An arc fault detection method based on fully adaptive random configuration network

Abstract

The application provides an arc fault detection method based on a completely self-adaptive random configuration network, which builds a network structure by self-adaptive learning of arc fault current characteristics, judges the redundancy of a newly added node to realize adjustment of the network width, judges the error level of the newly added node to realize adjustment of the network depth, and associates the network structure with the current signal, so that data driving is realized, the calculation complexity of the network is reduced, and accurate and rapid detection of the arc fault is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to current signal detection technology, in particular to an arc fault detection method based on a fully adaptive random configuration network. BACKGROUND

[0002] With the increasing number of equipment types in low-voltage power supply systems, the probability of electrical fires caused by arc faults due to line aging, loose contact, etc. has greatly increased. Therefore, it is particularly important to study an effective method for detecting arc faults.

[0003] Existing arc fault detection methods mainly fall into three categories: detection based on arc mathematical models, detection based on arc physical phenomena, and detection based on voltage or current waveforms. Among them, the detection method based on current waveform installs a monitoring point on the upstream of the line to achieve arc fault protection for the downstream branch, which has stronger applicability and flexibility, and is more favored by researchers compared with other detection methods.

[0004] The detection method based on current waveform inputs current signal data into the detection model for detection. Existing detection models mainly include BP neural network, convolutional neural network (CNN), etc., but the detection results depend on complex network structure and large-scale training data, with poor generalization ability and difficulty in realizing adaptive association between network structure and current data. Li et al. [Li J, Zou G, Wang W, et al. Low-voltage series arc fault detection based on ECMC and VB-SCN [J]. Electric Power Systems Research, 2023, 218: 109222.] proposed to use Stochastic Configuration Networks (SCN) to adaptively learn the features of current signals to build network structure, effectively realizing data-driven and universal approximation. However, the random configuration network uses the Sigmoid activation function to map the periodic current data to [0, 1], which reduces the expression ability of arc fault features in positive and negative periods. At the same time, when the current signal changes greatly or very small, it is easy to form a flat area due to function saturation, increasing the possibility of generating low-quality hidden layer nodes. In addition, although the random configuration network uses a supervised mechanism to control the random generation process of hidden layer nodes, it only has one hidden layer, which is difficult to learn all the features of arc fault current, inevitably produces linear redundant nodes, increases the computational complexity of the network and the sensitivity of overfitting, and even generates a pathological output matrix, ultimately affecting the generalization performance and stability of the network. SUMMARY

[0005] The purpose of the present application is to provide an arc fault detection method based on a fully adaptive random configuration network, which can learn the characteristics of current signals to build network structure, realize adaptive adjustment of network width and depth, solve the problem of adaptability of existing arc fault detection algorithms. At the same time, eliminate the negative impact of linear nodes of the network on the performance of arc fault detection, reduce the computational complexity of the detection algorithm, improve the generalization ability of the detection algorithm, realize fast and stable detection of arc fault, and ensure the safety of low-voltage power use. The purpose of the present application is achieved as follows:

[0006] A fully adaptive random configuration network comprises the following steps,

[0007] Step S1, collecting current signals;

[0008] Step S2, learning current signals by using a random configuration network improved by a Tanh activation function to generate hidden layer nodes;

[0009] Step S3, judging the redundancy of hidden layer nodes to realize adaptive adjustment of the width of the random configuration network;

[0010] Step S4, judging the error level of hidden layer nodes to realize adaptive adjustment of the depth of the random configuration network.

[0011] In an embodiment of the present application, the step S2 further comprises,

[0012] Step S21, learning current characteristics by using a Tanh activation function to generate hidden layer nodes of the network, build network structure, and the output F of the built network is:

[0013]

[0014]

[0015]

[0016] Wherein F0=0, τ(z)=(e z -e -z ) / (e z +e -z ) is a Tanh activation function, is the output of the l D th hidden layer node of the Dth hidden layer. When D=1, X D =[x1,x1,…,x n ] is a current data set; when D>1, X D =H D-1L-1 represents the output of the previous hidden layer, and A Tdenotes the transpose of A, rand() is a random sequence generating function, and n is the dimension of X, is the output weight, and m denotes the dimension of the output of each training set, and are the input weight and bias of the l D th node of the Dth hidden layer, respectively, λ = [υ min : Δυ: υ max ] is a scaling function, and T_max is the maximum number of random configurations.

