DC arc discharge fault detection method based on multi-scale feature extraction and clustering screening

Abstract

The invention discloses a DC arc discharge fault detection method based on multi-scale feature extraction and clustering screening, and relates to the technical field of artificial intelligence algorithms and photovoltaic fault diagnosis, and the method comprises the following steps: S100, signal collection; s200, carrying out multi-scale analysis; s300, time domain feature extraction is carried out; s400, frequency domain feature extraction and clustering screening are carried out; s500, generating a feature vector; s600, training a random forest model; and S700, analyzing a series direct current arc discharge fault. According to the invention, the identification rate of direct-current arc fault detection is effectively improved, the false alarm rate of direct-current arc fault detection is reduced, and the overhaul and maintenance cost of photovoltaic equipment is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence algorithms and photovoltaic fault diagnosis, and in particular to a DC arcing fault detection method using multi-scale feature extraction and clustering screening. Background Art

[0002] As one of the important sources of renewable energy, solar energy has been widely used in power generation. In recent years, a large number of industrial photovoltaic power plants and residential photovoltaic power stations (referred to as PV systems) have been built. PV systems are composed of various components, including photovoltaic modules, batteries, inverters, controllers, loads, and cables. The power generated by photovoltaic modules can be used to power loads through inverters or charge batteries. If the power generated exceeds the load demand and the power required for battery charging, the excess power can be delivered to the grid. Due to the large number of cables and connectors in PV systems, the possibility of arcing cannot be ignored. In order to prevent arc-related fires, the National Electrical Code (NEC) requires that all PV systems using a maximum DC voltage of more than 80 volts must detect arc faults in a timely manner.

[0003] Most of the existing series DC arc detection technologies rely on signal feature analysis, usually analyzing the collected line current signal to identify whether there is an arc. These methods can detect arc faults to a certain extent, but because the characteristics of the series DC arc signal are similar to those of other interference signals in the photovoltaic system, the false alarm rate is often high, affecting the accuracy and efficiency of the detection. Especially in complex working environments, the detection equipment is easily affected by external electromagnetic interference, environmental noise and other factors, further increasing the possibility of false alarms and missed alarms. The difficulty of series DC arc detection technology also lies in how to distinguish between electrical fluctuations in normal operation and series DC arc signals. In the increasingly complex and changeable working scenes of photovoltaic power stations, the existing technology still has a high accuracy detection, especially the false alarm rate. The number of series plugs in photovoltaic power stations is very large, and photovoltaic power stations are usually built in remote areas. The misjudgment of the series DC arc fault detection unit will greatly increase the maintenance cost.

[0004] The patent "A series fault arc detection method based on VMD Hilbert marginal spectrum multi-feature fusion" (publication number CN116577612A) proposes a fault arc detection method based on variational mode decomposition (VMD) and Hilbert marginal spectrum analysis (HHT) multi-feature fusion. The method first extracts multiple intrinsic mode functions of different frequencies by performing variational mode decomposition on the collected current signal, and then performs marginal spectrum analysis on each modal component. For the selection of modal component marginal spectrum, a modal component selection method based on multi-feature fault discrimination is proposed to select feature quantities with high fault discrimination of each modal component marginal spectrum. In order to reduce feature redundancy, linear component analysis is used for feature fusion, and finally a neural network is introduced for classification and verification of detection accuracy. However, the current data extracted from the photovoltaic system in the patent is a huge collection, which increases the workload of the machine, cannot well screen out valuable feature quantities, and remove redundant features, resulting in a high false alarm rate and missed alarm rate for series DC arc faults. The current existing technology has an accuracy rate of 96% to 98% for detecting series DC arc faults.

[0005] Therefore, technicians in this field are committed to developing a DC arc fault detection method based on multi-scale feature extraction and cluster screening. Summary of the invention

[0006] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to improve the recognition rate of DC arc fault detection and reduce the false alarm rate of DC arc fault detection.

