Electric shock safety event identification method based on wavelet multi-resolution analysis and neural network

Abstract

The invention discloses an electric shock safety event identification method based on wavelet multi-resolution analysis and a neural network, and relates to the technical field of electric power safety detection and intelligent diagnosis. Analyzing electric shock residual current signal wavelets; identifying an electric shock event based on the GRU neural network; and performing electric shock diagnosis based on wavelet analysis and a GRU neural network. The invention provides an electric shock safety event identification method based on wavelet multi-resolution analysis and a neural network. A time period of occurrence of an identified electric shock event is given based on residual current signal characteristics. The residual current protection device has the advantages that the problems of maloperation and operation refusal generally existing in the existing residual current protection device are solved, the electric shock event can be accurately distinguished from other events influencing the change of the residual current value, the intelligent level of an electricity safety system is improved, and the important significance is achieved for preventing the electric shock electricity safety accident.

Description

Technical Field

[0001] The present invention relates to the technical field of power safety detection and intelligent diagnosis, and in particular to a method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network. Background Art

[0002] With the increase in electricity demand and the intensification of leakage problems caused by line damage and electrical equipment failure in the distribution network, the incidence of fire and electric shock accidents has increased significantly, which has made power safety issues the focus of social attention. Residual current protection devices are widely used in low-voltage distribution networks to prevent such accidents. The core working principle of these devices is to determine whether there is a risk of electric shock in the line by judging whether the total residual current effective value reaches the preset rated action threshold.

[0003] However, the residual current protection devices currently used generally have the phenomenon of false operation and refusal to operate. They can only identify whether the current exceeds the set threshold, but cannot effectively distinguish between the increase of residual current caused by insulation failure of lines and electrical appliances, extreme weather, transient current surge caused by sudden large load, and increase or decrease of residual current caused by electric shock accidents. They are also unable to accurately determine the type and location of electric shock accidents. In order to effectively protect people from electric shock accidents, it is necessary to analyze the transient characteristics of electric shock accidents, extract the identification criteria of electric shock accidents, and develop new residual current protection devices based on electric shock current action. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing electric shock signal detection method has problems such as insufficient feature extraction and classification accuracy, insufficient real-time and timeliness of electric shock diagnosis, insufficient detection sensitivity and applicability, and lack of accurate classification capabilities for complex electric shock scenarios.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network, including obtaining an electric shock residual current signal; analyzing the wavelet of the electric shock residual current signal; identifying electric shock events based on a GRU neural network; and performing electric shock diagnosis based on wavelet analysis and a GRU neural network.

[0007] As a preferred solution of the electric shock safety event identification method based on wavelet multi-resolution analysis and neural network described in the present invention, wherein: the acquisition of the electric shock residual current signal includes collecting the residual current signal in the circuit;

[0008] Pre-process the residual current signal of electric shock, downsample it, and reduce the frequency of the current waveform to 10kHz;

[0009] The maximum overlap discrete wavelet transform was used to further reduce the noise of the data;

[0010] Intercept the residual current signal of electric shock.

[0011] As a preferred solution of the electric shock safety event identification method based on wavelet multi-resolution analysis and neural network described in the present invention, wherein: the wavelet for analyzing the electric shock residual current signal includes selecting the 11th order wavelet in the Daubechies series wavelet when performing wavelet multi-resolution analysis on the intercepted residual current signal;

[0012] The sampling frequency of the residual current signal of electric shock is 10kHz after processing, the main frequency component frequency of the residual current is 50Hz, and the residual current is decomposed by k-layer wavelet;

[0013] Perform wavelet multi-resolution analysis on the residual current signal to output a high-frequency coefficient distribution diagram, perform statistics on the maximum and minimum amplitudes of the signals in each period of each layer of wavelet high-frequency coefficient distribution, and output the difference to reflect the degree of fluctuation and describe the degree of amplitude mutation;

[0014] From the properties of multi-resolution analysis, we can get:

[0015]

[0016] Among them, {V j} is the scale space, {W i} is the wavelet space;

[0017] Assume that I is the actual signal to be processed, and the measured signal Ij is I in the scale space {V j}, the multi-resolution representation of Ij is:

[0018]

[0019] d j,k = <I(t),Ψ j,k (t)>

[0020] Assume that the function I(t) moves to the wavelet space {W j The signal obtained after projection is d j (t)∈W j ,but:

[0021]

[0022] Among them, j represents the scale transformation factor, k represents the time translation factor, represents the scale factor signal at different scales, ψ j,k (t) represents the scale factor signal at different scales The orthogonal signal is a wavelet signal, I(t) represents the source signal, Ψ j,k (t) represent the rough information and fine information of the source signal, c j,k represents the expansion coefficient, d j,k Indicates the distribution state of the corresponding high-frequency components;

[0023] The source signal is represented as:

[0024]

[0025] When the scaling function and the wavelet function form a normalized orthogonal basis, the open coefficient of the signal can be calculated by the inner product of the source signal and the corresponding function signal, expressed as:

[0026]

[0027] As a preferred solution of the electric shock safety event identification method based on wavelet multi-resolution analysis and neural network described in the present invention, wherein: the analysis of the electric shock residual current signal wavelet includes describing the electric shock characteristics by using the ratio of the cumulative sum of the amplitude mutation amount of the corresponding stage of the dimensionally normalized wavelet high-frequency signal;

[0028] The j-layer high-frequency signal d of the residual current signal j (t) After dimension normalization, we get d j ′(t) is expressed as:

[0029]

