DC arc fault detection method and device, medium and product

Abstract

The invention provides a direct-current arc fault detection method and device, a medium and a product, and relates to the technical field of arc fault detection, and the method comprises the steps: collecting a current signal of a time sequence in a direct-current loop through a current sensor; processing the current signal by adopting a discrete wavelet transform method, inputting a generated wavelet coefficient matrix into a predetermined 1D-CNN network to extract a depth time-frequency feature, calculating a statistical feature of a wavelet coefficient, and performing data fusion on the depth time-frequency feature and the statistical feature to obtain a multi-scale time-frequency fusion feature; and inputting the multi-scale time-frequency fusion features into a pre-trained RF classifier model, and outputting an arc fault occurrence probability. The problem that the arc fault is difficult to detect accurately in the prior art because the difference between the normal current signal and the arc fault current signal is small is solved; and the occurrence probability of the arc fault is difficult to accurately judge.

Description

Technical Field

[0001] This invention relates to the field of arc fault detection technology, and in particular to a DC arc fault detection method, device, medium, and product. Background Technology

[0002] During the operation of a DC power supply system, the occurrence of arc faults can seriously threaten the safe and stable operation of the system, and may even lead to safety accidents such as equipment damage and fires. Therefore, accurate detection of arc faults in DC circuits is of great practical significance.

[0003] In existing technologies, the detection of DC arc faults largely relies on the time-domain characteristic analysis of current signals, such as collecting parameters like amplitude, RMS value, and harmonic content, and determining whether an arc fault has occurred by setting thresholds. However, in actual DC circuit operating conditions, the difference between normal current signals and arc fault current signals is often small, especially in the early stages of an arc fault, when the amplitude change of the fault current is not significant. Relying solely on a single time-domain characteristic parameter is insufficient to effectively distinguish between normal and fault currents, easily leading to missed or false detections. Furthermore, traditional arc fault detection methods mostly only output binary classification results of fault or normal, failing to quantitatively assess the probability of arc fault occurrence. This qualitative detection approach is ill-suited to the fault early warning requirements of high-precision power supply systems, exhibiting poor detection accuracy and low efficiency. Summary of the Invention

[0004] The embodiments of the present invention provide a DC arc fault detection method, device, medium and product, which aims to solve the problem that the existing technology is difficult to accurately detect arc faults due to the small difference between normal current signals and arc fault current signals, and is also difficult to accurately determine the probability of arc fault occurrence.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for detecting DC arc faults, comprising the following steps: The current signal of the DC circuit in a time series is acquired by a current sensor; The current signal is processed using the discrete wavelet transform method. The generated wavelet coefficient matrix is ​​input into a predetermined 1D-CNN network to extract deep time-frequency features. At the same time, the statistical features of the wavelet coefficients are calculated. The deep time-frequency features and the statistical features are fused to obtain multi-scale time-frequency fusion features. The multi-scale time-frequency fusion features are input into a pre-trained RF classifier model, which outputs the probability of arc fault occurrence. When the probability of an arc fault occurring exceeds a predetermined probability threshold, an arc fault is determined to have occurred, and an early warning mechanism is triggered.

[0006] Furthermore, the current signal is captured using a sliding window of fixed length, and the adjacent windows of the sliding window overlap for a predetermined time length.

[0007] Furthermore, it also includes: normalizing the multi-scale time-frequency fusion features.

[0008] Furthermore, the statistical characteristics include mean, standard deviation, variance, maximum value, minimum value, median, energy, quantile, kurtosis and / or skewness.

[0009] Furthermore, in the step of processing the current signal using the discrete wavelet transform method, the discrete wavelet transform uses the 'db4' wavelet basis, and the current signal is decomposed into four levels using the 'db4' wavelet basis to generate five levels of wavelet transform coefficients, and the five levels of wavelet transform coefficients are organized into a multi-channel matrix form.

[0010] Furthermore, the discrete wavelet transform processes the current signal using the following mathematical expression: , In the formula, These are wavelet transform coefficients, representing the signal at different scales. and location The intensity of the time-frequency characteristics at that location; This is a scaling factor used to control the frequency of the wavelet; This is the translation factor, used to control the time position of the wavelet; It is a current signal. For time; Mother wavelet function The function after scaling, translation, and taking the complex conjugate.

