Multi-channel direct current arc detection method and device and electronic equipment

Abstract

The invention discloses a multi-channel direct current arc detection method and device and electronic equipment. The method comprises the following steps: acquiring original normal data and original arc data of a multi-channel DC circuit; preprocessing the data, generating time-domain and frequency-domain features corresponding to the original normal data, and generating time-domain and frequency-domain features corresponding to the original arc data; constructing a classification and identification model of multi-channel direct current arc detection; the method comprises the following steps: acquiring a to-be-detected multi-channel direct-current electric signal, preprocessing the multi-channel direct-current electric signal, and generating a target time domain feature and a target frequency domain feature corresponding to the multi-channel direct-current electric signal; and inputting the target time domain feature and the target frequency domain feature into a classification and identification model of multi-channel direct current arc detection to obtain a classification and identification result of each channel of direct current signals in the multi-channel direct current signals. According to the method, the direct-current arc of each channel can be accurately identified, the detection accuracy is improved, and the rapid and efficient completion of the detection task is also realized.

Description

Technical Field

[0001] This invention belongs to the field of power electronics technology, specifically relating to a multi-channel DC arc detection method, device, and electronic equipment. Background Technology

[0002] In photovoltaic energy storage systems, DC-side arcing is characterized by stable current, sustained intensity, and difficulty in self-extinguishing, posing a significantly higher fire risk than AC-side arcing. The high temperatures generated by arcing can melt conductors, burn insulation materials, and even trigger electrical fires, seriously threatening the safe operation of the system and the safety of the surrounding environment. Since photovoltaic energy storage systems typically consist of multiple DC branches (i.e., multiple channels), arcing faults can occur in any branch. Without precise monitoring capabilities for each channel, potential risks cannot be detected in a timely manner. Therefore, there is an urgent need to develop a high-precision, high-reliability multi-channel DC arc detection method to achieve real-time, accurate monitoring and location of arcs in each branch.

[0003] Existing DC arc detection methods mainly rely on fixed thresholds or frequency domain features, but they have significant drawbacks in complex multi-channel environments. On one hand, the fixed threshold method struggles to adapt to load differences and environmental interference across different branches. When the operating conditions of a channel change (e.g., sudden load changes), the fixed threshold cannot be dynamically adjusted, easily leading to a higher false alarm rate and difficulty in accurately identifying DC arcs in each channel. On the other hand, the frequency domain feature method is susceptible to noise interference and has weak robustness. In multi-channel scenarios, it cannot balance real-time performance and accuracy, easily resulting in missed or false alarms. These limitations prevent existing methods from meeting the practical needs of multi-channel DC arc detection. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by proposing a multi-channel DC arc detection method, device and electronic equipment. This method can accurately identify the DC arc in each channel, thereby effectively improving the detection accuracy of the DC arc detection method.

[0005] In a first aspect, the present invention provides a multi-channel DC arc detection method, the method comprising the following steps:

[0006] Acquire the multi-channel DC signal to be detected, preprocess the multi-channel DC signal, and generate the target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal;

[0007] The target time-domain features and target frequency-domain features are input into the classification and recognition model of multi-channel DC arc detection to obtain the classification and recognition results of each channel of DC signal in the multi-channel DC signal, thereby realizing multi-channel DC arc detection;

[0008] Among them, the classification and recognition model for multi-channel DC arc detection is a convolutional neural network learning model built based on multi-channel DC circuit sample label data.

[0009] As a specific implementation method, the multi-channel DC signal is preprocessed to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal, specifically including:

[0010] Step A1: Remove abnormal data from the multi-channel DC signal to obtain the multi-channel DC signal with abnormalities removed;

[0011] Abnormal data in multi-channel DC signals include multi-channel DC signals with mixed data sampling or incorrect labeling.

[0012] Step A2: Extract time-domain data from the anomaly-free multi-channel DC signal according to the preset third fixed frame length to obtain the time-domain features of the third process; and extract frequency-domain data from the anomaly-free multi-channel DC signal according to the preset third fixed frame length to obtain the frequency-domain features of the third process.

[0013] Step A3: Standardize the time-domain features of the third process to generate the third time-domain features corresponding to the multi-channel DC signal; and standardize the frequency-domain features of the third process to generate the third frequency-domain features corresponding to the multi-channel DC signal.

[0014] As a specific implementation method, after obtaining the classification and identification results of each channel of the multi-channel DC signal, the method further includes:

[0015] The sliding window counting method is used to perform time-domain statistics and fault early warning on the classification and recognition results.

[0016] As a specific implementation method, before inputting the target time-domain features and target frequency-domain features into the classification and recognition model for multi-channel DC arc detection, the method further includes:

[0017] Acquire sample label data for multi-channel DC circuits; the sample label data includes raw normal data and raw arc data;

[0018] The original normal data is preprocessed to generate a first time-domain feature and a first frequency-domain feature corresponding to the original normal data; and the original electric arc data is preprocessed to generate a second time-domain feature and a second frequency-domain feature corresponding to the original electric arc data.

[0019] Based on the original normal data, the original arc data, the first time domain feature, the first frequency domain feature, the second time domain feature, and the second frequency domain feature, a classification and recognition model for multi-channel DC arc detection is constructed.

