CN115879044A - CNN network-based GIS switching-on/off state current detection method and device - Google Patents

Abstract

The invention discloses a GIS switching-on and switching-off state current detection method and device based on a CNN network, wherein the method comprises the following steps: collecting the opening and closing states of a disconnecting link and a grounding switch and corresponding coupling capacitance current data sets in all states of GIS equipment; preprocessing the acquired current data, and constructing a training set of the network; constructing a CNN network; optimizing the CNN network parameters by using a particle swarm optimization algorithm to obtain optimal CNN network parameters; acquiring fault current data of a disconnecting link and a grounding switch of GIS equipment in real time, preprocessing the fault current data, inputting the preprocessed data into an optimal CNN network, and outputting a GIS switching-on and switching-off state judgment result by using the optimal CNN network; the invention has the advantages that: the execution speed is fast, the real-time performance is high, and after the model is trained, the output result is more accurate, and the reliability is strong.

Description

GIS opening and closing state current detection method and device based on CNN network

Technical Field

The present invention relates to the field of detection of knife switch opening and closing states in GIS equipment, and more specifically to a method and device for detecting current in the opening and closing states of GIS based on a CNN network.

Background Art

Gas-Insulated Metal-enclosed Switchgear (GIS) is a metal-enclosed switchgear that uses gas in whole or in part as the insulating medium instead of air at atmospheric pressure. Compared with traditional open-type distribution devices, GIS has the advantages of small footprint, fully sealed components that are not affected by the environment, easy operation, small maintenance workload, easy equipment installation, and short construction period.

The 500kV GIS of a power station was jointly produced by Mitsubishi Corporation of Japan and Xi'an Switch Factory. It was manufactured in 1990 and put into operation in 1992. The model is 500-GNS-MS50, the main wiring is one and a half times wiring, a total of 4 series and 12 intervals, the switch, knife switch (isolation knife switch), and earth knife (grounding knife switch) are all pneumatic operating mechanisms, using centralized air supply, the rated operating air pressure is 1.5MPa, including 33 knife switches (99 phases) and 43 earth knives (129 phases). During the period from 2015 to 2017, the 500kV GIS of the power station exposed 4 defects of knife switches and earth knives that failed to open and close due to internal mechanism failures, which seriously affected the safe and stable operation of the system. In order to avoid equipment accidents, it is necessary to study the opening and closing of knife switches and earth knives.

Chinese patent publication number CN113484741A discloses a GIS disconnector working state monitoring device and method, which uses an open-type current transformer to obtain the current value of the GIS disconnector operating mechanism driving motor that changes with time during the operation process, and indirectly judges the operation state of the disconnector according to the current value of the operating mechanism driving motor that changes with time during the operation process; uses an angular displacement sensor to obtain the rotation displacement value of the GIS disconnector operating mechanism driving motor that changes with time, and judges the position state of the GIS disconnector knife switch; accumulates the number of GIS disconnector opening and closing operations, predicts the mechanical life of the GIS disconnector, and effectively improves the accuracy of disconnector working state monitoring. However, this patent application relies on the current value and sensor data during the operation of the operating mechanism driving motor to judge the knife switch action and position, which is not real-time. The current knife switch action and position may actually be the historical action and position of the knife switch calculated by the current and sensor data at the historical moment, resulting in the actual judgment result being inaccurate and unreliable.

Summary of the invention

The technical problem to be solved by the present invention is that the existing method for judging the working state of a GIS disconnector is not real-time, the judgment result is not accurate enough, and the reliability is not high.

The present invention solves the above technical problems by the following technical means: a GIS opening and closing state current detection method based on a CNN network, the method comprising:

Step 1: Collect the opening and closing status of the switch and ground switch in various states of the GIS equipment and the corresponding coupling capacitor current data set;

Step 2: Preprocess the collected current data and construct a network training set;

Step 3: Construct a CNN network, whose input is the training set and output is the opening and closing status of the knife switch and the ground knife;

Step 4: Use particle swarm optimization algorithm to optimize CNN network parameters to obtain the optimal CNN network parameters;

Step 5: Set the CNN network according to the optimal CNN network parameters to obtain the optimal CNN network. Collect the fault current data of the knife switch and ground switch of the GIS equipment in real time, pre-process them and input them into the optimal CNN network. Use the optimal CNN network to output the GIS opening and closing status judgment results.

Beneficial effect: The present invention obtains the optimal CNN network parameters by constructing a CNN network and optimizing the CNN network parameters using a particle swarm optimization algorithm, thereby obtaining the optimal CNN network. Finally, the optimal CNN network is used to judge the real-time collected knife switch and ground switch fault current data, and output the GIS opening and closing status. Compared with the existing sensor measurement method, the trained neural network is used for result judgment, which has fast execution speed and high real-time performance. After training, the output results are more accurate and reliable.

Furthermore, the step 1 comprises:

表示热备状态下第k个地刀电流；When the GIS equipment is in hot standby state, the opening and closing status of each group of knife switches and earth switches is knife switch closed and earth switch opened. The earth switch opening and closing status Y ^H = {y ₁ ^H ,y ₂ ^H …y _k ^H …y _K ^H } is recorded, and the current data set I ^H = {I ₁ ^H ,I ₂ ^H …I _k ^H …I _K ^H } is collected at the same time.

Indicates the kth ground switch current in hot standby state;

表示冷备用状态下第k个地刀电流,

表示冷备用状态下第m个刀闸电流；When the GIS equipment is in cold standby state, the opening and closing status of each group of knife switches and earth switches is knife switch open and earth switch open. The opening and closing status of earth switches and knife switches Y ^C = {y ₁ ^C ,y ₂ ^C …y _k ^C …y _K ^C ,y ₁ ^C' ,y ₂ ^C' …y _m ^C' …y _M ^C' } is recorded. At the same time, the current data set of earth switches and knife switches I ^C = {I ₁ ^C ,I ₂ ^C …I _k ^C …I _K ^C ,I ₁ ^C' ,I ₂ ^C' …I _m ^C' …I _M ^C' } is collected, where

Indicates the kth ground switch current in cold standby state,

Indicates the current of the mth switch in the cold standby state;

表示检修状态下第m个边界刀闸电流。When the GIS equipment is under maintenance, the opening and closing status of each group of switch and earth switch is switch open and earth switch closed. The switch opening and closing status is recorded, Y ^R = {y ₁ ^R ,y ₂ ^R …y _m ^R …y _M ^R }, and the switch current data set ^IR = {I ₁ ^R ,I ₂ ^R …I _m ^R …I _M ^R } is collected at the same time, where

Indicates the current of the mth boundary switch in the maintenance state.

