Self-adaptive DC fault arc detection method based on gated convolution pruning

Abstract

The invention discloses a self-adaptive direct current fault arc detection method based on gated convolution pruning, and belongs to the field of power system fault diagnosis. According to the method, firstly, collected direct current data are continuously sliced, self-adaptive local Z-Score standardization processing is independently executed on each slice, and the data are mapped to a unified numerical value distribution interval by utilizing statistical characteristics of the slices, so that interference of different voltage grades and load power changes on current amplitude is eliminated. Then, inputting the normalized data into a lightweight model based on gating convolution pruning; according to the model, firstly, dimension reduction and time sequence compression are performed on data through a feature embedding layer, and then high-frequency texture and long time sequence dependence features of a fault arc are accurately extracted under an extremely low dimension by utilizing a double-branch gating architecture and through cooperation of one-dimensional deep convolution and a Sigmoid gating mechanism. And finally, outputting a fault probability through global average pooling and a linear classifier. The method is simple in model structure and has excellent cross-domain generalization ability and detection precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system fault diagnosis, in particular to an adaptive DC fault arc detection method based on gated convolution pruning suitable for photovoltaic, energy storage and DC power distribution systems. BACKGROUND

[0002] DC fault arc is the main hidden danger of electrical fire in photovoltaic power generation, energy storage system and DC microgrid. Since DC arc has no natural zero point, it is difficult to extinguish itself once it arcs, and its current waveform characteristics are often covered by normal fluctuations of the load, making detection difficult. Existing detection methods mainly fall into two categories: one is the traditional method based on artificial feature extraction (such as Fourier transform, wavelet transform), which relies on fixed threshold criteria and is prone to false positives under load switching or inverter noise interference; the second is the existing deep learning method (such as Transformer or standard Mamba), which has high detection accuracy, but its model structure contains complex self-attention mechanisms or recursive state space equations (SSM), resulting in a large number of parameters and a complicated calculation process, making it difficult to deploy on low-cost, low-power embedded microcontrollers (MCU). Therefore, there is an urgent need for a lightweight detection method that can maintain high-dimensional feature extraction capability while significantly reducing computational complexity. SUMMARY

[0003] The purpose of the present application is to provide an adaptive DC fault arc detection method based on gated convolution pruning, which introduces adaptive local Z-Score standardization and a Mamba model based on front-end downsampling and pruning mechanism to solve the problem of weak model generalization ability and high computational complexity in the prior art.

[0004] To achieve the above purpose, the present application provides the following technical scheme: An adaptive DC fault arc detection method based on gated convolution pruning, comprising the following steps:

[0005] S1: Obtain real-time current sampling data of a DC circuit and continuously slice according to a preset time window length;

[0006] S2: Perform adaptive local Z-Score standardization on the sliced data to eliminate amplitude differences caused by voltage levels using the statistical characteristics of the sliced data;

[0007] S3: Construct a lightweight model containing a feature embedding layer (Tokenizer) and a dual-branch gated convolution core unit; the feature embedding layer is used for data dimensionality reduction and downsampling, and the core unit is used for feature extraction;

[0008] S4: Input the normalized data into the model and extract fault features after downsampling;

[0009] S5: Calculate the failure probability using global average pooling and a linear classifier.

[0010] S6: When the probability value exceeds the preset judgment threshold, it is determined that there is a fault arc in the DC circuit.

[0011] Furthermore, step S1 specifically includes:

[0012] The DC circuit bus current signal is acquired through a current acquisition unit. The sampling rate was set to a frequency sufficient to capture the high-frequency characteristics of the electric arc, ensuring that the high-frequency noise characteristics of the fault arc could be captured. A sliding window technique was used to slice the continuous signal, i.e., each current slice sequence... Include There are 1 sampling point, denoted as _ . .

[0013] Furthermore, step S2 specifically includes:

[0014] This step involves each slice. Independently perform adaptive local Z-score normalization.

[0015] First, calculate the mean of the current slice. and standard deviation :

[0016] ,

[0017]

[0018] in For the first The current values ​​at each sampling point were then used. Subsequently, statistics were used to map the original sequence to a normalized sequence. , of which element The calculation is as follows:

[0019]

[0020] In the formula, Values Used to prevent pure DC signals ( This leads to a division-by-zero error. Through this step, regardless of whether the original circuit is under heavy current load or light current load, the normalized data distribution is stretched to a uniform scale with zero mean and unit variance. This allows subsequent models to focus on waveform distortion caused by fault arcs without needing to consider the absolute amplitude of the signal, thus giving the algorithm extremely strong adaptability across voltage levels.

[0021] Furthermore, step S3 specifically includes:

[0022] This step is typically completed during the system initialization phase. The lightweight model built in this step mainly consists of two parts: a feature embedding layer (Tokenizer) and a core unit.