[0020] As a specific implementation method, the original arc data is obtained through artificially simulated arc fault experiments or extracted from historical arc fault records.

[0021] As a specific implementation method, a classification and recognition model for multi-channel DC arc detection is constructed based on raw normal data, raw arc data, a first time-domain feature, a first frequency-domain feature, a second time-domain feature, and a second frequency-domain feature. Specifically, this includes:

[0022] The original normal data is matched with the first time domain feature and the first frequency domain feature to obtain the first matched data; and the original arc data is matched with the second time domain feature and the second frequency domain feature to obtain the second matched data.

[0023] The first and second matching data are combined to obtain the sample dataset;

[0024] The sample dataset is divided into training and test sets according to a preset ratio;

[0025] Based on the training set, the initial convolutional neural network learning model is iteratively trained to obtain the initial classification and recognition model;

[0026] Input the test set into the initial classification and recognition model to obtain the confusion matrix corresponding to the initial classification and recognition model;

[0027] The performance of the initial classification and recognition model is evaluated based on the confusion matrix, and the evaluation results are obtained.

[0028] Based on the evaluation results, the parameters of the initial classification and recognition model were adjusted to obtain a classification and recognition model for multi-channel DC arc detection.

[0029] As a specific implementation method, the initial convolutional neural network learning model is iteratively trained based on the training set to obtain an initial classification and recognition model, specifically including:

[0030] Step B1: Preset the initial convolutional neural network learning model and set the initial parameters of the initial convolutional neural network learning model; the initial parameters include batch size, early stopping condition, and number of iterations;

[0031] Step B2: Input the training set into the initial convolutional neural network learning model in batches according to the batch size;

[0032] Step B3: During the initial convolutional neural network learning model operation, extract multi-channel shared features and generate independent binary classification results for each channel's features;

[0033] Step B4: Update the model parameters of the convolutional neural network learning model and output the updated model parameters;

[0034] Step B5: Repeat steps B2 to B4 until the maximum number of iterations is reached or the early stopping condition is met;

[0035] Step B6: Construct the initial classification and recognition model based on the updated model parameters from the final output.

[0036] As a specific implementation method, the performance of the initial classification and recognition model is evaluated based on the confusion matrix to obtain the evaluation results, which specifically include:

[0037] The accuracy and false alarm rate of arc identification are obtained by calculating the confusion matrix.

[0038] Among them, the arc recognition accuracy rate is the proportion of samples correctly identified by the initial classification model out of the total test samples; the arc recognition false alarm rate is the proportion of samples that were actually arced and correctly identified by the initial classification model.

[0039] The evaluation results of the initial classification and recognition model are obtained based on the accuracy and false negative rate of arc recognition.

[0040] In a second aspect, the present invention provides a multi-channel DC arc detection device, the device comprising:

[0041] The acquisition unit is used to acquire the multi-channel DC signal to be detected;

[0042] The preprocessing unit, connected to the acquisition unit, is used to preprocess the multi-channel DC signal to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal.

[0043] The input unit, connected to the preprocessing unit, is used to input the target time-domain features and target frequency-domain features into the classification and recognition model of multi-channel DC arc detection, so as to obtain the classification and recognition results of each channel DC signal in the multi-channel DC signal, thereby realizing multi-channel DC arc detection;

[0044] Among them, the classification and recognition model for multi-channel DC arc detection is a convolutional neural network learning model built based on multi-channel DC circuit sample label data.

[0045] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the multi-channel DC arc detection method according to the first aspect.

[0046] This invention constructs a classification and recognition model by fusing multi-dimensional features, achieving accurate detection of multi-channel DC arcs. This effectively improves the accuracy and reliability of arc identification, reduces false alarm and false negative rates caused by load fluctuations and noise interference, and enables precise location of faulty channels. Specifically, it includes:

[0047] (1) Real-time detection and channel positioning: The multi-channel DC signal to be detected is preprocessed to generate the corresponding target time domain features and target frequency domain features. After being input into the classification and recognition model, the classification and recognition results of each channel can be output, thereby realizing the real-time detection of multi-channel DC arc and the accurate positioning of fault channels.

[0048] (2) Since the method of the present invention combines the frequency domain and the time domain, it is not a simple frequency domain feature method. Therefore, it is not affected by noise, has strong robustness, and can balance real-time performance and accuracy in multi-channel scenarios. It is not easy to have missed or false alarms. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the multi-channel DC arc detection method in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the multi-channel DC arc detection process in an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of the structure of the classification and recognition model for multi-channel DC arc detection in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the sliding window counting method for multi-channel DC arc detection in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of a multi-channel DC arc detection device in an embodiment of the present invention;

[0054] Figure 6 This is an architectural diagram of an electronic device according to an embodiment of the present invention.

[0055] Reference numerals: 10, acquisition unit; 20, preprocessing unit; 30, input unit; 100, processor; 200, memory. Detailed Implementation

[0056] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0057] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.

[0058] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.

[0059] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.

[0060] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0061] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.

[0062] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.

[0063] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.

[0064] In modern industrial production, multi-channel DC arc detection technology is crucial for ensuring stable equipment operation, timely fault diagnosis, and system reliability. However, existing detection methods, such as fixed threshold methods and single-frequency domain feature analysis, often fail to meet practical needs in complex and ever-changing industrial environments due to a lack of sufficient adaptability and robustness. These methods are prone to false alarms or missed alarms, negatively impacting equipment safety and production continuity.

