A series arc fault detection method in a low voltage power distribution system

Abstract

The application discloses a series arc fault detection method in a low-voltage power distribution system, real current data and labels are obtained through the construction of normal and arc data collection platforms, public data sets are collected to verify the generalization ability of the model, three artificial experiences of flat shoulder, inter-period similarity and current mutation are summarized by observing the current waveform, and an interpretable model capable of processing one-dimensional time series data is constructed, the prediction results of the three artificial experiences are compared with the real labels, the loss weight of inconsistent samples is increased, so that the experience can be embedded into the model and assist the model in training; in this way, the problem of training the interpretable neural tree in the small sample environment with fewer labeled samples can be solved through the mechanism embedding mode, so that the application has the advantages of interpretability, good generalization ability, high prediction accuracy and the like.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, and more specifically, relates to a method for detecting series arc faults in low-voltage power distribution systems. Background Technology

[0002] Series arcing faults are one of the leading causes of electrical fires in low-voltage power distribution systems. Each year, arcing faults cause more than 10,000 household fires, resulting in serious casualties and property damage. Arcing faults are typically caused by factors such as loose wiring, prolonged circuit overload, and insulation aging, accompanied by high-temperature spattering from localized electrodes, which can ignite nearby flammable materials, creating a fire hazard. To prevent potential losses, national electrical regulations have incorporated arcing fault detection into emerging early fire detection technologies. However, the concealed, random, and diverse nature of arcing faults makes their detection extremely challenging.

[0003] Currently, machine learning and deep learning methods are two main directions in research on series arc fault detection. Machine learning classification methods, such as decision trees and random forests, have high interpretability and demonstrate good classification ability in small sample situations. However, these machine learning-based methods require extracting arc features before training the model, and their accuracy is inferior to deep learning methods when dealing with more complex data. Deep learning methods do not require prior feature extraction and have excellent classification capabilities, making them a research hotspot in this field. Researchers have proposed various arc detection methods based on neural networks, but these studies mainly focus on improving model performance and reducing the number of model parameters and computational cost, neglecting the black-box nature of neural networks. Furthermore, the performance of deep learning models degrades when the amount of data is insufficient. Therefore, how to integrate arc fault experience summarized in the laboratory into an interpretable model to improve the model's reliability remains an unsolved problem. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for detecting series arc faults in low-voltage power distribution systems, which quickly and accurately identifies arc faults by combining decision trees and neural networks.

[0005] To achieve the above-mentioned objective, the present invention provides a method for detecting series arc faults in a low-voltage power distribution system, characterized by comprising the following steps:

[0006] (1) Data acquisition and preprocessing to obtain the current sample of each load;

[0007] (2) Extract the features of the current sample, including shoulder, period similarity and current abrupt change;

[0008] (3) Construct a binary tree composed of internal nodes and leaf nodes as a one-dimensional neural tree serial arc fault detection model.

[0009] (4) Train the one-dimensional neural tree serial arc fault detection model until it converges;

[0010] (5) Detect arc faults based on the trained one-dimensional neural tree series arc fault detection model.

[0011] The objective of this invention is achieved as follows:

[0012] This invention discloses a method for detecting series arc faults in low-voltage power distribution systems. It obtains real current data and labels by building a normal and arc data acquisition platform, collects publicly available datasets to verify the model's generalization ability, and summarizes three human experiences—flat shoulders, periodic similarity, and current abrupt changes—by observing current waveforms. An interpretable model capable of processing one-dimensional time series data is then constructed. The prediction results combining these three human experiences are compared with the real labels, and inconsistent samples are weighted to increase their loss. This allows the experiences to be embedded into the model and assist in its training. This mechanism-embedding approach solves the problem of training interpretable neural trees in small-sample environments with limited labeled samples, thus offering advantages such as interpretability, good generalization ability, and high prediction accuracy.

[0013] Meanwhile, the series arc fault detection method of the present invention, which integrates an interpretable model based on experience, also has the following beneficial effects:

[0014] (1) When training traditional machine learning methods such as decision trees and random forests, additional feature extraction steps are required. This process may not only consume time, but may also reduce classification accuracy and cause the loss of important information due to improper feature extraction. However, neural trees directly use raw current data for end-to-end training. This method can extract complex and useful features from raw data without human intervention, thereby simplifying the whole process and providing a solid foundation for the deployment of embedded devices.