[0007] This application analyzes that although the existing technology uses variational mode decomposition in the frequency domain to extract the frequency domain features of useful frequency bands, it only analyzes specific frequency bands and is not universal. In addition, the existing technology extracts less frequency domain feature information and does not further screen the extracted frequency domain features. The accuracy of fault detection will be greatly reduced in scenarios with complex and changeable working conditions and large external disturbances.

[0008] This application adopts a multi-scale feature extraction method to extract more time domain and frequency domain information. First, a multi-scale analysis method is used to obtain the coarse-grained current sequence at different scales, and then the coarse-grained current sequence at different scales is further processed in the time domain and frequency domain to obtain time domain and frequency domain features, and finally these features are fused as the decision input of the model.

[0009] Multi-scale analysis can reveal the complexity of signals at different scales. Compared with single-scale, the features extracted from multiple scales contain more valuable information, which helps to improve the accuracy of subsequent classification. Assume that the original time series current data is ,in N Represents the sequence length. The current sequence after multi-scale transformation is: in , is the scale factor, Usually a positive integer between 1 and 10. To take downward The integer part of When , it is represented as the original time series current data. , then the sequence , , The lengths are 4096, 2048 and 1365 respectively.

[0010] After multi-scale transformation of the time series current data, time domain and frequency domain analysis are performed to obtain features containing time domain and frequency domain information. Since the frequency domain information is a high-dimensional feature vector Since the feature vector has a high dimension, directly using it as the model input will result in a slow training speed. Redundant and invalid features will also reduce the model's fault recognition accuracy. Therefore, it is necessary to use the frequency domain feature vector C Perform dimensionality reduction processing.

[0011] KNN (K-Nearest Neighbor) or k nearest neighbor algorithm is the most widely used clustering algorithm. It clusters similar samples together based on distance calculation and divides data points into K The main idea of ​​the algorithm is that each cluster has a center point, and the data points are assigned to the cluster with the closest center point. The center point of each cluster is updated iteratively to minimize the sum of the distances from all data points to the center point of the cluster to which they belong. This application uses the KNN algorithm to cluster features, clustering similar features together, and obtaining a total of K clusters (classes), and then select the best features from each cluster (class), retaining the best features of each cluster to form a feature vector .

[0012] In one embodiment of the present invention, a DC arc fault detection method with multi-scale feature extraction and cluster screening is provided, comprising: S100, signal acquisition, under normal working conditions and DC arc fault conditions, collect current signals to obtain M The sequence length is N The timing current data, M and N is a positive integer; S200, multi-scale analysis, using multi-scale methods to process the time series current data to obtain coarse-grained current sequences at different scales; S300, time domain feature extraction, calculate the standard deviation and approximate entropy of the current coarse-grained sequence at different scales, and obtain the time domain feature vector; S400, frequency domain feature extraction and clustering screening, discrete Fourier analysis is performed on the coarse-grained current sequences at different scales, and then the results of discrete Fourier analysis are clustered to screen out effective features and obtain frequency domain feature vectors; S500, generating a feature vector, combining the time domain feature vector and the frequency domain feature vector to obtain a current feature vector; S600, random forest model training, using normal current feature vectors and current feature vectors with arc faults to train the random forest model; S700, series DC arcing fault analysis, input the current feature vector into the trained random forest model for analysis, if it is judged to be a DC arcing fault, it will be reported.

[0013] Optionally, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, M The value range is [50000, 150000].

[0014] Preferably, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, M =138848.

[0015] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, N The value range is [512, 4096].

[0016] Preferably, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, N =1024.

[0017] Optionally, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, it is assumed in step S200 that the time series current data is ,in N Represents the sequence length, and the coarse-grained sequences of currents at different scales are as follows: in , is the scale factor, is a positive integer between 1 and 10, To take downward The integer part of , which is represented by the original time series current data.