[0030] The sampling signals of three consecutive cycles before and after the electric shock are decomposed by multi-layer wavelet, and then the d j ′(t) is divided into three stages: T-1 before electric shock, T0 during electric shock, and T1 after electric shock. The cumulative sum of the normalized amplitude mutation output in each stage is expressed as:

[0031]

[0032] Among them, N is the number of sampling points included in each stage;

[0033] The cumulative sum of normalized amplitude mutations in the three stages of T-1, T0, and T1 is expressed as ΔD j-1 , ΔD j0 , ΔD j1 , and use the characteristic parameter η of the j-layer electric shock accident j-1 , η j1 Define ΔD j0 With ΔD j-1 , ΔD j1 The ratio is expressed as:

[0034]

[0035] The wavelet high-frequency coefficient distribution of each layer of the residual current signal is extracted and output according to the electric shock characteristics, and the output wave high-frequency signal amplitude mutation amount is accumulated and ΔD j The characteristic parameter η of electric shock with multi-layer wavelet j-1 , η j1 ;

[0036] If the multi-layer wavelet high-frequency signals all have the T0 stage amplitude mutation amount cumulative sum is the highest among the three stages, that is, ΔD j0 >D j-1 And ΔD j0 >ΔD j1 , and the ratio of the amplitude mutation and the extracted characteristic parameter η j If both are greater than 1, it is determined that the analyzed residual current signal has the characteristics of electric shock to a living organism.

[0037] As a preferred solution of the electric shock safety event recognition method based on wavelet multi-resolution analysis and neural network described in the present invention, wherein: the electric shock event recognition based on GRU neural network includes the forward propagation of GRU represented as:

[0038] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0039] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0040]

[0041] Among them, r t 、z t 、h t 、h t represents the reset gate, update gate, candidate hidden state and current hidden state; b represents weight and bias; σ(·) and tanh(·) represent the sigmoid function and tanh function. The sigmoid function limits the data to the range of [0,1], and the tanh function limits the data to the range of [-1,1].

[0042] Reset Gate t Controls the previous state history information to be written into the current candidate set h tThe reset gate is proportional to the information written in the previous state. The information of the previous moment and the current moment are right-multiplied by the weight matrix and added. The data added with the bias information is sent to the reset gate, that is, multiplied by the sigmoid function, and the output value is between [0,1];

[0043] Update gate z t It is used to control the degree to which the state information of the previous moment is brought into the current state. The value of the update gate is proportional to the state information brought in at the previous moment. Similar to the data processing of the reset gate, the information of the previous moment and the current moment are respectively right-multiplied by the weight matrix and added together. The data after adding the bias information is sent to the update gate, which is multiplied by the sigmoid function. The output value is between [0,1].

[0044] As a preferred solution of the electric shock safety event identification method based on wavelet multi-resolution analysis and neural network described in the present invention, the electric shock event identification based on GRU neural network includes using the mean absolute error value to quantitatively evaluate the model prediction accuracy expressed as:

[0045]

[0046] Among them, y i ,y i Represent the predicted value and the true value respectively, and n represents the number of tests;

[0047] GRU neural network model optimization includes data set partitioning and model optimization;

[0048] Divide the data set; including training set, validation set and test set;

[0049] The training set is used to train the model parameters and accounts for 70% of the total data set;

[0050] The validation set is used to adjust the model hyperparameters and accounts for 15% of the total dataset;

[0051] The test set is used to finally evaluate the model performance, accounting for 15% of the total dataset;

[0052] Model optimization involves calculating the MAE of the network layers to determine the model accuracy corresponding to different numbers of network layers.

[0053] As a preferred solution of the electric shock safety event identification method based on wavelet multi-resolution analysis and neural network described in the present invention, wherein: the electric shock diagnosis based on wavelet analysis and GRU neural network includes that when no electric shock accident occurs, there is no sudden change in the residual current of the power grid, so the residual current signal characteristics when no electric shock accident occurs are divided into category 1;

[0054] Considering the randomness of electric shock accidents in actual situations, electric shock accidents occur in the T-1, T0, and T1 cycles. When the electric shock occurs in the T-1 cycle, the corresponding signal feature is classified as category 2. When the electric shock occurs in the T0 cycle, the extracted signal feature is classified as category 3. When the electric shock occurs in the T1 cycle, the extracted signal feature is classified as category 4.

[0055] Randomly select the test electric shock signals, intercept 600 points of each signal according to the classification, and extract the electric shock features according to the biological body electric shock feature determination method. The signal interception method includes:

[0056] Category 1 residual current signal, intercepting 600 continuous points of normal residual current signal;

[0057] The second type of residual current signal, that is, the electric shock occurs at any sampling point within the T-1 period, and the corresponding residual current signal sampling point is recorded as t1(i);

[0058] The third type of residual current signal, that is, the electric shock occurs at any sampling point within the T0 period, and the corresponding residual current signal sampling point is recorded as t2(i);

[0059] The fourth type of residual current signal, that is, the electric shock occurs at any sampling point within the T1 period, and the corresponding residual current signal sampling point is recorded as t3(i);

[0060] Building an electric shock recognition model includes:

[0061] The residual current transformer is connected to the three-phase power supply of A, B and C respectively, and the residual current signal is obtained by using pork as the electric shock body;

[0062] The acquired data were downsampled and denoised using MODWT, and 600 sampling points were intercepted and divided into three stages: T-1, T0, and T1 for subsequent analysis;

[0063] The 11th-order wavelet in the Daubechies series of wavelets is selected for multi-layer wavelet decomposition, and the multi-layer high-frequency signals of the residual current signal are dimensionally normalized.