Furthermore, the step 2 comprises:

S21, collecting the fault current data of the switch and the ground switch, performing wavelet transform on the fault current signal, thereby obtaining the wavelet coefficients _Wi (a, b) and the selected frequency _fins ;

得到零矩阵，其中，设(Δa)_k＝a_k-a_k-1，Δf＝f_k-f_k-1，a_k为第k个瞬时分量的尺度因子，f_k是第k个瞬时分量的中心频率，取k∈[0，n_a]，f_s为信号采样频率；由S21中的f_ins，根据式f_ins＝2^kΔf·f_s/n_a计算得到k值，其中

S22, through the formula

The zero matrix is obtained, where (Δa) _k = _ak - _ak-1 , Δf = f _k - _{f k-1} , _ak is the scale factor of the kth instantaneous component, f _k is the center frequency of the kth instantaneous component, k∈[0, _na ], f _s is the signal sampling frequency; from _fins in S21, the k value is calculated according to the formula _fins = 2 ^kΔf ·f _s / _na , where

Determine whether k∈[0, n _a ] satisfies. If so, then

Repeat the above steps for a preset number of times to make the signal thinner and clearly present it on the time-frequency diagram.

重建信号及其各个分量，其中

R_e为取实部。Passing

Reconstruct the signal and its individual components, where

R _e is the real part.

Furthermore, the S21 includes:

Select a wavelet mother function ψ(t) so that its Fourier transform Ψ(ω) satisfies the following admissibility condition of the wavelet function:

Select the scaling factor a and the translation factor b, scale and translate the wavelet mother function ψ(t), and you can get the wavelet basis function:

Perform wavelet transform on the current signal according to the following formula to obtain the wavelet coefficients _Wi (a, b):

where ψ ^* (t) is the complex conjugate of ψ(t).

Furthermore, the CNN network includes a sequentially connected input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer. The convolution layer has three layers, and the convolution kernel extracts the characteristics of the current spectrum. The pooling layer also has three layers, and the output data of the three convolution layers are respectively reduced in dimension; the output layer is composed of a Softmax module and a classification module.

得到当前神经元

式中，M_j为输入到神经元j的向量总数，

为前一层输入神经元的特征图；ki_j ^l为卷积核的权重；b_j为第j种特征图的偏置；σ为ReLU激活函数；Furthermore, the convolutional layer is constructed by the formula

Get the current neuron

Where _Mj is the total number of vectors input to neuron j,

is the feature map of the input neuron in the previous layer; ki _j ^l is the weight of the convolution kernel; b _j is the bias of the jth feature map; σ is the ReLU activation function;

Softmax module output

Where _xj is the jth output of the fully connected layer, k is the number of inputs to the Softmax module, k = 2, then P1 and P2 are

y_i为事件标签值，y₁＝1，y₂＝0。P1 and P2 represent the probabilities of outputting y=1 and y=0 respectively; the probability values P1 and P2 are input into the classification module, and the cross loss entropy function value L is calculated in the classification module, where

_yi is the event label value, _y1 = 1, _y2 = 0.

Furthermore, the CNN network also includes an Inception-v3 structure, which uses filters of size 1×n and n×1 to replace the n×n convolution in the CNN network.

Furthermore, the step four includes: setting the loss function L _loss , when the loss value L of the CNN network is less than L _loss , the network training is reliable, and the label value with a larger value in P1 and P2 is used as the network output; otherwise, it indicates that the network weight distribution is improper, and the particle swarm optimization algorithm is used to reallocate the weights and repeat the above process until the network output conditions are met to obtain the optimal CNN network parameters.

Furthermore, the particle swarm optimization algorithm reallocates weights, including:

S41. After the CNN network calculates the error between the expected value and the actual value, each particle takes the batch size of samples in the CNN network, the training data discard rate, the number of convolution kernels N _c , the convolution kernel size _Mc and the network initial bias b _j as the hyperparameters to be optimized, and the hyperparameters to be optimized are used as the particle dimensions;

S42, using the error between the expected value and the actual value of the CNN network as the fitness function, calculating the fitness value of each particle, saving the optimal point of each particle in n iterative updates as the individual optimal solution P _best , and then comparing it with the particles of the entire population, selecting the global optimal value G _best and saving it;

S43, comparing the fitness value of the particle with the optimal fitness value of the individual, if the fitness value of the particle is larger, replacing it with the optimal fitness value of the individual;

S44, comparing the individual optimal fitness values of all particles with the global optimal fitness value of the group, if the individual optimal fitness value is larger, it is replaced by the global optimal fitness value;

S45, update the speed and position of the particle, and return to execute step S42;

S46. Compare the current number of iterations with the maximum number of iterations. When n＜N, the iteration continues; when n＞N, the iteration ends, or the iteration ends when the minimum error is reached. If both iteration end conditions are not met, return to step S42, otherwise exit the loop to obtain the global optimal fitness value. The parameters of the particles corresponding to the global optimal fitness value are the optimal CNN network parameters obtained by the particle swarm optimization.

The present invention also provides a GIS opening and closing state current detection device based on a CNN network, the device comprising:

The data acquisition module is used to collect the opening and closing status of the knife switch and the ground knife in various states of the GIS equipment and the corresponding coupling capacitor current data set;

A training set construction module is used to pre-process the collected current data and construct a training set for the network;

The network construction module is used to construct a CNN network, whose input is the training set and output is the opening and closing status of the knife switch and the ground knife;

Parameter optimization module, used to optimize CNN network parameters using particle swarm optimization algorithm to obtain the optimal CNN network parameters;

The result output module is used to set the CNN network according to the optimal CNN network parameters to obtain the optimal CNN network, collect the switch and ground switch fault current data of the GIS equipment in real time, and input them into the optimal CNN network after preprocessing, and use the optimal CNN network to output the GIS opening and closing status judgment result.

Furthermore, the data acquisition module is also used for:

表示热备状态下第k个地刀电流；When the GIS equipment is in hot standby state, the opening and closing status of each group of knife switches and earth switches is knife switch closed and earth switch opened. The earth switch opening and closing status Y ^H = {y ₁ ^H ,y ₂ ^H …y _k ^H …y _K ^H } is recorded, and the current data set I ^H = {I ₁ ^H ,I ₂ ^H …I _k ^H …I _K ^H } is collected at the same time.