[0018] Optionally, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in any of the above embodiments, step S300 includes: S310, variance calculation, for coarse-grained current sequences at different scales , calculate the standard deviation , using the following formula: in, is the coarse-grained sequence of current at different scales The average value of the data is as follows: ; S320, approximate entropy calculation, calculation of coarse-grained current sequences at different scales The approximate entropy of ,Approximate entropy is a nonlinear dynamic parameter that measures the regularity and unpredictability of fluctuations in time series data (time series current data is a type of time series data). It uses a non-negative value to represent the complexity of the time series, reflecting the possibility of the emergence of new information. The more complex the time series, the higher the approximate entropy.

[0019] Further, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, step S320 includes: S321. Construct vector space from coarse-grained current sequences at different scales Extract P A set of vector spaces is constructed by continuous elements. k The vectors are represented as follows: in, P is a positive integer; S322, calculate the distance between vectors, for , calculate the k vector and the rest The distance between two vectors is given by the following formula: in, ,and ; S323, calculate similarity, calculate and The similarity between them is expressed as follows: in in, Represents a coefficient between 0 and 1, which can be adjusted as needed; S324, calculate the approximate entropy and calculate the coarse-grained sequence of current at different scales The approximate entropy of , the formula is as follows: in , ln (*) The function is an exponential function.

[0020] Further, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, in step S321 P Is 2 or 3.

[0021] Preferably, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, in step S321 P =2.

[0022] Optionally, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in any of the above embodiments, step S400 includes: S410, discrete Fourier transform, for coarse-grained current sequences at different scales Perform discrete Fourier transform to obtain the corresponding frequency domain vector , the formula is as follows: in i represents the imaginary part of a complex number, For scale The frequency domain step factor under Rounding to get, is a discrete time domain signal, is a discrete frequency domain signal, s Represents the frequency corresponding to each discrete frequency domain signal, is the sampling frequency; S420, frequency domain vector merging, merging frequency domain vectors of all scales together to obtain a frequency domain vector , the formula is as follows: in Represents the maximum transformation scale, the frequency domain vector The data length is The formula is as follows: ; S430, frequency domain feature clustering. Since the frequency domain features after multi-scale transformation have a high dimension, directly using them as input of the model will reduce the accuracy of feature recognition. Therefore, redundant features are removed from the frequency domain features after multi-scale transformation, and then effective features are extracted. S440, frequency domain feature screening, using a method to measure the average difference between the characteristics of samples in normal operation and those with arc faults (referred to as average difference measurement ( ADM ) method) Find the optimal features of each cluster in the frequency domain feature clustering results. The formula is as follows: in, For each sample j frequency domain features, is the total number of all samples, is the number of normal samples, is the number of samples with arc faults, Indicates i The normal sample j Features, Indicates i The first sample with arc fault j The numerator of the above formula represents the first feature of all normal samples. j The average of the features minus the first feature of all arc fault samples j The absolute value of the mean of the features, The larger the value of represents the better the discrimination of the feature, the better the feature is selected from each cluster. The feature composition with the highest value K The optimal frequency domain eigenvector .

[0023] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the maximum transformation scale in step S420 is The range is 1-10.

[0024] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the maximum transformation scale The value of is 5.

[0025] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the frequency domain feature clustering in step S430 uses the KNN clustering algorithm to divide the data points into K Clusters, is the frequency domain feature obtained from a single current data sample. Assuming M samples, the frequency domain feature vectors of all samples are expressed as the following matrix: in, Indicates i The frequency domain eigenvector of samples is The matrix, the columns of which are the frequency domain features of each sample, has Column vectors, all column vectors are expressed as follows: in, i Indicates i column vector, , each column vector has M elements, Indicates the first sample i frequency domain features; Each cluster in the KNN algorithm has a center point. The feature points are assigned to the cluster where the nearest center point is located. The center point of each cluster is iteratively updated to minimize the sum of the distances from all feature points to the center point of the cluster to which they belong. The sum of squared errors within the cluster is minimized. The formula is as follows: in, is the center point of the cluster, , It is a cluster points within, From the point in the cluster to the center of the cluster The Euclidean distance of .

