Arc detection method based on an interpretable deep learning model and related devices

Abstract

The application relates to the field of detection, and discloses an arc detection method based on an interpretable deep learning model and related equipment thereof. The method comprises the following steps: acquiring a current signal under a direct-current arc fault state and image data corresponding to the current signal; performing time-frequency conversion processing on the image data corresponding to the current signal to obtain image time-frequency features; inputting the image time-frequency features into a preset interpretable deep learning model to generate feature density scatter plots corresponding to each observation frequency band; based on model interpretability values in the feature density scatter plots corresponding to each observation frequency band, calculating bandwidth ranges corresponding to the image time-frequency features in each observation frequency band; and based on the bandwidth ranges and the corresponding image time-frequency features, detecting a direct-current arc. According to the application, the image data obtained by a monitoring device is processed to obtain feature density scatter plots corresponding to the image data, the most relevant observation frequency band and the corresponding image features are screened, and a direct-current arc fault is accurately detected.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of detection, in particular to an arc detection method based on an interpretable deep learning model and a related device thereof. BACKGROUND

[0002] In recent years, the upstream and downstream market scale of China's photovoltaic industry is gradually expanding. Due to the increasing policy support, the proportion of photovoltaic power generation is increasing. Due to the long-term exposure of photovoltaic power station equipment to harsh environments, long service life of equipment, and difficulty in daily maintenance, the insulation layer of the line is damaged, faults occur frequently, and the incidence of electrical accidents is increasing. The electrical fire caused by arc fault in the direct current system of the photovoltaic power station seriously threatens the safety of users and the safety of the power grid. Due to the damage of the wire, the line will have a breakpoint under the action of voltage, which will cause the air to be broken down to produce an arc. The arc fault may damage the electrical equipment itself or endanger life safety due to external factors or manufacturing problems of the equipment. The traditional leakage protection device can only work when a short circuit occurs in the line, and cannot locate and determine the fault, resulting in large power loss and consuming a lot of time and labor cost. This also makes the market demand for arc fault detection technology in the direct current system large.

[0003] Most of the detection systems on the market today only use data regularity induction methods, and the features lack fault explanation and have weak correlation, which may cause detection failure due to unreliable features. SUMMARY

[0004] The main purpose of the present application is to solve the technical problem of low reliability of direct current arc fault detection.

[0005] The present application provides an arc detection method based on an interpretable deep learning model, which comprises: obtaining a current signal under a direct current arc fault state and at least one image data corresponding to the current signal; performing time-frequency transformation processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; inputting the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot comprises a model interpretability value corresponding to the at least one image time-frequency feature; based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band, calculating the bandwidth range corresponding to each observation frequency band of the at least one image time-frequency feature; and detecting the direct current arc based on the bandwidth range and the corresponding image time-frequency feature.

[0006] Optionally, in a first implementation of the first aspect of the present invention, after acquiring the current signal under the DC arc fault state and at least one image data corresponding to the current signal, the method further includes: acquiring the sampling frequencies of the current signal and the image data respectively, wherein the sampling frequencies include the signal sampling frequency and the image sampling frequency; determining whether there is a frequency doubling relationship between the signal sampling frequency and the image sampling frequency; if so, processing the image data based on the frequency doubling relationship; if not, calculating the greatest common divisor of the signal sampling frequency and the image sampling frequency, and updating the current signal and the image data based on the greatest common divisor.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the step of calculating the greatest common divisor of the signal sampling frequency and the image sampling frequency, and updating the current signal and the image data based on the greatest common divisor, includes: calculating the greatest common divisor of the signal sampling frequency and the image sampling frequency; determining the processing frequency of the signal sampling frequency and the image sampling frequency based on the greatest common divisor; dividing the time window of the current signal and the image data into equal parts based on the processing frequency; calculating the average data value of the current signal and the image data in the time window respectively, and updating the current signal and the image data respectively based on the average value.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band includes: performing wavelet transform on the at least one image time-frequency feature to obtain corresponding wavelet decomposition data; inputting the wavelet decomposition data into a preset interpretable deep learning model to obtain corresponding model interpretability values; and generating a feature density scatter plot corresponding to each observation frequency band based on the model interpretability values.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing wavelet transform on at least one image time-frequency feature to obtain corresponding wavelet decomposition data includes: determining the number of decomposition layers applied to the image time-frequency feature; dividing the frequency bands according to the number of decomposition layers to obtain multiple observation frequency bands; and performing wavelet transform on at least one image time-frequency feature based on a preset number of decomposition layers to obtain wavelet decomposition data segmented into observation frequency bands corresponding to the number of decomposition layers.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the model interpretability value includes a maximum interpretability value and a minimum interpretability value; the step of calculating the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band includes: traversing and calculating the maximum interpretability value and the minimum interpretability value of the at least one image time-frequency feature in each observation frequency band; and calculating the bandwidth range of each image time-frequency feature based on the maximum interpretability value and the minimum interpretability value.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the step of detecting the DC arc based on the bandwidth range and the corresponding image time-frequency features includes: selecting the observation frequency band with the largest bandwidth range in the image time-frequency features as the detection frequency band; monitoring image data corresponding to the image time-frequency features based on the detection frequency band and the image time-frequency features corresponding to the detection frequency band, thereby realizing the detection of the DC arc.

[0012] Optionally, in the seventh implementation of the first aspect of the present invention, the step of selecting the observation frequency band with the largest bandwidth range among the image time-frequency features as the detection frequency band includes: selecting the observation frequency band with the largest bandwidth range among the image time-frequency features; determining whether the largest bandwidth ranges are the same; if not, then applying the observation frequency band with the largest bandwidth range as the detection frequency band; if so, then removing image time-frequency features with the same largest bandwidth range.

[0013] Optionally, in an eighth implementation of the first aspect of the present invention, after removing the image time-frequency features with the same bandwidth range and the largest bandwidth, the method includes: if all the image time-frequency features are removed, then reselecting a second decomposition layer applied to the image time-frequency features; and processing the at least one image data obtained based on the second decomposition layer.

