An arc fault detection method
By framing the DC distribution network current signal, eliminating the DC component, and applying Hilbert-Huang transform, combined with a deep anomaly detection model, the problem of DC arc fault detection was solved, achieving reliable detection and wide application across the entire length range.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID JIANGSU ELECTRIC POWER CO LTD
- Filing Date
- 2022-02-28
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies are insufficient for effectively detecting DC arc faults, especially along the entire length of the line, and traditional methods require extensive professional knowledge and experience, which limits their application and promotion.
The current signal after framing and eliminating DC components is subjected to Hilbert-Huang transform to generate a time-frequency spectrum. Fault detection is then performed using a trained deep anomaly detection model. Features are automatically extracted using a deep neural network, and the model training is optimized by combining Hamming window function and loss function.
It enables fault detection across the entire length of the DC distribution network, improving the reliability and applicability of the detection, enhancing the model's generalization ability, and enabling the identification of insignificant arc faults.
Smart Images
Figure CN114527357B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an arc fault detection method, belonging to the field of power distribution fault detection technology. Background Technology
[0002] Arc faults are one of the most common fault types in distribution networks, usually caused by aging insulation and faulty connections. If not detected and eliminated in time, they can lead to catastrophic consequences such as fires. DC arc faults do not exhibit the 'zero-downtime' phenomenon of AC arc faults, making them more difficult to detect. Currently, commonly used detection methods include physical characteristics, time-domain characteristics, frequency-domain characteristics, time-frequency-domain characteristics, and intelligent detection algorithms. Physical characteristic methods utilize the physical phenomena of arc light and sound produced when an arc occurs to detect it. While the principle is intuitive, this method is limited to local protection of key equipment and cannot monitor the entire length of the line. Furthermore, designing time-domain, frequency-domain, and time-frequency-domain characteristic detection methods suitable for DC arc fault detection requires extensive professional knowledge and rich engineering experience, which also limits the application and promotion of such methods. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide an arc fault detection method for fault monitoring in DC distribution networks. Its applicability is not limited by traditional detection methods, providing a reliable basis for fault detection and effectively improving the reliability of DC distribution networks.
[0004] To achieve the above objectives, the present invention employs the following technical solution:
[0005] On one hand, the present invention provides an arc fault detection method, the method comprising:
[0006] Acquire current signals from DC power distribution networks;
[0007] The acquired current signal is segmented into frames and the DC component of each frame is eliminated.
[0008] Perform Hilbert-Huang transform on each frame of the signal after eliminating the DC component to obtain the time spectrum of the current signal;
[0009] The time-spectrum graph is input into the trained deep anomaly detection model to obtain the fault detection results.
[0010] Specifically, the training method for the deep anomaly detection model includes the following steps:
[0011] Training samples are collected to construct a training sample set, which includes: normal current signals and arc current signals under fault conditions;
[0012] The training samples in the training sample set are divided into frames and the DC component of the signal in each frame is eliminated.
[0013] Perform Hilbert-Huang transform on each frame of signal after eliminating the DC component to obtain the corresponding time-spectrum.
[0014] The time-spectral graphs of each training sample are input into a pre-established deep neural network until the deep neural network converges, thereby obtaining a trained deep anomaly detection model.
[0015] Specifically, the arc current signal is generated by an arc experimental device.
[0016] Specifically, the normal current signal includes the current generated in the circuit when loads such as refrigerators, computers, televisions, range hoods, and fluorescent lights are operating normally.
[0017] Specifically, the sampling frequency fs of the training samples is not less than 500kHz, and the sampling time is 5s.
[0018] Specifically, the frame length N of the framing is... f Take 0.1×fs, frame shift N d Taking 0.05×fs, K current samples i are obtained through frame segmentation. k (n), 1≤k≤K, where k is the k-th current sample, n is the count of the sampling points, and i k (n) represents the current signal at the nth sampling point under the kth current sample.
[0019] On the one hand, the present invention provides a method for detecting arc faults, the method further comprising performing windowing processing on each frame signal using a Hamming window function after framing.
[0020] Specifically, the DC component of each frame of signal is eliminated using the following formula:
[0021]
[0022] In the formula, ω(n,α)=(1-α)-αcos(2πn / (N f -1), where n is the count of sampling points, N f Let be the frame length, α be a defined parameter (here, 0.46), and k be the k-th current sample obtained from the frame division. The current signal acquired after windowing processing. To eliminate the DC component from the current signal, ω and ω(n, a) are represented in the form of Hamming window functions.