[0015] (2) The arc fault detection method using experience fusion can significantly reduce the amount of data required for model training, making it possible to apply deep learning for arc fault detection in a small sample environment; experience-guided model learning can not only improve classification accuracy, but also show higher stability and generalization ability when dealing with new data and dealing with uncertainty; experience-based models are easier to interpret and understand than purely data-driven methods because they rely on known rules and theories.

[0016] (3) Deep neural networks typically lack interpretability, and traditional machine learning methods may lose interpretability due to model complexity when dealing with large or complex data. Interpretable neural tree models combine the architecture of decision trees with the convolutional layers of neural networks, achieving the advantages of shallow model depth, few trainable parameters, and low model complexity. Furthermore, the interpretability of such models is further enhanced due to the embedding of mechanisms. In practical application environments, the model can be fine-tuned according to specific needs, which provides a great guarantee for the reliability of the invention. Attached Figure Description

[0017] Figure 1 This is a flowchart of a series arc fault detection method in a low-voltage power distribution system according to the present invention;

[0018] Figure 2 This is the experimental circuit diagram for an AC series arc fault.

[0019] Figure 3 This is a schematic diagram of the load connection method;

[0020] Figure 4 yes Figure 3 The diagram shows the current waveform under the load connection method shown.

[0021] Figure 5 This is a schematic diagram of the training loss of a one-dimensional neural tree serial arc fault detection model;

[0022] Figure 6 This is a schematic diagram illustrating the training accuracy of a one-dimensional neural tree-based series arc fault detection model.

[0023] Figure 7 This is a schematic diagram of a one-dimensional neural tree serial arc fault detection model after training.

[0024] Figure 8 This is a schematic diagram illustrating the model training accuracy at 10kHz in the IEEE public dataset;

[0025] Figure 9 This is a schematic diagram of the model training accuracy at 6400Hz in the IEEE public dataset. Detailed Implementation

[0026] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0027] Example

[0028] Figure 1 This is a flowchart of a series arc fault detection method in a low-voltage power distribution system according to the present invention.

[0029] In this embodiment, as Figure 1 As shown, the present invention provides a method for detecting series arc faults in a low-voltage power distribution system, comprising the following steps:

[0030] S1. Data acquisition and preprocessing;

[0031] S1.1. Current signals of AC lines under different loads during normal operation and arc fault conditions are collected at equal intervals. For each load, m current sampling points are collected, and the collected current signals are denoted as I = {I1, I2, ..., I...}. j ,…,I h}, I j ={i1,i2,…,i m}, where I j Let i represent the current signal collected under the j-th load. m This represents the current value collected at the m-th sampling point, and h represents the number of loads;

[0032] In this embodiment, an AC series arc fault test circuit is constructed according to the requirements of the series carbonization path arc experiment, such as... Figure 2 As shown in the diagram, the circuit consists of an AC power supply, loads, an arc generator, an air switch, a current transformer, and an oscilloscope. The arc generator is selected and designed entirely according to standard requirements, and its main structure comprises a carbon fixed electrode, a copper moving electrode, and a screw regulator. There are a total of 18 types of loads, including single loads and multiple loads in parallel, connected as shown in the diagram. Figure 3 As shown, 18,000 current sampling points were collected for each type of load, including current under normal and fault conditions. Figure 4 This is a schematic diagram of the normal and fault current waveforms under three different loads at 3000 sampling points;

[0033] S1.2, Current signal I collected for each load j Perform frame division;

[0034] Assuming each frame contains w current sampling points, then each load has Frame current sample, denoted The current sample for each load is then represented as: {I j1 ,I j2 ,…,I jl …,I jd In this example, each frame of current samples contains 600 sampling points, for a total of 540 samples.

[0035] Then, each frame of current sample is labeled with 0 or 1, where 0 indicates normal and 1 indicates arc fault.