Indicates the kth ground switch current in hot standby state;

表示冷备用状态下第k个地刀电流,

表示冷备用状态下第m个刀闸电流；When the GIS equipment is in cold standby state, the opening and closing status of each group of knife switches and earth switches is knife switch open and earth switch open. The opening and closing status of earth switches and knife switches Y ^C = {y ₁ ^C ,y ₂ ^C …y _k ^C …y _K ^C ,y ₁ ^C' ,y ₂ ^C' …y _m ^C' …y _M ^C' } is recorded. At the same time, the current data set of earth switches and knife switches I ^C = {I ₁ ^C ,I ₂ ^C …I _k ^C …I _K ^C ,I ₁ ^C' ,I ₂ ^C' …I _m ^C' …I _M ^C' } is collected, where

Indicates the kth ground switch current in cold standby state,

Indicates the current of the mth switch in the cold standby state;

表示检修状态下第m个边界刀闸电流。When the GIS equipment is under maintenance, the opening and closing status of each group of knife switches and earth switches is knife switch open and earth switch closed. The knife switch opening and closing status is recorded, Y ^R = {y ₁ ^R ,y ₂ ^R …y _m ^R …y _M ^R }, and the knife switch current data set ^IR = {I ₁ ^R ,I ₂ ^R …I _m ^R …I _M ^R } is collected at the same time, where

Indicates the current of the mth boundary switch in the maintenance state.

Furthermore, the training set construction module is also used for:

S21, collecting the fault current data of the switch and the ground switch, performing wavelet transform on the fault current signal, thereby obtaining the wavelet coefficients _Wi (a, b) and the selected frequency _fins ;

得到零矩阵，其中，(Δa)_k＝a_k-a_k-1，Δf＝f_k-f_k-1，a_k为第k个瞬时分量的尺度因子，f_k是第k个瞬时分量的中心频率，取k∈[0，n_a]，f_s为信号采样频率；由S21中的f_ins，根据式f_ins＝2^kΔf·f_s/n_a计算得到k值，其中

S22, through the formula

The zero matrix is obtained, where (Δa) _k = _ak - _ak-1 , Δf = f _k - _{f k-1} , _ak is the scale factor of the kth instantaneous component, f _k is the center frequency of the kth instantaneous component, k∈[0, _na ], and f _s is the signal sampling frequency; the k _value is calculated from _{fins in S21 according to the formula fins} = 2 ^kΔf ·f _s / _na , where

Determine whether k∈[0, n _a ] satisfies. If so, then

Repeat the above steps for a preset number of times to make the signal thinner and clearly present it on the time-frequency diagram.

重建信号及其各个分量，其中

R_e为取实部。Passing

Reconstruct the signal and its individual components, where

R _e is the real part.

Furthermore, the S21 includes:

Select a wavelet mother function ψ(t) so that its Fourier transform Ψ(w) satisfies the following admissibility condition of the wavelet function:

Select the scaling factor a and the translation factor b, scale and translate the wavelet mother function ψ(t), and you can get the wavelet basis function:

Perform wavelet transform on the current signal according to the following formula to obtain the wavelet coefficients _Wi (a, b):

where ψ ^* (t) is the complex conjugate of ψ(t).

Furthermore, the CNN network includes a sequentially connected input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer. The convolution layer has three layers, and the convolution kernel extracts the characteristics of the current spectrum. The pooling layer also has three layers, and the output data of the three convolution layers are respectively reduced in dimension; the output layer is composed of a Softmax module and a classification module.

得到当前神经元

式中，M_j为输入到神经元j的向量总数，

为前一层输入神经元的特征图；k_ij ^l为卷积核的权重；b_j为第j种特征图的偏置；σ为ReLU激活函数；Furthermore, the convolutional layer is constructed by the formula

Get the current neuron

Where _Mj is the total number of vectors input to neuron j,

is the feature map of the input neuron in the previous layer; k _ij ^l is the weight of the convolution kernel; b _j is the bias of the jth feature map; σ is the ReLU activation function;

Softmax module output

Where _xj is the jth output of the fully connected layer, k is the number of inputs to the Softmax module, k = 2, then P1 and P2 are

y_i为事件标签值，y₁＝1，y₂＝0。P1 and P2 represent the probabilities of outputting y=1 and y=0 respectively; the probability values P1 and P2 are input into the classification module, and the cross loss entropy function value L is calculated in the classification module, where

_yi is the event label value, _y1 = 1, _y2 = 0.

Furthermore, the CNN network also includes an Inception-v3 structure, which uses filters of size 1×n and n×1 to replace the n×n convolution in the CNN network.

Furthermore, the parameter optimization module is also used to: set the loss function L _loss , when the loss value L of the CNN network < L _loss , the network training is reliable, and the label value with a larger value in P1 and P2 is used as the network output; otherwise, it means that the network weight distribution is improper, and the particle swarm optimization algorithm is used to reallocate the weights and repeat the above process until the network output conditions are met to obtain the optimal CNN network parameters.

Furthermore, the particle swarm optimization algorithm reallocates weights, including:

S41. After the CNN network calculates the error between the expected value and the actual value, each particle takes the batch size of samples in the CNN network, the training data discard rate, the number of convolution kernels N _c , the convolution kernel size _Mc and the network initial bias b _j as the hyperparameters to be optimized, and the hyperparameters to be optimized are used as the particle dimensions;

S42, using the error between the expected value and the actual value of the CNN network as the fitness function, calculating the fitness value of each particle, saving the optimal point of each particle in n iterative updates as the individual optimal solution P _best , and then comparing it with the particles of the entire population, selecting the global optimal value G _best and saving it;

S43, comparing the fitness value of the particle with the optimal fitness value of the individual, if the fitness value of the particle is larger, replacing it with the optimal fitness value of the individual;

S44, comparing the individual optimal fitness values of all particles with the global optimal fitness value of the group, if the individual optimal fitness value is larger, it is replaced by the global optimal fitness value;

S45, update the speed and position of the particle, and return to execute step S42;

S46. Compare the current number of iterations with the maximum number of iterations. When n＜N, the iteration continues; when n＞N, the iteration ends, or the iteration ends when the minimum error is reached. If both iteration end conditions are not met, return to step S42, otherwise exit the loop to obtain the global optimal fitness value. The parameters of the particles corresponding to the global optimal fitness value are the optimal CNN network parameters obtained by the particle swarm optimization.