[0014] Optionally, in a ninth implementation of the first aspect of the present invention, the step of detecting the DC arc based on the bandwidth range and the corresponding image time-frequency features further includes: if the number of image data is greater than one, selecting the observation frequency band with the largest bandwidth range for each image time-frequency feature corresponding to the image data as a detection frequency band to obtain a detection frequency band set, wherein the detection frequency band set includes detection frequency bands corresponding to the number of image data; comparing several bandwidth ranges corresponding to the detection frequency bands in the detection frequency band set, and applying the detection frequency band with the largest bandwidth range and the image time-frequency features corresponding to the detection frequency bands to monitor the image data corresponding to the image time-frequency features to achieve the detection of the DC arc.

[0015] A second aspect of the present invention provides an arc detection device based on an interpretable deep learning model. The arc detection device includes: a data acquisition module for acquiring a current signal under a DC arc fault state and at least one image data corresponding to the current signal; a time-frequency transformation module for performing time-frequency transformation processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; a density scatter plot generation module for inputting the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes a model interpretability value corresponding to the at least one image time-frequency feature; a bandwidth range calculation module for calculating the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band; and a detection module for detecting the DC arc based on the bandwidth range and the corresponding image time-frequency feature.

[0016] Optionally, in a first implementation of the second aspect of the present invention, the arc detection device based on an interpretable deep learning model further includes a sampling frequency update module. The sampling frequency update module includes: a sampling frequency acquisition unit, which acquires the sampling frequencies of the current signal and the image data respectively, wherein the sampling frequency includes the signal sampling frequency and the image sampling frequency; a frequency judgment unit, which determines whether there is a frequency doubling relationship between the signal sampling frequency and the image sampling frequency; a frequency doubling processing unit, which processes the image data based on the frequency doubling relationship if the relationship exists; and a common divisor calculation unit, which calculates the greatest common divisor of the signal sampling frequency and the image sampling frequency if the relationship does not exist, and updates the current signal and the image data based on the greatest common divisor.

[0017] Optionally, in a second implementation of the second aspect of the present invention, the common divisor calculation unit includes: calculating the greatest common divisor of the signal sampling frequency and the image sampling frequency; determining a processing frequency for the signal sampling frequency and the image sampling frequency based on the greatest common divisor; dividing the time window of the current signal and the image data into equal parts based on the processing frequency; calculating the average data value of the current signal and the image data in the time window respectively, and updating the current signal and the image data respectively based on the average value.

[0018] Optionally, in a third implementation of the second aspect of the present invention, the density scatter plot generation module includes: a wavelet transform unit, which performs wavelet transform on at least one image time-frequency feature to obtain corresponding wavelet decomposition data; a model interpretability numerical calculation unit, which inputs the wavelet decomposition data into a preset interpretable deep learning model to obtain corresponding model interpretability values; and a density scatter plot generation unit, which generates feature density scatter plots corresponding to each observation frequency band based on the model interpretability values.

[0019] Optionally, in a fourth implementation of the second aspect of the present invention, the wavelet transform unit includes: determining the number of decomposition layers applied to the image time-frequency features; dividing the frequency bands according to the number of decomposition layers to obtain multiple observation frequency bands; and performing wavelet transform on at least one image time-frequency feature based on a preset number of decomposition layers to obtain wavelet decomposed data segmented into observation frequency bands corresponding to the number of decomposition layers.

[0020] Optionally, in a fifth implementation of the second aspect of the present invention, the bandwidth range calculation module includes: traversing and calculating the maximum and minimum interpretability values ​​of the at least one image time-frequency feature under each of the observed frequency bands; and calculating the bandwidth range of each of the image time-frequency features based on the maximum and minimum interpretability values.

[0021] Optionally, in a sixth implementation of the second aspect of the present invention, the detection module includes: a detection frequency band selection unit, which selects the observation frequency band with the largest bandwidth range in the image time-frequency features as the detection frequency band; and a DC arc detection unit, which monitors image data corresponding to the image time-frequency features based on the detection frequency band and the image time-frequency features corresponding to the detection frequency band, thereby realizing the detection of DC arc.

[0022] Optionally, in a seventh implementation of the second aspect of the present invention, the detection frequency band selection unit includes: selecting the observation frequency band with the largest bandwidth range among the image time-frequency features; determining whether the largest bandwidth ranges are the same; if not, then using the observation frequency band with the largest bandwidth range as the detection frequency band; if so, then removing image time-frequency features with the same largest bandwidth range.

[0023] Optionally, in an eighth implementation of the second aspect of the present invention, the detection module further includes a decomposition layer reselection unit, the decomposition layer reselection unit comprising: if all the image time-frequency features are removed, then reselecting a second decomposition layer applied to the image time-frequency features; and processing the at least one image data obtained based on the second decomposition layer.

[0024] Optionally, in a ninth implementation of the second aspect of the present invention, the detection module further includes: if the number of image data is greater than one, selecting the observation frequency band with the largest bandwidth range as the detection frequency band for the image time-frequency features corresponding to the image data, thereby obtaining a detection frequency band set, wherein the detection frequency band set includes detection frequency bands corresponding to the number of image data; comparing several bandwidth ranges corresponding to the detection frequency bands in the detection frequency band set, and applying the detection frequency band with the largest bandwidth range and the image time-frequency features corresponding to the detection frequency bands to monitor the image data corresponding to the image time-frequency features, thereby realizing the detection of DC arc.

[0025] A third aspect of the present invention provides an arc detection device based on an interpretable deep learning model, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the arc detection device based on the interpretable deep learning model to perform the various steps of the above-described arc detection method based on the interpretable deep learning model.

[0026] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps of the above-described arc detection method based on an interpretable deep learning model.

[0027] The technical solution provided by this invention involves acquiring a current signal under a DC arc fault state and at least one image data corresponding to the current signal; performing time-frequency transformation processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; inputting the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes a model interpretability value corresponding to the at least one image time-frequency feature; calculating the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band; and detecting the DC arc based on the bandwidth range and the corresponding image time-frequency feature. This application processes image data obtained from monitoring equipment to obtain a feature density scatter plot corresponding to the image data, filters the observation frequency band with the strongest correlation and the corresponding image features, and accurately detects the occurring DC arc fault. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the first embodiment of the arc detection method based on an interpretable deep learning model in this invention.

[0029] Figure 2This is a schematic diagram of a second embodiment of the arc detection method based on an interpretable deep learning model in this invention.