[0065] In view of this, the inventors of this application propose a multi-channel DC arc detection method, device, and electronic device. This method comprehensively captures and reflects various characteristics of the arc signal and employs a deep learning model to effectively fuse the time-domain and frequency-domain features of each channel, constructing an end-to-end autonomous feature extraction and fault identification system. This invention can be widely applied in industrial and energy fields that require simultaneous monitoring of multiple DC channels to ensure circuit safety and rapid fault detection and location. It not only improves the accuracy and robustness of detection but also enhances the system's anti-interference capabilities and adaptability to different operating conditions, thereby effectively reducing the probability of false positives and false negatives.

[0066] To ensure the effectiveness of this technical solution in practical applications, specific embodiments were designed. Embodiment 1 provides a method for multi-channel DC arc detection, which achieves high-precision arc fault identification through a convolutional neural network learning model. Embodiment 2 provides a corresponding multi-channel DC arc detection device, embedding the method from Embodiment 1 into the hardware system to realize the hardware application of the method. Embodiment 3 provides supporting electronic equipment, forming a complete application solution based on this detection method by integrating hardware and software systems. These embodiments (details below) achieve high-precision, real-time detection of multi-channel DC arcs and accurate location of fault channels through a classification-based convolutional neural network learning model, significantly improving the accuracy and robustness of identification.

[0067] Example 1:

[0068] like Figure 1 As shown in the figure, this embodiment provides a multi-channel DC arc detection method, which includes the following steps K1 to K3.

[0069] Step K1: Acquire the multi-channel DC signal to be detected.

[0070] To acquire the multi-channel DC signals to be tested, this embodiment employs two or more high-precision analog-to-digital converters (ADCs) to simultaneously acquire the voltage and current signals in the system. This acquisition process aims to comprehensively capture the electrical characteristics under diverse scenarios—including signal characteristics corresponding to different voltage levels, current magnitudes, impedance network structures, cable lengths, and fault locations—thereby obtaining multi-channel DC signals covering a wide range of operating conditions. These signals can be experimental data obtained by simulating different operating conditions, or field data collected from actual sites that reflect various real-world operating conditions.

[0071] Step K2: Preprocess the multi-channel DC signal to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal.

[0072] As a specific implementation method, the multi-channel DC signal is preprocessed to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal, specifically including steps A1 to A3.

[0073] Step A1: Remove abnormal data from the multi-channel DC signal to obtain the multi-channel DC signal with abnormalities removed; abnormal data in the multi-channel DC signal includes multi-channel DC signals with mixed data sampling or incorrect labeling.

[0074] Step A2: Extract time-domain data from the anomaly-free multi-channel DC signal according to the preset third fixed frame length to obtain the time-domain features of the third process; and extract frequency-domain data from the anomaly-free multi-channel DC signal according to the preset third fixed frame length to obtain the frequency-domain features of the third process.

[0075] Step A3: Standardize the time-domain features of the third process to generate the third time-domain features corresponding to the multi-channel DC signal; and standardize the frequency-domain features of the third process to generate the third frequency-domain features corresponding to the multi-channel DC signal.

[0076] Step K3: Input the target time-domain features and target frequency-domain features into the classification and recognition model for multi-channel DC arc detection to obtain the classification and recognition results of each channel of the multi-channel DC signal, thereby realizing multi-channel DC arc detection; wherein, the classification and recognition model for multi-channel DC arc detection is a convolutional neural network learning model built based on multi-channel DC circuit sample label data.

[0077] As a specific implementation method, before inputting the target time-domain features and target frequency-domain features into the classification and recognition model for multi-channel DC arc detection, the method further includes steps K31 to K33.

[0078] Step K31: Obtain sample label data for the multi-channel DC circuit; the sample label data includes raw normal data and raw arc data.

[0079] As a specific implementation method, the original arc data is obtained through artificially simulated arc fault experiments or extracted from historical arc fault records.

[0080] More specifically, there are two main approaches to acquiring raw arc data: one is to conduct targeted experimental data collection through artificially simulating arc faults, and the other is to screen and extract relevant data from existing arc fault historical records. During the acquisition process, two or more high-precision analog-to-digital converters (ADCs) are used to simultaneously capture voltage and current signals, aiming to comprehensively cover the electrical characteristics under various complex scenarios. Regardless of different voltage levels and current magnitudes, or differences in impedance network structures, cable lengths, and fault locations, accurate signal capture can be obtained. This design ensures that the acquired multi-channel DC signals have rich feature dimensions, covering various operating conditions, and that the data includes both normal operating conditions and arc fault conditions. This ensures the diversity and representativeness of the data during the acquisition process. Furthermore, after each acquisition, the data is meticulously labeled, clearly identifying its corresponding operating condition category, providing a clear basis for the subsequent training of the classification and recognition model for multi-channel DC arc detection. The acquired signals include not only experimental data generated by simulating different operating conditions, but also real data collected from actual sites reflecting various operating conditions. The specific method for obtaining real data is to screen out fault records related to DC arcs from the fault logs of the sites in recent years, and extract the original monitoring data of the time period near the records. This data is then combined with the actual environmental conditions on site. This rich basic data lays a solid foundation for training the classification and recognition model of multi-channel DC arc detection. The wider the collection range and the more diverse the scenarios, the more comprehensive the samples provided, which helps the classification and recognition model of multi-channel DC arc detection adapt to complex and changing actual operating conditions, and improve its recognition accuracy and robustness.

