};

[0113] Where SignalAug' is the normalized input data, min(SignalAug{i,1}) is the minimum value of the running parameters for each data packet, and max(SignalAug{i,1}) is the maximum value of the running parameters for each data packet.

[0114] By employing the above technical solution, normalization maps each element value in the data packet to the interval [0,1], eliminating dimensional differences between different data packets. This helps the model learn and process data better, improving the model's convergence speed and stability. Normalization scales the data packet based on its minimum and maximum values, ensuring that all element values ​​fall within the same numerical range, thus avoiding the influence of numerical range on model training. The shuffling operation disrupts the order of the data packets, increasing the randomness and diversity of the data. This helps the model learn the data distribution and characteristics better, improving its generalization ability.

[0115] In one embodiment of this application, the diagnostic network model of the superconducting fusion device is constructed using both the Inception and LSTM algorithms.

[0116] Specifically, model training consists of five main parts: Inception feature extraction, LSTM feature processing and classification, network construction, setting training parameters, and network training. Inception feature extraction extracts multi-scale features of the input signal through multiple convolutional and pooling layers, and processes them through dimensionality reduction and non-linear activation functions.

[0117] The LSTM neural network module utilizes long short-term memory units to model and process the extracted feature sequences, and achieves classification tasks through appropriate activation functions. The following is a description of the complete network structure:

[0118] 1) Input layer: Sequence input layer, with the input dimension set to 3.

[0119] 2) The first convolutional layer: has 3 filters, each with a filter size of 1, and is filled with "same".

[0120] 3) ReLU activation function layer: introduces nonlinear properties.

[0121] 4) DropoutLayer: Used to prevent overfitting in neural networks. During training, it randomly sets the output of some neurons to zero, thereby reducing the dependencies between neurons and increasing the network's generalization ability.

[0122] 5) The second convolutional layer: has 3 filters, each with a filter size of 1 and padding mode of "same".

[0123] 6) ReLU activation function layer: Introduces nonlinear properties.

[0124] 7) DropoutLayer: Used to prevent overfitting in neural networks. During training, it randomly sets the output of a portion of neurons to zero, thereby reducing the dependencies between neurons and increasing the network's generalization ability.

[0125] 8) The second convolutional layer has 3 filters, each with a filter size of 1 and a padding method of "same".

[0126] 9) ReLU activation function layer: Introduces nonlinear properties.

[0127] 10) DropoutLayer: Used to prevent overfitting in neural networks. During training, it randomly sets the output of some neurons to zero, thereby reducing the dependencies between neurons and increasing the network's generalization ability.

[0128] 11) Three Inception modules: Each module consists of three convolutional layers for more complex feature extraction.

[0129] 12) Concatenation Layer: A layer in deep learning used to connect or concatenate feature maps from multiple inputs along a certain dimension. Concatenation layers are commonly used to build neural network models with multiple branches or multiple inputs.

[0130] 13) BiLSTM layer: has 66 hidden units.

[0131] 14) DropoutLayer: Used to prevent overfitting in neural networks. During training, it randomly sets the output of some neurons to zero, thereby reducing the dependencies between neurons and increasing the network's generalization ability.

[0132] 15) Fully connected layer: connects all input and output units.

[0133] 16) Softmax layer: used for probability output in classification problems.

[0134] 17) Classification layer: Classifies the network output.

[0135] 18) These layers are connected in sequence to form a complete neural network structure.

[0136] The training process is as follows:

[0137] (1) The optimizer is Adam, which is an optimization algorithm commonly used to train neural networks.

[0138] (2) The maximum number of training rounds is 300, which means that the neural network will be trained on the entire dataset for 300 rounds.

[0139] (3) The size of each mini-batch is 150, which means that 150 samples are selected from the training data for training in each iteration.

[0140] (4) The initial learning rate is 0.005, which means the learning rate used in the initial stage of training.

[0141] (5) The gradient threshold is set to 1 to help stabilize the training process.

[0142] (6) The operating environment is set to automatic selection.

[0143] (7) The training progress graphical display interface is opened.

[0144] (8) Output training progress information.

[0145] Next, the neural network structure is defined and constructed using the Inception and LSTM algorithms.

[0146] The neural network uses three one-dimensional convolutional layers (Convolution1DLayer) to extract features from the sequence data, and then these features are passed to an LSTM layer for further processing of the sequence data. The following is the workflow of this model:

[0147] One-dimensional convolutional layers: These convolutional layers are used to extract features from the input sequence. Convolutional operations can detect local patterns and features in the input sequence. Each convolutional layer consists of multiple convolutional kernels, each used to detect a different feature. Convolutional layers generate new feature representations by sliding the convolutional kernels across the sequence and performing convolutional operations. These convolutional layers gradually reduce the sequence length while increasing the dimensionality of the features. By adding LeakyReluLayer rectified linear units between the convolutional layers as activation functions, the network is helped to learn complex features.

[0148] LSTM Layer: Once features are extracted and the sequence length is reduced, these features are passed to the LSTM layer. An LSTM layer is a recurrent neural network layer used to capture long-term dependencies in a sequence. The internal structure of an LSTM layer includes forget gates, input gates, cell states, and output gates, allowing it to efficiently process information at different time steps.