[0030] Figure 3 This is a schematic diagram of the third embodiment of the arc detection method based on an interpretable deep learning model in this invention.

[0031] Figure 4 This is a schematic diagram of one embodiment of the arc detection device based on an interpretable deep learning model according to the present invention;

[0032] Figure 5 This is a schematic diagram of another embodiment of the arc detection device based on an interpretable deep learning model according to the present invention;

[0033] Figure 6 This is a schematic diagram of an embodiment of an arc detection device based on an interpretable deep learning model according to the present invention;

[0034] Figure 7 This is a schematic diagram of the feature density scatter plot of image time-frequency features in an embodiment of the present invention;

[0035] Figure 8 This is another schematic diagram of the feature density scatter plot of the time-frequency features of an image in an embodiment of the present invention. Detailed Implementation

[0036] This invention provides an arc detection method based on an interpretable deep learning model. The method includes: acquiring a current signal under a DC arc fault state and at least one image data corresponding to the current signal; performing time-frequency transformation processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; inputting the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes a model interpretability value corresponding to the at least one image time-frequency feature; calculating the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band; and detecting the DC arc based on the bandwidth range and the corresponding image time-frequency feature. This application processes image data obtained from a monitoring device to obtain a feature density scatter plot corresponding to the image data, filters the observation frequency bands with the strongest correlation and the corresponding image features, and accurately detects the occurring DC arc fault.

[0037] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0038] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The first embodiment of the arc detection method based on an interpretable deep learning model in this invention includes:

[0039] 101. Acquire the current signal under DC arc fault conditions and at least one image data corresponding to the current signal;

[0040] In this embodiment, by acquiring the current signal generated and the corresponding time of the current signal generation under the condition of DC arc fault, image data of the DC arc fault state are captured from different angles.

[0041] 102. Perform time-frequency transformation on at least one image data corresponding to the current signal to obtain at least one image time-frequency feature;

[0042] In this embodiment, the acquired image data corresponding to the current signal is subjected to time-frequency transformation processing to obtain at least one image time-frequency feature. The number of image time-frequency features is determined by the number of image data. The number of acquired image data is at least one, that is, at least one set of image data of the DC arc fault state taken at any angle, and the acquisition time of the image data is equal to the acquisition time of the current signal when the DC arc fault occurs.

[0043] 103. Input at least one time-frequency feature of the image into a preset interpretable deep learning model to generate a scatter plot of the feature density corresponding to each observation frequency band;

[0044] In this embodiment, the preset interpretable deep learning model is the XGboost algorithm. Based on at least one image time-frequency feature, the image time-frequency feature is input into the XGboost model to calculate the shap value of the image feature after wavelet transform, and the feature density scatter plot of each frequency band obtained by time-frequency transform is obtained.

[0045] Specifically, the basic idea of ​​the XGBoost algorithm is to construct a series of weak classifiers (i.e., base classifiers) and combine them using gradient boosting to obtain a powerful classifier.

[0046] On the one hand, by dividing the image time-frequency features into different observation frequency bands, the resulting feature density scatter plots are also different. The feature density scatter plot represents the distribution of shap values ​​within an observation frequency band. That is, if the image video features are divided into N observation frequency bands, then the corresponding feature density scatter plots under N observation frequency bands are obtained.

[0047] 104. Based on the model interpretability values ​​in the feature density scatter plots corresponding to each observation frequency band, calculate the bandwidth range of at least one image time-frequency feature in each observation frequency band.

[0048] In this embodiment, the corresponding model interpretability value is found in the feature density scatter plot according to different observation frequency bands. A feature density scatter plot typically projects time-frequency feature information onto a two-dimensional plane, using feature density as color or brightness to visually observe the distribution of time-frequency features. The model interpretability value typically refers to a model trained using machine learning or other methods, used to identify the category or label to which different time-frequency features belong.

[0049] Specifically, the model interpretability value can be a shap value, which is a method used to interpret the predictions of interpretable deep learning models. Based on the concept of shapley values ​​in cooperative game theory, it assesses the importance of each feature by calculating its contribution to the prediction result. The shap value can help us understand the degree to which the model contributes to the prediction of each feature, as well as the interactions between different features.

[0050] Specifically, for a specific time-frequency feature of an image, the formula for calculating the shap value of feature v is as follows:

[0051]

[0052] Where fx represents the predicted value of the image time-frequency feature in the decision tree, N is the set of all features in the training set with dimension T, and S is a subset extracted from N with dimension |S|.

[0053] On one hand, features are ranked according to the average absolute value of the shap (score). A wider shap indicates a large cluster of samples, representing a greater influence of the feature. A shap value skewed to the right indicates a stronger positive correlation between features, while a shap value skewed to the left indicates a stronger negative correlation. By comparing the bandwidth range of feature density scatter plots obtained using different time-frequency analysis methods within the same frequency band, the superiority of different time-frequency analysis methods can be compared. This allows for the selection of the optimal time-frequency analysis method for arc fault features. A wider bandwidth range represents the weight of the arc motion characteristics' influence on the arc current characteristics. In other words, a larger length value corresponds to a higher average shap value for the arc motion characteristics, better illustrating the magnitude of the arc motion characteristics' influence on the current characteristics.

[0054] 105. Detection of DC arc based on bandwidth range and corresponding image time-frequency features.

[0055] In this embodiment, the device is trained using pre-acquired current data and corresponding image data to obtain the bandwidth range and corresponding image time-frequency features that best represent the current characteristics in a DC arc. By applying the bandwidth range and corresponding image time-frequency features, the device is used to detect DC arcs.

[0056] In this embodiment of the invention, a current signal under a DC arc fault state and at least one image data corresponding to the current signal are acquired; time-frequency transformation processing is performed on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; the at least one image time-frequency feature is input into a preset interpretability deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes a model interpretability value corresponding to the at least one image time-frequency feature; based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band, the bandwidth range of the at least one image time-frequency feature in each observation frequency band is calculated; the DC arc is detected based on the bandwidth range and the corresponding image time-frequency feature. This application processes the image data obtained by the monitoring device to obtain a feature density scatter plot corresponding to the image data, filters the observation frequency band with the strongest correlation and the corresponding image features, and accurately detects the occurred DC arc fault.