[0023] Furthermore, the present invention provides a method for detecting arc faults, wherein the loss function of the deep anomaly detection model is:
[0024]
[0025] In the formula, R and w are the parameters to be optimized; k is the k-th current sample obtained from the frame segmentation; K p K n These represent the number of normal current signals and arc current signals, respectively; φ(D) k ;w) is D k A deep neural network with input D and parameters w, where D k This is the time-frequency spectrum of the current signal obtained from the Hilbert-Huang transform, with time and frequency parameters attached, which can be expressed as D. k (i, j), 1 ≤ i ≤ N f , 1≤j≤fs / 2, where fs is the sampling frequency of the sample, N f The frame length is denoted by ν, and i and j represent time and frequency, respectively; c is the center of the training samples in the deep neural network mapping space, given as the origin; ν p ν n λ is a hyperparameter; ||·|| F It is the F-norm;
[0026] Deep Neural Network φ(D) k The system has a total of 7 layers, of which layers 1, 3, and 5 are convolutional layers, layers 2 and 4 are pooling layers, and layers 7 and 8 are fully connected layers. The activation function for the convolutional layers is ReLU, and the activation function for the fully connected layers is Sigmoid.
[0027] If ||φ(D';w)-c||>R, then the line is determined to have an arc fault; otherwise, the line is determined to be normal. Here, D' is the time-frequency matrix of the line signal to be tested, obtained by the Hilbert-Huang transform. Substituting the time-frequency characteristics, it can be expressed as D'(i,j).
[0028] Specifically, the SGD algorithm is used to train the deep anomaly detection model. Each training cycle includes 200 gradient updates. In the first 180 gradient updates, the parameter R is fixed and the parameter w is updated. In the last 20 gradient updates, the parameter w is fixed and the parameter R is updated. This process is repeated until the deep neural network converges.
[0029] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0030] 1. Before detecting the current signal, this invention performs frame segmentation and eliminates the DC component of the collected DC distribution network current signal. Based on this, a time spectrum is obtained according to the Hilbert-Huang transform. Then, the time spectrum is substituted into the trained deep anomaly detection model to obtain the fault detection result. This method provides a reliable basis for fault detection in DC distribution networks, can conveniently realize the monitoring of the entire line length, has a wide range of applications and is conducive to application and promotion.
[0031] 2. Based on the single-class deep anomaly detection model, this invention introduces negative sample (fault current sample) loss, which allows fault current samples to participate in training, thereby improving the model's generalization ability.
[0032] 3. This invention builds a deep anomaly detection model by constructing a deep neural network, which integrates the two originally separate processes of feature extraction and classifier training. Compared with the method of manually designing features, the features automatically extracted by the deep neural network are more conducive to the training of the classifier.
[0033] 4. The loss function of the deep anomaly detection model provided by this invention compresses the normal sample space into a hypersphere with the smallest possible volume, thus making it easier to detect arc faults with insignificant phenomena. Attached Figure Description
[0034] Figure 1 This is a flowchart of an arc fault detection method provided by an embodiment of the present invention;
[0035] Figure 2 This is a block diagram illustrating the structural principle of the deep neural network upon which the deep anomaly detection model in this embodiment of the invention is based. Detailed Implementation
[0036] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0037] Example 1:
[0038] This invention provides an arc fault detection method. First, it acquires the current signal from a DC distribution network. Then, it segments the acquired current signal into frames and eliminates the DC component of each frame. Next, it performs a Hilbert-Huang transform on each frame after DC component elimination to obtain the time-frequency matrix (time-frequency matrix) of the current signal. The time-frequency matrix is then input into a trained deep anomaly detection model to obtain the fault detection result. This detection method obtains the effective feature value of the current by eliminating the DC component from the acquired DC current signal. It then performs a Hilbert-Huang transform on these effective feature values to obtain the time-frequency matrix (time-frequency matrix) of the current signal. By substituting the time-frequency matrix into the deep anomaly detection model, it determines whether a fault has occurred in the line current. This invention utilizes a trained deep anomaly detection model to directly transform and detect the acquired current, providing a reliable means for troubleshooting faults in DC power grids. Since current acquisition can be performed at various locations within the distribution network, it can easily achieve full-length line monitoring without being limited by the local limitations of traditional detection methods, which is beneficial for the widespread application of this method.
[0039] An arc fault detection method provided in this embodiment of the invention includes the following steps in training the deep anomaly detection model to ensure its reliability:
[0040] Training samples are collected to construct a training sample set, which includes: normal current signals and arc current signals under fault conditions. The deep anomaly detection model is trained to identify normal current samples and abnormal current samples respectively.