[0036] S2. Extract the characteristics of the current sample;

[0037] S2.1 Extract each current sample I jl The shoulder;

[0038] During a series arc fault, the loop current exhibits a "zero-sink" phenomenon near zero current. Compared to the loop current under normal conditions, the current shoulder duration during an arc fault is longer. To mathematically represent this empirical finding, an effective dual-threshold method is designed to detect the shoulder duration in a power frequency periodic waveform, as follows:

[0039] Record current sample I jl ={i j1 i j2 ,…,i jn ,…,i jw}, n∈[1,w], determine each current i jn Does it meet the following requirements:

[0040] |i jn |<th1&|i jn+1 |<th1&|i jn+2 |<th1&|i jn+3 |<th1&|i jn+3 -i jn |<th2

[0041]

[0042] th2 = 7% × I rms

[0043] Among them, I rms Indicates the effective value of the current;

[0044] If the above conditions are met, then sampling point i will be... jn As current sample I jl The shoulder area, traversing the current sample I jl All current values ​​are calculated, and then the total number of shoulders is counted and denoted as the current shoulder N_shoulders;

[0045] S2.2 Extract each current sample I jl Periodic similarity;

[0046] When the load is operating normally, the current is relatively stable, and the period is highly similar to the cycle period. However, during an arc fault, due to the high-frequency randomness, there will be a certain amount of abrupt change in the current signal, resulting in a low similarity between the cycle periods. Therefore, the Pearson correlation coefficient can be used to reflect the autocorrelation between the current signal periods. The specific extraction process is as follows:

[0047] S2.2.1 Set up a notch filter with the following transfer function H(z):

[0048]

[0049] Where a is the notch filter coefficient, and z represents the Z-domain;

[0050] S2.2.2, Using a notch filter to sample current I jl Perform notch filtering to obtain the filtered current sample.

[0051] S2.2.3, Assume current sample If there are t = 2 periods, meaning each period has η = w / t sampling points, then the Pearson correlation coefficient between the c-th period and the (c+1)-th period is denoted as r. c,c+1 ;

[0052]

[0053] in, This represents the current value at the k-th sampling point within the c-th period after filtering. This represents the current value at the k-th sampling point within the (c+1)-th cycle after filtering. and These are the average current values ​​of all sampling points in the c-th and c+1-th cycles, respectively.

[0054] S2.2.4. Using the Pearson correlation coefficient between periods to represent the similarity between periods, the current sample is calculated using the following formula. The total period similarity is r, denoted as period similarity P_correlation;

[0055]

[0056] S2.3, Extraction current sudden change;

[0057] When an arc fault occurs, the harmonic content of the loop current increases significantly. Compared to the loop current under normal conditions, the loop current waveform during an arc fault exhibits noticeable spikes, especially pronounced with resistive loads. Therefore, we can extract current spikes using the following method:

[0058] S2.3.1, Remove each current sample I jl DC component:

[0059]

[0060] Among them, i jn_dc Indicate i jn The value after removing the DC component;

[0061] S2.3.2, Let I be the current sample after removing the DC component. jl_dc ={i j1_dc i j2_dc ,…,i jn_dc ,…,i jw_dc}; then for the current sample I jl_dc The spectrum X is obtained by performing a fast Fourier transform. jl [k]:

[0062]

[0063] S2.3.3, Regarding the spectrum X jl [k] is normalized:

[0064]

[0065] S2.3.4, Calculate the current sample I jl Total Harmonic Distortion (THD) jl :

[0066]

[0067] Let the total harmonic distortion be denoted as the current mutation Thd_values;

[0068] S3. Construct a one-dimensional neural tree-based series arc fault detection model;

[0069] A one-dimensional neural tree model is a binary tree structure consisting of internal nodes and leaf nodes. A solver is deployed on each leaf node, and a router is deployed on each internal node. A converter is deployed on each decision path of the binary tree.

[0070] In this embodiment, the converters are distributed on each decision path of the binary tree, and each path has one or more converters. These converters are non-linear functions used to transform sample data and pass it to the next node. The converters specifically include convolutional layers, pooling layers, and ReLU activation functions, which take the output of the previous module as input and transform it before passing it to the next module.

[0071] Routers are distributed among the internal nodes of a binary tree, distributing input data to its left and right child nodes based on routing decisions; the routing decisions are based on random sampling according to the Bernoulli distribution.

[0072] The solver is distributed across each leaf node in the binary tree. The input to the solver is the data processed by the transformer, and the output is an estimate of the conditional distribution. The solver consists of one-dimensional convolution, ReLU activation function, max pooling, flattening operation, Dropout operation, and fully connected layers.

[0073] S4. Train a one-dimensional neural tree series arc fault detection model;

[0074] S4.1 Select a batch of current samples I from the acquired current signals I. jl The input is fed into a one-dimensional neural tree series arc fault detection model, and then the current sample I is processed by a converter. jl Feature extraction is performed as follows:

[0075] For each current sample I jl Feature extraction is performed using a converter, and the current is denoted as the characteristic current.