[0057] Please see Figure 2 The second embodiment of the arc detection method based on an interpretable deep learning model in this invention includes:

[0058] 201. Acquire the current signal under DC arc fault conditions and at least one image data corresponding to the current signal;

[0059] 202. Perform time-frequency transformation on at least one image data corresponding to the current signal to obtain at least one image time-frequency feature;

[0060] 203. Determine the number of decomposition layers applied to the time-frequency features of the image;

[0061] In this embodiment, the number of wavelet decomposition layers applied to the image time-frequency features is determined by determining the number of wavelet decomposition layers performed on the image time-frequency features.

[0062] Specifically, wavelet decomposition is a signal processing technique that decomposes a signal into sub-signals of different scales to better understand and analyze its characteristics. Wavelet decomposition uses a specific set of functions called wavelet basis functions to achieve signal decomposition. These wavelet basis functions are localization functions of the original signal at different frequencies and time scales. Wavelet decomposition provides the ability to simultaneously locate signal features in both time and frequency. The definition of the continuous wavelet transform is as follows:

[0063]

[0064] f(t) represents the original time-domain signal; a is the scaling factor and a cannot be assigned a value of 0; b is the shift factor; It is the wavelet transform of f(t); R represents the set of real numbers.

[0065] 204. Divide the frequency bands according to the number of decomposition layers to obtain multiple observation frequency bands;

[0066] Specifically, the choice of the number of decomposition layers determines the number of observation bands. For example, choosing 6 decomposition layers results in 2 to the power of 6 observation bands, or 64 different observation bands.

[0067] 205. Perform wavelet transform on at least one image time-frequency feature based on a preset number of decomposition layers to obtain wavelet decomposition data segmented into observation frequency bands corresponding to the number of decomposition layers;

[0068] In this embodiment, wavelet transform uses wavelet basis functions to decompose the signal. Wavelet basis functions are a set of special functions with local properties and local support (finite time and frequency range). Wavelet basis functions can switch between the time and frequency domains and analyze different frequencies and time scales of the signal through translation and scaling operations. The wavelet transform process can be divided into two steps: decomposition and reconstruction. Decomposition: Decomposition processes the signal through successive low-pass filters (blurring) and high-pass filters (details). The low-pass filter preserves the low-frequency components within the signal, while the high-pass filter extracts the high-frequency details. Through iterative processing, the signal is decomposed into multiple frequency bands of different scales. Reconstruction: Based on the frequency bands obtained from decomposition, inverse wavelet transform is used to restore the original signal. The reconstruction process gradually restores the signal's details and low-frequency information by merging the various frequency bands.

[0069] 206. Input the wavelet decomposition data into the preset interpretability deep learning model to obtain the corresponding model interpretability value;

[0070] In this embodiment, the preset interpretable deep learning model uses the XGBoost algorithm. XGBoost (eXtreme Gradient Boosting) is a tree-based ensemble learning algorithm proposed by Chen Tianqi in 2006 and has quickly gained popularity. It boasts advantages such as efficiency, flexibility, and scalability, while demonstrating powerful performance in many machine learning tasks. XGBoost employs gradient boosting, which trains a new decision tree model in each iteration and combines it with the previous model. During training, XGBoost approximates the objective function using a second-order Taylor expansion, enabling it to dynamically select the optimal split point and adjust the leaf node weights to improve the accuracy and generalization performance of the interpretable deep learning model.

[0071] Specifically, by inputting wavelet decomposition data into the XGboost algorithm, a feature density scatter plot consisting of several shap values ​​corresponding to the image data under different observation frequency bands is obtained.

[0072] 207. Based on model interpretability, numerically generate feature density scatter plots corresponding to each observation frequency band;

[0073] 208. Based on the model interpretability values ​​in the feature density scatter plots corresponding to each observation frequency band, calculate the bandwidth range of at least one image time-frequency feature in each observation frequency band.

[0074] 209. Select the observation frequency band with the largest bandwidth range in the time-frequency features of the image;

[0075] In this embodiment, the bandwidth range represents the length value. The larger the length value, the better it indicates the correlation between the arc motion characteristics and the current characteristics. The length value is calculated as: length value = maximum interpretability - minimum interpretability. By calculating the length value under different observation frequency bands, the observation frequency band corresponding to the largest length value is selected, and it is determined whether the largest bandwidth range is the same.

[0076] Specifically, for example, in the actual monitoring process, if image data A has a maximum bandwidth range of 100 in all observation frequency bands and appears in both the [6, 2] and [6, 3] frequency bands, then the image time-frequency features corresponding to image data A with the same maximum bandwidth range are removed, and it is determined whether the maximum bandwidth range of the image time-frequency features corresponding to image data B is the same, until all image data have been judged.

[0077] Specifically, if all image time-frequency features have been determined and all image time-frequency features corresponding to all image data have been removed, then a new second decomposition layer for the image time-frequency features is selected. For example, if the first selected decomposition layer is 6, the second selected decomposition layer can be 5, 7, or another custom decomposition layer. Then, the selected second decomposition layer is updated, and the process restarts from step 204 based on the updated decomposition layer until a maximum bandwidth range exists, and these maximum bandwidth ranges are not identical. The observation frequency band with the largest bandwidth range is then used as the detection frequency band.

[0078] 210. Determine if the maximum bandwidth range is the same;

[0079] 211. If not, then the observation frequency band with the largest bandwidth range shall be used as the detection frequency band;

[0080] 212. If so, then remove the time-frequency features of images with the largest bandwidth range that are also the same.

[0081] 213. If all time-frequency features of the image are removed, then the second decomposition layer applied to the time-frequency features of the image is reselected;

[0082] 214. At least one image data obtained based on the second decomposition layer processing;

[0083] 215. Based on the detection frequency band and the corresponding image time-frequency characteristics of the detection frequency band, image data corresponding to the image time-frequency characteristics are monitored to realize the detection of DC arc.

[0084] In this embodiment, the device is trained using pre-acquired current data and corresponding image data to obtain the bandwidth range and corresponding image time-frequency features that best represent the current characteristics in a DC arc. By applying the bandwidth range and corresponding image time-frequency features, the device is used to detect DC arcs.

