Converter fault current identification method based on knowledge and data double driving

Abstract

The application relates to a knowledge and data double-driven converter fault inrush current identification method, and belongs to the technical field of power systems, which comprises the following steps: step one: collecting one-dimensional original current signals and performing pretreatment; step two: constructing a knowledge driving layer and extracting 8-dimensional time-frequency physical features; step three: constructing a data driving layer and extracting deep features; and step four: based on an attention mechanism double-path feature fusion and classification model, fusing the 8-dimensional time-frequency physical features and the deep features, and outputting three classification probability results of the current signal belonging to a normal state, a fault inrush current or a non-fault inrush current.

Description

Technical Field

[0001] This invention belongs to the field of power system technology and relates to a method for identifying faulty inrush current in converters based on a dual-drive approach of knowledge and data. Background Technology

[0002] The converter transformer is a core component in a flexible DC transmission system (MMC-HVDC). When a single-phase ground fault occurs on the valve side of the converter transformer, a fault inrush current containing a large amount of DC bias and second harmonics will be generated due to the blocking of the MMC submodule and the unidirectional conduction characteristics of the diodes. This inrush current not only has a severely distorted and extremely asymmetrical waveform, but it also easily causes the converter transformer differential protection to malfunction during faults outside the protection zone and fail to operate during faults within the protection zone, seriously threatening the safe and stable operation of the DC power grid.

[0003] Currently, methods for identifying transformer inrush current mainly fall into two categories: One approach is the traditional method based on physical characteristics, such as second harmonic braking, discontinuity angle principle, and waveform symmetry. These methods have clear physical meanings, but due to the complex and variable waveforms of faulty inrush currents, a single dimension of characteristics is insufficient to fully characterize their patterns. Furthermore, the setting of the criterion threshold is highly dependent on human experience, and these methods are prone to failure under different operating conditions, such as high transition resistance and strong noise, lacking robustness.

[0004] Second, there are purely data-driven deep learning methods, such as using standard convolutional neural networks (CNNs) to directly process current waveforms. While these methods have strong feature extraction capabilities, they suffer from a typical black-box problem, lacking physical interpretability. Furthermore, when processing multiphase currents, standard convolutions force cross-channel fusion, which not only easily masks subtle time-domain features such as discontinuity angles and spikes within a single channel, but also leads to a quadratic increase in the number of network parameters, resulting in excessively high computational complexity and making it difficult to meet the real-time requirements of relay protection. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for identifying faulty inrush current in converters based on a dual-driven approach of knowledge and data. Addressing the difficulty of threshold tuning in traditional physical methods, this invention extracts eight-dimensional time-frequency features with clear physical meaning through wavelet transform in the knowledge-driven layer. To address the problem of large parameter counts and obscuring subtle features in pure data-driven methods, this invention designs a lightweight, deeply separable convolutional network in the data-driven layer to extract local morphological distortions. Finally, an attention mechanism is used to fuse the two approaches. This invention aims to achieve both physical interpretability and deep feature mining capabilities, enabling accurate differentiation between faulty inrush current, non-faulty inrush current, and normal states, thus ensuring the stable operation of DC transmission systems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying faulty inrush current in converter transformers based on a dual-driven approach of knowledge and data includes the following steps: Step 1: Acquire the one-dimensional raw current signal and perform preprocessing; Step 2: Construct a knowledge-driven layer and extract 8-dimensional time-frequency physical features; Step 3: Construct a data-driven layer and extract deep features; Step 4: The dual-path feature fusion and classification model based on the attention mechanism fuses the 8-dimensional time-frequency physical features and depth features, and outputs a three-classification probability result indicating whether the current signal belongs to a normal state, a faulty inrush current, or a non-faulty inrush current.

[0007] Furthermore, step one specifically includes the following steps: Collect three-phase current signals from each side of the converter transformer to construct a one-dimensional current time series; Data of a fixed window length before and after the fault occurred were extracted as input samples and subjected to zero-mean standardization to eliminate the influence of data units under different operating conditions.