[0081] Step K32: Preprocess the original normal data to generate a first time-domain feature and a first frequency-domain feature corresponding to the original normal data; and preprocess the original arc data to generate a second time-domain feature and a second frequency-domain feature corresponding to the original arc data.

[0082] As a specific implementation method, the original normal data is preprocessed to generate a first time-domain feature and a first frequency-domain feature corresponding to the original normal data, specifically including steps C1 to C3.

[0083] Step C1: Remove abnormal data from the original normal data to obtain the original normal data after removing abnormalities; the abnormal data in the original normal data includes the original normal data with mixed data collection or labeling errors.

[0084] Step C2: Extract time-domain data from the original normal data after removing anomalies according to a preset first fixed frame length to obtain the first process time-domain features; and extract frequency-domain data from the original normal data after removing anomalies according to a preset first fixed frame length to obtain the first process frequency-domain features.

[0085] Step C3: Standardize the time-domain features of the first process to generate the first time-domain features corresponding to the original normal data; and standardize the frequency-domain features of the first process to generate the first frequency-domain features corresponding to the original normal data.

[0086] As a specific implementation method, the original electric arc data is preprocessed to generate a second time-domain feature and a second frequency-domain feature corresponding to the original electric arc data, specifically including steps D1 to D3.

[0087] Step D1: Remove abnormal data from the original arc data to obtain normal arc data after removing abnormalities; abnormal data in the original arc data includes original arc data with mixed data collection or incorrect labeling.

[0088] Step D2: Extract time-domain data from the original arc data after removing anomalies according to the preset second fixed frame length to obtain the second process time-domain features; and extract frequency-domain data from the original arc data after removing anomalies according to the preset second fixed frame length to obtain the second process frequency-domain features.

[0089] Step D3: Standardize the time-domain features of the second process to generate second time-domain features corresponding to the original arc data; and standardize the frequency-domain features of the second process to generate second frequency-domain features corresponding to the original arc data.

[0090] This embodiment presets three fixed frame lengths: a first fixed frame length, a second fixed frame length, and a third fixed frame length. These preset values ​​can be the same or different values ​​as needed. The first, second, and third fixed frame lengths are all preset to 1024 or other powers of 2. This preset is based on the computational efficiency of the Fast Fourier Transform (FFT). When the input data length is a power of 2, FFT can significantly reduce multiplication and addition operations in the calculation, improve processing speed, and ensure the accuracy of spectral feature extraction. Choosing a frame length such as 1024 allows for rapid frequency domain analysis, providing efficient support for signal interpretation. More importantly, for convolutional neural networks, a uniform input length is crucial to ensuring the stable operation of the classification and recognition model for multi-channel DC arc detection. A fixed frame length, such as 1024, ensures that all input data dimensions are consistent, avoiding feature alignment errors and convolution kernel matching problems caused by inconsistent data lengths. This not only simplifies network design but also maintains the consistency of feature extraction during training and prediction, reduces errors caused by inconsistent data formats, and thus improves the stability and accuracy of the classification and recognition model output for multi-channel DC arc detection.

[0091] Step K33: Construct a classification and recognition model for multi-channel DC arc detection based on the original normal data, original arc data, first time-domain features, first frequency-domain features, second time-domain features, and second frequency-domain features. This multi-channel DC arc detection classification and recognition model is a machine learning model with a convolutional neural network (CNN) as its core architecture. Its construction process is based on sample label data from multi-channel DC circuits. Specifically, the training and optimization of this model rely on labeled sample data collected from multi-channel DC circuits—data that clearly distinguishes between electrical signals under normal operating conditions (normal data) and electrical signals during arc faults (arc data). The characteristics of CNNs enable them to automatically extract hidden multi-dimensional features (such as signal fluctuation patterns and changes in specific frequency components) from the sample data through multi-layer convolution and pooling operations, without the need for manually preset feature rules. During training, the multi-channel DC arc detection classification and recognition model continuously adjusts network parameters (such as convolutional kernel weights and bias values) by learning the feature differences between normal data and arc data from a large number of labeled samples, gradually optimizing its classification ability for the two types of data. The final classification and recognition model for multi-channel DC arc detection can independently analyze the electrical signals of each channel in a multi-channel DC circuit. It utilizes learned feature patterns to determine whether the input signal belongs to a normal state or an arc state, thereby achieving accurate identification and classification of multi-channel DC arcs. Its core advantage lies in its ability to adapt to the complex signal environment in multi-channel scenarios, improving classification accuracy through deep mining of data features, and overcoming the limitations of existing methods that rely on fixed thresholds or manually designed features.

[0092] As a specific implementation method, a classification and recognition model for multi-channel DC arc detection is constructed based on the original normal data, the original arc data, the first time domain feature, the first frequency domain feature, the second time domain feature, and the second frequency domain feature, specifically including steps K331 to K337.