[0026] Further, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, step S430 includes: S431, initialization, from all points , Random selection K as the center point, and initialize the maximum number of iterations to Imax ; S432, assign data points and calculate each point and K The distance between the center points is as follows: Will Assign to the cluster with the closest center point; S433, update the center point, recalculate the center point of each cluster, the formula is as follows: in It is a cluster The number of all points in ; S434, iterative calculation, repeating steps S432 to S434 until the maximum number of iterations is reached Imax ,get K The frequency domain feature clustering results of clusters and K center points; S435, output results, return K The frequency domain feature clustering results of clusters and K A center point.

[0027] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the preset number of iterations is Imax The range is 100-1000.

[0028] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the preset number of iterations is Imax =300 。

[0029] Optionally, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in any of the above embodiments, step S600 includes: S610, training a random forest model, where the random forest model includes a set number of CART decision trees, each CART decision tree is trained separately, the input is a current feature vector, and the output is 0 (indicating normal) or 1 (indicating a DC arc fault); S620, set a training sample set, for each CART decision tree, set , is labeled with a class label(x) A set of training samples, class labels -1 and +1 represent two different classes. x express D dimensional current feature vector; after the CART decision tree training is completed, the final set of all leaf nodes is the set of all training samples, which is expressed as follows: in, Represents the number of leaf nodes in the CART decision tree. is the set of all leaf nodes, Indicates leaf nodes, Indicates leaf nodes; S630, train the CART decision tree, use The coefficient is used as the loss function to evaluate the classification error, and the The characteristic split leaf node with the smallest coefficient value, The coefficient is calculated as follows: , in, and The explanation is as follows. For a binary classification problem, suppose there are two categories, A and B. For any leaf node in the CART decision tree, suppose the proportion (probability) of category A is , then the proportion (probability) of another category B is 1- , that is, the probability that the leaf node can correctly classify category A is ; The probability of correctly classifying category B is ,visible ; S640, CART decision tree training ends. When the criterion for stopping training of the CART decision tree is met, the training ends. S650, complete the random forest model training, and loop step S620-step S650 until all CART decision trees are trained, thus completing the random forest model training.

[0030] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the number range set in step S610 is 5-100.

[0031] Preferably, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the set number is 15.

[0032] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the criterion for stopping training of the CART decision tree in step S640 is: when the internal node is a pure leaf node, that is, it only contains faulty samples or only contains normal samples, or the number of samples in the leaf node has reached the set minimum number.

[0033] Furthermore, in the DC arc fault detection method of multi-scale feature extraction and cluster screening in the above embodiment, the minimum number is set to 5.

[0034] This application uses an artificial intelligence clustering algorithm to cluster similar frequency domain features, remove redundant features with similar classification effects, and select the optimal features from each class, effectively improving the recognition rate of DC arc fault detection, and reducing the false alarm rate of DC arc fault detection, thereby reducing the inspection and maintenance costs of photovoltaic equipment.

[0035] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart illustrating a DC arc fault detection method using multi-scale feature extraction and cluster screening according to an exemplary embodiment. DETAILED DESCRIPTION

[0037] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0038] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is schematically exaggerated in some places in the drawings.

[0039] The inventors have designed a DC arc fault detection method based on multi-scale feature extraction and cluster screening, such as Figure 1 As shown, the following steps are included: S100, signal acquisition, under normal working conditions and DC arc fault conditions, collect current signals to obtain M The sequence length is N The timing current data, M and N is a positive integer, M =138848, N =1024.

[0040] S200, multi-scale analysis, using multi-scale methods to process the time series current data to obtain coarse-grained current sequences at different scales; assuming that the time series current data is ,in N Represents the sequence length, and the coarse-grained sequences of currents at different scales are as follows: in , is the scale factor, is a positive integer between 1 and 10, To take downward The integer part of , which is represented by the original time series current data.