[0076] First, the current sample I... jl Perform one-dimensional convolution operations:

[0077]

[0078] Where c(n) is the output of the one-dimensional convolution operation, representing the convolution result at position n, I jl (n+kp) is the value of the input signal at position n+k, ω(k) is the weight of the convolution kernel at position k, p is the padding length, and K is the length of the convolution kernel;

[0079] Then, the convolution result c(n) is activated by the RuLU function:

[0080] ReLU(c(n)) = max(0, c(n))

[0081] Finally, max pooling is used to reduce the amount of data.

[0082]

[0083] Where H is the size of the pooling window, ξ is the step size, and h is the position index within the pooling window;

[0084] S4.2, Characteristic current Input to the router, based on Bernoulli distribution of characteristic current Random sampling yields three routing decisions: split, hold, and stop. Split means splitting the input data into two child nodes (left and right), hold means continuing to generate the next child node as a single node, and stop means not generating any more child nodes.

[0085] The router performs three routing decisions on the characteristic current. Perform routing output, denoted as... r = 1, 2, 3 represent the three routing decisions: split, maintain, and stop, respectively.

[0086] S4.3, Output the route The input is fed into the solver to predict the conditional probabilities corresponding to the three route outputs;

[0087]

[0088] in, Indicates that given input Given the model parameters Θ, the model predicts the output Y. jl The probability distribution is given by Θ = (θ, ψ, φ), which is the parameter set of the model, including router parameters θ, converter parameters ψ, and solver parameters φ. Indicates input The probability of being assigned to the f-th leaf node. This represents a specific prediction distribution after selecting the f-th leaf node, where z∈{0,1} L , z f =1 means that leaf node f was used, and L is the total number of leaf nodes;

[0089] The obtained conditional probability distribution This corresponds to the probability of each category, that is, the probability of generating a 0 or 1 label. The category with the highest probability is the model's predicted result.

[0090] S4.4 Calculate the current sample I jl The probability of being assigned to a leaf node f

[0091]

[0092] Among them, P f Let f represent the set of paths from the root node to the leaf node f. This represents the probability of a router's decision on a path. It is a binary relation that is true when the leaf node f is the left subtree of the internal node o;

[0093] S4.5, Update current sample I jl The weights;

[0094] Set an empirical value P based on current shoulder N_Shoulders, current cycle similarity P_correlation, and current abrupt change Thd_values;

[0095] P=(N_shoulders<=th3)&(P_correlation>=th4)&(Thd_values<th5)*(predicted≠0)

[0096] If the current sample I jl If the characteristics satisfy the above empirical value P, then the weights of the current samples are updated to...

[0097] S4.6. Use the experience-guided negative log-likelihood as the loss function to calculate the loss value under the three route outputs;

[0098]

[0099] Wherein, p(Y|I weight ,Θ) represents the conditional probability of the model predicting the label given the input and parameters, and d is the number of samples in a batch. Y = {Y j1 ,Y j2 ,…,Y jl ,…,Y jd} represents the set of predicted outputs for a batch of samples. This represents the probability that a batch of inputs will be assigned to the f-th leaf node, i.e., the probability of path selection. The model output Y at the f-th leaf node jl The probability of;

[0100] S4.7 Select the route output with the minimum loss as the final route output for this batch of current samples;

[0101] S4.8 Repeat steps S4.1-S4.7, iterating repeatedly until the one-dimensional neural tree serial arc fault detection model converges.

[0102] S5, Series Arc Fault Detection;

[0103] In step S1, the current signal of the AC line is collected in real time, and then the current sample is extracted after preprocessing and input into the trained one-dimensional neural tree series arc fault detection model to detect whether an arc fault has occurred.

[0104] Training loss, accuracy, and binary tree structure, such as Figure 5 As shown in 6 and 7, Figure 5 The dashed and solid lines represent the losses on the training and validation sets, respectively, and both losses converge to below 0.2. Figure 6 The dashed and solid lines represent the accuracy of the validation set and the accuracy of the test set, respectively. Figure 5 and Figure 6 It can be seen that the model has converged and has good performance. Figure 7 The binary tree shown has 3 levels, with two internal nodes and three leaf nodes, and each decision path uses one or two transformers.