[0093] Step K331: Match the original normal data with the first time domain feature and the first frequency domain feature to obtain the first matched data; and match the original arc data with the second time domain feature and the second frequency domain feature to obtain the second matched data.

[0094] Step K332: Integrate the first matching data and the second matching data to obtain the sample dataset.

[0095] Step K333: Divide the sample dataset into training set and test set according to a preset ratio.

[0096] Step K334: Based on the training set, iteratively train the initial convolutional neural network learning model to obtain the initial classification and recognition model.

[0097] In a specific implementation, the initial convolutional neural network learning model is iteratively trained based on the training set to obtain an initial classification and recognition model, which includes steps B1 to B6.

[0098] Step B1: Preset the initial convolutional neural network learning model and set the initial parameters of the initial convolutional neural network learning model; the initial parameters include batch size, early stopping condition, and number of iterations.

[0099] Step B2: Input the training set into the initial convolutional neural network learning model in batches according to the batch size.

[0100] Step B3: During the initial convolutional neural network learning model operation, extract multi-channel shared features and generate independent binary classification results for each channel's features.

[0101] Step B4: Update the model parameters of the convolutional neural network learning model and output the updated model parameters.

[0102] Specifically, this embodiment uses the Adam optimizer to update the model parameters of the convolutional neural network. It combines the advantages of momentum and adaptive learning rate, effectively accelerating training and improving model performance. However, besides the Adam optimizer, this embodiment can also employ various other optimization strategies, such as stochastic gradient descent (SGD), which minimizes error by gradually adjusting parameters and sometimes offers more flexible control during training; Adagrad, which adjusts the learning rate based on frequent parameter updates, suitable for handling sparse data; RMSProp, which achieves more stable convergence by smoothing gradients; and Adadelta, an adaptive optimizer that helps address the learning rate decay problem. Each optimizer performs differently in different application scenarios and data distributions; choosing a suitable optimization strategy can effectively improve the model's training performance and generalization ability.

[0103] Step B5: Repeat steps B2 to B4 until the maximum number of iterations is reached or the early stopping condition is met.

[0104] Step B6: Construct the initial classification and recognition model based on the updated model parameters from the final output.

[0105] Step K335: Input the test set into the initial classification and recognition model to obtain the confusion matrix corresponding to the initial classification and recognition model.

[0106] Step K336: Evaluate the performance of the initial classification and recognition model based on the confusion matrix to obtain the evaluation results.

[0107] As a specific implementation method, the performance of the initial classification and recognition model is evaluated based on the confusion matrix to obtain the evaluation result, specifically including steps K3361 to K3362.

[0108] Step K3361: Calculate the arc recognition accuracy Acc and the arc recognition false negative rate Recall using the confusion matrix; the arc recognition accuracy Acc is the proportion of samples correctly identified by the initial classification model out of all test set samples; the arc recognition false negative rate Recall is the proportion of samples that were correctly identified by the initial classification model out of all samples that actually had an arc.

[0109] Step K3362: Based on the arc recognition accuracy Acc and the arc recognition false negative rate Recall, obtain the evaluation results of the initial classification and recognition model.

[0110] The following example will be used to illustrate step K336 in detail. First, set the indicators, specifically: arc recognition accuracy is set to 100%, arc recognition false alarm rate is set to 0%, and arc recognition delay time is 0.5s.

[0111] Next, set the confusion matrix:

[0112]

[0113] In the confusion matrix,

[0114] True Negative (TN): The model predicts the true negative class as negative as the actual negative class.

[0115] False Positive (FP): The actual value is negative, but the model incorrectly predicts it as positive.

[0116] False Negative (FN): The actual value is positive, but the model incorrectly predicts it as negative.

[0117] True Positive (TP): The actual value is positive, and the model prediction is also positive.

[0118] Arc detection accuracy Acc = (TN + FP) / (TN + FP + FN + TP)

[0119] Arc detection false negative rate Recall = TN / (TN+FP).

[0120] Then, based on the arc recognition accuracy (Acc) and the arc recognition false negative rate (Recall), the evaluation results of the initial classification and recognition model are obtained.

[0121] In some implementations, such as Figure 2As shown in step S4, this embodiment constructs a multi-channel DC arc detection model. The main structure includes a single-input multiple-output (SIMO) model and a multiple-input multiple-output (MIMO) model. It also involves convolutional neural networks (CNN), long short-term memory neural networks (LSTM), recurrent neural networks (RNN), a Transformer based on self-attention, and various machine learning models. Given the high correlation between the data from each PV branch (photovoltaic branch in a photovoltaic power generation system, referring to the electrical lines or branches used to connect solar photovoltaic modules to the inverter or distribution network), this scheme selects the CNN model from the SIMO structure to simplify the model and improve efficiency. A classification model based on amplitude data in the 3kHz to 100kHz frequency band is established to distinguish between normal and arc fault states. This SIMO model adopts a single-input multiple-output design, taking the input data from multiple PV branches (each channel size is 1024) as the overall input. Through layer-by-layer feature extraction, the classification results for each channel are finally obtained in different output layers. The model's input layer has a shape of (batchsize, 1024, n), where batchsize represents the number of samples in one training iteration, and n is the number of PV branches collected simultaneously. The feature extraction part consists of four one-dimensional convolutional modules, each containing a convolutional layer, a max-pooling layer, and a ReLU activation function. The convolutional kernel sizes are 4, 2, 2, and 4, respectively. Max pooling is used in all pooling layers to enhance the model's feature capture capability. After shared feature extraction, each PV branch can output a corresponding binary classification result (normal or arc fault). To avoid overfitting, the model can flexibly adjust the convolutional kernel size and pooling method, and regularization techniques such as Dropout are also introduced.