[0064] Calculate the cumulative sum of the normalized amplitude mutation of the wavelet high-frequency signal at each layer of the three stages T-1, T0, and T1 and the characteristic parameters of the electric shock accident, and output the characteristic data for identifying the electric shock of the organism;

[0065] Build a GRU electric shock diagnosis model, and divide the feature data set into training set, test set and validation set as model input for training and validation;

[0066] Using the trained model, input any set of data to diagnose the type of electric shock and determine whether electric shock occurs and whether the electric shock occurs in the T-1, T0 or T1 stage.

[0067] Another object of the present invention is to provide an electric shock safety event identification system based on wavelet multi-resolution analysis and neural network, which can identify electric shock events based on GRU neural network, thereby solving the problem of insufficient real-time and timeliness of electric shock diagnosis in current electric shock signal detection methods.

[0068] As a preferred solution of the electric shock safety event identification system based on wavelet multi-resolution analysis and neural network described in the present invention, it includes: a current signal acquisition module, a wavelet analysis module, an electric shock event identification module, and an electric shock diagnosis module; the current signal acquisition module is used to obtain the residual current signal of electric shock; the wavelet analysis module is used for wavelet analysis of the residual current signal of electric shock; the electric shock event identification module is used for electric shock event identification based on GRU neural network; the electric shock diagnosis module is used for electric shock diagnosis based on wavelet analysis and GRU neural network.

[0069] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of a method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network.

[0070] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural networks.

[0071] Beneficial effects of the present invention: The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network provided by the present invention obtains the residual current signal of the three-phase power supply through the residual current transformer, and uses the analytical ability of wavelet multi-resolution analysis in different scale windows to extract the amplitude change characteristics of the residual current signal in each frequency band, and then mathematically characterizes the amplitude surge characteristics of the multi-layer wavelet high-frequency band, and provides a criterion for the analysis of residual current signals that hide the electric shock characteristics of biological bodies. The GRU neural network is used to learn and identify the category of electric shock events, and the time period of the identified electric shock event is given based on the residual current signal characteristics. It solves the common problems of false operation and refusal to operate in the residual current protection devices currently used, and can accurately distinguish electric shock events from other events that affect the change of residual current values, thereby improving the intelligence level of the power safety system, which is of great significance for preventing electric shock safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0073] Figure 1 This is the overall flowchart of a method for identifying electric shock safety events based on wavelet multi - resolution analysis and neural network provided for the first embodiment of the present invention.

[0074] Figure 2 This is the overall flowchart of a system for identifying electric shock safety events based on wavelet multi - resolution analysis and neural network provided for the third embodiment of the present invention. Detailed implementation manners

[0075] To make the above - mentioned objects, features and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0076] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, provides a method for identifying electric shock safety events based on wavelet multi - resolution analysis and neural network, including:

[0077] S1: Acquisition of electric shock residual current signal.

[0078] Furthermore, collect the residual current signal in the circuit. Take the three - phase mains power passing through the isolating switch as the input three - phase power to ensure the safety of the signal acquisition process. Take three resistors with relatively large resistance values (such as 200Ω) and form loops with the A, B, and C - phase power supplies through the residual current transformer respectively. Take another resistor with a relatively large resistance value (such as 280Ω) and connect it in parallel with the residual current transformer. Obtain the residual current signal through the oscilloscope connected in parallel with the resistor. The residual current transformer measures the total three - phase residual current. The electric shock body is connected in parallel to the A, B, and C phases to obtain electric shock signals of different phases respectively. To ensure safety, pork can be selected as the electric shock body to obtain the residual current signal.

[0079] Pre - processing of the electric shock residual current signal. The sampling frequency of the oscilloscope is relatively high, and the collected original residual current signal is affected by noise interference. Therefore, it is necessary to pre - process the original residual current signal. First, perform down - sampling processing to reduce the frequency of the current waveform to 10kHz; then use the maximum overlap discrete wavelet transform (MODWT) to further denoise the data. The result of denoising is to preserve the low - frequency components of the signal.

[0080] Considering the real-time requirements of electric shock accident identification, it is necessary to intercept the residual current signal for subsequent feature identification. In order to extract the relevant complete information before and after the electric shock as much as possible, 600 sampling points are selected for wavelet analysis. The signal is divided into three stages according to the number of sampling points: T-1 (before electric shock), T0 (during electric shock), and T1 (after electric shock). Each stage contains 200 sampling signals.

[0081] S2: Analyze the electric shock residual current signal wavelet.

[0082] Furthermore, for the intercepted residual current signal, the 11th order wavelet in the Daubechies series wavelet is selected for wavelet multi-resolution analysis. The sampling frequency of the electric shock residual current signal is 10kHz after processing, and the main frequency component frequency of the residual current is 50Hz. The residual current can be decomposed by k layers (k=3,4,5,...) of wavelets.

[0083] The residual current signal is subjected to wavelet multi-resolution analysis to obtain the high-frequency coefficient distribution diagram. The maximum and minimum amplitudes of the signals in each period of the wavelet high-frequency coefficient distribution of each layer are counted, and the difference is calculated to reflect the degree of fluctuation, thereby describing the degree of amplitude mutation.

[0084] Wavelet multi-resolution analysis can analyze the signal step by step by transforming the scale factor. When the scale is large, it provides a wider time window, the analyzed frequency is low, and an overview analysis can be performed; when the scale is small, the frequency window is small, and a detailed analysis can be performed. Therefore, by smoothly decreasing the scale factor, the signal can be analyzed from coarse to fine at its corresponding scale.