[0008] Furthermore, step two specifically includes the following steps: By leveraging the multi-resolution characteristics of discrete wavelet transform, physical criteria reflecting the fault-related inrush mechanism are extracted in a targeted manner. S21: Perform a 3-level discrete wavelet transform on the preprocessed one-dimensional current signal, using Sym2 as the wavelet basis function; Discretize the parameters:

[0009]

[0010] In the formula: j, k∈Z are discretization ordinal numbers.

[0011] Substituting the above parameters into the wavelet basis expression, we can obtain the discretized wavelet basis functions:

[0012] Therefore, the fault signal x ( t The discrete wavelet transform of ) can be defined as follows:

[0013] The analysis of fault transient signals all employs binary discrete wavelet transform, taking the scale basis. s 0=2, translation step size τ When 0=1, the discrete wavelet transform of the fault signal can be simplified to:

[0014] S22: Extract the third-level high-frequency detail component D3 after wavelet decomposition; S23: Calculate and extract 8 key physical feature parameters for the D3 component from both time and frequency domains, including 4 time domain features: mean, standard deviation, skewness, and kurtosis, and 4 frequency domain features: peak frequency, average power frequency, power spectral entropy, and total energy. S24: The eight key physical feature parameters are Z-score standardized to construct an 8-dimensional feature vector.

[0015] Furthermore, in step three, a lightweight one-dimensional convolutional neural network based on depthwise separable convolutions is constructed to automatically learn deep morphological features, specifically including the following steps: S31: Initial large-scale convolution: Perform large-scale convolution operation on the original signal, which is equivalent to low-pass filtering, filtering out high-frequency white noise and preserving the DC bias trend; To design a network structure suitable for fault inrush detection, it is necessary to clarify the relationship between the input matrix size h1×m1×d1, the convolution kernel size a×b, the number of convolution kernels K, the stride S, and the output matrix size h2×m2×d2. The relationship is as follows:

[0016] S32: Deep Convolution: Assigns an independent one-dimensional time-domain convolution kernel to each phase current input channel, slides only in the time axis direction, does not perform cross-channel fusion, and captures local micro-distortions within a single channel; Let the height of the output feature map be... h out The number of parameters in depthwise convolution P dw Computational complexity F dw They are respectively:

[0017]

[0018] S33: Pointwise convolution: reconstructs the spatial coupling relationship between different frequency components or different phase sequence currents; Let the number of output channels be C out The number of parameters in pointwise convolution is... P pw Computational complexity F pw for:

[0019]

[0020] S34: Global Average Pooling: Extracts global features with translation invariance through a global average pooling layer, and finally outputs a 256-dimensional deep convolutional feature vector after dimensionality reduction.

[0021] Furthermore, step four specifically includes the following steps: S41: The 8-dimensional mechanism features output by the data-driven layer and the 256-dimensional depth features output by the knowledge-driven layer are concatenated to form a 264-dimensional composite feature vector. S42: Input the composite vector into the fully connected layer of the attention mechanism, and use the explicit physical meaning of the physical features to guide the network to dynamically weight the deep features;

[0022] In the formula, x l Indicates the first l-1 After the layer undergoes a fully connected operation, it is mapped to the 1st layer. l The first layer j The output vector of each neuron n express l-1 The number of neurons in a layer. Indicates the first l-1 The first layer i The nth neuron is mapped to the nth neuron through a fully connected operation. l The first layer j The weights used by each neuron The input value for a fully connected circuit is the nth... l-1 The first layer i Number of neurons, Indicates the first l-1 Layers are fully connected up to the first layer. l The first in the layer j The bias used for each neuron. After this linear combination, nonlinearity is introduced through the activation function and finally passed to the output layer.

[0023] S43: After passing through the ReLU activation function, it is input to the Softmax output layer, which outputs the three-class probability results of whether the current signal belongs to the normal state, fault inrush current or non-fault inrush current.