[0017] In an embodiment of the present application, the step S2 further comprises,

[0018] a residual error e of the randomly configured network is:

[0019]

[0020] where f is a target output, is an output matrix.

[0021] In an embodiment of the present application, the step S2 further comprises,

[0022] a supervision mechanism of the randomly configured network is:

[0023]

[0024] where 0 < r < 1, is a non-negative real sequence, is the output of the l D th node of the Dth hidden layer, r is a learning parameter, and is an increasing sequence gradually approaching 1, q = 1, 2, …, m.

[0025] In an embodiment of the present application, the step S2 further comprises,

[0026] an output weight β of the randomly configured network is:

[0027]

[0028] where T is a label matrix, and ||·||F F denotes the Frobenius norm, denotes the pseudo-inverse,

[0029] In an embodiment of the present application, the step S3 further comprises,

[0030]

[0031] where, and respectively represent the training accuracy and training error of the Lth hidden layer node, and the minimum node contribution degree is set as δ. D If the newly generated node satisfies formula (7), that is, there is a redundant linear node pair, the training accuracy of the network has a negative effect, and the effect on the network convergence process can be ignored, the node is deleted and output Otherwise, keep

[0032] In an embodiment of the present application, the step S4 further comprises,

[0033] Let X D = H DL-1 , and build a new hidden layer according to formulas (1)-(6) until D = Depth or and output the network model, where Depth is the maximum number of hidden layers set, and tol is the tolerance level set.

[0034] In an embodiment of the present application, the arc fault detection method based on the completely adaptive stochastic configuration network further comprises,

[0035] Step S5, using the completely adaptive stochastic configuration network network built to realize the detection of arc fault.

[0036] The present application provides an arc fault detection method based on a completely adaptive stochastic configuration network (CASCN), which realizes adaptive adjustment of the network width and depth, realizes data-driven, and reduces the complexity of network calculation. At the same time, the Tanh activation function is used to map the current signal to [-1, 1], which retains the characteristics of the current signal in the positive and negative cycles to the greatest extent, is more sensitive to the small changes of the current data when the arc fault occurs, reduces the possibility of function saturation, and further reduces the flat area that may cause redundant linear nodes, further simplifies the network structure, avoids the problems of gradient decline and poor generalization ability caused by too deep network, and realizes accurate and rapid detection of arc fault. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 Flowchart of the arc fault detection method based on the completely adaptive stochastic configuration network;

[0038] Figure 2 Arc fault current waveform diagram;

[0039] Figure 3 Activation function curve;

[0040] Figure 4 Completely adaptive stochastic configuration network model when D = 2.​ DETAILED DESCRIPTION

[0041] The present application is described in greater detail by the specific embodiments below, from which other advantages and effects of the present application can be easily understood by those skilled in the art from the content disclosed in the specification.

[0042] It should be understood that the drawings attached to the specification are only used to understand and read the content disclosed in the specification by those skilled in the art, and do not have technical substantive significance, and therefore do not define the limitations of the implementation of the present application.

[0043] As shown in Figure 1 , the present application provides a flow chart of an arc fault detection method based on a fully adaptive stochastic configuration network:

[0044] Step S1, collecting current signals. The current signal is denoted as x(t), as shown in Figure 2 in combination with Figure 4 , and then a current data set X=[x1, x1, …, x n ] is constructed.

[0045] Step S2, using a Tanh activation function to improve the stochastic configuration network to learn the current signal, and generating hidden layer nodes.

[0046] Step S3, judging the redundancy of the hidden layer nodes to realize adaptive adjustment of the width of the stochastic configuration network.

[0047] Step S4, judging the error level of the hidden layer nodes to realize adaptive adjustment of the depth of the stochastic configuration network.

[0048] The fully adaptive stochastic configuration network is built by the above steps.

[0049] Step S5, using the fully adaptive stochastic configuration network built to realize detection of arc faults.

[0050] The fully adaptive stochastic configuration network of the present application has the following characteristics,

[0051] 1. Using a Tanh activation function to enhance the ability of the network to acquire current characteristics;

[0052] 2. Judging the linearity of the hidden layer nodes to realize adaptive adjustment of the width of the network;

[0053] 3. Judging the error level of the hidden layer nodes to realize adaptive adjustment of the depth of the network.