[0085] Multi-resolution analysis is to use different resolutions to gradually approximate the function to be analyzed. From the properties of multi-resolution analysis, we can get:

[0086]

[0087] Among them, {V j} is the scale space, {W i} is the wavelet space.

[0088] It should be noted that, assuming that I is the actual signal to be processed, the measured signal Ij is I in the scale space {V j}, the multi-resolution representation of Ij can be expressed as follows:

[0089] I j =I j-1 +ω j-1 =I j-2 +ω j-2 +ω j-1 =…

[0090] n j =Ij0 +ω j0 +ω j0+1 +…+ω j-1

[0091] in,

[0092]

[0093] d j,k = <I(t),Ψ j,k (t)

[0094] Assume that the function I(t) moves to the wavelet space {W j The signal obtained after projection is d j (t)∈W j ,but:

[0095]

[0096] Let j be the scale transformation factor and k be the time translation factor. Then the scale factor signal at different scales is The orthogonal signal is the wavelet signal ψ j,k (t), then the multi-resolution analysis of the source signal I(t) can expand the signal into a scale signal and a wavelet signal, as shown in formula (6). ψ j,k (t) corresponds to the coarse information and fine information of the source signal, respectively, and the expansion coefficient c j,k The distribution of low-frequency components of the corresponding source signal, d j,k Corresponding to the distribution state of high-frequency components, the expansion coefficient is the discrete wavelet transform of the signal.

[0097] The source signal is represented as:

[0098]

[0099] When the scaling function and the wavelet function form a normalized orthogonal basis, the expansion coefficient of the signal can be calculated by the inner product of the source signal and the corresponding function signal:

[0100]

[0101] Furthermore, the electric shock feature extraction uses the ratio of the cumulative sum of the amplitude mutations of the corresponding stage of the dimensionally normalized wavelet high-frequency signal to describe the electric shock feature. j (t) After dimension normalization, we get d j '(t), the calculation formula is:

[0102]

[0103] The sampling signals of three consecutive cycles before and after the electric shock are decomposed by multi-layer (j layers, j = 1, 2, ...) wavelet decomposition, and then the d j '(t) is divided into three stages: T-1 (before electric shock), T0 (during electric shock), and T1 (after electric shock). The cumulative sum of normalized amplitude mutations in each stage is calculated according to the following formula:

[0104]

[0105] Where N is the number of sampling points included in each stage.

[0106] Finally, the cumulative sum of the normalized amplitude mutations in the three stages of T-1, T0, and T1 is expressed as ΔD j-1 , ΔD j0 , ΔD j1 , and use the characteristic parameter η of the j-layer electric shock accident j-1 , η j1 Define ΔD j0 With ΔD j-1 , ΔD j1 The ratio is as shown in the following formula:

[0107]

[0108] It should be noted that the wavelet high-frequency coefficient distribution of each layer of the calculated residual current signal is calculated according to equations (6) to (8), and the cumulative sum of the amplitude mutation of the wavelet high-frequency signal ΔD is obtained. j The characteristic parameter η of electric shock with multi-layer wavelet j-1 , η j1 If the multi-layer wavelet high-frequency signals all have the cumulative sum of the amplitude mutation in the T0 stage being the highest among the three stages, that is, ΔD j0 >ΔD j-1 And ΔD j0 >ΔD j1 ; and the ratio of the extracted characteristic parameter amplitude mutation and η j If both are greater than 1, it is determined that the analyzed residual current signal has the characteristics of electric shock to a living organism.

[0109] S3: Identify electric shock events based on GRU neural network.

[0110] Furthermore, the GRU neural network, the gated recurrent unit (GRU) is a type of recurrent neural network (RNN), and a variant of the long short-term memory network (LSTM). Although LSTM can solve the gradient vanishing and gradient exploding problems caused by long-term dependence in recurrent neural networks, three gate functions are introduced in LSTM: input gate, forget gate, and output gate to control input values, memory values, and output values. There are many parameters and it is difficult to train. The GRU model has only two gates, the reset gate and the update gate. The gate structure can effectively solve the gradient vanishing and gradient exploding problems of recurrent neural networks, and it converges faster than the LSTM network. The GRU network connects repeated nuclei into a chain structure, and performs calculations and data propagation through the reset gate and update gate in each nucleus.

[0111] The forward propagation formula of GRU is as follows:

[0112] r t =σ(W r ·[h t-1 ,x t ]·+b r )

[0113] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0114]

[0115] Among them, r t 、z t 、h t 、h t Represents reset gate, update gate, candidate hidden state and current hidden state; b represents weight and bias; σ(·) and tanh(·) represent sigmoid function and tanh function. The sigmoid function limits the data to the range of [0,1], and the tanh function limits the data to the range of [-1,1]. The definitions are as follows:

[0116]

[0117] It should be noted that the reset gate r t Controls how much historical information of the previous state is written into the current candidate set h tThe smaller the reset gate is, the less information of the previous state is written. The information of the previous moment and the current moment are right-multiplied by the weight matrix and added together. The data after adding the bias information is sent to the reset gate, that is, multiplied by the sigmoid function, and the resulting value is between [0,1].

[0118] Update gate z t It is used to control the degree to which the state information of the previous moment is brought into the current state. The larger the value of the update gate, the more state information of the previous moment is brought into the current state. Just like the data processing of the reset gate, the information of the previous moment and the current moment are respectively multiplied by the weight matrix on the right and then added. The data after adding the bias information is sent to the update gate, that is, multiplied by the sigmoid function, and the value obtained is between [0,1]. It's just that the values ​​and uses of the weight matrix and bias are different twice.