[0122] like Figure 3 As shown, this embodiment employs a typical one-dimensional convolutional neural network (1D-CNN) architecture, designed to process multiple channels (e.g., PV1, PV2, up to PV...). n This architecture performs parallel classification on the input time-series signals. It consists of multiple convolutional blocks, followed by fully connected dense layers, and finally outputs the classification result for each channel. Each input channel (PV1, PV2, ..., PV...)... nThese represent signal data from different measurement points or devices, input into the network as continuous time series data. Each convolutional block (Convblock) consists of three main parts: The first part is a one-dimensional convolutional layer (Conv1D), used to extract local features and patterns from the input sequence. The parameter "Filter:n" represents the number of convolutional kernels, i.e., the number of output channels. It scans the entire input using a sliding window to capture key information in the sequence. The second part is the activation function, using ReLU (Rectified Linear Unit), defined as f(x) = max(0,x). Its function is to introduce non-linearity and enhance the model's expressive ability when learning complex features. The third part is a max pooling layer (MaxPooling1D), which reduces the feature dimension and computational load by downsampling over time, while also enhancing the model's robustness to features. The parameter "Kernel:m" represents the window size covered by the pooling operation. The above three parts, arranged sequentially, constitute a standard Convblock. Stacking multiple Convblocks allows for the extraction of higher-order, more abstract feature representations layer by layer, thereby increasing the model's expressive power and achieving the goal of capturing long-range dependencies. At the back end of the model, a fully connected (Dense Layer) is used to map the extracted feature vectors to the final classification space, for each input channel (PV1, PV2, ..., PV...). n This generates independent classification prediction results. These outputs (such as PV1 classification results, PV2 classification results, and so on) generate independent classification prediction results. n This model is suitable for multi-target recognition scenarios, such as detecting the working status or fault type of each PV unit. The main features of the overall architecture include: the input layer receives raw signal data from multiple channels; Conv1D is responsible for extracting local time-series features; the ReLU activation function adds nonlinearity; MaxPooling1D achieves feature dimensionality reduction and enhances anti-interference capabilities; Convblock serves as the basic unit for feature extraction; and the Dense layer maps features to classification output. Ultimately, this deep learning model is not only compact and complete, capable of efficiently processing multi-channel signals, but also possesses excellent multi-target recognition capabilities and powerful feature extraction capabilities, adapting to various complex signal variation environments.

[0123] As a specific implementation method, this embodiment configures complete parameter settings during model training to ensure training effectiveness. The optimizer uses the Adam optimizer with a learning rate of 0.01. An early stopping strategy is employed: training stops early if the validation set loss remains unchanged for 20 consecutive epochs. To improve training stability and efficiency, a dynamic learning rate mechanism is used: if the validation set loss does not improve within 6 consecutive epochs, the current learning rate is multiplied by 0.9 for adjustment. During training, the batch size is set to 256, and the maximum number of iterations is 100 epochs. Through reasonable parameter settings and model structure design, this embodiment achieves effective identification and classification of DC arc signals.

[0124] Step K337: Adjust the parameters of the initial classification and recognition model based on the evaluation results to obtain the classification and recognition model for multi-channel DC arc detection.

[0125] As a more preferred implementation, after obtaining the classification and identification results of each channel of the multi-channel DC signal, that is, after step K3, the method further includes step K4: using the sliding window counting method to perform time-domain statistics and fault warning on the classification and identification results.

[0126] In practice, each PV channel, after being inferred by the multi-channel DC arc detection classification and recognition model, will obtain two values: a normal probability value and an arc probability value. After setting the window size, the average normal probability value = the sum of normal probability values ​​within the window / window size; the average arc probability value = the sum of arc probability values ​​within the window / window size. The final judgment result is obtained based on whether the average normal probability value and the average arc probability value are greater than their respective thresholds. Specifically, as follows... Figure 4 As shown, the window size N is set to 10. An alarm is triggered when the probability of a given event exceeds a threshold multiple times within the window. The arc probability threshold is 0.95, and the normal state threshold is 0.5. The system determines whether to trigger an alarm by sliding and analyzing consecutive data points in a fixed-size window (e.g., 10) within the data stream. Each analysis includes the 10 most recent data results, allowing for rapid decision-making based on the latest data. A high probability threshold of 0.95 is set for arc occurrence. That is, if the average arc probability calculated within the selected window exceeds 0.95, the system determines that an arc event has occurred and triggers an alarm. A normal state threshold of 0.5 is also set. If the average normal probability calculated within the selected window exceeds 0.5, the system determines that no arc has occurred and maintains a normal state. This helps the system remain silent when there is no arc, avoiding unnecessary alarms. In specific implementation, as shown... Figure 4As shown, the data sequence contains 14 normal / arc results. Each result corresponds to a probability value representing the likelihood of a normal / arc occurrence. The figure marks a continuous result sequence with a window size of 10. When data is added to the window sequence, the average probability value of the first 5 results in the normal sequence is below 0.5, and the average probability value of the first 9 results in the arc sequence is below 0.95, so no results are reported. As the sequence continues to increase, the corresponding normal / arc results are reported. This analysis method allows the system to maintain high sensitivity while avoiding false alarms caused by occasional abnormal data. The sliding window counting method can effectively monitor and predict the occurrence of events such as arcs by dynamically analyzing the latest results in the data stream. Appropriately setting the window size and threshold helps balance the system's sensitivity and stability, reducing false alarms and missed alarms.