[0011] Furthermore, when an arc fault occurs, the arc event data is entered into a pre-built arc fingerprint database. The arc event data includes timestamps, probability values, waveform data, and / or multi-scale time-frequency fusion features. The arc fingerprint database also uses a predetermined incremental learning algorithm to update the RF classifier model based on newly added arc event data samples.

[0012] Secondly, the present invention provides a DC arc fault detection device, including a memory and a processor, wherein the memory stores at least one program, and the at least one program is executed by the processor to implement the DC arc fault detection method as described above.

[0013] Thirdly, the present invention provides a computer-readable storage medium storing at least one program, wherein the at least one program is executed by a processor to implement the DC arc fault detection method as described above.

[0014] Fourthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the DC arc fault detection method as described above.

[0015] The above technical solution has the following technical effects: A time-series current signal in a DC circuit is acquired using a current sensor. The current signal is processed using discrete wavelet transform, and the generated wavelet coefficient matrix is ​​input into a pre-defined 1D-CNN network to extract deep time-frequency features. Simultaneously, statistical features of the wavelet coefficients are calculated. The deep time-frequency features and the statistical features are then fused to obtain multi-scale time-frequency fusion features. These multi-scale time-frequency fusion features are input into a pre-trained RF classifier model to output the probability of arc fault occurrence. This invention solves the problem that existing technologies struggle to accurately detect arc faults and accurately determine their probability of occurrence due to the small difference between normal and arc fault current signals. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of a DC arc fault detection method according to an embodiment of the present invention; Figure 2 This is a schematic diagram comparing the characteristic coefficients of a normal current signal and an arc fault current signal after wavelet transform according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a DC arc fault detection device according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0018] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0019] Example 1: Figure 1 This is a flowchart illustrating a DC arc fault detection method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method of this embodiment includes the following steps: The current signal of the DC circuit in a time series is acquired by a current sensor; In one specific implementation, the current signal is captured using a sliding window of fixed length, with adjacent windows overlapping for a predetermined time length to balance real-time detection and data integrity.

[0020] In one specific implementation, this embodiment is applied to a 100kW / 215kWh lithium-ion battery energy storage cabinet. The current sensing unit uses a TMR7559-B tunnel magnetoresistive current sensor, installed at the positive or negative terminal of the DC bus of the energy storage battery cluster. Its sensitivity is 6.25mV / A, and its bandwidth is DC-150kHz, capable of accurately capturing microampere-level current changes. The data acquisition unit uses a USB3133A data acquisition card, with a sampling frequency set to 10kHz to ensure complete acquisition of the high-frequency components of the electric arc. An Intel Core i7-8700T industrial-grade control computer running Windows 10 is used to handle the detection algorithm software. The algorithm's software implementation is based on the Python programming language, mainly relying on libraries such as pywt, scikit-learn, and numpy. Data preparation for model training is conducted on an experimental platform using an arc generator conforming to the UL 1699B standard to simulate a series arc. A total of 150,000 normal current sample data points and 100,000 arc current sample data points were collected. Each sample data point is 1000 points long and corresponds to a duration of 50ms.

[0021] The discrete wavelet transform method is used to process the current signal. The generated wavelet coefficient matrix is ​​input into a predetermined 1D-CNN network to extract the depth time-frequency features. At the same time, the statistical features of the wavelet coefficients are calculated. The depth time-frequency features and statistical features are fused to obtain multi-scale time-frequency fusion features. In one specific implementation, the discrete wavelet transform uses the 'db4' wavelet basis.

[0022] In one specific implementation, the discrete wavelet transform uses mathematical expressions to process current signals: , In the formula, These are wavelet transform coefficients, representing the signal at different scales. and location The intensity of the time-frequency characteristics at that location; This is a scaling factor used to control the frequency of the wavelet; This is the translation factor, used to control the time position of the wavelet; It is a current signal. For time; Mother wavelet function The function after scaling, translation, and taking the complex conjugate. Multi-resolution analysis is achieved by discretizing a and b; in one specific implementation, the discretization of a and b is taken... , , j, k∈Z.