[0119] Furthermore, the GRU neural network model evaluation uses the mean absolute error (MAE) to quantitatively evaluate the model prediction accuracy. The smaller the MAE, the higher the model accuracy. The MAE calculation formula is as follows:

[0120]

[0121] Among them, y i ,y i Represent the predicted value and the true value respectively, and n represents the number of tests.

[0122] GRU neural network model optimization, data set division, and reasonable data set division ensure that the model has sufficient data support in each stage of training, tuning, and testing, thereby improving the final detection accuracy and generalization ability.

[0123] Training set: used to train model parameters, accounting for 70% of the total data set.

[0124] Validation set: used to tune model hyperparameters, accounting for 15% of the total dataset.

[0125] Test set: used for final evaluation of model performance, accounting for 15% of the total dataset

[0126] Model optimization, the GRU neural network model consists of an input layer, a GRU network layer, a fully connected layer, and an output layer. In theory, increasing the number of network layers can enhance the fitting ability and improve the effect. However, in practice, too many network layers will lead to overfitting and increased training difficulty, making it difficult for the model to converge.

[0127] By calculating the MAE corresponding to the network layer number, determine whether the models corresponding to different network layer numbers have higher accuracy. Fix the number of neurons (e.g., set to 64), change the different network layer numbers, calculate the MAE under different network layer numbers, and set the network layer number when the MAE is the smallest as the optimal network layer number of the GRU neural network model.

[0128] The number of neurons will also affect the performance of the model. A small number of neurons in the GRU layer will cause underfitting, while too many neurons will cause overfitting problems. Therefore, it is important to select a suitable number of neurons. Fix the GRU network layer number (e.g., set to 2), change the different numbers of neurons, calculate the MAE under different numbers of neurons, and set the number of neurons when the MAE is the smallest as the optimal number of neuron parameters of the GRU neural network model.

[0129] The learning rate is used to determine the step size of each iteration and belongs to the tuning parameter of the optimization algorithm. A too small learning rate will prolong the training period, while a too large learning rate will hinder convergence. Setting a suitable learning rate can make the loss function converge to the minimum value, that is, make the gap between the predicted value and the true value reach the minimum. The smaller the loss value, the better the prediction effect of the model.

[0130] S4: Conduct electric shock diagnosis based on wavelet analysis and GRU neural network.

[0131] Furthermore, for the classification of electric shock categories, when there is no electric shock accident, there is no sudden change in the residual current of the power grid. Therefore, the signal characteristics of the residual current when there is no electric shock accident are classified as category 1. Considering the randomness of electric shock accidents in actual situations, electric shock accidents may occur in any cycle of T-1, T0, and T1. When the electric shock occurs in the T-1 cycle, the corresponding signal characteristics are classified as category 2. When the electric shock occurs in the T0 cycle, the extracted signal characteristics are classified as category 3. When the electric shock occurs in the T1 cycle, the extracted signal characteristics are classified as category 4.

[0132] Randomly select several groups (it is recommended to be >100 groups) of test electric shock signals, and intercept 600 points of signals for each group according to the above classification. Extract the electric shock characteristics according to the method for determining the characteristics of biological electric shock proposed in Section 5.2. The signal interception method is as follows:

[0133] For the first type of residual current signal, intercept 600 consecutive points of normal residual current signals.

[0134] For the second type of residual current signal, that is, when the electric shock occurs at any sampling point within the T-1 cycle, the corresponding sampling point of the residual current signal is denoted as t1(i).

[0135] For the third type of residual current signal, that is, when the electric shock occurs at any sampling point within the T0 cycle, the corresponding sampling point of the residual current signal is denoted as t2(i).

[0136] Category 4 residual current signal, that is, electric shock occurs in the T1 period

[0137] It should be noted that the process of establishing the electric shock identification model is that in the first step, the residual current transformer is connected to the three-phase power supply A, B, and C respectively, and the pork is used as the electric shock body to obtain the residual current signal;

[0138] In the second step, the acquired data were downsampled and denoised using MODWT, and 600 sampling points were intercepted and divided into three stages: T-1, T0, and T1 for subsequent analysis;

[0139] The third step is to select the 11th order wavelet in the Daubechies series of wavelets for multi-layer wavelet decomposition, and perform dimension normalization on the multi-layer high-frequency signals of the residual current signal;

[0140] The fourth step is to calculate the cumulative sum of the normalized amplitude mutation of the wavelet high-frequency signal of each layer in the three stages T-1, T0, and T1 and the characteristic parameters of the electric shock accident, and obtain the characteristic data for identifying the electric shock of the organism;

[0141] The fifth step is to build a GRU electric shock diagnosis model, and divide the feature data set into a training set, a test set, and a validation set as model input for training and validation;

[0142] Finally, the trained model is used to input any set of data to diagnose the type of electric shock and determine whether electric shock occurs and whether the electric shock occurs in the T-1, T0 or T1 stage.