[0127] Figure 2This document presents a complete flowchart of the multi-channel DC arc detection method in this embodiment, demonstrating closed-loop management from data acquisition to model deployment. First, in step S1, the system collects normal DC power data under different operating conditions through experimental setup or on-site monitoring. This data covers various load states, voltage levels, line lengths, and impedance conditions, constructing a "normal" sample library covering a wide range of operating environments to lay the foundation for subsequent model training. Next, step S2 cleans and preprocesses the data, including removing outliers, filling in missing data, and performing standardization or normalization. It also synchronizes the sampling time axis of the multi-channel ADC to ensure the consistency and stability of the input data, providing high-quality training material for the deep learning model. In step S3, the system further collects arc and non-arc fault state data, including experimental data from artificially simulated arcs and actual data from historical fault records. This data is then precisely labeled and classified to form a labeled training set. Subsequently, step S4 constructs features from the preprocessed samples, extracting high-dimensional feature vectors such as current mutation rate, voltage drop depth, and signal spectrum features. The dataset is then divided into training, validation, and test sets to ensure proper model training and performance evaluation. Step S5 trains the training set using a deep neural network (such as CNN), dynamically adjusting model parameters through optimizers like Adam, RMSProp, or SGD to maximize the model's performance on the validation set, ensuring the model fully learns the characteristic differences between arc and non-arc signals. Step S6 evaluates key metrics such as accuracy, recall, and F1 score using the validation or test set. If the recognition rate reaches or exceeds 95%, the model is deemed satisfactory; otherwise, step S7 proceeds to iterative training by adjusting sample composition, optimizing network structure, and adjusting hyperparameters such as learning rate or loss function until the preset standard is met. Finally, in step S8, the trained model is deployed to the actual detection system to collect and analyze voltage and current signals in real time, output the judgment result of the arc state, and trigger the corresponding early warning or recording mechanism, thereby achieving rapid and accurate detection of arc faults. This closed-loop process not only has the advantages of rich data, multi-condition adaptability, and deep learning intelligent recognition, but also significantly improves the accuracy of arc detection and system stability through continuous optimization.

[0128] The multi-channel DC arc detection method provided in this embodiment achieves high-precision, real-time detection and accurate localization of multi-channel DC arc faults by deeply fusing multi-source, multi-domain features and constructing an intelligent classification model based on convolutional neural networks. This method effectively improves data quality through a standardized preprocessing workflow and employs a strategy combining multi-channel shared features and independent binary classification in the training of the classification and recognition model for multi-channel DC arc detection. This significantly enhances the anti-interference capability and robustness of the classification and recognition model under complex operating conditions, greatly reducing the probability of false alarms and missed alarms. Furthermore, it achieves continuous tracking and timely early warning of arc faults through a sliding window counting method, thus providing efficient, reliable, and intelligent technical support for the safe operation of DC circuits in industrial production.

[0129] Example 2:

[0130] like Figure 5 As shown, this embodiment provides a multi-channel DC arc detection device, which includes:

[0131] Acquisition unit 10 is used to acquire the multi-channel DC signal to be detected;

[0132] The preprocessing unit 20, connected to the acquisition unit 10, is used to preprocess the multi-channel DC signal to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal.

[0133] The input unit 30, connected to the preprocessing unit 20, is used to input the target time-domain features and target frequency-domain features into the classification and recognition model of multi-channel DC arc detection, so as to obtain the classification and recognition results of each channel DC signal in the multi-channel DC signal, thereby realizing multi-channel DC arc detection;

[0134] Among them, the classification and recognition model for multi-channel DC arc detection is a convolutional neural network learning model built based on multi-channel DC circuit sample label data.

[0135] The apparatus in this embodiment is capable of performing the method in Embodiment 1.

[0136] Example 3:

[0137] like Figure 6 As shown, this embodiment provides an electronic device, which includes a memory 200 and a processor 100. The memory 200 stores a computer program. When the processor 100 runs the computer program stored in the memory 200, the processor 100 executes the multi-channel DC arc detection method according to Embodiment 1.

[0138] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-channel DC arc detection method, characterized in that, The method includes the following steps: Acquire the multi-channel DC signal to be detected, preprocess the multi-channel DC signal, and generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal; The target time-domain features and the target frequency-domain features are input into the classification and recognition model for multi-channel DC arc detection to obtain the classification and recognition results of each channel of the multi-channel DC signal, thereby realizing multi-channel DC arc detection. The classification and recognition model for multi-channel DC arc detection is a convolutional neural network learning model built based on multi-channel DC circuit sample label data.