[0085] Specifically, for example, during actual monitoring, image data at the same monitoring angle corresponding to image data A is continuously acquired, and the same image time-frequency characteristics are detected in the [6, 2] detection frequency band of the image data as a method to determine whether a DC arc occurs.

[0086] Similarly, for example, in the actual monitoring process, image data of the same monitoring angle corresponding to image data B is continuously acquired, and the same image time-frequency characteristics are detected in the [6, 3] detection frequency band of the image data as a method to determine whether a DC arc occurs.

[0087] In this embodiment of the invention, the process involves selecting the observation frequency band with the largest bandwidth range among the image time-frequency features; determining whether the largest bandwidth ranges are the same; if not, applying the observation frequency band with the largest bandwidth range as the detection frequency band; if so, eliminating image time-frequency features with the same largest bandwidth range. Compared to the prior art, this application iterates through all image time-frequency features, and after all image time-frequency features corresponding to all image data have been eliminated, a second decomposition layer for the image time-frequency features is reselected. For example, if the first selected decomposition layer is 6, the second selected decomposition layer can be 5, 7, or other custom decomposition layers. The selected second decomposition layer is then updated, and the process restarts from step 204 based on the updated decomposition layer until a largest bandwidth range exists and the largest bandwidth ranges are not the same. In this case, the observation frequency band with the largest bandwidth range is applied as the detection frequency band. Based on different decomposition layers, the most relevant detection frequency band is selected, and then the DC arc is monitored based on the most relevant detection frequency band and the corresponding image time-frequency features, thereby improving the relevance and accuracy of monitoring the DC arc state.

[0088] Please see Figure 3 The third embodiment of the arc detection method based on an interpretable deep learning model in this invention includes:

[0089] 301. Acquire the current signal under DC arc fault conditions and at least one image data corresponding to the current signal;

[0090] 302. Obtain the sampling frequencies of the current signal and image data respectively;

[0091] In this embodiment, the current signal and image data under DC arc fault conditions are sampled separately. Both the current signal and image data are one-dimensional signals, and the calculation of the shap value requires that the number of input points be consistent. When the sampling frequency is the same, the current signal and image data are captured within the same time period.

[0092] On one hand, when the current signal and image sampling frequencies differ, the acquisition duration is processed to ensure that corresponding data are captured within the same time period. When different sampling frequencies are harmonics, the average of the data values ​​with more values ​​is calculated according to the harmonics, ensuring that the number of acquired current data and image data is consistent. When different sampling frequencies are not harmonics, the greatest common divisor of different sampling frequencies is found using methods such as short division, prime factorization, and Euclidean algorithm. The average of the different data values ​​is then summed, ensuring that the number of acquired current data and image data is consistent. In this way, the current data and arc image data at the time of an arc fault correspond in the time domain.

[0093] 303. Determine whether there is a harmonic relationship between the signal sampling frequency and the image sampling frequency;

[0094] 304. If it exists, then process the image data based on the harmonic relationship;

[0095] 305. If it does not exist, calculate the greatest common divisor of the signal sampling frequency and the image sampling frequency;

[0096] 306. Determine the processing frequency of signal sampling frequency and image sampling frequency based on the greatest common divisor;

[0097] Specifically, assuming the sampling frequency of the current data is 50kHz, 90,599 raw data points were collected. The sampling frequency of the image data is 15kHz, with 25,548 raw data points collected. Both sets of data were collected simultaneously. The time required for current data acquisition was measured to be 1.81198s, while the time required for image data acquisition was 1.7032s. Since the image data acquisition time is shorter than the current data acquisition time, the image acquisition time was selected as the benchmark. 25,500 image data points were extracted within 1.7s, corresponding to 85,000 current data points acquired within the same timeframe. The greatest common divisor of 50kHz and 15kHz is 5k. The average of every 10 data points from the 50kHz sampling rate and the average of every 3 data points from the 15kHz sampling rate were taken, ensuring that the number of current and image data points acquired is consistent. This allows for a time-domain correspondence between the current data and image data during an arc fault, enabling the study of the correlation mechanism between the arc fault current signal and image data.

[0098] 307. Divide the time window of current signal and image data based on the average processing frequency;

[0099] 308. Calculate the average values ​​of the current signal and image data within the time window, and update the current signal and image data based on the average values ​​respectively;

[0100] 309. Perform time-frequency transformation processing on at least one image data corresponding to the current signal to obtain at least one image time-frequency feature;

[0101] 310. Input at least one time-frequency feature of an image into a preset interpretable deep learning model to generate a scatter plot of feature density corresponding to each observation frequency band;

[0102] 311. Calculate the maximum and minimum interpretability of at least one image time-frequency feature under each observation frequency band.

[0103] In this embodiment, the maximum and minimum interpretability values ​​of the image time-frequency features in the feature density scatter plot are calculated for each observation frequency band. The maximum interpretability value is the maximum shap value of the image / video features corresponding to the same observation frequency band, and the minimum interpretability value is the maximum shap value of the image / video features corresponding to the same observation frequency band.

[0104] 312. Calculate the bandwidth range of time-frequency features for each image based on the maximum and minimum interpretability values;

[0105] In this embodiment, the bandwidth range represents the length value. The larger the length value, the better it indicates the correlation between the arc motion characteristics and the current characteristics. The length value is calculated as: length value = maximum interpretability - minimum interpretability. By calculating the length value under different observation frequency bands, the observation frequency band corresponding to the largest length value is selected as the detection frequency band.

[0106] 313. If the number of image data is greater than one, the observation frequency band with the largest bandwidth range for the time-frequency features of the image data is selected as the detection frequency band to obtain the detection frequency band set, wherein the detection frequency band set contains the detection frequency bands corresponding to the number of image data.

[0107] In this embodiment, if the number of image data is greater than one, it is necessary to select the observation frequency band with the largest bandwidth range as the detection frequency band for each image data corresponding to the image feature, and obtain the detection frequency band set.

[0108] Specifically, for example, if there are image data A and image data B, where the observation frequency band corresponding to the maximum bandwidth range of image data A is [6, 2], and the observation frequency band corresponding to the maximum bandwidth range of image data B is [6, 3], then image data A and the corresponding [6, 2] detection frequency band, and image data B and the corresponding [6, 3] detection frequency band are respectively used as detection frequency band sets for monitoring DC arc.