[0149] Output layer: One or more fully connected layers are added after the LSTM layer to generate the final output.

[0150] The entire model will be trained using the backpropagation algorithm to optimize the weights of the convolutional kernels, LSTM layers, and fully connected layers, enabling the model to learn from the input data and extract useful features to adapt to the requirements of the task.

[0151] This architecture, which combines convolutional layers, non-linear activation, Dropout, and LSTM layers, is very useful when processing sequence data because it can simultaneously capture local features, long-term dependencies, and reduce overfitting during training.

[0152] The Inception module contains three layers, each employing the annealing algorithm and shortcut connections to improve the network's training and optimization process. Shortcut connections are a connection method that directly connects the outputs of certain layers to the inputs of subsequent layers, allowing for more efficient information transfer during information propagation and gradient backpropagation.

[0153] In deep neural networks, as network depth increases, gradients may vanish or explode during backpropagation, making the network difficult to train. Introducing shortcut connections can alleviate these problems and improve the training effectiveness and performance of the network.

[0154] In the annealing algorithm, shortcut connections are implemented by adding the output of some layers to the input of subsequent layers. The advantage of this is that detailed information from lower layers can be directly passed to higher layers, allowing the higher layers to better understand and utilize the feature representations of the lower layers. This direct connection provides additional paths, making information flow more easily and gradients propagate more readily, thereby accelerating the network's convergence and training process.

[0155] Shortcut connections have been widely used in some successful network architectures, such as Residual Networks and DenseNet. By introducing shortcut connections, they effectively solve the problems of vanishing gradients and training difficulties in deep networks, thereby improving the performance and expressive power of the networks.

[0156] Adding an LSTM layer to a model can be done by specifying the input dimension, the number of hidden units, and the activation function.

[0157] As needed, add fully connected layers, Dropout layers, etc., to enhance the network's expressive power and prevent overfitting.

[0158] Define the input and output layers: Determine the shape of the input layer, i.e., the dimensions of the input data. This depends on the number and structure of the features in the dataset. Define the shape of the output layer, i.e., the number of categories in the classification task. Generally, a fully connected layer is used as the output layer, and an appropriate activation function, Softmax, is chosen.

[0159] Connect the network layers: connect the input layer to the first hidden layer, the hidden layer to the next layer, and so on, until the last hidden layer. Ensure that the input dimension of each layer matches the output dimension of the layer above it.

[0160] Compile the network model: Configure the model's optimizer Adam and specify the learning rate and other parameters. Choose an appropriate loss function, such as cross-entropy loss (for classification tasks) or mean squared error loss (for regression tasks).

[0161] Choose evaluation metrics, such as accuracy, precision, and recall.

[0162] Precision, recall, and F1 score are used as performance validation metrics, as shown in the table. Accuracy refers to the number of correctly classified data, and it is a very commonly used evaluation metric. However, when the sample distribution of each class is highly uneven, high accuracy does not necessarily mean a good classification algorithm. Precision is the proportion of actual positive data to the predicted positive data. Recall is the proportion of predicted positive data to the actual positive data. Precision and recall are contradictory metrics. When precision is high, recall tends to be low, and when precision is low, recall tends to be high. To comprehensively consider these two metrics, the F1 score is proposed, which is calculated as the harmonic mean of the two metrics.

[0163] The specific formulas for calculating accuracy, precision, recall, and F1 score are as follows:

[0164]

[0165] Combining the results of single-switch faults and multi-switch faults, the accuracy in 22*3 sets of test data is as follows: Figure 2 As shown, the recall rate is as follows: Figure 3 As shown.

[0166] In the switching model, the minimum precision of the Tc-C phase is 0.9914. The minimum precision, recall, and F1 score are 0.9841, 0.9673, and 0.9784, respectively, indicating good model performance. The test set error is 1.652893e-02. The test set accuracy is 9.834711e+01%.

[0167] By adopting the above technical solution, the model can be built and trained, and fault prediction can be achieved.

[0168] In one embodiment of this application, generating a diagnostic network model for a superconducting fusion device specifically includes:

[0169] The Inception module is defined with n parallel branches; the mathematical expression for the i-th branch can be represented as:

[0170] Conv i =Conv(X,W i );

[0171] Wherein, Conv(X,W) i This indicates that a convolution is performed on the input X, using a convolution kernel of W. i Connect the outputs of all branches to obtain the output of the Inception module:

[0172] Y=Concatenate([Pool1,Pool2,...,Pool n ]);

[0173] Here, Concatenate means connecting multiple feature maps together, and Y represents the output.

[0174] Specifically, using convolution kernel W i Perform a convolution operation on the input X to obtain the output Conv of branch i. i Convolution operations can use a sliding window to perform local perception on the input and extract feature information from the input.

[0175] Conv the output of all branches i These are connected together to form the output Y of the Inception module. The connection operation can concatenate the feature maps of the branches along the depth dimension to obtain multi-scale feature representations.