[0041] The training samples in the training sample set are divided into frames, and the DC component of each frame signal is removed. Then, the Hilbert-Huang transform is performed on each frame signal after removing the DC component to obtain the corresponding time-spectrum D. k ;
[0042] The time-spectral graphs D of each training sample k The input is fed into a pre-built deep neural network until the deep neural network converges, thereby obtaining a trained deep anomaly detection model.
[0043] The training method for a deep anomaly detection model provided by this invention trains the deep anomaly detection model by building a deep neural network and introducing current signals from normal and abnormal samples. The features automatically extracted by the deep neural network are more conducive to the training of the classifier. In addition, negative samples can also participate in the training, which improves the model's adaptability to fresh samples.
[0044] The present invention provides an arc fault detection method. To ensure the introduction of negative sample current signals, the arc current signals are simulated and generated by an arc experimental device. Through the simulation of the experimental device, fault current signals with different characteristics can be used to train a deep anomaly detection model. The experimental device can generate sufficient negative samples to ensure the accuracy of the detection model.
[0045] An arc fault detection method provided by this invention involves collecting normal current to identify and train normal power consumption current. The normal current signal can include the current generated in the circuit when loads such as refrigerators, computers, televisions, range hoods, and fluorescent lights are operating normally.
[0046] The present invention provides an arc fault detection method, which, in order to ensure the reliability of the collected current data information and fully reflect the detailed characteristics of the current, sets the sampling frequency fs of the training samples to be no less than 500kHz and the sampling time to 5s.
[0047] This invention provides an arc fault detection method, which, in order to rationally acquire and utilize the current signal, sets the frame length N for framing. f Take 0.1×fs, frame shift N dTaking 0.05×fs, K current samples i are obtained through frame segmentation. k (n), 1≤k≤K, where k is the k-th current sample, n is the count of the sampling points, and i k (n) means the current signal at the nth sampling point under the kth current sample.
[0048] Example 2:
[0049] The arc fault detection method provided in this embodiment of the invention differs from that in Embodiment 1 in that, in order to eliminate the picket fence effect caused by framing and avoid the loss of important or characteristic frequency components that would render signal processing meaningless, the method further includes applying a Hamming window function to each frame of signal after framing.
[0050] An arc fault detection method provided in this invention eliminates the DC component of each frame of signal using the following formula:
[0051]
[0052] In the formula, ω(n,a)=(1-α)-αcos(2πn / (N f -1), where n is the count of sampling points, N f Let be the frame length, α be a defined parameter (here, 0.46), and k be the k-th current sample obtained from the frame division. The current signal acquired after windowing processing. To eliminate the DC component from the current signal, ω and ω(n, a) are represented in the form of Hamming window functions.
[0053] Example 3:
[0054] This invention provides an arc fault detection method, which differs from Embodiment 2 in that it specifically provides a loss function for a deep anomaly detection model, providing a reliable basis for training and using sampled current signals. The loss function of the deep anomaly detection model is as follows:
[0055]
[0056] The difference between the predicted and actual results is "penalized" by the above loss function, where R and w are the parameters to be optimized; k is the k-th current sample obtained from the frame segmentation; K p K n These represent the number of normal current signals and arc current signals, respectively; φ(D) k ;w) is D k A deep neural network with input D and parameters w, where D kIt is a time-frequency spectrum obtained from the Hilbert-Huang transform, with time and frequency parameters represented as D. k (i, j), D k (i, j), 1 ≤ i ≤ N f , 1≤j≤fs / 2, where fs is the sampling frequency of the sample, N f The frame length is denoted by ν, and i and j represent time and frequency, respectively; c is the center of the training samples in the deep neural network mapping space, given as the origin; ν p ν n λ is a hyperparameter; ||·|| F For the F-norm, in the formula, by training to determine the value of parameter R, the normal sample space can be compressed into a hypersphere that is as small as possible, making it easier to detect insignificant arc faults;
[0057] Here, to ensure the deep neural network φ(D) k The learning ability of φ(D) is used to set up a deep neural network φ(D). k The deep neural network (d) consists of 7 layers: layers 1, 3, and 5 are convolutional layers; layers 2 and 4 are pooling layers; and layers 7 and 8 are fully connected layers. This structure allows the network to utilize the two-dimensional structure of the input data. Compared to other deep or feedforward neural networks, the convolutional neural network requires fewer parameters and delivers better results in image and speech recognition. The activation function for the convolutional layers is ReLU (Rectified Linear Unit, also known as a modified linear unit, a commonly used activation function in artificial neural networks), and the activation function for the fully connected layers is the Sigmoid activation function (used to map variables to 0 and 1). k The relationship can be referenced in the following text. Figure 2 As shown;
[0058] When performing fault detection, substitute the characteristic value obtained by the detection current. If ||φ(D';w)-c||>R, it is determined that the line has an arc fault. Otherwise, the line is determined to be normal. Here, D' is the time-frequency matrix of the signal to be detected obtained according to the Hilbert-Huang transform (substituting the time and frequency characteristics can also be expressed as D'(i,j)).