[0109] Specifically, for example, during actual monitoring, image data at the same monitoring angle corresponding to image data A is continuously acquired, and the same image time-frequency characteristics are detected in the [6, 2] detection frequency band of the image data as a method to determine whether a DC arc occurs.

[0110] Similarly, for example, in the actual monitoring process, image data of the same monitoring angle corresponding to image data B is continuously acquired, and the same image time-frequency characteristics are detected in the [6, 3] detection frequency band of the image data as a method to determine whether a DC arc occurs.

[0111] In practical applications, such as Figure 7 As shown, there are image data M, image data Y, and image data X. On one of the observation frequency bands a, a scatter plot of the feature density corresponding to image data M, image data Y, and image data X is obtained. Since the comparison is made on the same observation frequency band, it can be intuitively seen that the bandwidth range corresponding to image data M is the largest on observation frequency band a.

[0112] In practical applications, such as Figure 8 As shown, there are image data M, image data Y, and image data X. On one of the observation frequency bands b, a scatter plot of the feature density corresponding to image data M, image data Y, and image data X is obtained. Since the comparison is made on the same observation frequency band, it can be intuitively seen that the bandwidth range corresponding to image data M is the largest on the observation frequency band b.

[0113] If the bandwidth of image data M in observation frequency band a is the largest among all observation frequency bands, then the observation frequency band a and the image time-frequency features corresponding to observation frequency band a can be extracted from image data M as a method to determine whether a DC arc has occurred.

[0114] 314. Compare several bandwidth ranges corresponding to the detection frequency band in the detection frequency band set, apply the detection frequency band with the largest bandwidth range and the image time-frequency characteristics corresponding to the detection frequency band, monitor the image data corresponding to the image time-frequency characteristics, and realize the detection of DC arc.

[0115] In this embodiment of the invention, if the number of image data is greater than one, the observation frequency band with the largest bandwidth range corresponding to the image time-frequency features of each image data is selected as the detection frequency band, resulting in a detection frequency band set. The detection frequency band set includes detection frequency bands corresponding to the number of image data. Several bandwidth ranges corresponding to the detection frequency bands in the detection frequency band set are compared, and the detection frequency band with the largest bandwidth range is applied to the image time-frequency features corresponding to the detection frequency bands to monitor the image data corresponding to the image time-frequency features, thereby realizing the detection of DC arcs. Compared to the prior art, this application, when the number of image data is greater than one, requires selecting the observation frequency band with the largest bandwidth range corresponding to the image time-frequency features of different image data as the detection frequency band, resulting in a detection frequency band set. The method of determining whether a DC arc has occurred is used by monitoring the detection frequency bands in the corresponding monitoring frequency band set to see if the same image time-frequency features occur.

[0116] The above describes the arc detection method based on an interpretable deep learning model in the embodiments of the present invention. The following describes the arc detection device based on an interpretable deep learning model in the embodiments of the present invention. Please refer to [link / reference]. Figure 4 One embodiment of the arc detection device based on an interpretable deep learning model in this invention includes:

[0117] Data acquisition module 401 is used to acquire current signals under DC arc fault conditions and at least one image data corresponding to the current signals;

[0118] The time-frequency conversion module 402 is used to perform time-frequency conversion processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature;

[0119] The density scatter plot generation module 403 is used to input the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes the model interpretability value corresponding to the at least one image time-frequency feature.

[0120] The bandwidth range calculation module 404 is used to calculate the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band.

[0121] The detection module 405 is used to detect the DC arc based on the bandwidth range and the corresponding image time-frequency features.

[0122] In this embodiment of the invention, a current signal under a DC arc fault state and at least one image data corresponding to the current signal are acquired; time-frequency transformation processing is performed on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; the at least one image time-frequency feature is input into a preset interpretability deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes the model interpretability value corresponding to the at least one image time-frequency feature; based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band, the bandwidth range of the at least one image time-frequency feature in each observation frequency band is calculated; and the DC arc is detected based on the bandwidth range and the corresponding image time-frequency feature. This application processes the image data obtained by the monitoring device to obtain a feature density scatter plot corresponding to the image data, filters the observation frequency band with the strongest correlation and the corresponding image features, and accurately detects the occurred DC arc fault.

[0123] Please see Figure 5 Another embodiment of the arc detection device based on an interpretable deep learning model in this invention includes:

[0124] Data acquisition module 401 is used to acquire current signals under DC arc fault conditions and at least one image data corresponding to the current signals;

[0125] The time-frequency conversion module 402 is used to perform time-frequency conversion processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature;

[0126] The density scatter plot generation module 403 is used to input the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes the model interpretability value corresponding to the at least one image time-frequency feature.

[0127] The bandwidth range calculation module 404 is used to calculate the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band.

[0128] The detection module 405 is used to detect the DC arc based on the bandwidth range and the corresponding image time-frequency features.

[0129] Furthermore, the arc detection device based on an interpretable deep learning model also includes a sampling frequency update module 406, which comprises:

[0130] The sampling frequency acquisition unit 4061 acquires the sampling frequencies of the current signal and the image data, respectively, wherein the sampling frequencies include the signal sampling frequency and the image sampling frequency; the frequency judgment unit 4062 determines whether there is a frequency doubling relationship between the signal sampling frequency and the image sampling frequency; the frequency doubling processing unit 4063 processes the image data based on the frequency doubling relationship if it exists; the greatest common divisor calculation unit 4064 calculates the greatest common divisor of the signal sampling frequency and the image sampling frequency if they do not exist, and updates the current signal and the image data based on the greatest common divisor.

[0131] Furthermore, the common divisor calculation unit 4064 includes:

[0132] Calculate the greatest common divisor of the signal sampling frequency and the image sampling frequency; determine the processing frequency of the signal sampling frequency and the image sampling frequency based on the greatest common divisor; divide the time window of the current signal and the image data into equal parts based on the processing frequency; calculate the average data value of the current signal and the image data in the time window respectively, and update the current signal and the image data respectively based on the average value.