[0023] (1) Feature Embedding Layer Construction: In order to map the one-dimensional current data of a single channel to a high-dimensional feature space and reduce the subsequent computational load, the model first constructs a feature embedding layer. Let the normalized slice output in step S2 be... ( (where the number of sampling points is ), the calculation process for this layer can be represented as:

[0024]

[0025] in, The first convolutional layer (large kernel, stride = 5) is used to compress the temporal length to... ; Use the GELU activation function; This is the second convolutional layer, used to map features to the hidden layer dimensions of the model. The final output feature sequence .

[0026] (2) Core Unit Construction: After the feature embedding layer, a core unit based on a gated convolution mechanism is constructed. This unit removes the self-attention mechanism of the traditional Transformer and the SSM recursive equation of Mamba, retaining only the efficient feedforward and gated structure. The specific calculation process is as follows:

[0027] Input projection and channel segmentation: the sequence output by the feature embedding layer. The feature dimension is expanded to 2ED by performing a linear mapping through the input projection layer. The calculation formula is as follows:

[0028]

[0029] in Subsequently, in the channel dimension, It is divided into two parts: main feature branches and gated branch Both dimensions are .

[0030] Dual-branch processing:

[0031] Main characteristic branch: for Perform a one-dimensional depthwise convolution operation to extract local temporal textures, and then process them using the SiLU activation function to obtain... :

[0032]

[0033] Gated branch: for Perform the Sigmoid activation operation to generate a value in Gating weights between :

[0034]

[0035] Feature Reconstruction and Output: A gating mechanism is used to adaptively filter effective features. Specifically, the content of the "main branch" is controlled by the strength of the "gated branch," which then maps the output back to the original dimension. The residuals are then superimposed to obtain the core unit output. :

[0036]

[0037] in, This represents element-wise multiplication. . This refers to the time-series features rich in fault information extracted by the core unit.

[0038] Furthermore, step S4 specifically includes:

[0039] The normalized sequence after step S2 The data is input into the model. The data stream is sequentially downsampled and compressed through the feature embedding layer (to obtain...). ) and dual-branch gating screening of the core unit (to obtain Because the core unit uses depthwise separable convolutions and removes fully connected matrix multiplications (except for the projection layer), the computational complexity of this process is only [value missing]. Far lower than the self-attention mechanism It is suitable for real-time processing in embedded systems.

[0040] Furthermore, step S5 specifically includes:

[0041] Temporal feature sequences output by the model Global average pooling is performed to compress the temporal dimension and obtain a global feature vector. :

[0042]

[0043] Then, Input a fully connected classifier layer and calculate the output logits. The probability value of the current slice belonging to the "fault arc" category is calculated using the Softmax function. :

[0044]

[0045] Furthermore, step S6 specifically includes:

[0046] The calculated probability value Compared with the preset judgment threshold Comparison: If a fault is detected, a prompt will be output; otherwise, the detection will continue in the next time window.

[0047] The beneficial effects of this invention are as follows:

[0048] (1) Improved anti-interference capability and detection accuracy under complex working conditions: This invention designs a dual-branch parallel gating architecture. By utilizing the Sigmoid gating branch, which is independent of the feature extraction branch, an "attention mask" can be dynamically generated based on the instantaneous energy intensity of the input signal. This mechanism endows the model with adaptive filtering capabilities, enabling it to keenly capture the weak high-frequency ripples of the fault arc, while automatically suppressing steady-state interference caused by inverter switching noise or load mutations, thereby significantly reducing the false alarm rate and false negative rate.

[0049] (2) Good generalization ability: By adaptive local Z-Score normalization, the current signals under different voltage levels and power are mapped to the same distribution space, so that the algorithm can be applied without retraining for specific loads. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0052] Figure 2 This is a schematic diagram of the overall architecture of the lightweight model in this invention.

[0053] Figure 3 This is a schematic diagram comparing the current waveforms before and after local Z-Score normalization. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the relevant technical solutions will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments described herein are only a part of this application and not all embodiments. All other embodiments that can be obtained by those skilled in the art based on the embodiments in this application without creative effort should be included within the protection scope of this application.

[0055] Detailed Description of the Embodiments: The present invention will be further described in detail below with reference to the accompanying drawings.

[0056] like Figure 1 As shown, an adaptive DC fault arc detection method based on gated convolution pruning specifically includes the following steps:

[0057] Step S1: Signal Acquisition and Slicing

[0058] As one embodiment, step S1 includes: acquiring the bus current signal of the DC circuit through the current acquisition unit. The sampling rate is set to a frequency sufficient to capture the high-frequency characteristics of the electric arc, such as... This is to ensure that the high-frequency noise characteristics of the fault arc can be captured. A sliding window technique is used to slice the continuous signal, with a window length of 50ms, i.e., each current slice sequence... Include There are 1 sampling point, denoted as _ . .