[0023] In one specific implementation, the `pywt.wavedec` function is called for each sample data, performing a 4-layer decomposition using the 'db4' wavelet basis to generate a 5-layer coefficient matrix. A dual-branch feature fusion strategy is adopted: the CNN branch takes the 5-layer wavelet transform coefficients as multi-channel input and automatically extracts 32-dimensional depth time-frequency features through the 1D-CNN network; the statistical feature branch calculates 11 statistics for each layer of coefficients, including mean, standard deviation, variance, maximum, minimum, median, energy, quantile, kurtosis, and skewness, forming 55-dimensional statistical features; finally, the depth time-frequency features and statistical features are fused into an 87-dimensional multi-scale time-frequency fusion feature. The 1D-CNN network contains two convolutional blocks, employing a 32→64 filter, global average pooling, and fully connected layers.

[0024] In this embodiment, wavelet transform, with its inherent denoising capabilities, effectively separates signal from noise. 1D-CNN, through weight sharing and local connectivity of convolutional kernels, further enhances its robustness against noise. The random forest algorithm, by voting on the results of multiple decision trees, naturally possesses advantages in resisting overfitting and reducing variance. This triple-technology combination enables the proposed solution to exhibit exceptional robustness against common energy storage system malfunctions such as inverter switching noise, load switching disturbances, and electromagnetic radiation interference. Experimental results show that, even under strong noise conditions, its area under the receiver operating characteristic (AUC) remains consistently above 0.99, a performance metric far exceeding any existing threshold-based or simple model-based detection scheme.

[0025] By optimizing feature dimensions, limiting the wavelet decomposition layer to four layers, employing a lightweight 1D-CNN network structure, selecting a computationally efficient random forest model, and using a sliding window overlapping sampling mechanism at the software level, the entire process from signal acquisition to fault determination was successfully completed within a 50ms response time, meeting the stringent requirements of energy storage systems for protection action speed. Simultaneously, its key indicators such as detection accuracy and recall rate are all better than 99.9%, perfectly resolving the technical contradiction between high-precision detection and rapid response.

[0026] In one specific implementation, the method further includes: normalizing the multi-scale time-frequency fusion features to the [0,1] interval to eliminate the influence of dimensions.

[0027] The multi-scale time-frequency fusion features are input into a pre-trained RF classifier model, which outputs the probability of arc fault occurrence. In this embodiment, Discrete Wavelet Transform (DWT), 1D-CNN deep learning, and Random Forest (RF) ensemble learning algorithms are deeply integrated to form a complete technology chain of "multi-scale feature extraction - deep feature fusion - intelligent decision-making." DWT provides a "mathematical microscope" function for signals, accurately stripping away the high-frequency, burst, and non-stationary features of electric arcs; 1D-CNN automatically learns deep nonlinear patterns from the spatiotemporal relationships of wavelet coefficients; and the RF algorithm excels at handling high-dimensional, nonlinear feature spaces. This fusion of three technologies produces a powerful synergistic effect: DWT provides the optimal basis function for feature extraction, CNN automates and deepens feature representation, and RF ensures the accuracy and stability of classification decisions, ultimately achieving a qualitative leap in detection performance.

[0028] In one specific implementation, the model outputs a probability value P representing the existence of an arc fault in the current data segment. The probability value P(arc fault) = (number of decision trees that predict "arc fault") / (total number of decision trees). A random forest consists of 100 decision trees (n_estimators=100). For an input sample: if 95 trees vote that it is an "arc fault" and 5 trees vote that it is "normal", then the probability value output by the RF model is: P = 95 / 100 = 0.95. This probability value P ∈ [0, 1] directly represents the model's confidence or certainty in judging that "the sample is an arc fault". The closer the P value is to 1, the more certain the model is that a fault has occurred; the closer it is to 0, the more certain it is that it is operating normally.

[0029] The core bottleneck of existing technologies lies in their reliance on fixed thresholds. This invention completely abandons the traditional approach of determining thresholds based on current, di / dt, or high-frequency energy, and adopts a novel, data-driven, end-to-end intelligent decision-making paradigm. The system's final output is a probability value calculated based on a deep learning model. The decision threshold T is merely a post-parameter used to adjust sensitivity and specificity, rather than a pre-set physical quantity threshold that requires experience. This greatly enhances the system's adaptability and generalization capabilities to different application scenarios, solving a long-standing technical problem that has plagued those skilled in the art.