[0133] Furthermore, the density scatter plot generation module 403 includes:

[0134] Wavelet transform unit 4031 performs wavelet transform on at least one image time-frequency feature to obtain corresponding wavelet decomposition data; model interpretability numerical calculation unit 4032 inputs the wavelet decomposition data into a preset interpretable deep learning model to obtain corresponding model interpretability values; density scatter plot generation unit 4033 generates feature density scatter plots corresponding to each observation frequency band based on the model interpretability values.

[0135] Furthermore, the wavelet transform unit 4031 includes:

[0136] The number of decomposition layers applied to the time-frequency features of the image is determined; frequency bands are divided according to the number of decomposition layers to obtain multiple observation frequency bands; wavelet transform is performed on at least one time-frequency feature of the image based on a preset number of decomposition layers to obtain wavelet decomposition data segmented into observation frequency bands corresponding to the number of decomposition layers.

[0137] Furthermore, the bandwidth range calculation module 404 includes:

[0138] The maximum and minimum interpretability values ​​of at least one image time-frequency feature are calculated for each of the observed frequency bands; the bandwidth range of each image time-frequency feature is calculated based on the maximum and minimum interpretability values.

[0139] Furthermore, the detection module 405 includes:

[0140] The detection frequency band selection unit 4051 selects the observation frequency band with the largest bandwidth range in the image time-frequency features as the detection frequency band; the DC arc detection unit 4052 monitors the image data corresponding to the image time-frequency features based on the detection frequency band and the image time-frequency features corresponding to the detection frequency band, thereby realizing the detection of DC arc.

[0141] Furthermore, the detection frequency band selection unit 4051 includes:

[0142] Select the observation frequency band with the largest bandwidth range among the image time-frequency features; determine whether the largest bandwidth ranges are the same; if not, use the observation frequency band with the largest bandwidth range as the detection frequency band; if so, remove the image time-frequency features with the same largest bandwidth range.

[0143] Furthermore, the detection module 405 also includes a decomposition layer reselection unit 4053, which includes:

[0144] If all the time-frequency features of the image are removed, a second decomposition layer is selected to apply to the time-frequency features of the image; the at least one image data is processed based on the second decomposition layer.

[0145] Furthermore, the detection module 405 includes:

[0146] If the number of image data is greater than one, the observation frequency band with the largest bandwidth range corresponding to the image time-frequency features of the image data is selected as the detection frequency band to obtain a detection frequency band set. The detection frequency band set contains detection frequency bands corresponding to the number of image data. Several bandwidth ranges corresponding to the detection frequency bands in the detection frequency band set are compared. The detection frequency band with the largest bandwidth range and the image time-frequency features corresponding to the detection frequency bands are applied to monitor the image data corresponding to the image time-frequency features to realize the detection of DC arc.

[0147] In this embodiment of the invention, a current signal under a DC arc fault state and at least one image data corresponding to the current signal are acquired; time-frequency transformation processing is performed on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; the at least one image time-frequency feature is input into a preset interpretability deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes the model interpretability value corresponding to the at least one image time-frequency feature; based on the model interpretability value in the feature density scatter plot corresponding to each observation frequency band, the bandwidth range of the at least one image time-frequency feature in each observation frequency band is calculated; and the DC arc is detected based on the bandwidth range and the corresponding image time-frequency feature. This application processes the image data obtained by the monitoring device to obtain a feature density scatter plot corresponding to the image data, filters the observation frequency band with the strongest correlation and the corresponding image features, and accurately detects the occurred DC arc fault.

[0148] above Figure 4 and Figure 5 The arc detection device based on an interpretable deep learning model in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The arc detection equipment based on an interpretable deep learning model in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0149] Figure 6 This is a schematic diagram of the structure of an arc detection device 600 based on an interpretable deep learning model according to an embodiment of the present invention. The arc detection device 600 based on the interpretable deep learning model can vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 610 (e.g., one or more processors) and a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing application programs 633 or data 632. The memory 620 and storage media 630 can be temporary or persistent storage. The program stored in the storage media 630 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the arc detection device 600 based on the interpretable deep learning model. Furthermore, the processor 610 may be configured to communicate with the storage media 630 and execute the series of instruction operations in the storage media 630 on the arc detection device 600 based on the interpretable deep learning model.

[0150] The arc detection device 600 based on an interpretable deep learning model may further include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input / output interfaces 660, and / or one or more operating systems 631, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 6 The arc detection device structure shown is not intended to limit the arc detection device based on the interpretable deep learning model. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0151] The present invention also provides an arc detection device based on an interpretable deep learning model. The computer device includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs each step of the arc detection method based on the interpretable deep learning model in the above embodiments.

[0152] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the various steps of the arc detection method based on an interpretable deep learning model.

[0153] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0154] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0156] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An arc detection method based on an interpretable deep learning model, characterized in that, The arc detection method based on an interpretable deep learning model includes: Acquire the current signal under DC arc fault conditions and at least one image data corresponding to the current signal; Perform time-frequency transformation processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; The at least one image time-frequency feature is input into a preset interpretability deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes the maximum and minimum interpretability values ​​corresponding to the at least one image time-frequency feature; Based on the maximum and minimum interpretability values ​​in the feature density scatter plots corresponding to each observation frequency band, the bandwidth range of the at least one image time-frequency feature in each observation frequency band is calculated. The DC arc is detected based on the observation frequency band with the largest bandwidth and the corresponding image time-frequency features.

2. The arc detection method based on an interpretable deep learning model according to claim 1, characterized in that, After acquiring the current signal under the DC arc fault state and at least one image data corresponding to the current signal, the method further includes: The sampling frequencies of the current signal and the image data are obtained respectively, wherein the sampling frequency includes the signal sampling frequency and the image sampling frequency; Determine whether there is a harmonic relationship between the signal sampling frequency and the image sampling frequency; If it exists, then the image data is processed based on the frequency multiplication relationship; If it does not exist, the greatest common divisor of the signal sampling frequency and the image sampling frequency is calculated, and the current signal and the image data are updated based on the greatest common divisor.

3. The arc detection method based on an interpretable deep learning model according to claim 2, characterized in that, The step of calculating the greatest common divisor (GCD) of the signal sampling frequency and the image sampling frequency, and updating the current signal and the image data based on the GCD, includes: Calculate the greatest common divisor of the signal sampling frequency and the image sampling frequency; The processing frequency for determining the signal sampling frequency and the image sampling frequency is based on the greatest common divisor. The time windows of the current signal and the image data are divided based on the average processing frequency; The average values ​​of the current signal and the image data within the time window are calculated respectively, and the current signal and the image data are updated based on the average values ​​respectively.