2. The multi-channel DC arc detection method according to claim 1, characterized in that, The preprocessing of the multi-channel DC signal to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal specifically includes: Step A1: Remove abnormal data from the multi-channel DC signal to obtain a multi-channel DC signal with abnormalities removed; The abnormal data of the multi-channel DC signal includes multi-channel DC signals with mixed data sampling or incorrect labeling. Step A2: Extract time-domain data from the anomaly-removed multi-channel DC signal according to a preset third fixed frame length to obtain the third process time-domain features; and extract frequency-domain data from the anomaly-removed multi-channel DC signal according to a preset third fixed frame length to obtain the third process frequency-domain features. Step A3: Standardize the time-domain features of the third process to generate a third time-domain feature corresponding to the multi-channel DC signal; and standardize the frequency-domain features of the third process to generate a third frequency-domain feature corresponding to the multi-channel DC signal.

3. The multi-channel DC arc detection method according to claim 1, characterized in that, After obtaining the classification and identification result of each channel of the multi-channel DC signal, the method further includes: The sliding window counting method is used to perform time-domain statistics and fault warning on the classification and recognition results.

4. The multi-channel DC arc detection method according to any one of claims 1 to 3, characterized in that, Before inputting the target time-domain features and the target frequency-domain features into the classification and recognition model for multi-channel DC arc detection, the method further includes: Acquire sample tag data of a multi-channel DC circuit; the sample tag data includes raw normal data and raw arc data; The original normal data is preprocessed to generate a first time-domain feature and a first frequency-domain feature corresponding to the original normal data; and the original electric arc data is preprocessed to generate a second time-domain feature and a second frequency-domain feature corresponding to the original electric arc data. Based on the original normal data, the original arc data, the first time-domain feature, the first frequency-domain feature, the second time-domain feature, and the second frequency-domain feature, a classification and recognition model for multi-channel DC arc detection is constructed.

5. The multi-channel DC arc detection method according to claim 4, characterized in that, The original arc data was obtained through artificially simulated arc fault experiments or extracted from historical arc fault records.

6. The multi-channel DC arc detection method according to claim 4, characterized in that, The step of constructing a classification and recognition model for multi-channel DC arc detection based on the original normal data, the original arc data, the first time-domain feature, the first frequency-domain feature, the second time-domain feature, and the second frequency-domain feature specifically includes: The original normal data is matched with the first time-domain feature and the first frequency-domain feature to obtain first matched data; and the original arc data is matched with the second time-domain feature and the second frequency-domain feature to obtain second matched data. The first matching data and the second matching data are integrated to obtain a sample dataset; The sample dataset is divided into a training set and a test set according to a preset ratio; Based on the training set, the initial convolutional neural network learning model is iteratively trained to obtain the initial classification and recognition model; Input the test set into the initial classification and recognition model to obtain the confusion matrix corresponding to the initial classification and recognition model; The performance of the initial classification and recognition model is evaluated based on the confusion matrix to obtain the evaluation results; Based on the evaluation results, the parameters of the initial classification and recognition model are adjusted to obtain a classification and recognition model for multi-channel DC arc detection.

7. The multi-channel DC arc detection method according to claim 6, characterized in that, The step of iteratively training the initial convolutional neural network learning model based on the training set to obtain the initial classification and recognition model specifically includes: Step B1: Preset the initial convolutional neural network learning model and set the initial parameters of the initial convolutional neural network learning model; the initial parameters include batch size, early stopping condition, and number of iterations; Step B2: Input the training set into the initial convolutional neural network learning model in batches according to batch size; Step B3: During the operation of the initial convolutional neural network learning model, extract multi-channel shared features and generate independent binary classification results for the features of each channel; Step B4: Update the model parameters of the convolutional neural network learning model and output the updated model parameters; Step B5: Repeat steps B2 to B4 until the maximum number of iterations is reached or the early stopping condition is met; Step B6: Construct the initial classification and recognition model based on the updated model parameters from the final output.

8. The multi-channel DC arc detection method according to claim 6, characterized in that, The step of evaluating the performance of the initial classification and recognition model based on the confusion matrix to obtain the evaluation result specifically includes: The arc identification accuracy and arc identification false alarm rate are calculated using the confusion matrix. Wherein, the arc recognition accuracy rate is the proportion of samples correctly identified by the initial classification and recognition model out of all test set samples; the arc recognition false negative rate is the proportion of samples that were correctly identified by the initial classification and recognition model out of all samples that actually experienced an arc. The evaluation results of the initial classification and recognition model are obtained based on the arc recognition accuracy and the arc recognition false negative rate.

9. A multi-channel DC arc detection device, characterized in that, include: The acquisition unit is used to acquire the multi-channel DC signal to be detected; A preprocessing unit, connected to the acquisition unit, is used to preprocess the multi-channel DC signal to generate target time-domain features and target frequency-domain features corresponding to the multi-channel DC signal. An input unit, connected to the preprocessing unit, is used to input the target time-domain features and the target frequency-domain features into the classification and recognition model for multi-channel DC arc detection, so as to obtain the classification and recognition result of each channel DC signal in the multi-channel DC signal, thereby realizing multi-channel DC arc detection; The classification and recognition model for multi-channel DC arc detection is a convolutional neural network learning model built based on multi-channel DC circuit sample label data.

10. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the multi-channel DC arc detection method according to any one of claims 1 to 8.

Images