[0105] To demonstrate the universality of the method of this invention, we also conducted experiments on the public dataset IAED, and the experimental results are as follows: Figure 8 As shown in Figure 9, Figure 8 and Figure 9 The training accuracy of the model on 10kHz and 6400Hz data from the IEEE public dataset is shown in the figures. As can be seen from the figures, the invention performs well on large datasets. The parameters of the public dataset are shown in Table 1.

[0106]

[0107]

[0108] Table 1

[0109] As shown in Table 1, the public dataset provides two data resolution settings. Data was collected by a 16-bit, 10kHz high-resolution (HR) recorder (standalone measurement device) and a 12-bit, 6400Hz low-resolution (LR) sensor (embedded in a real-time identification system). Arcing accidents can occur on a branch of a parallel circuit or on the main branch of a series circuit; three different appliance connection types were designed, such as... Figure 3 As shown. The characteristics of the electric arc also depend on the type of appliance; different load types can introduce different current waveforms. We simulated our electric arc. There are ten different household appliances in total, according to... Figure 3 The combinations shown have 15 different loading configurations for the HR data portion and 10 different loading configurations for the LR data portion. The series fault arc detection method of this invention takes into account the detection accuracy of datasets of different sizes.

[0110] In the actual results, as shown in Table 2, the mechanism-based embedding models in this invention outperform traditional machine learning and deep learning methods.

[0111] Performance indicators Decision Tree SVM 1DCNN ANT This invention Accuracy 0.8889 0.7222 0.8623 0.9375 0.9583 F1 0.8889 0.8026 0.8500 0.9365 0.9574 Recall 0.8889 0.7222 0.8744 0.9474 0.9649 Precision 0.8889 0.7257 0.8567 0.9333 0.9535