[0059] Step S2: Adaptive Local Z-Score Standardization

[0060] As one embodiment, step S2 includes: Figure 3 As shown, the amplitude of the original current signal varies with load power and system voltage level. To eliminate the interference of this amplitude difference on the model, this step performs a step on each slice. Independently perform adaptive local Z-score normalization.

[0061] First, calculate the mean of the current slice. and standard deviation :

[0062] ,

[0063]

[0064] in For the first The current values ​​at each sampling point were then used. Subsequently, statistics were used to map the original sequence to a normalized sequence. , of which element The calculation is as follows:

[0065]

[0066] In the formula, Values Used to prevent pure DC signals ( This leads to a division-by-zero error. Through this step, regardless of whether the original circuit is under heavy current load or light current load, the normalized data distribution is stretched to a uniform scale with zero mean and unit variance. This allows subsequent models to focus on waveform distortion caused by fault arcs without needing to consider the absolute amplitude of the signal, thus giving the algorithm extremely strong adaptability across voltage levels.

[0067] Step S3: Construct a lightweight model based on gated convolution pruning

[0068] As one embodiment, step S3 includes: This step is typically completed during the system initialization phase. The lightweight model constructed in this step mainly consists of two parts: a feature embedding layer and a core unit.

[0069] (1) Feature Embedding Layer Construction: In order to map the one-dimensional current data of a single channel to a high-dimensional feature space and reduce the subsequent computational load, the model first constructs a feature embedding layer. Let the normalized slice output in step S2 be... ( (where the number of sampling points is ), the calculation process for this layer can be represented as:

[0070]

[0071] in, The first convolutional layer (large kernel, stride = 5) is used to compress the temporal length to... ; Use the GELU activation function; This is the second convolutional layer, used to map features to the hidden layer dimensions of the model. (In this embodiment) The final output feature sequence .

[0072] (2) Core Unit Construction: After the feature embedding layer, a core unit based on a gated convolution mechanism is constructed. This unit removes the self-attention mechanism of the traditional Transformer and the SSM recursive equation of Mamba, retaining only the efficient feedforward and gated structure. The specific calculation process is as follows:

[0073] Input projection and channel segmentation: the sequence output by the feature embedding layer. By performing linear mapping through the input projection layer, the feature dimension is expanded to 2ED (expansion factor). (That is, the middle dimension is 16). The calculation formula is as follows:

[0074]

[0075] in Subsequently, in the channel dimension, It is divided into two parts: main feature branches and gated branch Both dimensions are .

[0076] Dual-branch processing:

[0077] Main characteristic branch: for Perform a one-dimensional depthwise convolution operation (kernel size in this embodiment) Local temporal textures are extracted and processed using the SiLU activation function to obtain... :

[0078]

[0079] Gated branch: for Perform the Sigmoid activation operation to generate a value in Gating weights between :

[0080]

[0081] Feature Reconstruction and Output: A gating mechanism is used to adaptively filter effective features. Specifically, the content of the "main branch" is controlled by the strength of the "gated branch," which then maps the output back to the original dimension. The residuals are then superimposed to obtain the core unit output. :

[0082]

[0083] in, This represents element-wise multiplication. . This refers to the time-series features rich in fault information extracted by the core unit.

[0084] Step S4: Feature Extraction

[0085] As one embodiment, step S4 includes: processing the normalized sequence after step S2. The data is input into the model. The data stream is sequentially downsampled and compressed through the feature embedding layer (to obtain...). ) and dual-branch gating screening of the core unit (to obtain Because the core unit uses depthwise separable convolutions and removes fully connected matrix multiplications (except for the projection layer), the computational complexity of this process is only [value missing]. Far lower than the self-attention mechanism It is suitable for real-time processing in embedded systems.

[0086] Step S5: Failure Probability Calculation

[0087] As one embodiment, step S5 includes: processing the temporal feature sequence output by the model. Global average pooling is performed to compress the temporal dimension and obtain the global feature vector. :

[0088]

[0089] Then, Input a fully connected classifier layer and calculate the output logits. The probability value of the current slice belonging to the "fault arc" category is calculated using the Softmax function. :

[0090]

[0091] Step S6: Fault Judgment

[0092] As one embodiment, step S6 includes: calculating the probability value Compared with the preset judgment threshold (This embodiment is set as) Compare: If a fault is detected, a prompt will be output; otherwise, the detection will continue in the next time window.