The advantages of the present invention are as follows: the present invention obtains the optimal CNN network parameters by constructing a CNN network and optimizing the CNN network parameters using a particle swarm optimization algorithm, thereby obtaining the optimal CNN network, and finally using the optimal CNN network to judge the real-time collected knife switch and ground switch fault current data, and output the GIS opening and closing status. Compared with the sensor measurement method in the prior art, the present invention uses a trained neural network to judge the results, with fast execution speed and high real-time performance, and after training, the output results are more accurate and reliable.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of three situations of knife switches and ground switches in the prior art;

FIG2 is a diagram of the CNN network architecture in the GIS opening and closing state current detection method based on the CNN network disclosed in Example 1 of the present invention;

FIG3 is a schematic diagram of the structure of the inception module in the GIS opening and closing state current detection method based on the CNN network disclosed in Example 1 of the present invention;

FIG4 is a flow chart of a particle swarm optimization algorithm in a GIS opening and closing state current detection method based on a CNN network disclosed in Example 1 of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Example 1

The inventor studied the opening and closing conditions of knife switches and earth switches, and conducted a statistical analysis of the failures of knife switches and earth switches. The statistical analysis found that the common problem of GIS equipment is that the knife switches and earth switches fail to complete the action according to the predetermined instructions due to the failure of the internal transmission mechanism, and there is no half-open or half-closed state. Adjacent parallel lines are in a strong electric field, which will generate an induced voltage and generate an electric coupling capacitor current through the earth switch grounding plate when the earth switch is in the closed state, while there is no line coupling capacitor current in the grounding plate in the open state. The voltage-equalizing covers on both sides of the knife switch break can be regarded as the two plates of the capacitor, and the break can be equivalent to a fixed capacitor. There is high voltage at one end of the break and the other end is grounded. There must be a coupling capacitor current passing through the break and flowing into the ground grid through the earth switch grounding plate. As shown in Figure 1, the knife switch and the earth switch have the following three situations:

1) The knife switch break is connected in series with the ground knife break (see Figure 1(a)), and the coupling capacitor current calculation formula is as follows:

Where j is the imaginary unit.

2) When the knife switch is in the open state and the ground switch is in the closed state (see Figure 1(b)), the coupling capacitor current calculation formula is as follows:

I _c =j2πfC _g U

3) When the earth switch is in the disconnected state and the knife switch is in the closed state (see Figure 1(c)), the coupling capacitor current calculation formula is as follows:

I _c =j2πfC _j U

For the fracture of GIS knife switch and earth switch, since it is a purely capacitive structure, the resistive current is 0, and the total leakage current is the coupling capacitor current; therefore, by detecting the characteristic changes of the coupling capacitor current during the switching operation of the knife switch and earth switch, the opening and closing conditions of the knife switch and earth switch can be judged.

Based on the above analysis, the present invention provides a GIS opening and closing state coupling capacitor current detection method based on CNN network, which realizes the judgment of GIS opening and closing by collecting the characteristic changes of coupling capacitor current during the operation of knife switch and ground knife. The method process of the present invention is described in detail below.

S1. Collect the opening and closing status of the knife switch and the ground knife in each state of the GIS equipment and the coupling capacitor current data set; the specific process of this step is as follows:

表示热备状态下第k个地刀电流；处于冷备用状态时，每组刀闸、地刀开合情况为：刀闸分闸、地刀分闸；记录地刀和刀闸分合状态Y^C＝{y₁ ^C,y₂ ^C…y_k ^C…y_K ^C，y₁ ^C',y₂ ^C'…y_m ^C'…y_M ^C'},同时采集地刀和刀闸电流数据集I^C＝{I₁ ^C,I₂ ^C…I_k ^C…I_K ^C,I₁ ^C',I₂ ^C'…I_m ^C'…I_M ^C'},中

表示冷备用状态下第k个地刀电流,中

表示冷备用状态下第m个刀闸电流；处于检修状态时，每组刀闸、地刀开合情况为：刀闸分闸、地刀合闸；记录刀闸分合状态，Y^R＝{y₁ ^R,y₂ ^R…y_m ^R…y_M ^R},同时采集刀闸电流数据集I^R＝{I₁ ^R,I₂ ^R…I_m ^R…I_M ^R},其中

表示检修状态下第m个边界刀闸电流。When the GIS equipment is powered off and restored, the current status of each device is collected; when in hot standby state, the opening and closing status of each group of knife switches and earth switches is: knife switch closed, earth switch open; record the earth switch opening and closing status Y ^H = {y ₁ ^H ,y ₂ ^H …y _k ^H …y _K ^H }, and collect the current data set I ^H = {I ₁ ^H ,I ₂ ^H …I _k ^H …I _K ^H } at the same time

Indicates the kth earth switch current in the hot standby state; in the cold standby state, the opening and closing conditions of each group of switch and earth switch are: switch open, earth switch open; record the opening and closing status of earth switch and switch Y ^C = {y ₁ ^C ,y ₂ ^C …y _k ^C …y _K ^C ,y ₁ ^C' ,y ₂ ^C' …y _m ^C' …y _M ^C' }, and collect the earth switch and switch current data set I ^C = {I ₁ ^C ,I ₂ ^C …I _k ^C …I _K ^C ,I ₁ ^C' ,I ₂ ^C' …I _m ^C' …I _M ^C' } at the same time,

Indicates the kth ground switch current in cold standby state,

Indicates the current of the mth knife switch in the cold standby state; when in the maintenance state, the opening and closing status of each set of knife switches and earth switches is: knife switch open, earth switch closed; record the knife switch opening and closing status, Y ^R = {y ₁ ^R ,y ₂ ^R ...y _m ^R ...y _M ^R }, and collect the knife switch current data set ^IR = {I ₁ ^R ,I ₂ ^R ...I _m ^R ...I _M ^R }, where

Indicates the current of the mth boundary switch in the maintenance state.