4. The arc detection method based on an interpretable deep learning model according to claim 1, characterized in that, The step of inputting the at least one image time-frequency feature into a preset interpretable deep learning model to generate a feature density scatter plot corresponding to each observation frequency band includes: Perform wavelet transform on at least one time-frequency feature of the image to obtain the corresponding wavelet decomposition data; The wavelet decomposition data is input into a preset interpretability deep learning model to obtain the corresponding model interpretability value. Based on the interpretability of the model, a scatter plot of the feature density corresponding to each observation frequency band is generated.

5. The arc detection method based on an interpretable deep learning model according to claim 4, characterized in that, The step of performing wavelet transform on at least one time-frequency feature of an image to obtain the corresponding wavelet decomposition data includes: Determine the number of decomposition layers applied to the time-frequency features of the image; The frequency bands are divided according to the number of decomposition layers to obtain multiple observation frequency bands; Wavelet transform is performed on at least one image time-frequency feature based on a preset number of decomposition levels to obtain wavelet decomposition data segmented into observation frequency bands corresponding to the number of decomposition levels.

6. The arc detection method based on an interpretable deep learning model according to claim 1, characterized in that, The calculation of the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the maximum and minimum interpretability values ​​in the feature density scatter plot corresponding to each observation frequency band includes: The maximum and minimum interpretability values ​​of the at least one image time-frequency feature are calculated for each of the observed frequency bands. The bandwidth range of each of the image time-frequency features is calculated based on the maximum and minimum interpretability values.

7. The arc detection method based on an interpretable deep learning model according to claim 5, characterized in that, The detection of the DC arc based on the observation frequency band with the largest bandwidth and the corresponding image time-frequency features includes: The observation frequency band with the largest bandwidth range among the time-frequency features of the image is selected as the detection frequency band; Based on the detection frequency band and the corresponding image time-frequency features, image data corresponding to the image time-frequency features is monitored to achieve the detection of DC arc.

8. The arc detection method based on an interpretable deep learning model according to claim 7, characterized in that, The step of selecting the observation frequency band with the largest bandwidth range in the time-frequency features of the image as the detection frequency band includes: Select the observation frequency band with the largest bandwidth range among the time-frequency features of the image; Determine if the maximum bandwidth range is the same; If not, the observation frequency band with the largest bandwidth range shall be used as the detection frequency band; If so, then remove the time-frequency features of images with the largest bandwidth range.

9. The arc detection method based on an interpretable deep learning model according to claim 8, characterized in that, After removing image time-frequency features with the same bandwidth range as the largest bandwidth, the process includes: If all the time-frequency features of the image are removed, then a new second decomposition layer is selected to apply to the time-frequency features of the image; The at least one image data obtained based on the second decomposition layer processing.

10. The arc detection method based on an interpretable deep learning model according to claim 6, characterized in that, The method of detecting the DC arc based on the observation frequency band with the largest bandwidth and the corresponding image time-frequency features further includes: If the number of image data is greater than one, then the observation frequency band with the largest bandwidth range for the image time-frequency features corresponding to the image data is selected as the detection frequency band to obtain a detection frequency band set, wherein the detection frequency band set contains detection frequency bands corresponding to the number of image data. By comparing several bandwidth ranges corresponding to the detection frequency bands in the detection frequency band set, and applying the detection frequency band with the largest bandwidth range and the image time-frequency characteristics corresponding to the detection frequency band, image data corresponding to the image time-frequency characteristics is monitored to achieve the detection of DC arc.

11. An arc detection device based on an interpretable deep learning model, applied to an autonomous driving system, characterized in that, The arc detection device based on an interpretable deep learning model includes: The data acquisition module is used to acquire the current signal under DC arc fault conditions and at least one image data corresponding to the current signal; The time-frequency conversion module is used to perform time-frequency conversion processing on the at least one image data corresponding to the current signal to obtain at least one image time-frequency feature; The density scatter plot generation module is used to input the at least one image time-frequency feature into a preset interpretability deep learning model to generate a feature density scatter plot corresponding to each observation frequency band, wherein the feature density scatter plot includes the maximum and minimum interpretability values ​​corresponding to the at least one image time-frequency feature; The bandwidth range calculation module is used to calculate the bandwidth range of the at least one image time-frequency feature in each observation frequency band based on the maximum and minimum interpretability values ​​in the feature density scatter plot corresponding to each observation frequency band. The detection module is used to detect the DC arc based on the observation frequency band with the largest bandwidth and the corresponding image time-frequency features.

12. The arc detection device based on an interpretable deep learning model according to claim 11, characterized in that, The arc detection device based on an interpretable deep learning model further includes a sampling frequency update module, which includes: The sampling frequency acquisition unit acquires the sampling frequencies of the current signal and the image data, respectively, wherein the sampling frequency includes the signal sampling frequency and the image sampling frequency; The frequency determination unit determines whether there is a harmonic relationship between the signal sampling frequency and the image sampling frequency; If a frequency multiplication processing unit exists, it processes the image data based on the frequency multiplication relationship. If the greatest common divisor calculation unit does not exist, it calculates the greatest common divisor of the signal sampling frequency and the image sampling frequency, and updates the current signal and the image data based on the greatest common divisor.

13. The arc detection device based on an interpretable deep learning model according to claim 12, characterized in that, The common divisor calculation unit includes: Calculate the greatest common divisor of the signal sampling frequency and the image sampling frequency; The processing frequency for determining the signal sampling frequency and the image sampling frequency is based on the greatest common divisor. The time windows of the current signal and the image data are divided based on the average processing frequency; The average values ​​of the current signal and the image data within the time window are calculated respectively, and the current signal and the image data are updated based on the average values ​​respectively.

14. An arc detection device based on an interpretable deep learning model, characterized in that, The arc detection device based on an interpretable deep learning model includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the arc detection device based on the interpretable deep learning model to perform the steps of the arc detection method based on the interpretable deep learning model as described in any one of claims 1-10.

15. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the arc detection method based on an interpretable deep learning model as described in any one of claims 1-10.

Images