[0143] Example 2, an embodiment of the present invention, provides a method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0144] First, the first step is to obtain the residual current signal IM; the second step is to obtain the residual current signal IM' after downsampling and MODWT denoising of the acquired data, and intercept 600 sampling points, which are evenly divided into three stages of T-1, T0, and T1 for subsequent analysis; the third step is to select the 11th-order wavelet in the Daubechies series of wavelets for multi-layer wavelet decomposition, and perform dimensional normalization on the multi-layer decomposition high-frequency signal coefficients of the residual current signal; the fourth step is to use the dimensional normalization results calculated in the third step to calculate the cumulative sum of the normalized amplitude mutation of the high-frequency signal coefficients of each layer of wavelets in the three stages of T-1, T0, and T1 and the characteristic parameters of electric shock accidents, and obtain the mathematical representation data for identifying electric shock of biological bodies; the fifth step is to build a GRU electric shock diagnosis model, and divide the feature data set into training set, test set, and verification set as model input for training and verification; finally, the trained model is used to input the residual current data to diagnose the electric shock category and determine whether electric shock occurs and the time period of the electric shock event.

[0145] In the specific implementation process, for the residual current signal collected during electric shock, when performing MODWT noise reduction processing on the residual current raw data in the second step, the following steps can be adopted:

[0146] (1) Perform MODWT decomposition on the original signal to obtain wavelet coefficients of different scales. You can use the "sym4" wavelet in the MATLAB operating environment and directly use the modwt function to decompose the original signal into 4 layers.

[0147] (2) According to the characteristics of the signal and the statistical characteristics of the noise, select a suitable threshold function to perform threshold processing on the wavelet coefficients. In MATLAB, wavelet signal denoising can be achieved by calling the wdenoise function.

[0148] (3) Perform an inverse transform on the wavelet coefficients after threshold processing to obtain the denoised signal. In MATLAB, the built-in wavelet analysis function imodwt can be used to reconstruct the signal after wavelet threshold denoising.

[0149] In the second step, the residual current signal IM' is intercepted to 600 sampling points and divided into three stages of T-1, T0 and T1 according to the time average for subsequent analysis.

[0150] The third step is to perform multi-layer wavelet decomposition on the residual current signal IM' (the decomposition methods of the three stages T-1, T0, and T1 are the same), select the 11th order wavelet (referred to as db11 wavelet) in the Daubechies series of wavelets, and perform 7-layer wavelet decomposition to decompose the original signal into I = a 7 +d 7 +d 6 +d 5 +d4 +d 3 +d 2 +d 1 From formula (3), we can get a 7 The value of d can be obtained from equations (4) and (5) 1 d 2 ……d 7 The value of .

[0151] is the scale function of the db11 wavelet, and ψ(t) is the db11 wavelet function. In this way, we obtain a series of wavelet decomposition signals of IM' on scale j (j = 1, 2, ..., 7), among which the wavelet high-frequency coefficients (i.e., d j,k ) constitutes the discrete wavelet transform matrix of the residual current signal IM'. In order to better compare the mutation degree of the wavelet high-frequency coefficients, it is necessary to normalize them. The processing method is shown in formula (9). After normalization, the wavelet high-frequency coefficients at different scales j (j = 1, 2, ..., 7) are recorded as DIj.

[0152] The fourth step is to quantify the normalized wavelet high-frequency coefficients DIj of different scales j, and define the amplitude mutation accumulation and ΔD respectively. j and characteristic parameter η of electric shock accident j The calculation methods are detailed in formula (10) and formula (11). The cumulative sum of the normalized amplitude mutations in the three stages of T-1, T0, and T1 is expressed as ΔD j-1 , ΔD j0 , ΔD j1 , and use the characteristic parameter η of the j-layer electric shock accident j-1 , η j1 Define ΔD j0 With ΔD j-1 , ΔD j1 The following are four situations of when electric shock occurs:

[0153] Category 1: The electric shock event occurs outside the three stages of T-1, T0, and T1. At this time, ΔD j-1 , ΔD j0 , ΔD j1 There is little difference in the values ​​between them, and η j-1 , η j1 Both are approximately equal to 1.

[0154] Category 2: The electric shock event occurs in the T-1 stage. At this time, max(ΔD j-1 ,ΔD j0 ,ΔD j1 )=ΔD j-1 And η j-1 <1.

[0155] Category 3: The electric shock event occurs at T0. At this time, max(ΔD j-1 ,ΔD j0 ,ΔD j1 )=ΔD j0 And η j-1 >1,η j >1.

[0156] Category 4: The electric shock event occurs at stage T1. At this time, max(ΔD j-1 ,ΔD j0 ,ΔD j1 )=ΔD j1 And η j1 <1.

[0157] By accumulating the amplitude mutation of the normalized wavelet high-frequency coefficient DIj at three stages T-1, T0, and T1, ΔD j and characteristic parameter η of electric shock accident j By analyzing, the time period when the electric shock incident occurred can be determined.

[0158] In the fifth step, when building the GRU electric shock diagnosis model, it is necessary to determine the GRU neural network parameters, including the number of GRU network layers, the number of neurons, and the iteration step size, and use the mean absolute error (MAE) to judge the model accuracy. The learning rate can be used to determine the step size of each iteration and is a tuning parameter of the optimization algorithm. Setting an appropriate learning rate can make the loss function converge to the minimum value, that is, minimize the difference between the predicted value and the true value. The smaller the loss value, the better the prediction effect of the model.

[0159] The number of iterations is set to 30, the number of neurons is fixed to 64, the number of network layers is changed, and the MAE results are calculated as shown in Table 1.

[0160] Table 1 GRU network layer MAE

[0161] Current number of GRU network layers Number of neurons MAE 1 64 0.2000 2 64 0.1333 3 64 0.2333 4 64 0.2000

[0162] The number of iterations is set to 30, the number of GRU network layers is fixed to 2, the number of neurons is changed, and the MAE results are calculated as shown in Table 2.