S2. Preprocess the collected current signal, use the wavelet decomposition algorithm to obtain the frequency band information of the signal, remove some clutter signals, complete the noise reduction processing of the signal, and construct the network training set; the specific process is as follows:

S21, collect the fault current data I(t) of the switch and the ground switch, perform wavelet transform on the fault current signal, thereby obtaining the wavelet coefficients _Wi (a, b) and the selected frequency _fins , and select a wavelet mother function ψ(t) so that its Fourier transform Ψ(ω) satisfies the admissibility condition of the wavelet function as follows:

Select the scaling factor a and the translation factor b, and perform scaling and translation operations on the wavelet mother function ψ(t) to obtain the wavelet basis function:

Perform wavelet transform on the current signal I(t) as follows to obtain _Wi (a, b):

where ψ ^* (t) is the complex conjugate of ψ(t).

S22, reconstruct the components of the multi-component signal by inverse transformation of synchronous compressed wavelet transform SWT, convert the time-scale plane (a, b) to the time-frequency plane (f _ins , b), and obtain S _i (f _k , b) by the following formula;

Wherein, (Δa) _k = _ak - _ak-1 , Δf = f _k - _{f k-1} , _ak is the scale factor of the kth instantaneous component, f _k is the center frequency of the kth instantaneous component; the instantaneous frequency is compressed at (f _k - 0.5(Δf) _k , f _k - 0.5(Δf) _k ), which sharpens the time-frequency diagram; k∈[0, _na ], S _i (f _k , b) is the zero matrix, the signal sampling frequency is f _s , so the signal frequency is [f _s / _na , f _s / 2]; from _fins in S21, the k value is calculated according to the following formula;

f _ins = 2 ^kΔf ·f _s /n _a

判断k∈[0，n_a]是否满足，若满足则有in

Determine whether k∈[0, n _a ] satisfies. If so, then

Repeat the above steps to make the signal thinner so that it can be clearly presented on the time-frequency diagram. The signal and its components are reconstructed by the following SWT inverse transform:

R_e为取实部；in

R _e is the real part;

S3. Construct a CNN network model. The network model is mainly divided into input layer, CNN layer, and output layer. The training set is input into the model through the input layer, and features are extracted through the CNN layer. The feature vector is generated and the model results are output through the output layer. The specific process is as follows:

并以此组成不同的特征图，激活函数选取ReLU函数，计算公式如下：S31. Construct the initial CNN network. The structure of CNN includes input layer, convolution layer, pooling layer, fully connected layer and output layer, among which convolution layer, pooling layer and fully connected layer constitute CNN layer. The input layer receives the current time-frequency spectrum after wavelet decomposition processing, and takes each pixel in the time-frequency spectrum as input. The convolution layer acts on the current time-frequency spectrum through different numbers and sizes of convolution kernels and extracts its features. In order to avoid the situation where different convolution kernels repeatedly extract the same feature of the current time-frequency spectrum, a three-layer convolution network structure is designed. The number of convolution kernels in each convolution layer is the best network parameter obtained by particle swarm optimization. The convolution kernel extracts the features of the current time-frequency spectrum, and the current neuron can be obtained by adding bias and activation functions.

Different feature maps are formed in this way, and the activation function selects the ReLU function. The calculation formula is as follows:

为前一层输入神经元的特征图；k_ij ^l为卷积核的权重；b_j为第j种特征图的偏置；σ为ReLU激活函数；Where _Mj is the total number of vectors input to neuron j,

is the feature map of the input neuron in the previous layer; k _ij ^l is the weight of the convolution kernel; b _j is the bias of the jth feature map; σ is the ReLU activation function;

Three pooling layer structures are designed, and the maximum pooling is used to reduce the dimension of the convolution layer output data to reduce the complexity of the data. In the fully connected layer, it is ensured that it has connections with all previous neurons, which is responsible for reducing the dimension of the matrix data of the pooling layer and extracting feature samples. The Softmax loss function is used as the conversion function of the fully connected layer. The CNN model structure is shown in Figure 2.

S32, add Inception module, directly replace the n×n convolution layer in the neural network with Inception module, so as to decompose the n×n convolution in the particle swarm optimization process into 1×n and n×1 filters to reduce the complexity of the CNN network structure. Generally, when n is relatively large, the Inception module is used to replace the n×n convolution layer. In practical applications, the value of n is set according to needs. The network structure of CNN is to stack convolution layers. In order to achieve better results, it is often necessary to set more layers to make the network structure deeper. Adding Inception combines different convolution layers in parallel, stretches the width of the network, reduces the depth of the network structure, and runs filters of multiple sizes on the same level network, reducing the complexity of CNN. The Inception-v3 structure is used. This structure uses filters of size 1×n and n×1 to replace n×n convolution. They perform convolution and pooling operations on the previous layer. This structure makes the filter smaller, reduces the number of parameters, and increases the calculation speed, which improves the training efficiency. The structure of the Inception module is shown in Figure 3.

S33, the output layer is composed of a Softmax module and a classification module, x1 and x2 are the outputs of the fully connected layer, and P1 and P2 are the outputs of the Softmax module of the output layer

Where k is the number of Softmax module inputs, k = 2, then P1 and P2 are

P1 and P2 represent the probabilities of the output being y=1 and y=0 respectively; the probability values P1 and P2 are input into the classification module, and the cross loss entropy function value L is calculated in the classification module;

y_i为事件标签值，y₁＝1，y₂＝0；设置损失函数L_loss＝1.2×10^-3。当L<L_loss,网络训练可靠，将P1、P2中数值较大的标签值作为网络输出；否则表示网络权重分配不当，应重新分配权值重复上述过程直至满足。其中，如图4所示，分配权值采用粒子群优化算法，算法过程参阅以下步骤。in

_Yi is the event label value, _y1 = 1, _y2 = 0; set the loss function _Lloss = 1.2× ^10-3 . When L< _Lloss , the network training is reliable, and the label value with the larger value in P1 and P2 is used as the network output; otherwise, it means that the network weight distribution is improper, and the weight should be redistributed and the above process is repeated until it is satisfied. As shown in Figure 4, the particle swarm optimization algorithm is used to allocate weights, and the algorithm process refers to the following steps.