[0041] S300, time domain feature extraction, calculates the standard deviation and approximate entropy of the current coarse-grained sequences at different scales to obtain the time domain feature vector; specifically includes: S310, variance calculation, for coarse-grained current sequences at different scales , calculate the standard deviation , using the following formula: in, is the coarse-grained sequence of current at different scales The average value of the data is as follows: ; S320, approximate entropy calculation, calculation of coarse-grained current sequences at different scales The approximate entropy of Approximate entropy is a nonlinear dynamic parameter that measures the regularity and unpredictability of fluctuations in time series data. It uses a non-negative value to represent the complexity of the time series, reflecting the possibility of new information. The more complex the time series, the higher the approximate entropy. Specifically, it includes: S321. Construct vector space from coarse-grained current sequences at different scales Extract P A set of vector spaces is constructed by continuous elements. k The vectors are represented as follows: in, P is a positive integer, P =2; S322, calculate the distance between vectors, for , calculate the k vector and the rest The distance between two vectors is given by the following formula: in, ,and ; S323, calculate similarity, calculate and The similarity between them is expressed as follows: in in, Represents a coefficient between 0 and 1, which can be adjusted as needed; S324, calculate the approximate entropy and calculate the coarse-grained sequence of current at different scales The approximate entropy (AE) of is as follows: in , ln (*) The function is an exponential function.

[0042] S400, frequency domain feature extraction and clustering screening, discrete Fourier analysis is performed on the coarse-grained current sequences at different scales, and then the results of discrete Fourier analysis are clustered to screen out effective features and obtain frequency domain feature vectors; specifically including: S410, discrete Fourier transform, for coarse-grained current sequences at different scales Perform discrete Fourier transform to obtain the corresponding frequency domain vector , the formula is as follows: in i represents the imaginary part of a complex number, For scale The frequency domain step factor under Rounding to get, is a discrete time domain signal, is a discrete frequency domain signal, s Represents the frequency corresponding to each discrete frequency domain signal, is the sampling frequency; S420, frequency domain vector merging, merging frequency domain vectors of all scales together to obtain a frequency domain vector , the formula is as follows: in represents the maximum transformation scale, =5, frequency domain vector The data length is The formula is as follows: ; S430, frequency domain feature clustering. Since the frequency domain features after multi-scale transformation have a high dimension, directly using them as the input of the model will reduce the accuracy of feature recognition. Therefore, the redundant features of the frequency domain features after multi-scale transformation are removed, and then effective features are extracted. Frequency domain feature clustering uses the KNN clustering algorithm to divide the data points into clusters by minimizing the sum of the distances from each data point to the center point of the cluster to which it belongs. K Clusters, is the frequency domain feature obtained from a single current data sample. Assuming M samples, the frequency domain feature vectors of all samples are expressed as the following matrix: in, Indicates i The frequency domain eigenvector of samples is the matrix The matrix, the columns of which are the frequency domain features of each sample, has Column vectors, all column vectors are expressed as follows: in, i Indicates i column vector, , each column vector has M elements, Indicates the first sample i frequency domain features; Each cluster in the KNN algorithm has a center point. The feature points are assigned to the cluster where the nearest center point is located. The center point of each cluster is iteratively updated to minimize the sum of the distances from all feature points to the center point of the cluster to which they belong. The sum of squared errors within the cluster is minimized. The formula is as follows: in, is the center point of the cluster, , It is a cluster points within, From the point within the cluster to the cluster center The Euclidean distance of ; specifically including: S431, initialization, from all points , Random selection K as the center point, and initialize the maximum number of iterations to Imax , Imax =300; S432, assign data points and calculate each point and K The distance between the center points is as follows: Will Assign to the cluster with the closest center point; S433, update the center point, recalculate the center point of each cluster, the formula is as follows: in It is a cluster The number of all points in ; S434, iterative calculation, repeating steps S432 to S434 until the maximum number of iterations is reached Imax , K The frequency domain feature clustering results of clusters and K center points; S435, output results, return K The frequency domain feature clustering results of clusters and K A center point.