[0093] Regarding the training method of the model: In this embodiment, the lightweight model based on gated convolution pruning is trained using a supervised learning method. First, a training dataset containing normal current samples and fault arc samples is constructed and divided into a training set and a validation set according to a preset ratio. All training and validation data must be strictly processed by adaptive local Z-score standardization as described in step S2 before being input into the model to ensure the consistency of data distribution. During the training process, the cross-entropy loss function is used as the objective function to measure the difference between the model's predicted values ​​and the true labels. A general optimization algorithm (such as AdamW) is used to iteratively update the learnable parameters in the model through backpropagation until the loss function converges. Finally, the model parameters with the highest accuracy on the validation set are saved as the detection model.

[0094] Finally, it should be noted that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An adaptive DC fault arc detection method based on gated convolution pruning, characterized in that, The process includes the following steps: Step S1: Acquire the real-time current signal of the DC circuit and continuously slice it according to a preset time window length to obtain the signal containing... Current slice sequence of sampling points ; Step S2: Slice the current sequence Adaptive local Z-score normalization is performed, which maps the data to a normalized sequence using the statistical characteristics of the current slice. To eliminate the current amplitude differences caused by variations in voltage levels and load power; Step S3: Construct a lightweight model based on gated convolution pruning, the model including a feature embedding layer (Tokenizer) and a dual-branch gated convolution core unit; the feature embedding layer is used to perform feature mapping and temporal downsampling on the input data; the core unit abandons the recursive structure of the state space model and adopts a dual-branch gated convolution architecture including an input projection layer, a one-dimensional deep convolution layer, gated branches, and an output projection layer; Step S4: Convert the normalized sequence The data is input into the lightweight model based on gated convolutional pruning. First, the single-channel time-series data is mapped to multi-channel features and compressed in length through a feature embedding layer. Then, it is input into the core unit, where a dual-branch gating mechanism is used to reconstruct the signal features and extract time-series features containing fault information. Step S5: The time-series features output by the lightweight model are subjected to global average pooling to obtain a global feature vector. Then, a linear classifier is used to calculate the probability value that the current slice belongs to a fault arc. Step S6: When the probability value When the preset threshold is exceeded, the DC circuit is determined to have a fault arc.

2. The adaptive DC fault arc detection method based on gated convolution pruning according to claim 1, characterized in that, The specific implementation method of step S2 is as follows: First, calculate the current current slice sequence. mean and standard deviation : , in, For the first slice Each sampling point is used, and then the normalized sequence is calculated according to the following formula. elements in : in, To prevent tiny constants with a denominator of zero, the processing forces the model to focus on the high-frequency texture and randomness characteristics of the waveform.

3. The adaptive DC fault arc detection method based on gated convolution pruning according to claim 1, characterized in that, The feature embedding layer in step S3 adopts a cascaded convolutional structure, specifically including: a first convolutional layer, which uses a large convolutional kernel and a stride greater than 1 to downsample the input sequence and perform preliminary feature extraction; a second convolutional layer, which uses a small convolutional kernel to map the feature dimension to the hidden layer dimension of the model; and a non-linear activation function is connected between the two convolutional layers.

4. The adaptive DC fault arc detection method based on gated convolution pruning according to claim 1, characterized in that, The specific operation process of the dual-branch gated convolutional core unit in step S3 is as follows: Let the feature sequence output by the feature embedding layer be... ,in For sequence length, The hidden layer dimension of the model; 1) Input projection and channel segmentation: The input is projected through an input projection layer. Linearly map to the extended dimension 2ED, and segment along the channel dimension into main feature branches. and gated branch : in, , For learnable weights and biases, The expansion factor; after segmentation 2) Processing of main characteristic branches: for Perform one-dimensional depthwise convolution and SiLU activation to obtain the processed main features. : 3) Gated branch processing: For Execute Sigmoid activation to generate gated weights : 4) Feature Reconstruction and Output: The two processed branches are multiplied element-wise, and then mapped back to the original dimensions through the output projection layer. Simultaneously, residual connections are introduced to obtain the final output. : in, This represents element-wise multiplication. , .

5. The adaptive DC fault arc detection method based on gated convolution pruning according to claim 1, characterized in that, The lightweight model in step S3 has the following specific architecture parameters: the stride of the first convolutional layer of the feature embedding layer is set to 5 to achieve 5x downsampling; the number of core unit stacking layers is set to 1; the hidden layer dimension is set to 8; and the state expansion factor is set to 2.

6. The adaptive DC fault arc detection method based on gated convolution pruning according to claim 1, characterized in that, The one-dimensional deep convolutional layer in step S3 is used to extract local short-time features of the current signal, and its convolutional kernel size is set to 4.

7. The adaptive DC fault arc detection method based on gated convolution pruning according to claim 1, characterized in that, The specific implementation method of step S5 is as follows: Temporal feature sequences output by the model Perform global average pooling to obtain the global feature vector. : vector Input a linear classifier and calculate the probability value of belonging to a fault arc using the Softmax function. : in, , For classifier parameters, The confidence score corresponds to the fault category; in step S6, if If so, it is determined to be a fault.

Images