S4, using particle swarm optimization algorithm to select and optimize the established CNN network structure and initial parameters;

S41. After the CNN network calculates the error between the expected value and the actual value, each particle takes the batch size of samples in the CNN network, the training data dropout rate, the number of convolution kernels N _c , the convolution kernel size _Mc and the network initial bias b _j as the hyperparameters to be optimized and as the particle dimensions;

S42, using the error between the expected value and the actual value of the CNN network as the fitness function, calculate the fitness value of each particle. Save the optimal point of each particle in the n-times iterative update as the individual optimal solution P _best , and then compare it with the particles of the entire population, and select the global optimal value G _best to save;

S43, comparing the fitness value of the particle with the optimal fitness value of the individual, if the fitness value of the particle is larger, replacing it with the optimal fitness value of the individual;

S44, comparing the individual optimal fitness values of all particles with the global optimal fitness value of the group, if the individual optimal fitness value is larger, it is replaced by the global optimal fitness value.

S45, update the velocity and position of the particle by the following formula, and return to execute step S42;

表示第i个粒子在第n次迭代时的速度，

为全局学习因子，

为局部学习因子，R₁和R₂为(0，1)间的随机数,

为第n次迭代后的个体极值,

为第n次迭代后的群体极值，

为第n次迭代后的位置；w _n is the inertia weight coefficient,

represents the velocity of the i-th particle at the n-th iteration,

is the global learning factor,

is the local learning factor, _R1 and _R2 are random numbers between (0, 1),

is the individual extreme value after the nth iteration,

is the population extreme value after the nth iteration,

is the position after the nth iteration;

At the same time, the inertia weight coefficient w _n is processed nonlinearly through the following formula to enhance the local and global search capabilities of multiple targets;

w _{o is} the initial inertia weight coefficient, w _f is the inertia weight value after the iteration ends, and N is the maximum number of iterations;

按下式进行更新；At the same time, the learning factor

Press the following formula to update;

为学习因子初始值。

is the initial value of the learning factor.

和惯性权重系数w_n的动态调整，平衡局部和全局搜索的能力，提高粒子群算法全局搜索水平与收敛速度。Through the learning factor

And the dynamic adjustment of the inertia weight coefficient w _n balances the local and global search capabilities and improves the global search level and convergence speed of the particle swarm algorithm.

S46. Compare the current number of iterations with the maximum number of iterations. When n＜N, the iteration continues; when n＞N, the iteration ends, or the iteration ends when the minimum error is reached. If both iteration end conditions are not met, return to step S42, otherwise exit the loop to obtain the global optimal fitness value. The parameters of the particles corresponding to the global optimal fitness value are the optimal CNN network parameters obtained by the particle swarm optimization.

S5. The CNN network is set according to the optimal CNN network parameters to obtain the optimal CNN network. The fault current data of the knife switch and the ground switch of the GIS equipment are collected in real time and input into the optimal CNN network after preprocessing. The optimal CNN network is used to output the judgment result of the GIS opening and closing status.

Example 2

Based on Example 1, Example 2 of the present invention further provides a GIS opening and closing state current detection device based on a CNN network, the device comprising:

The data acquisition module is used to collect the opening and closing status of the knife switch and the ground knife in various states of the GIS equipment and the corresponding coupling capacitor current data set;

A training set construction module is used to pre-process the collected current data and construct a training set for the network;

The network construction module is used to construct a CNN network, whose input is the training set and output is the opening and closing status of the knife switch and the ground knife;

Parameter optimization module, used to optimize CNN network parameters using particle swarm optimization algorithm to obtain the optimal CNN network parameters;

The result output module is used to set the CNN network according to the optimal CNN network parameters to obtain the optimal CNN network, collect the switch and ground switch fault current data of the GIS equipment in real time, and input them into the optimal CNN network after preprocessing, and use the optimal CNN network to output the GIS opening and closing status judgment result.

Specifically, the data acquisition module is also used for:

表示热备状态下第k个地刀电流；When the GIS equipment is in hot standby state, the opening and closing status of each group of knife switches and earth switches is knife switch closed and earth switch opened. The earth switch opening and closing status Y ^H = {y ₁ ^H ,y ₂ ^H …y _k ^H …y _K ^H } is recorded, and the current data set I ^H = {I ₁ ^H ,I ₂ ^H …I _k ^H …I _K ^H } is collected at the same time.

Indicates the kth ground switch current in hot standby state;

表示冷备用状态下第k个地刀电流,

表示冷备用状态下第m个刀闸电流；When the GIS equipment is in cold standby state, the opening and closing status of each group of knife switches and earth switches is knife switch open and earth switch open. The opening and closing status of earth switches and knife switches Y ^C = {y ₁ ^C ,y ₂ ^C …y _k ^C …y _K ^C ,y ₁ ^C' ,y ₂ ^C' …y _m ^C' …y _M ^C' } is recorded. At the same time, the current data set of earth switches and knife switches I ^C = {I ₁ ^C ,I ₂ ^C …I _k ^C …I _K ^C ,I ₁ ^C' ,I ₂ ^C' …I _m ^C' …I _M ^C' } is collected, where

Indicates the kth ground switch current in cold standby state,

Indicates the current of the mth switch in the cold standby state;

表示检修状态下第m个边界刀闸电流。When the GIS equipment is under maintenance, the opening and closing status of each group of knife switches and earth switches is knife switch open and earth switch closed. The knife switch opening and closing status is recorded, Y ^R = {y ₁ ^R ,y ₂ ^R …y _m ^R …y _M ^R }, and the knife switch current data set ^IR = {I ₁ ^R ,I ₂ ^R …I _m ^R …I _M ^R } is collected at the same time, where

Indicates the current of the mth boundary switch in the maintenance state.

Specifically, the training set construction module is also used for:

S21, collecting the fault current data of the switch and the ground switch, performing wavelet transform on the fault current signal, thereby obtaining the wavelet coefficients _Wi (a, b) and the selected frequency _fins ;

得到零矩阵，其中，(Δa)_k＝a_k-a_k-1，Δf＝f_k-f_k-1，a_k为第k个瞬时分量的尺度因子，f_k是第k个瞬时分量的中心频率，取k∈[0，n_a]，f_s为信号采样频率；由S21中的f_ins，根据式f_ins＝2^kΔf·f_s/n_a计算得到k值，其中

S22, through the formula

The zero matrix is obtained, where (Δa) _k = _ak - _ak-1 , Δf = f _k - _{f k-1} , _ak is the scale factor of the kth instantaneous component, f _k is the center frequency of the kth instantaneous component, k∈[0, _na ], and f _s is the signal sampling frequency; the k _value is calculated from _{fins in S21 according to the formula fins} = 2 ^kΔf ·f _s / _na , where

Determine whether k∈[0, n _a ] satisfies. If so, then

Repeat the above steps for a preset number of times to make the signal thinner and clearly present it on the time-frequency diagram.