[0024] The beneficial effects of this invention are as follows: (1) Extremely high recognition accuracy and reliability: This invention integrates 8-dimensional explicit physical time-frequency mechanism features and 256-dimensional deep morphological features, and introduces an attention mechanism for weight compounding, which completely avoids the problem of easy failure of manually tuned thresholds in traditional methods, and solves the black box blind spot of pure deep learning algorithms. Experiments have shown that the method has an accuracy of 100% in identifying faulty inrush flow and an overall macro average score of 0.9901. (2) High computational efficiency, meeting the requirements of relay protection speed: The data-driven layer innovatively adopts depth-separable convolution instead of standard convolution. Depth convolution is responsible for temporal abrupt changes, and pointwise convolution is responsible for multi-channel spatial correlation. This decoupling design removes a large number of invalid feature redundancies, reducing the number of network parameters and computational complexity to about 1 / N of standard convolution, which greatly improves computational efficiency. (3) Strong noise resistance and good generalization ability: The wavelet transform of the physical layer and the initial large-scale convolution of the data layer both have excellent filtering characteristics. Experimental data show that, under harsh conditions of strong 20dB Gaussian white noise interference and a transition resistance of up to 300Ω, the overall recognition accuracy of this method can still be maintained above 96.90% and 98.05%, demonstrating strong engineering practical robustness.

[0025] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 The flowchart shows a converter transformer fault inrush current identification method driven by both knowledge and data. Figure 2 The original data distribution; Figure 3 The distribution of data after training; Figure 4 This is the result of fault identification; Figure 5 This is the confusion matrix; Figure 6 This is the ROC curve. Detailed Implementation

[0027] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0029] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0030] Example 1: This invention provides a method for identifying faulty inrush current in converters based on a dual-driven approach of knowledge and data. The specific operation process and control steps are as follows: Figure 1 As shown, it includes the following steps: Step 1: Acquisition and Preprocessing of One-Dimensional Raw Current Signal Three-phase current signals from each side of the converter transformer are collected to construct a one-dimensional current time series. Data with a fixed window length before and after the fault occurs are extracted as input samples and subjected to zero-mean standardization to eliminate the influence of data dimensions under different operating conditions.

[0031] Step 2: Construction of the Knowledge-Driven Layer and Extraction of Eight-Dimensional Time-Frequency Physical Features By leveraging the multi-resolution characteristics of discrete wavelet transform, physical criteria reflecting the fault-related inrush mechanism are extracted in a targeted manner. (1) Perform a three-level discrete wavelet transform on the preprocessed one-dimensional current signal. To reduce phase distortion and adapt to transient change characteristics, Sym2 is used as the wavelet basis function, as follows: Discretize the parameters:

[0032]

[0033] In the formula: j, k∈Z are discretization ordinal numbers.

[0034] Substituting the above parameters into the wavelet basis expression, we can obtain the discretized wavelet basis functions:

[0035] Therefore, the fault signal x ( t The discrete wavelet transform of ) can be defined as follows:

[0036] The analysis of fault transient signals all employs binary discrete wavelet transform, taking the scale basis. s 0=2, translation step size τ When 0=1, the discrete wavelet transform of the fault signal can be simplified to:

[0037] (2) Extract the third high-frequency detail component D3 after wavelet decomposition. This component reflects the transient oscillation and sudden change information caused by the fault. (3) Eight key physical feature parameters are calculated and extracted for the D3 component from the two dimensions of time domain and frequency domain. The four time domain features are the mean, standard deviation, skewness, and kurtosis, and the four frequency domain features are the peak frequency, average power frequency, power spectral entropy, and total energy.

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] (4) The above 8 parameters are Z-score standardized to construct an 8-dimensional feature vector.

[0047] Step 3: Construction of the data-driven layer and deep feature extraction A lightweight one-dimensional convolutional neural network based on depthwise separable convolution is constructed to automatically learn deep morphological features: the model parameters are selected as shown in Table 1.