[0176] In convolutional operations, kernels of different sizes can be used to extract feature information at different scales. In the output connection, the `Concatenate` function is used to concatenate multiple feature maps along the depth dimension. For each parallel branch in the Inception module, convolutional operations are performed sequentially to obtain the branch's output. Then, the outputs of all branches are concatenated to form the module's output.

[0177] By employing the above technical solution and using convolutional kernels of different sizes, the Inception module can extract feature information from multiple scales of the input. This helps the model to effectively perceive and extract objects or features at different scales.

[0178] This application also discloses a fault diagnosis system for a thyristor converter, including: a thyristor converter;

[0179] The data acquisition unit is used to collect normal operating parameters and abnormal operating parameters of the thyristor converter; and

[0180] The control unit acquires normal operating parameters and abnormal operating parameters collected by the data acquisition unit, and predicts whether the superconducting fusion device will malfunction based on the thyristor converter fault diagnosis method described above.

[0181] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for fault diagnosis of a thyristor converter, characterized in that, Includes the following steps: The normal operating parameters and abnormal operating parameters of the superconducting fusion device under predetermined operating conditions are obtained to form the original dataset; The original dataset is preprocessed to generate a training dataset; Based on the training parameters, a diagnostic network model for a superconducting fusion device is constructed and trained. Real-time acquisition of operating parameters of the superconducting fusion device; prediction of whether the superconducting fusion device will malfunction using the diagnostic network model of the superconducting fusion device. The original dataset is represented as follows: Phase I i = { I i {j},i∈[A,B,C],j∈[M,N]}; Where A, B, and C represent one phase of the three-phase current, and M is obtained for each phase. The current data of all monitoring points at N / f time points, where f is the sampling frequency of the current data of the monitoring points, and M and N are the number of rows and columns of the current data, respectively. The process before preprocessing the original dataset also includes: The original dataset is packaged and encapsulated into a data package to form the training dataset; Each data packet is a dataset representing a single industrial frequency cycle, and the number of data packets is f. T, the number of data packets is 3 M N / f T, where T is the time of a single power frequency cycle; Preprocessing the original dataset includes: The data packets are labeled to generate sequence data packets with sequence numbers; White noise data with the same latitude as the sequence data packets is generated and added to the sequence data packet set to form an expanded training dataset; The expanded training set is normalized and mixed to generate a converged training dataset. Generating white noise data with the same latitude as the sequence data packets and adding it to the sequence data packet set to form an expanded training dataset includes: By iterating through each data packet in the sequence data packet set, Gaussian white noise with the same dimension as the data packet is added to the sequence data packet set, forming an expanded training dataset. The number of iterations is a multiple of the expanded dataset, and its mathematical expression is: w(t) = A randn(M, N) SignalAug{m,1}=dataset+ w(t),m∈{3 M N / f T} Where w(t) is the Gaussian white noise signal, A is the signal amplitude, and randn(M, N) is the generator that produces M signals following a standard normal distribution. A function for generating an N-dimensional random matrix, which is used to generate a matrix of size M. N is a Gaussian white noise matrix, dataset is the original dataset, and SignalAug is the expanded training dataset.

2. The thyristor converter fault diagnosis method as described in claim 1, characterized in that, The predetermined operating conditions include at least one of the following: rectification operating conditions, steady-state operating conditions, and inverter operating conditions.

3. The thyristor converter fault diagnosis method as described in claim 1, characterized in that, The expanded training set is normalized and mixed to generate a convergent training dataset, including: Normalize each element in SignalAug, mapping its value to the interval [0, 1]. The function expression is as follows: SignalAug’ = (SignalAug{i, 1} - min(SignalAug{i, 1})) / (max(SignalAug{i,1}) - min(SignalAug{i, 1})); m∈{3 M·N / f·T}; Where SignalAug' is the normalized input data, min(SignalAug{i, 1}) is the minimum value of the running parameters for each data packet, and max(SignalAug{i, 1}) is the maximum value of the running parameters for each data packet.

4. The thyristor converter fault diagnosis method as described in claim 1, characterized in that, The diagnostic network model for the superconducting fusion device is constructed using both the Inception and LSTM algorithms.

5. The thyristor converter fault diagnosis method as described in claim 4, characterized in that, The generation of a diagnostic network model for a superconducting fusion device specifically includes: The Inception module is defined with n parallel branches; the mathematical expression for the i-th branch can be represented as: Conv i = Conv(X, W i ) ; Wherein, Conv(X, W) i ) indicates that a convolution is performed on the input X, using a convolution kernel of W. i Connect the outputs of all branches to obtain the output of the Inception module: Y = Concatenate([Pool1, Pool2, ..., Pool n ]) ; Here, Concatenate means connecting multiple feature maps together, and Y represents the output.

6. A fault diagnosis system for a thyristor converter, characterized in that, include: Thyristor converter; The data acquisition unit is used to collect normal operating parameters and abnormal operating parameters of the thyristor converter; as well as The control unit acquires normal operating parameters and abnormal operating parameters collected by the data acquisition unit, and predicts whether the superconducting fusion device will malfunction according to the thyristor converter fault diagnosis method according to any one of claims 1 to 5.

Images