[0054] To this end, the fully adaptive random configuration network is improved by using the Tanh activation function to map the input signal to [-1, 1], which can maximize the characteristics of the current signal in the positive and negative cycles. At the same time, the Tanh function has a steeper gradient, making the random configuration network more sensitive to small changes in current data when arc fault occurs, reducing the possibility of function saturation, thereby reducing the flat area that may lead to the generation of redundant linear nodes, improving the detection rate while avoiding the problems of gradient decline and poor generalization ability caused by the excessive depth of the fully adaptive random configuration network.

[0055] The step S2 uses the random configuration network improved by the Tanh activation function to learn the current signal and generate the hidden layer nodes, including:

[0056] In step S21, the random configuration network improved by the Tanh activation function is used to learn the current characteristics and generate the hidden layer nodes of the network, and the network structure is built. The output F of the random configuration network is:

[0057]

[0058]

[0059]

[0060] wherein F0=0, τ(z)=(e z -e -z ) / (e z +e -z ) is the Tanh activation function, is the output of the l D th hidden layer node of the Dth hidden layer. When D=1, X D =[x1,x1,…,x n ] is the current data set; when D>1, X D =H D-1L-1 represents the output of the previous hidden layer. A T represents the transpose of A, rand() is a random sequence generation function, and n is the dimension of X. Please refer to Figure 4 is the fully adaptive random configuration network model when D=2, is the output weight, and m represents the dimension of the output of each training set. and are the input weight and bias of the l D th node of the Dth hidden layer, respectively. λ=[υ min :Δυ:υ max ] is a scale function, and T_max is the maximum number of random configurations. Figure 3The contrast curves of the activation functions Sigmoid and Tanh. It is not difficult to find that the gradient of the Sigmoid function is obvious near 0, but the gradient is small in the saturation region, and the curve is approximately a straight line. While the Tanh function maps the input signal to [-1, 1] while retaining the advantages of the Sigmoid function, it can retain the characteristics of the current signal in the positive and negative cycles to the greatest extent. At the same time, the Tanh function has a steeper gradient, which is more sensitive to small changes in current data when arc fault occurs, reduces the possibility of function saturation, and thus reduces the flat area that may lead to redundant linear nodes, improves the detection rate, and avoids the problems of gradient decline and poor generalization ability caused by deep network.

[0061] Step S22, define the residual error e as:

[0062]

[0063] where f is the target output, is the output of the network.

[0064] Step S23, the supervision mechanism of the randomly configured network is:

[0065]

[0066] where 0 < r < 1, μ LD is a non-negative real sequence, r is a learning parameter, which is an increasing sequence gradually approaching 1, q = 1, 2, …, m. is the output of the L D th node of the Dth hidden layer.

[0067] Step S24, calculate the output weight β of the randomly configured network:

[0068]

[0069] where T is the label matrix, and ||·||F F represents the Frobenius norm, represents the pseudo-inverse,

[0070] The step S3 judges the redundancy of the hidden layer nodes, and realizes the self-adaptive adjustment of the width of the randomly configured network, more specifically including,

[0071]

[0072] where, and respectively represent the L Dthe training accuracy and training error of the new generated node, δ is the set minimum node contribution. If the new generated node satisfies formula (7), i.e. there is a redundant linear node pair, the training accuracy of the network is negatively affected, the influence on the network convergence process can be ignored, then the node is deleted , and the output is Otherwise, the node is retained

[0073] Step S4 judges the error level of the hidden layer node, realizes the adaptive adjustment of the randomly configured network depth, and more specifically includes:

[0074] Let X D = H DL-1 , according to the formulas (1)-(6), build a new hidden layer until D = Depth or stop building and output the network model, Depth is the set maximum number of hidden layers, and tol is the set tolerance level. Figure 4 is a completely adaptive randomly configured network model structure diagram when D = 2, wherein satisfies formula (7), the redundant linear node is discarded, does not satisfy formula (7), and satisfies the network stops training.