[0048] Table 1

[0049] (1) Initial large-scale convolution: Perform large-scale convolution operation on the original signal, which is equivalent to low-pass filtering, filtering out high-frequency white noise and retaining the DC bias trend. (2) Depthwise convolution: Assign an independent one-dimensional time-domain convolution kernel to each phase current input channel, which only slides in the time axis direction and does not perform cross-channel fusion, thus keenly capturing local micro-distortions such as discontinuity angle and peak kurtosis within a single channel. (3) Pointwise convolution: Reconstruct the spatial coupling relationship between different frequency components or different phase sequence currents. (4) Global average pooling: Extract global features with translation invariance through global average pooling layers. The final output is a 256-dimensional deep convolution feature vector after dimensionality reduction.

[0050] Step 4: Dual-path feature fusion and classification based on attention mechanism: (1) The 8-dimensional mechanism features output by the data-driven layer and the 256-dimensional depth features output by the knowledge-driven layer are concatenated to form a 264-dimensional composite feature vector; (2) Input the composite vector into the fully connected layer of the attention mechanism, and use the explicit physical meaning of the physical features to guide the network to dynamically weight the deep features;

[0051] In the formula, x l Indicates the first l-1 After the layer undergoes a fully connected operation, it is mapped to the 1st layer. l The first layer j The output vector of each neuron n express l-1 The number of neurons in a layer. Indicates the first l-1 The first layer i The nth neuron is mapped to the nth neuron through a fully connected operation. l The first layer j The weights used by each neuron The input value for a fully connected circuit is the nth... l-1 The first layer i Number of neurons, Indicates the first l-1 Layers are fully connected up to the first layer. l The first in the layer jThe bias used for each neuron. After this linear combination, nonlinearity is introduced through the activation function and finally passed to the output layer.

[0052] (3) After passing through the ReLU activation function, the current signal is input to the Softmax output layer, and the output results are classified into three categories: normal state, fault inrush current or non-fault inrush current.

[0053] The output layer is the last layer of the neural network, responsible for generating the output result based on the features extracted and processed by the previous layers. It is often used in the Softmax function to transform a vector of real-valued values ​​into a probability distribution, ensuring that the sum of each element equals 1 and lies between 0 and 1. Its globally differentiable property allows for the use of optimization algorithms such as gradient descent to train the neural network. The relevant formula is:

[0054] To verify the effectiveness of the knowledge- and data-driven fault inrush current identification method proposed in this invention, a ±800kV DC transmission system model (Table 2) was built on PSCAD / EMTDC software. This model simulated normal operation, valve-side single-phase grounding fault, grid-side single-phase grounding fault, valve-side phase-to-phase fault, and grid-side phase-to-phase fault. Flux saturation points were set from 1.1 pu to 1.3 pu to acquire different samples. To avoid the impact of initial system instability on model training, the fault start time was set to 0.2 s during data acquisition, and four time offsets of 0, 5, 10, and 15 ms were set within a 20 ms period to ensure that the neural network's identification of fault types was not affected by the fault occurrence time. The sampling frequency was set to 20 kHz, and the data sampling window length was 200 ms. A total of 2190 valid samples were collected, which were strictly divided into a training set of 1533 sets, a validation set of 329 sets, and a test set of 328 sets.

[0055] Table 2

[0056] (1) Results and analysis of faulty inrush flow identification like Figure 2 Visualizing the input layer of the original current signal reveals that the data distribution of normal, faulty, and non-faulty inrush current samples is relatively concentrated, exhibiting overlap and a certain degree of similarity. This indicates that the original time-domain signal contains a large amount of redundant information, and the linear separability between different categories is poor. Figure 3 The three classes of samples output from features learned through the constructed knowledge and data dual-driven layer and attention mechanism show extremely high inter-class separability with almost no overlap and clear cluster boundaries. This indicates that the model demonstrates good performance in identifying faulty inrush flows.

[0057] like Figure 4 As shown, the black true value and the red predicted value completely overlap, intuitively demonstrating that this method has extremely high prediction accuracy for the three types of power signals. Figure 5 The accuracy rate for identifying faulty inrushes was 100%, the accuracy rate for normal states was 100%, and the accuracy rate for non-faulty inrushes was 99.17%. Only a few samples were misclassified as other categories, demonstrating a very strong feature discrimination capability. The overall average recognition rate was 99.39%.