[0059] This invention provides an arc fault detection method. The deep anomaly detection model is trained using the SGD algorithm (gradient descent algorithm). By using the SGD algorithm, the parameters w and R are gradually converged, eventually converging to a good extreme point. Since the fluctuation range of parameter w is larger than that of parameter R, parameters R and w can be determined separately according to a certain ratio. Here, each training cycle is set to include 200 gradient updates. In the first 180 gradient updates, parameter R is fixed and parameter w is updated. In the last 20 gradient updates, parameter w is fixed and parameter R is updated. This process is repeated until the deep neural network converges.
[0060] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. An arc fault detection method characterized by, The method includes: Acquire current signals from DC power distribution networks; The acquired current signal is segmented into frames and the DC component of each frame is eliminated. Perform Hilbert-Huang transform on each frame of the signal after eliminating the DC component to obtain the time spectrum of the current signal; The time-spectrum graph is input into the trained deep anomaly detection model to obtain fault detection results; The loss function of the depth anomaly detection model is: ； In the formula, and represents the parameters to be optimized; k represents the k-th current sample obtained from the frame segmentation. , These represent the number of normal current signals and the number of arc current signals, respectively. Therefore For input, A deep neural network with parameters; where This is the time-frequency spectrum of the current signal obtained from the Hilbert-Huang transform, with time and frequency parameters shown below. , 1≤i≤ , 1≤j≤fs / 2, where fs is the sampling frequency of the sample. The frame length, i, and j represent time and frequency, respectively. Let be the center of the training samples in the mapping space of the deep neural network, and let be the origin. , , For hyperparameters; It is the F-norm; Deep Neural Networks There are a total of 7 layers, of which layers 1, 3, and 5 are convolutional layers, layers 2 and 4 are pooling layers, and layers 6 and 7 are fully connected layers; the activation function of the convolutional layers is ReLU, and the activation function of the fully connected layers is Sigmoid. if If the line shows an arc fault, it is determined that the line is in normal condition; otherwise, the line is considered normal. The time-frequency matrix of the signal to be detected, obtained from the Hilbert-Huang transform, is substituted into the time-frequency characteristic representation as follows: ; The SGD algorithm was used to train the deep anomaly detection model. Each training epoch included 200 gradient updates, with the parameters fixed during the first 180 gradient updates. Update parameters The parameters were fixed during the last 20 gradient updates. Update parameters This process is repeated until the deep neural network converges.
2. The arc fault detection method of claim 1, wherein, The training method for the deep anomaly detection model includes the following steps: Training samples are collected to construct a training sample set, which includes: normal current signals and arc current signals under fault conditions; The training samples in the training sample set are divided into frames and the DC component of the signal in each frame is eliminated. Perform Hilbert-Huang transform on each frame of signal after eliminating the DC component to obtain the corresponding time-spectrum. The time-spectral graphs of each training sample are input into a pre-established deep neural network until the deep neural network converges, thereby obtaining a trained deep anomaly detection model.
3. The arc fault detection method of claim 2, wherein, The arc current signal is generated by an arc experimental device.
4. The arc fault detection method of claim 2, wherein, The normal current signal includes the current generated in the circuit when loads such as refrigerators, computers, televisions, range hoods, and fluorescent lights are operating normally.
5. The arc fault detection method of claim 2, wherein, The sampling frequency fs of the training samples is not less than 500kHz, and the sampling time is 5s.
6. The arc fault detection method of claim 5, wherein, The frame length of the segment Take 0.1×fs, frame shift Using 0.05×fs, K current samples are obtained through frame segmentation. , 1≤k≤K, where n is the count of sampling points. This represents the current signal at the nth sampling point under the kth current sample.
7. The arc fault detection method of any one of claims 1-6, wherein, The method also includes applying a Hamming window function to window each frame of signal after framing.
8. The arc fault detection method of claim 7, wherein, The DC component of each frame of signal is eliminated using the following formula: ； In the formula, Where n is the count of sampling points, The frame length of the subframe; To define the parameters, we take 0.46 here; k is the k-th current sample obtained from the frame segmentation. The current signal acquired after windowing processing. To eliminate the DC component from the current signal, , These are all represented in functional form as Hamming window functions.