[0163] Table 2. MAE of neuron number

[0164] Current number of GRU network layers Number of neurons MAE 2 8 0.4333 2 16 0.4000 2 32 0.2666 2 64 0.2000 2 128 0.3000

[0165] Experimental tests with different learning rates can obtain different loss values, and the results are shown in Table 3.

[0166] Table 3 Learning rate and loss value

[0167] Learning Rate 0.0008 0.0009 0.0010 0.0020 0.0030 Loss value 0.2655 0.2412 0.2403 0.2739 0.2483

[0168] In summary, when building the GRU electric shock diagnosis model, two GRU network layers were selected, each with 64 neurons, and the learning rate was 0.0010. During the training process, as the number of iterations of the GRU model increased, the test loss function rate continued to decrease until convergence. At the same time, the test accuracy continued to increase and approached 100%, indicating that the model can learn the input feature data well.

[0169] After the GRU electric shock diagnosis model is trained, the residual current data can be input for electric shock category diagnosis to determine whether electric shock occurred in the time period of the data and the time period when the electric shock event occurred. In addition, the model can also be further diagnosed. For the case where the electric shock event occurs in the Ti time period, the residual current in the Ti time period is continuously diagnosed using the model until it converges to the t(i) moment, at which time the specific moment when the electric shock event occurred can be obtained.

[0170] Example 3, reference Figure 2 , which is an embodiment of the present invention, provides an electric shock safety event identification system based on wavelet multi-resolution analysis and neural network, including a current signal acquisition module, a wavelet analysis module, an electric shock event identification module, and an electric shock diagnosis module.

[0171] The current signal acquisition module is used to obtain the residual current signal of electric shock, the wavelet analysis module is used for wavelet analysis of the residual current signal of electric shock, the electric shock event recognition module is used for electric shock event recognition based on the GRU neural network, and the electric shock diagnosis module is used for electric shock diagnosis based on wavelet analysis and the GRU neural network.

[0172] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0173] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0174] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0175] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0176] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network, characterized in that: include: Obtain electric shock residual current signal; Analyze the wavelet of residual current signal of electric shock; Identify electric shock events based on GRU neural network; Electric shock diagnosis based on wavelet analysis and GRU neural network.

2. The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in claim 1 is characterized in that: The obtaining of the electric shock residual current signal comprises collecting the residual current signal in the circuit; Pre-process the residual current signal of electric shock, downsample it, and reduce the frequency of the current waveform to 10kHz; The maximum overlap discrete wavelet transform was used to further reduce the noise of the data; Intercept the residual current signal of electric shock.

3. The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in claim 2 is characterized in that: The wavelet analysis of the electric shock residual current signal includes selecting the 11th order wavelet in the Daubechies series wavelet when performing wavelet multi-resolution analysis on the intercepted residual current signal; The sampling frequency of the residual current signal of electric shock is 10kHz after processing, the main frequency component frequency of the residual current is 50Hz, and the residual current is decomposed by k-layer wavelet; Perform wavelet multi-resolution analysis on the residual current signal to output a high-frequency coefficient distribution diagram, perform statistics on the maximum and minimum amplitudes of the signals in each period of each layer of wavelet high-frequency coefficient distribution, and output the difference to reflect the degree of fluctuation and describe the degree of amplitude mutation; From the properties of multi-resolution analysis, we can get: Among them, {V j } is the scale space, {W i } is the wavelet space; Assume that I is the actual signal to be processed, and the measured signal Ij is I in the scale space {V j }, the multi-resolution representation of Ij is: I j =I j-1 +oh j-1 =I j-2 +oh j-2 +oh j-1 =… n j =I j0 +oh j0 +oh j0+1 +…+oh j-1 d j,k =<I(t),Ψ j,k (t)> Assume that the function I(t) moves to the wavelet space {W j The signal obtained after projection is d j (t)∈W j ,but: Among them, j represents the scale transformation factor, k represents the time translation factor, represents the scale factor signal at different scales, Represents the scale factor signal at different scales The orthogonal signal is a wavelet signal, I(t) represents the source signal, Ψ j,k (t) represent the rough information and fine information of the source signal, c j,k represents the expansion coefficient, d j,k Indicates the distribution state of the corresponding high-frequency components; The source signal is represented as: When the scaling function and the wavelet function form a normalized orthogonal basis, the open coefficient of the signal can be calculated by the inner product of the source signal and the corresponding function signal, expressed as:

4. The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in claim 3 is characterized in that: The wavelet analysis of the residual current signal of electric shock includes describing the electric shock characteristics by using the ratio of the cumulative sum of the amplitude mutation amount of the corresponding stage of the dimensionally normalized wavelet high-frequency signal; The j-layer high-frequency signal d of the residual current signal j (t) After dimension normalization, we get d j ′(t) is expressed as: The sampling signals of three consecutive cycles before and after the electric shock are decomposed by multi-layer wavelet, and then the d j ′(t) is divided into three stages: T-1 before electric shock, T0 during electric shock, and T1 after electric shock. The cumulative sum of the normalized amplitude mutation output in each stage is expressed as: Among them, N is the number of sampling points included in each stage; The cumulative sum of normalized amplitude mutations in the three stages of T-1, T0, and T1 is expressed as ΔD j-1 , ΔD j0 , ΔD j1 , and use the characteristic parameter η of the j-layer electric shock accident j-1 , η j1 Define ΔD j0 With ΔD j-1 , ΔD j1 The ratio is expressed as: The wavelet high-frequency coefficient distribution of each layer of the residual current signal is extracted and output according to the electric shock characteristics, and the output wave high-frequency signal amplitude mutation amount is accumulated and ΔD j The characteristic parameter η of multi-layer wavelet electric shock j-1 , η j1 ; If the multi-layer wavelet high-frequency signals all have the T0 stage amplitude mutation amount cumulative sum is the highest among the three stages, that is, ΔD j0 >D j-1 And ΔD j0 >ΔD j1 , and the ratio of the amplitude mutation and the extracted characteristic parameter η j If both are greater than 1, it is determined that the analyzed residual current signal has the characteristics of electric shock to a living organism.