[0058] To evaluate the model's ability to recognize various signals from multiple dimensions, this paper further introduces accuracy, recall, macro-mean, and ROC curve as evaluation criteria. Specific results are shown in Table 3 and... Figure 6 As shown.

[0059] Table 3

[0060] Depend on Figure 6 As can be seen, the area under the ROC curves for the three states—faulty inrush, normal, and non-faulty inrush—is all equal to 1.000, which verifies from a probabilistic and statistical perspective that the dual-drive model has extremely high classification reliability.

[0061] (2) Performance Analysis and Discussion To verify the robustness of the model under actual working conditions, noise immunity tests and transition resistance tolerance tests were conducted.

[0062] 1) Noise Resistance Analysis In practical engineering, the signals acquired often contain varying degrees of noise, which can severely affect the accuracy of fault identification. This paper injects different levels of Gaussian white noise into the test data to test the noise resistance of the proposed method. The results are shown in Table 4.

[0063] The data in the table shows that although the recognition accuracy decreases slightly under noise interference, the overall recognition performance remains at a high level, indicating that the model has good anti-noise interference capability, can adapt to noisy signals in actual engineering, and has good engineering practical value.

[0064] Table 4

[0065] 2) Analysis of the ability to withstand transition resistance The Ω values ​​of 0.01, 50, 100, 200, and 300 were fed into the trained model for testing, and the results are shown in Table 5.

[0066] Table 5

[0067] The data in the table shows that the higher the transition resistance, the slightly lower the accuracy, but the overall recognition level remains high, indicating that this model has a strong ability to withstand transition resistance.

[0068] (3) Model Comparison Analysis To verify the effectiveness of the model constructed in this invention in identifying faulty inrush flows, traditional machine learning algorithms, including Naive Bayes, Decision Tree (DT), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and a single-data-driven traditional Convolutional Neural Network (CNN) model, were compared and analyzed. All of the above models used the same training and test sets. The results of faulty inrush flow identification are shown in Table 6.

[0069] Table 6

[0070] As shown in Table 5, the accuracy of traditional machine learning algorithms is generally low, ranging from 81% to 87%, due to limitations in the representation capabilities of human-made features. While a single data-driven network improves the accuracy to 90.20%, it still suffers from some misclassifications. In contrast, the knowledge-and-data dual-driven model proposed in this paper achieves an accuracy of 99.39%, which is 17.59% to 9.19% higher than the aforementioned basic models. This strongly demonstrates that the knowledge-and-data dual-driven method can more effectively mine feature information from fault signals and exhibits optimal classification performance and generalization ability in fault surge identification tasks.

[0071] Example 2: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0072] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0073] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0074] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0075] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0076] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0077] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0078] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0079] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0080] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0081] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying faulty inrush current in converter transformers based on a dual-driven approach of knowledge and data, characterized in that: Includes the following steps: Step 1: Acquire the one-dimensional raw current signal and perform preprocessing; Step 2: Construct a knowledge-driven layer and extract 8-dimensional time-frequency physical features; Step 3: Construct a data-driven layer and extract deep features; Step 4: The dual-path feature fusion and classification model based on the attention mechanism fuses the 8-dimensional time-frequency physical features and depth features, and outputs a three-classification probability result indicating whether the current signal belongs to a normal state, a faulty inrush current, or a non-faulty inrush current.

2. The method for identifying faulty inrush current in converter transformers based on knowledge and data dual-drive as described in claim 1, characterized in that: Step one specifically includes the following steps: Collect three-phase current signals from each side of the converter transformer to construct a one-dimensional current time series; Data of a fixed window length before and after the fault occurred were extracted as input samples and subjected to zero-mean standardization to eliminate the influence of data units under different operating conditions.