[0112] Table 2

[0113] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for detecting series arc faults in a low-voltage power distribution system, characterized in that, Includes the following steps: (1) Data acquisition and preprocessing; (1.1) Current signals of AC lines under different loads during normal operation and arc fault conditions are collected at equal intervals. For each load, m current sampling points are collected, and the collected current signals are denoted as I = {I1, I2, ..., I...}. j ,…,I h }, I j ={i1,i2,…,i m }, where I j Let i represent the current signal collected under the j-th load. m This represents the current value collected at the m-th sampling point, and h represents the number of loads; (1.2) Current signal I collected under each load j Perform frame division; Assuming each frame contains w current sampling points, then each load has Frame current sample, denoted The current sample for each load is then represented as: {I j1 ,I j2 ,…,I jl …,I jd }; Then, each frame of current sample is labeled with 0 or 1, where 0 indicates normal and 1 indicates arc fault. (2) Extracting features from current samples; (2.1) Extract each current sample I jl The shoulder; Record current sample I jl ={i j1 i j2 ,…,i jn ,…,i jw }, n∈[1,w], determine each current i jn Does it meet the following requirements: |i jn |<th1&|i jn+1 |<th1&|i jn+2 |<th1&|i jn+3 |<th1&|i jn+3 -i jn |<th2 th2=7%×I rms Among them, I rms Indicates the effective value of the current; If the above conditions are met, then sampling point i will be... jn As current sample I jl The shoulder area, traversing the current sample I jl All current values ​​are calculated, and then the total number of shoulders is counted and denoted as the current shoulder N_shoulders; (2.2) Extracting each current sample I jl Periodic similarity; (2.2.1) Set up a notch filter with the following transfer function H(z): Where a is the notch filter coefficient, and z represents the Z-domain; (2.2.2) Using a notch filter to sample current I jl Perform notch filtering to obtain the filtered current sample. n∈[1,w]; (2.2.3) Assume a current sample If there are t periods, meaning each period has m = w / t sampling points, then the Pearson correlation coefficient between the c-th period and the (c+1)-th period is denoted as r. c,c+1 ; in, This represents the current value at the k-th sampling point within the c-th period after filtering. This represents the current value at the k-th sampling point within the (c+1)-th cycle after filtering. and These are the average current values ​​of all sampling points in the c-th and c+1-th cycles, respectively. (2.2.4) Using the Pearson correlation coefficient between periods to represent the similarity between periods, the current sample is calculated using the following formula. The total period similarity is r, denoted as period similarity P_correlation; (2.3) Sudden change in extraction current; (2.3.1) Remove each current sample I jl DC component: Among them, i jn_dc Indicate i jn The value after removing the DC component; (2.3.2) Let the current sample after removing the DC component be denoted as . Then, the current sample The spectrum X is obtained by performing a fast Fourier transform. jl [k]: (2.3.3) Regarding the spectrum X jl [k] is normalized: (2.3.4) Calculate the current sample I jl Total Harmonic Distortion (THD) jl : Let the total harmonic distortion be denoted as the current mutation Thd_values; (3) Construct a one-dimensional neural tree series arc fault detection model; The one-dimensional neural tree model is a binary tree structure composed of internal nodes and leaf nodes. Each leaf node deploys a solver, which consists of one-dimensional convolution, ReLU activation function, max pooling, flattening operation, Dropout operation, and fully connected layer. Each internal node deploys a router. Each decision path of the binary tree deploys a transformer, which includes convolutional layer, pooling layer, and ReLU activation function. (4) Training a one-dimensional neural tree series arc fault detection model; (4.1) Select a batch of current samples I from the acquired current signals I. jl The input is fed into a one-dimensional neural tree series arc fault detection model, and then the current sample I is processed by a converter. jl Feature extraction is performed as follows: For each current sample I jl Feature extraction is performed using a converter, and the current is denoted as the characteristic current. First, the current sample I... jl Perform one-dimensional convolution operations: Where c(n) is the output of the one-dimensional convolution operation, representing the convolution result at position n, I jl (n+kp) is the value of the input signal at position n+k, ω(k) is the weight of the convolution kernel at position k, p is the padding length, and K is the length of the convolution kernel; Then, the convolution result c(n) is activated by the RuLU function: ReLU(c(n)) = max(0, c(n)) Finally, max pooling is used to reduce the amount of data. Where H is the size of the pooling window, ξ is the step size, and h is the position index within the pooling window; (4.2) Characteristic current Input to the router, based on Bernoulli distribution of characteristic current Random sampling yields three routing decisions: split, hold, and stop. The router performs three routing decisions on the characteristic current. Perform routing output, denoted as... r = 1, 2, 3 represent the three routing decisions: split, maintain, and stop, respectively. (4.3) Output the route The input is fed into the solver to predict the conditional probabilities corresponding to the three route outputs; in, Indicates that given input Given the model parameters Θ, the model predicts the output Y. jl The probability distribution is given by Θ = (θ, ψ, φ), which is the parameter set of the model, including router parameters θ, converter parameters ψ, and solver parameters φ. Indicates input The probability of being assigned to the f-th leaf node. This represents a specific prediction distribution after selecting the f-th leaf node, where z∈{0,1} L , z f =1 means that leaf node f was used, and L is the total number of leaf nodes; The obtained conditional probability distribution This corresponds to the probability of each category, and the category with the highest probability is the predicted result of the model. (4.4) Calculate the current sample I jl The probability of being assigned to a leaf node f Among them, P f Let f represent the set of paths from the root node to the leaf node f. This represents the probability of a router's decision on a path. It is a binary relation that is true when the leaf node f is the left subtree of the internal node o; (4.5) Update current sample I jl The weights; Set an empirical value P based on current shoulder N_Shoulders, current cycle similarity P_correlation, and current abrupt change Thd_values; P=(N_shoulders<=th3)&(P_correlation>=th4)&(Thd_values<th5)*(predicted≠0) If the current sample I jl If the characteristics satisfy the above empirical value P, then the weights of the current samples are updated to... (4.6) Use the experience-guided negative log-likelihood as the loss function to calculate the loss value under the three route outputs; Wherein, p(Y|I weight ,Θ) represents the conditional probability of the model predicting the label given the input and parameters, and d is the number of samples in a batch. Y = {Y j1 ,Y j2 ,…,Y jl ,…,Y jd } represents the set of predicted outputs for a batch of samples. This represents the probability that a batch of inputs will be assigned to the f-th leaf node, i.e., the probability of path selection. The model output Y at the f-th leaf node jl The probability of; (4.7) Select the route output with the minimum loss as the final route output for this batch of current samples; (4.8) Repeat steps (4.1)-(4.7) and iterate until the one-dimensional neural tree serial arc fault detection model converges. (5) Detection of series arc faults; According to step (1), the current signal of the AC line is collected in real time, and then the current sample is extracted after preprocessing and input into the trained one-dimensional neural tree series arc fault detection model to detect whether an arc fault has occurred.

Images