[0075] Step S5, using the current data set as input, the completed adaptive randomly configured network is used to realize arc fault detection. First, the arc fault current data set X is obtained through experiment as input sample, M1 samples are randomly selected from X as training sample, and the remaining M2 samples are used as test sample to realize the optimal selection of the adaptive randomly configured network parameters λ, T_max and δ; then, on the basis of the optimal parameters, the adaptive randomly configured network structure is built to realize the rapid and stable detection of arc fault. Finally, the progress and rate of arc fault detection are obtained through the test of the test set, and the generalization accuracy is reported.

[0076] The arc fault detection method based on the completely adaptive randomly configured network can realize the adaptive adjustment of the network structure, effectively eliminate the negative influence of the linear redundant node on the detection performance of the network, at the same time, the Tanh function is used as the activation function of the completely adaptive randomly configured network, the feature expression ability of the arc fault is retained to the greatest extent, the network structure is further simplified, the generalization ability of the network is improved, and the rapid and stable detection of the arc fault can be realized.

[0077] The application provides a low-voltage arc fault detection method based on a fully self-adaptive random configuration network. The method first adopts a Tanh improved random configuration network to learn arc fault current data to generate hidden layer nodes, and improves the expression ability of arc fault features. Then, the redundancy degree of the hidden layer nodes is judged to realize self-adaptive adjustment of the network width, and the error level of the hidden layer nodes is judged to realize self-adaptive adjustment of the network depth, so that the negative influence of linear redundant nodes on the detection performance of the network is effectively eliminated. The method can realize self-adaptive adjustment of the network structure, further simplify the network structure, improve the generalization ability of the network, and realize rapid and stable detection of arc faults.

Claims

1. A fully adaptive randomly configured network, characterized in that, The method comprises the following steps: Step S1, sampling a current signal; Step S2, using a Tanh activation function to improve a random configuration network to learn the current signal, and generating hidden layer nodes; Step S3, judging the redundancy of the hidden layer nodes, and realizing self-adaptive adjustment of the width of the random configuration network; Step S4, judging the error level of the hidden layer nodes, and realizing self-adaptive adjustment of the depth of the random configuration network; The step S3 further comprises, (7) wherein, and respectively represent the training accuracy and the training error of the i-th hidden layer j-th hidden layer node, is the set minimum node contribution degree, if the newly generated node satisfies formula (7), that is, there is a redundant linear node pair, which has a negative impact on the training accuracy of the network, and the impact on the network convergence process can be ignored, then the node is deleted and output ; otherwise, keep .​​​ 2. The fully adaptive random configuration network of claim 1, wherein, The step S2 further comprises, Step S21, in the random configuration network, the current characteristics are learned by using the random configuration network improved by the Tanh activation function, the hidden layer nodes of the network are generated, the network structure is built, and the output of the built network is : (1) (2) (3) wherein , is a Tanh activation function, , in equation (1) , is the output of the th hidden layer th hidden layer node when , is the current data set; when , denotes the output of the preceding hidden layer, denotes the transpose of , is a random sequence generating function, is the dimension of , is the output weight, denotes the dimension of the output of each training set, and are the input weight and bias of the th hidden layer th node, respectively, is a scaling function, is the maximum number of random configurations.

3. A fully adaptive randomly configured network as claimed in claim 2, wherein, The step S2 further comprises, Step S22, the residual of the random configuration network is: (4) wherein is the target output, is the output matrix.

4. A fully adaptive randomly configured network as claimed in claim 3, wherein, The step S2 further comprises, The step S2 further comprises, (5) wherein, , is a non-negative real sequence, , , is the output of the th node of the th hidden layer, is a learning parameter, is an increasing sequence gradually approaching 1, .

5. A fully adaptive randomly configured network as claimed in claim 4, wherein, The step S2 further comprises, Step S24, output the weights of the randomly configured network is: (6) wherein is a label matrix, denotes the Frobenius norm, denotes the pseudo-inverse, .

6. A fully adaptive randomly configured network as claimed in claim 5, wherein, The step S4 further comprises, Let , build new hidden layers according to formulas (1)-(6) until or stop building and output the network model, is the maximum number of hidden layers set, is the tolerance level set.

7. Arc fault detection method based on fully adaptive random configuration network, using the fully adaptive random configuration network as claimed in claims 1-6, characterized in that, The step S4 further comprises, The step S4 further comprises, Step S5, using the completely self-adaptive random configuration network built to realize arc fault detection.

Images