[0030] In one specific implementation, the 1D-CNN network architecture includes: Input layer: Receives a wavelet coefficient matrix of shape (5, N); First convolutional block: 32 filters, kernel size 3, ReLU activation, batch normalization, max pooling (pooling size 2), dropout rate 0.2; Second convolutional block: 64 filters, kernel size 3, ReLU activation, batch normalization, max pooling (pooling size 2), dropout rate 0.2; Global average pooling layer: replaces flattening operations, reducing the number of parameters; Fully connected layer: 64 neurons, ReLU activation, L2 regularization (0.01), batch normalization, dropout rate 0.3; Feature output layer: 32 neurons, outputting deep feature vectors.

[0031] In one specific implementation, the RF (Random Forest Classifier Model) is trained using the following optimized combination of hyperparameters: the number of base learners (n_estimators) is 100; the maximum tree depth (max_depth) is 15; the minimum number of samples required for internal node splitting (min_samples_split) is 10; the minimum number of samples required for leaf nodes (min_samples_leaf) is 4; the maximum number of features (max_features) is 'sqrt'; whether bootstrap sampling is used; the class weight (class_weight) is 'balanced'; and whether out-of-bag (OOB) scores (oob_score) are calculated.

[0032] When the probability of an arc fault occurring exceeds a predetermined probability threshold, an arc fault is determined to have occurred, and an early warning mechanism is triggered. In one specific implementation, an adaptive threshold decision mechanism is used, with a default probability threshold T of 0.6, which can be dynamically adjusted by the user. In another specific implementation, the value of T is within the range of 0.1 ≤ T ≤ 0.9 for comparison; if P ≥ T, an arc fault is determined to have occurred, and the early warning mechanism is immediately triggered; otherwise, the system is considered to be operating normally.

[0033] In one specific implementation, when an arc fault occurs, the system immediately performs three operations: a red warning window pops up on the software interface and an alarm is issued; a tripping signal is sent to the DC circuit breaker; event details, including timestamp, probability value, waveform data, and feature vector, are recorded in the SQLite database; and the fingerprint database update process is triggered. In another specific implementation, the system integrates an arc fingerprint database, automatically recording the feature data and metadata of the detected event. When the number of new samples reaches a threshold (≥20) and occurs during system idle time (e.g., 22:00-06:00), an incremental learning process is automatically triggered: new samples are loaded and data quality is verified; incremental training is performed based on the existing model; after performance verification, a hot model update is performed; update logs are recorded and version rollback is supported. Through a self-evolution mechanism, the system can continuously optimize model performance and adapt to changes in the operating environment.

[0034] This invention introduces a self-evolving system architecture for the first time in the field of arc detection. By continuously accumulating real-world operational data through an arc fingerprint database, it automatically triggers incremental learning during system downtime, enabling hot updates and continuous optimization of the model. This design gives the system the characteristic of "becoming smarter with use," adapting to feature drift caused by equipment aging and environmental changes, thus overcoming the limitations of traditional detection methods that require "one-time training and lifelong use." The system supports seamless online upgrades, significantly reducing maintenance costs and ensuring long-term reliability and accuracy.

[0035] In one specific implementation, the arc fingerprint database adopts a table structure design, including: The `arc_fingerprints` table stores fingerprint ID, timestamp, tag, source, confidence level, verification status, and remarks. The raw_data table stores raw current data and is associated with the fingerprint ID; The feature_data table stores feature vectors and supports feature-level queries. The `model_updates` table records the path, performance metrics, and timestamp for each model update.

[0036] In one specific implementation, Figure 2 This is a schematic diagram comparing the characteristic coefficients of a normal current signal and an arc fault current signal after wavelet transform according to an embodiment of the present invention; as shown. Figure 2 As shown, the normal current (represented by the blue curve) fluctuates relatively smoothly and periodically, while the arc current (represented by the red curve) exhibits typical fault characteristics such as drastic amplitude changes, irregular spikes, and pulses.

[0037] The wavelet coefficient comparison chart below more clearly reveals the essential differences between the two signals: The wavelet coefficients of normal current have a much smaller range and fluctuation amplitude than those of arc current across all scales, from low-frequency cA to high-frequency cD.

[0038] Especially at the high-frequency detail coefficients such as cD1 and cD2, the arc current exhibits a significant and dispersed coefficient distribution, which directly corresponds to its high-frequency abrupt change characteristics in the time domain; while the high-frequency coefficients of the normal current are closely clustered near zero.