[0043] S440, frequency domain feature screening, using a method to measure the average difference between the characteristics of samples in normal operation and those with arc faults (referred to as average difference measurement ( ADM ) method) Find the optimal features of each cluster in the frequency domain feature clustering results. The formula is as follows: in, For each sample j frequency domain features, is the total number of all samples, is the number of normal samples, is the number of samples with arc faults, Indicates i The normal sample j Features, Indicates i The first sample with arc fault j The numerator of the above formula represents the first feature of all normal samples. j The average of the features minus the first feature of all arc fault samples j The absolute value of the mean of the features, The larger the value of represents the better the discrimination of the feature, the better the feature is selected from each cluster. The feature composition with the highest value K The optimal frequency domain eigenvector ; S500, generating a feature vector, combining the time domain feature vector and the frequency domain feature vector to obtain a current feature vector; S600, random forest model training, using normal and arc fault current feature vectors to train the random forest model; specifically including: S610, training a random forest model, where the random forest model includes 15 CART decision trees, each of which is trained separately, with the input being a current feature vector, and the output being 0 (indicating normal) or 1 (indicating a DC arc fault); S620, set a training sample set, for each CART decision tree, set , is labeled with a class label(x) A set of training samples, class labels -1 and +1 represent two different classes. x express D dimensional current feature vector; after the CART decision tree training is completed, the final set of all leaf nodes is the set of all training samples, which is expressed as follows: in, Represents the number of leaf nodes in the CART decision tree. is the set of all leaf nodes, Indicates leaf nodes, Indicates leaf nodes; S630, train the CART decision tree, use The coefficient is used as the loss function to evaluate the classification error, and the The characteristic split leaf node with the smallest coefficient value, The coefficient is calculated as follows: , in, and The explanation is as follows. For a binary classification problem, suppose there are two categories, A and B. For any leaf node in the CART decision tree, suppose the proportion (probability) of category A is , then the proportion (probability) of another category B is 1- , that is, the probability that the leaf node can correctly classify category A is ; The probability of correctly classifying category B is ,visible ; S640, the CART decision tree training ends. When the stopping training criteria of the CART decision tree are met, the training ends. The stopping training criteria are: when the internal node is a pure leaf node, that is, it contains only faulty samples or only normal samples, or the number of samples in the leaf node has reached the set minimum number, and the set minimum number is 5; S650, complete the random forest model training, and loop step S620-step S650 until all CART decision trees are trained, thus completing the random forest model training.

[0044] S700, series DC arcing fault analysis, input the current feature vector into the trained random forest model for analysis, if it is judged to be a DC arcing fault, it will be reported.

[0045] In order to verify the technical effect of the present application, the applicant conducted a comparative test using the test set current data, and used the above embodiment to detect DC arc faults, with an accuracy rate of 99.6%, of which the missed alarm rate was 0.3% and the false alarm rate was 0.1%; while the accuracy rate of the prior art for detecting DC arc faults was 96% to 98%. It can be seen that the present application effectively improves the recognition rate of DC arc fault detection and reduces the missed alarm rate and false alarm rate of DC arc fault detection.

[0046] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A DC arc fault detection method based on multi-scale feature extraction and cluster screening, characterized in that: The following steps are involved: S100, signal acquisition, under normal working conditions and DC arc fault conditions, collect current signals to obtain M The sequence length is N Timing current data, M and N is a positive integer; S200, multi-scale analysis, using a multi-scale method to process the time series current data to obtain coarse-grained current sequences at different scales; S300, extracting time domain features, calculating standard deviations and approximate entropies of the current coarse-grained sequences at different scales, respectively, to obtain time domain feature vectors; S400, frequency domain feature extraction and cluster screening, performing discrete Fourier analysis on the coarse-grained current sequences at different scales, clustering the results of the discrete Fourier analysis, screening out effective features, and obtaining frequency domain feature vectors; S500, generating a feature vector, combining the time domain feature vector and the frequency domain feature vector to obtain a current feature vector; S600, training a random forest model, using the normal current feature vector and the current feature vector with an arc fault to train a random forest model; S700, analyzing a series DC arcing fault, inputting the current feature vector into the trained random forest model for analysis, and reporting if it is determined to be a DC arcing fault.