5. The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in claim 4, characterized in that: The GRU neural network-based electric shock event recognition includes the forward propagation of GRU represented as: r t =σ(W r ·[h t-1 ,x t ]·+b r ) z t =σ(W z ·[h t-1 ,x t ]+b z ) Among them, r t 、z t 、h t 、h t Represents reset gate, update gate, candidate hidden state and current hidden state; b represents weight and bias; σ(·) and tanh(·) represent sigmoid function and tanh function. The sigmoid function limits the data to the range of [0,1], and the tanh function limits the data to the range of [-1,1]. Reset Gate t Controls the previous state history information to be written into the current candidate set h t The reset gate is proportional to the information written in the previous state. The information of the previous moment and the current moment are right-multiplied by the weight matrix and added. The data added with the bias information is sent to the reset gate, that is, multiplied by the sigmoid function, and the output value is between [0,1]; Update gate z t It is used to control the degree to which the state information of the previous moment is brought into the current state. The value of the update gate is proportional to the state information brought in at the previous moment. Similar to the data processing of the reset gate, the information of the previous moment and the current moment are respectively right-multiplied by the weight matrix and added together. The data after adding the bias information is sent to the update gate, which is multiplied by the sigmoid function. The output value is between [0,1].

6. The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in claim 5, characterized in that: The method of identifying electric shock events based on the GRU neural network includes using the mean absolute error value to quantitatively evaluate the prediction accuracy of the model, which is expressed as: Among them, y i ,y i Represent the predicted value and the true value respectively, and n represents the number of tests; GRU neural network model optimization includes data set partitioning and model optimization; Divide the dataset; Includes training set, validation set and test set; The training set is used to train the model parameters and accounts for 70% of the total data set; The validation set is used to adjust the model hyperparameters and accounts for 15% of the total dataset; The test set is used to finally evaluate the model performance, accounting for 15% of the total dataset; Model optimization involves calculating the MAE of the network layers to determine the model accuracy corresponding to different numbers of network layers.

7. The electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in claim 6, characterized in that: The electric shock diagnosis based on wavelet analysis and GRU neural network includes that when no electric shock accident occurs, there is no sudden change in the residual current of the power grid, so the residual current signal characteristics when no electric shock accident occurs are divided into category 1; Considering the randomness of electric shock accidents in actual situations, electric shock accidents occur in the T-1, T0, and T1 cycles. When the electric shock occurs in the T-1 cycle, the corresponding signal feature is classified as category 2. When the electric shock occurs in the T0 cycle, the extracted signal feature is classified as category 3. When the electric shock occurs in the T1 cycle, the extracted signal feature is classified as category 4. Randomly select the test electric shock signals, intercept 600 points of each signal according to the classification, and extract the electric shock features according to the biological body electric shock feature determination method. The signal interception method includes: Category 1 residual current signal, intercepting 600 continuous points of normal residual current signal; The second type of residual current signal, that is, the electric shock occurs at any sampling point within the T-1 period, and the corresponding residual current signal sampling point is recorded as t1(i); The third type of residual current signal, that is, the electric shock occurs at any sampling point within the T0 period, and the corresponding residual current signal sampling point is recorded as t2(i); The fourth type of residual current signal, that is, the electric shock occurs at any sampling point within the T1 period, and the corresponding residual current signal sampling point is recorded as t3(i); Building an electric shock recognition model includes: The residual current transformer is connected to the three-phase power supply of A, B and C respectively, and the residual current signal is obtained by using pork as the electric shock body; The acquired data were downsampled and denoised using MODWT, and 600 sampling points were intercepted and divided into three stages: T-1, T0, and T1 for subsequent analysis; The 11th-order wavelet in the Daubechies series of wavelets is selected for multi-layer wavelet decomposition, and the multi-layer high-frequency signals of the residual current signal are dimensionally normalized. Calculate the cumulative sum of the normalized amplitude mutation of the wavelet high-frequency signal at each layer of the three stages T-1, T0, and T1 and the characteristic parameters of the electric shock accident, and output the characteristic data for identifying the electric shock of the organism; Build a GRU electric shock diagnosis model, and divide the feature data set into training set, test set and validation set as model input for training and validation; Using the trained model, input any set of data to diagnose the type of electric shock and determine whether electric shock occurs and whether the electric shock occurs in the T-1, T0 or T1 stage.

8. A system using the electric shock safety event identification method based on wavelet multi-resolution analysis and neural network as claimed in any one of claims 1 to 7, characterized in that: It includes a current signal acquisition module, a wavelet analysis module, an electric shock event recognition module, and an electric shock diagnosis module; The current signal acquisition module is used to acquire the electric shock residual current signal; The wavelet analysis module is used for wavelet analysis of electric shock residual current signal; The electric shock event recognition module is used for electric shock event recognition based on the GRU neural network; The electric shock diagnosis module is used to perform electric shock diagnosis based on wavelet analysis and GRU neural network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying electric shock safety events based on wavelet multi-resolution analysis and neural network as described in any one of claims 1 to 7 are implemented.

Images