[0039] This comparative result strongly demonstrates that the multi-scale time-frequency features extracted through discrete wavelet transform can effectively and clearly amplify the differences between normal and arc fault states, providing highly separable feature inputs for accurate classification by subsequent machine learning models. This is the key reason for its superiority over traditional time-domain or frequency-domain methods. In one specific implementation, the fusion model exhibits excellent performance on the test set: accuracy 99.97%, precision 99.88%, recall 100.0%, and F1 score 99.94%. Compared with a single random forest model, the fusion model further improves precision and F1 score while maintaining high recall, demonstrating the technical advantages of deep feature fusion.

[0040] Example 2: Figure 3 This is a schematic diagram of the structure of a DC arc fault detection device according to an embodiment of the present invention, as shown below. Figure 3 As shown, the device includes a processor 301, a memory 302, a bus 303, and a computer program stored in the memory 302 and executable on the processor 301. The processor 301 includes one or more processing cores. The memory 302 is connected to the processor 301 via the bus 303. The memory 302 is used to store program instructions. When the processor executes the computer program, it implements the steps in the above-described method embodiment of Embodiment 1 of the present invention.

[0041] Furthermore, as an executable solution, the DC arc fault detection device can be a computer unit, which can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer unit may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described computer unit structure is merely an example and does not constitute a limitation on the computer unit. It may include more or fewer components, or combine certain components, or use different components. For example, the computer unit may also include input / output devices, network access devices, buses, etc., and this embodiment of the invention does not limit this.

[0042] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or other general-purpose processors, 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, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the computer unit, connecting various parts of the entire computer unit via various interfaces and lines.

[0043] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer unit by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0044] Example 3: The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the methods described in the embodiments of the present invention.

[0045] If the modules / units integrated in the computer unit are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0046] Example 4: The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the DC arc fault detection method as described above.

[0047] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for detecting DC arc faults, characterized in that, Includes the following steps: The current signal of the DC circuit in a time series is acquired by a current sensor; The current signal is processed using the discrete wavelet transform method. The generated wavelet coefficient matrix is ​​input into a predetermined 1D-CNN network to extract deep time-frequency features. At the same time, the statistical features of the wavelet coefficients are calculated. The deep time-frequency features and the statistical features are fused to obtain multi-scale time-frequency fusion features. The multi-scale time-frequency fusion features are input into a pre-trained RF classifier model, which outputs the probability of arc fault occurrence. When the probability of an arc fault occurring exceeds a predetermined probability threshold, an arc fault is determined to have occurred, and an early warning mechanism is triggered.

2. The DC arc fault detection method according to claim 1, characterized in that, The current signal is captured using a sliding window of fixed length, and the adjacent windows of the sliding window overlap for a predetermined time length.

3. The DC arc fault detection method according to claim 1, characterized in that, Also includes: The multi-scale time-frequency fusion features are normalized.

4. The DC arc fault detection method according to claim 1, characterized in that, The statistical characteristics include mean, standard deviation, variance, maximum value, minimum value, median, energy, quantile, kurtosis and / or skewness.

5. The DC arc fault detection method according to claim 1, characterized in that, In the step of processing the current signal using the discrete wavelet transform method, the discrete wavelet transform uses the 'db4' wavelet basis. The current signal is decomposed into four levels using the 'db4' wavelet basis to generate five levels of wavelet transform coefficients, and the five levels of wavelet transform coefficients are organized into a multi-channel matrix form.

6. The DC arc fault detection method according to claim 1, characterized in that, The discrete wavelet transform processes the current signal using the following mathematical expression: ， In the formula, These are wavelet transform coefficients, representing the signal at different scales. and location The intensity of the time-frequency characteristics at that location; This is a scaling factor used to control the frequency of the wavelet; This is the translation factor, used to control the time position of the wavelet; It is a current signal. For time; It is the function obtained by scaling, translating and taking the complex conjugate of the mother wavelet function.

7. The DC arc fault detection method according to claim 1, characterized in that, When an arc fault occurs, the arc event data is entered into a pre-built arc fingerprint database. The arc event data includes timestamps, probability values, waveform data, and / or multi-scale time-frequency fusion features. The arc fingerprint database also uses a predetermined incremental learning algorithm to update the RF classifier model based on newly added arc event data samples.

8. A DC arc fault detection device, characterized in that, The method includes a memory and a processor, wherein the memory stores at least one program, which is executed by the processor to implement the DC arc fault detection method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The storage medium stores at least one program, which is executed by a processor to implement the DC arc fault detection method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the DC arc fault detection method as described in any one of claims 1-7.

Images