3. The method for identifying faulty inrush current in converter transformers based on knowledge and data dual-drive as described in claim 1, characterized in that: Step two specifically includes the following steps: By leveraging the multi-resolution characteristics of discrete wavelet transform, physical criteria reflecting the fault-related inrush mechanism are extracted in a targeted manner. S21: Perform a 3-level discrete wavelet transform on the preprocessed one-dimensional current signal, using Sym2 as the wavelet basis function; the discrete wavelet transform is calculated as follows: Discretize the parameters: In the formula, j, k∈Z are discretization ordinal numbers; Substituting the parameters into the wavelet basis expression, we obtain the discretized wavelet basis functions: Therefore, the fault signal x ( t The discrete wavelet transform of ) is defined as: The analysis of fault transient signals all employs binary discrete wavelet transform, taking the scale basis. s 0=2, translation step size τ When 0=1, the discrete wavelet transform of the fault signal can be simplified to: S22: Extract the third-level high-frequency detail component D3 after wavelet decomposition; S23: Calculate and extract 8 key physical feature parameters for the D3 component from both time and frequency domains, including 4 time domain features: mean, standard deviation, skewness, and kurtosis, and 4 frequency domain features: peak frequency, average power frequency, power spectral entropy, and total energy. S24: The eight key physical feature parameters are Z-score standardized to construct an 8-dimensional feature vector.

4. The method for identifying faulty inrush current in converter transformers based on knowledge and data dual-drive as described in claim 1, characterized in that: In step three, a lightweight one-dimensional convolutional neural network based on depthwise separable convolution is constructed to automatically learn deep morphological features. This includes the following steps: S31: Initial large-scale convolution: Perform large-scale convolution operation on the original signal, which is equivalent to low-pass filtering, filtering out high-frequency white noise and preserving the DC bias trend; To design a network structure suitable for fault inrush detection, it is necessary to clarify the relationship between the input matrix size h1×m1×d1, the convolution kernel size a×b, the number of convolution kernels K, the stride S, and the output matrix size h2×m2×d2. The relationship is as follows: S32: Deep Convolution: Assigns an independent one-dimensional time-domain convolution kernel to each phase current input channel, slides only in the time axis direction, does not perform cross-channel fusion, and captures local micro-distortions within a single channel; Let the height of the output feature map be... h out The number of parameters in depthwise convolution P dw Computational complexity F dw They are respectively: S33: Pointwise convolution: reconstructs the spatial coupling relationship between different frequency components or different phase sequence currents; Let the number of output channels be C out The number of parameters in pointwise convolution is... P pw Computational complexity F pw for: S34: Global Average Pooling: Extracts global features with translation invariance through a global average pooling layer, and finally outputs a 256-dimensional deep convolutional feature vector after dimensionality reduction.

5. The method for identifying faulty inrush current in converter transformers based on knowledge and data dual-drive as described in claim 1, characterized in that: Step four specifically includes the following steps: S41: The 8-dimensional mechanism features output by the data-driven layer and the 256-dimensional depth features output by the knowledge-driven layer are concatenated to form a 264-dimensional composite feature vector. S42: Input the composite vector into the fully connected layer of the attention mechanism, and use the explicit physical meaning of the physical features to guide the network to dynamically weight the deep features; In the formula, x l Indicates the first l-1 After the layer undergoes a fully connected operation, it is mapped to the 1st layer. l The first layer j The output vector of each neuron n express l-1 The number of neurons in a layer. Indicates the first l-1 The first layer i The nth neuron is mapped to the nth neuron through a fully connected operation. l The first layer j The weights used by each neuron The input value for a fully connected circuit is the nth... l-1 The first layer i Number of neurons, Indicates the first l-1 Layers are fully connected up to the first layer. l The first in the layer j The bias used for each neuron; after this linear combination, nonlinearity is introduced through the activation function and finally passed to the output layer; S43: After passing through the ReLU activation function, it is input to the Softmax output layer, which outputs the three-class probability results of whether the current signal belongs to the normal state, fault inrush current or non-fault inrush current.

6. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the converter transformer fault inrush current identification method as described in any one of claims 1-5, which is based on a dual-drive approach of knowledge and data.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the converter transformer fault inrush current identification method based on knowledge and data dual-drive as described in any one of claims 1-5.

8. A computer program product, characterized in that: The method includes a computer program that, when executed by a processor, implements the knowledge- and data-driven method for identifying faulty inrush currents in converters as described in any one of claims 1-5.

Images