重建信号及其各个分量，其中

R_e为取实部。Passing

Reconstruct the signal and its individual components, where

R _e is the real part.

More specifically, the S21 includes:

Select a wavelet mother function ψ(t) so that its Fourier transform Ψ(ω) satisfies the following admissibility condition of the wavelet function:

Select the scaling factor a and the translation factor b, scale and translate the wavelet mother function ψ(t), and you can get the wavelet basis function:

Perform wavelet transform on the current signal according to the following formula to obtain the wavelet coefficients _Wi (a, b):

where ψ ^* (t) is the complex conjugate of ψ(t).

Specifically, the CNN network includes a sequentially connected input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer. The convolution layer has three layers, and the convolution kernel extracts the characteristics of the current spectrum. The pooling layer also has three layers, and the output data of the three convolution layers are respectively reduced in dimension; the output layer is composed of a Softmax module and a classification module.

得到当前神经元

式中，M_j为输入到神经元j的向量总数，

为前一层输入神经元的特征图；k_ij ^l为卷积核的权重；b_j为第j种特征图的偏置；σ为ReLU激活函数；More specifically, the convolutional layer is constructed by the formula

Get the current neuron

Where _Mj is the total number of vectors input to neuron j,

is the feature map of the input neuron in the previous layer; k _ij ^l is the weight of the convolution kernel; b _j is the bias of the jth feature map; σ is the ReLU activation function;

Softmax module output

Where _xj is the jth output of the fully connected layer, k is the number of inputs to the Softmax module, k = 2, then P1 and P2 are

y_i为事件标签值，y₁＝1，y₂＝0。P1 and P2 represent the probabilities of outputting y=1 and y=0 respectively; the probability values P1 and P2 are input into the classification module, and the cross loss entropy function value L is calculated in the classification module, where

_yi is the event label value, _y1 = 1, _y2 = 0.

More specifically, the CNN network also includes an Inception-v3 structure, which uses filters of size 1×n and n×1 to replace the n×n convolution in the CNN network.

More specifically, the parameter optimization module is also used to: set the loss function L _loss , when the loss value L of the CNN network < L _loss , the network training is reliable, and the label value with a larger value in P1 and P2 is used as the network output; otherwise, it means that the network weight distribution is improper, and the particle swarm optimization algorithm is used to reallocate the weights and repeat the above process until the network output conditions are met to obtain the optimal CNN network parameters.

More specifically, the particle swarm optimization algorithm reallocates weights, including:

S41. After the CNN network calculates the error between the expected value and the actual value, each particle takes the batch size of samples in the CNN network, the training data discard rate, the number of convolution kernels N _c , the convolution kernel size _Mc and the network initial bias b _j as the hyperparameters to be optimized, and the hyperparameters to be optimized are used as the particle dimensions;

S42, using the error between the expected value and the actual value of the CNN network as the fitness function, calculating the fitness value of each particle, saving the best point of each particle in n iterative updates as the individual optimal solution P _best , and then comparing it with the particles of the entire population, selecting the global optimal value G _best and saving it;

S43, comparing the fitness value of the particle with the optimal fitness value of the individual, if the fitness value of the particle is larger, replacing it with the optimal fitness value of the individual;

S44, comparing the individual optimal fitness values of all particles with the global optimal fitness value of the group, if the individual optimal fitness value is larger, it is replaced by the global optimal fitness value;

S45, update the speed and position of the particle, and return to execute step S42;

S46. Compare the current number of iterations with the maximum number of iterations. When n＜N, the iteration continues; when n＞N, the iteration ends, or the iteration ends when the minimum error is reached. If both iteration end conditions are not met, return to step S42, otherwise exit the loop to obtain the global optimal fitness value. The parameters of the particles corresponding to the global optimal fitness value are the optimal CNN network parameters obtained by the particle swarm optimization.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. GIS switching-on and switching-off state current detection method based on CNN network, characterized in that the method comprises:

the method comprises the following steps: collecting the opening and closing states of a disconnecting link and a grounding switch and corresponding coupling capacitance current data sets in all states of GIS equipment;

step two: preprocessing the acquired current data, and constructing a training set of the network;

step three: constructing a CNN network, wherein the input of the CNN network is a training set, and the output of the CNN network is the opening and closing state of a disconnecting link and a grounding switch;

step four: optimizing CNN network parameters by using a particle swarm optimization algorithm to obtain optimal CNN network parameters;

step five: and setting the CNN network according to the optimal CNN network parameters to obtain the optimal CNN network, acquiring fault current data of a disconnecting link and a grounding switch of the GIS equipment in real time, preprocessing the fault current data, inputting the fault current data into the optimal CNN network, and outputting a GIS switching-on and switching-off state judgment result by using the optimal CNN network.

2. The method for detecting GIS opening and closing state current based on CNN network according to claim 1, wherein the step one includes:

when the GIS equipment is in a hot standby state, the switching conditions of each group of disconnecting link and grounding switch are disconnecting link switching-on and grounding switch switching-off, and the switching state Y of the grounding switch is recorded ^H ＝{y ₁ ^H ,y ₂ ^H …y _k ^H …y _K ^H }, simultaneously acquiring a current data set I ^H ＝{I ₁ ^H ,I ₂ ^H …I _k ^H …I _K ^H Therein of

Representing the kth earth-knife current in a hot standby state;

when the GIS equipment is in a cold standby state, the opening and closing conditions of each group of disconnecting link and earthing switch are disconnecting link and earthing switch, and the opening and closing states Y of the earthing switch and the disconnecting link are recorded ^C ＝{y ₁ ^C ,y ₂ ^C …y _k ^C …y _K ^C ，y ₁ ^C' ,y ₂ ^C' …y _m ^C' …y _M ^C' Simultaneously collecting a ground knife current data set I and a knife switch current data set I ^C ＝{I ₁ ^C ,I ₂ ^C …I _k ^C …I _K ^C ,I ₁ ^C' ,I ₂ ^C' …I _m ^C' …I _M ^C' Therein of

Indicates the kth ground switch current in the cold standby state, < > is present>

Indicating the m-th knife switch current in a cold standby state;

when the GIS equipment is in the maintenance state, the switching conditions of each group of disconnecting link and grounding switch are disconnecting link and grounding switch switching, and the switching states of the disconnecting links and the grounding switch are recorded, namely Y ^R ＝{y ₁ ^R ，y ₂ ^R …y _m ^R …y _M ^R Simultaneously acquiring a disconnecting link current data set I ^R ＝{I ₁ ^R ，I ₂ ^R …I _m ^R …I _M ^R Therein of

Indicating the mth boundary knife switch current in the maintenance state.