2. The DC arc fault detection method of multi-scale feature extraction and cluster screening according to claim 1, characterized in that: Said M The value range is [50000, 150000].

3. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 1, characterized in that: Said N The value range is [512, 4096].

4. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 1, characterized in that: In step S200, it is assumed that the timing current data is , the coarse-grained sequence of currents at different scales is as follows: 。 5. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 4, characterized in that: The step S300 includes: S310, variance calculation, for the current coarse-grained sequence at different scales , calculate the standard deviation , using the following formula: ； S320, approximate entropy calculation, calculating the coarse-grained sequence of current at different scales The approximate entropy of .

6. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 5, characterized in that: The step S320 includes: S321, constructing a vector space from the coarse-grained sequences of currents at different scales Extract P A set of vector spaces is constructed by continuous elements. k The vectors are represented as follows: ； S322, calculate the distance between vectors, for , calculate the k vector and the rest The distance between two vectors is given by the following formula: ； S323, calculate similarity, calculate and The similarity between them is expressed as follows: ； S324, calculating approximate entropy, calculating the coarse-grained sequence of current at different scales The approximate entropy of , the formula is as follows: 。 7. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 5, characterized in that: The step S400 includes: S410, discrete Fourier transform, for the current coarse-grained sequences at different scales Perform discrete Fourier transform to obtain the corresponding frequency domain vector , the formula is as follows: ； S420, frequency domain vector merging, merging frequency domain vectors of all scales together to obtain a frequency domain vector , the formula is as follows: ； S430, frequency domain feature clustering, removing redundant features from the frequency domain features after multi-scale transformation, and then extracting effective features; S440, frequency domain feature screening, using the method of measuring the average difference between the normal operating sample features and the arc fault sample features to find the optimal features of each cluster in the frequency domain feature clustering results, the formula is as follows: 。 8. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 7, characterized in that: The frequency domain feature clustering in step S430 uses the KNN clustering algorithm to divide the data points into K Clusters, is the frequency domain feature obtained from a single current data sample. Assuming M samples, the frequency domain feature vectors of all samples are expressed as the following matrix: 。 9. The DC arc fault detection method of multi-scale feature extraction and cluster screening as claimed in claim 8, characterized in that: The step S430 includes: S431, initialization, from all points , Random selection K as the center point, and initialize the maximum number of iterations to Imax ; S432, assign data points and calculate each point and K The distance between the center points is as follows: Will Assign to the cluster with the closest center point; S433, update the center point, recalculate the center point of each cluster, the formula is as follows: ； S434, iterative calculation, repeating steps S432 to S434 until the maximum number of iterations is reached Imax, get K The frequency domain feature clustering results of clusters and K center points; S435, output the result, return the K The frequency domain feature clustering results of the clusters and the K A center point.

10. The DC arc fault detection method of multi-scale feature extraction and cluster screening according to claim 7, characterized in that: The step S600 includes: S610, training a random forest model, wherein the random forest model includes a set number of CART decision trees, each CART decision tree is trained separately, the input is a current feature vector, and the output is 0 or 1; S620, set a training sample set, for each CART decision tree, set , the CART decision tree training is completed, and the final set of all leaf nodes is the set of all training samples, which is expressed as follows: ； S630, train the CART decision tree, use The coefficient is used as the loss function to evaluate the classification error, and the Gini The characteristic split leaf node with the smallest coefficient value, Gini The coefficient is calculated as follows: ； S640, CART decision tree training ends. When the criterion for stopping training of the CART decision tree is met, the training ends. S650, completing the random forest model training, looping the steps S620 to S650 until all CART decision tree trainings are completed, thus completing the training of the random forest model.

Images