3. The method for detecting GIS opening and closing state current based on CNN network according to claim 1, wherein the second step comprises:

s21, collecting fault current data of the disconnecting link and the grounding switch, and performing wavelet transformation on a fault current signal to obtain a wavelet coefficient W _i (a, b) and a candidate frequency f _ins ；

S22, passing formula

A zero matrix is obtained, where (Δ a) _k ＝a _k -a _k-1 ，Δf＝f _k -f _k-1 ，a _k Is the scale factor of the k-th transient component, f _k Is the center frequency of the kth transient component, taking k e [0, n _a ]，f _s A signal sampling frequency; from f in S21 _ins According to the formula f _ins ＝2 ^kΔf ·f _s /n _a Calculating a k value, wherein &>

Judging k belongs to [0, n ] _a ]Whether it is satisfied, if so, then

Repeating the steps for preset times, making the signal thin and clearly showing on the time-frequency diagram,

passing through type

Reconstruct the signal and its respective component, wherein>

R _e The real part is taken. />

4. The method for detecting GIS opening/closing state current based on CNN network according to claim 3, wherein S21 includes:

a mother wavelet function psi (t) is selected such that its Fourier transform psi (omega) satisfies the admissibility conditions of the following wavelet functions

Selecting a scaling factor a and a shifting factor b, and performing scaling and shifting operations on the wavelet mother function psi (t) to obtain a wavelet basis function:

wavelet transformation is carried out on the current signal according to the following formula to obtain a wavelet coefficient W _i (a，b)

Wherein psi ^* (t) is the complex conjugate of ψ (t).

5. The method for detecting GIS switching-on/off state current based on CNN network according to claim 1, wherein CNN network comprises sequentially connected input layer, convolution layer, pooling layer, full-connection layer and output layer, the convolution layer is 3 layers, convolution kernel extracts current time-frequency spectrogram feature, the pooling layer is also 3, 3 convolution layers output data are respectively reduced in dimension; the output layer is composed of a Softmax module and a classification module.

6. The GIS switching-on/off state current detection method based on CNN network of claim 5, characterized in that the convolution layer passes through a formula

Obtain the current neuron>

In the formula, M _j For the total number of vectors input to neuron j, <' >>

Is a characteristic diagram of the previous layer of input neurons; k is a radical of formula _ij ¹ Weight of the convolution kernel; b is a mixture of _j Is the bias of the jth profile; σ is a ReLU activation function;

softmax module output

In the formula, x _j J is the j output of the full connection layer, k is the input number of the Softmax module, and k =2, then P1 and P2 are

P1, P2 represent probabilities of outputs of y =1 and y =0, respectively; inputting the probability values P1, P2 into a classification module, and calculating a cross-loss entropy function value L in the classification module, wherein

y _i Is an event tag value, y ₁ ＝1，y ₂ ＝0。

7. The method of claim 6, wherein the CNN network further comprises an inclusion-v 3 structure that uses filters of 1 × n and n × 1 sizes to replace the convolution of n × n in the CNN network.

8. The GIS switching-on/off state current detection method based on CNN network according to claim 6, characterized by that, the fourth step includes: setting a loss function L _loss When the loss value L of the CNN network is less than L _loss The network training is reliable, and the label value with a larger value in P1 and P2 is taken as the network output; otherwise, indicating that the network weight is not properly distributed, and redistributing the weight by adopting a particle swarm optimization algorithm to repeat the process until the network output condition is met to obtain the optimal CNN network parameter.

9. The GIS switching-on/off state current detection method based on CNN network of claim 8, wherein the weight redistribution by the particle swarm optimization comprises:

s41, after the error between the expected value and the actual value is calculated by the CNN network, the batch processing sample number batchsize, the training data discarding rate dropout and the convolution kernel number N of the CNN network are calculated by each particle _c Size of convolution kernel M _c And network initial bias b _j The parameters are used as hyper-parameters to be optimized, and the hyper-parameters to be optimized are used as particle dimensions;

s42, calculating the fitness value of each particle by taking the error between the expected value and the actual value of the CNN as a fitness function, and taking the optimal point of each particle in the n times of iterative updating as the individual optimal solution P _best Storing, comparing with the whole population particles, and selecting the global optimum value G _best Storing;

s43, comparing the fitness value of the particle with the individual optimal fitness value, and replacing the fitness value of the particle with the individual optimal fitness value if the fitness value of the particle is larger;

s44, comparing the individual optimal fitness values of all the particles with the group global optimal fitness value, and replacing the individual optimal fitness value with the global optimal fitness value if the individual optimal fitness value is larger;

s45, updating the speed and the position of the particles, and returning to execute the step S42;

s46, comparing the current iteration times with the maximum iteration times, and when N is less than N, continuing the iteration; and when N is larger than N, finishing iteration or finishing iteration when the minimum error is reached, returning to the step S42 if the two iteration finishing conditions are not met, otherwise, exiting the loop to obtain a global optimal fitness value, wherein each parameter of the particles corresponding to the global optimal fitness value is the optimal CNN network parameter obtained by particle swarm optimization.

10. GIS divide-shut brake state current detection device based on CNN network, its characterized in that, the device includes:

the data acquisition module is used for acquiring the opening and closing states of a disconnecting link and a grounding switch and corresponding coupling capacitance current data sets in all states of the GIS equipment;

the training set construction module is used for preprocessing the acquired current data and constructing a training set of the network;

the network construction module is used for constructing a CNN network, the input of the CNN network is a training set, and the output of the CNN network is the opening and closing states of a disconnecting link and a grounding switch;

the parameter optimization module is used for optimizing the CNN network parameters by using a particle swarm optimization algorithm to obtain optimal CNN network parameters;

and the result output module is used for setting the CNN network according to the optimal CNN network parameters to obtain the optimal CNN network, acquiring fault current data of a disconnecting link and a grounding switch of the GIS equipment in real time, preprocessing the fault current data, inputting the preprocessed data into the optimal CNN network, and outputting a GIS switching-on/off state judgment result by using the optimal CNN network.

Images