Power device big data intelligent monitoring method and internet of things system thereof

Abstract

The invention relates to the technical field of automation equipment of electric power device big data intelligent monitoring and internet of things systems, and discloses an electric power device big data intelligent monitoring method and an internet of things system thereof, and the system is composed of a data collector, a centralized controller, a remote monitoring terminal, a cloud platform and a mobile phone APP. The data collector and the integrated controller communicate through the LoRa communication module in a self-organizing network mode, communication between the data collector and the remote monitoring terminal, the cloud platform and the mobile phone APP is achieved through the integrated controller, and the remote monitoring terminal and the cloud platform conduct big data monitoring, storage and processing on the power device. And the mobile phone APP and the remote monitoring terminal are in two-way communication with the centralized controller through the cloud platform. According to the invention, through research and discussion of the intelligent power device, the power distribution management system can realize safe and reliable power management, remote monitoring, control and management of power equipment are realized, and the level and efficiency of power management are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring of big data of electric power devices and automation equipment technology of Internet of Things systems, and in particular to an intelligent monitoring method of big data of electric power devices and an Internet of Things system thereof. Background Art

[0002] The Internet of Things technology, big data and artificial intelligence are combined, and sensor technology is used to monitor various parameter information in the power device in real time and collect, transmit, monitor thresholds and intelligently identify data, so as to process, analyze, predict and maintain the data in the power device. When the power device is in an abnormal state, the voltage, current and power parameters of the power device will be abnormal. The changes of these parameters require monitoring personnel to track and monitor in time, so as to provide users with more comprehensive monitoring and management services. The possible faults of the power device can be found by analyzing the abnormal parameters of real-time data. The historical data can be analyzed by deep learning and other technologies, and solutions to solve the faults of the power device can be given in advance to avoid the occurrence and impact of the faults, so as to provide users with more accurate electricity usage suggestions and energy-saving measures. The power device big data intelligent monitoring method and its Internet of Things system provide users with more convenient and efficient power device management services through all management and services such as the power device background management system, user interface and related App. Users can more conveniently perform real-time monitoring and data management of power devices, safety monitoring, energy-saving management and remote control functions, which greatly improves the reliability, stability and operation convenience of the distribution system. Summary of the invention

[0003] Purpose of the invention: Through the seamless connection of the Internet of Things technology, a safer, more efficient and reliable power supply management solution is provided for enterprises. The present invention discloses a method for intelligent monitoring of power device big data and its Internet of Things system. Through the research and discussion of intelligent power devices, the power distribution management system can achieve safe and reliable power management, improve the efficiency and stability of power transmission, and the application prospects of intelligent power devices are very broad. People can realize remote monitoring, control and management of power equipment, and improve the level and efficiency of power management.

[0004] Technical solution: The present invention discloses a method for intelligent monitoring of big data of electric power equipment, comprising the following steps:

[0005] Step 1: Construct a parameter detection module. The parameter detection module includes AANN autoassociative neural network model, ensemble empirical mode decomposition model, BiLSTM neural network model-ARIMA model, BiLSTM neural network model, ARIMA model, GCN graph convolutional network model, TCN temporal convolutional network model-WOA fuzzy wavelet neural network model, GCN graph convolutional network model-WOA fuzzy wavelet neural network model and WOA fuzzy recursive neural network model; the structure diagram of the parameter detection module can be found in Figure 1 .

[0006] Step 2: Construct a power device big data intelligent processing subsystem, which includes 3 parameter detection modules, 4 TDL beat delayers, 2 BiLSTM neural network models and 2 GCN graph convolutional network models-hopfield neural network classifiers;

[0007] The output of the current sensor, the output of the temperature sensor and the output of the vibration sensor are respectively used as the corresponding inputs of the parameter detection module 1, the parameter detection module 2 and the parameter detection module 3 and the corresponding inputs of the GCN graph convolutional network model-hopfield neural network classifier 2. The output of the parameter detection module 1 is respectively used as the input of the TDL beat delay device 1, the corresponding input of the BiLSTM neural network model 1 and the corresponding input of the BiLSTM neural network model 2. The output of the parameter detection module 2 is used as the corresponding input of the BiLSTM neural network model 1, the output of the parameter detection module 3 is used as the corresponding input of the BiLSTM neural network model 2, and the output of the BiLSTM neural network model 1 is used as the input of the TDL beat delay device 2 and the BiLSTM neural network The output of BiLSTM neural network model 2 is used as the input of TDL beat delay 3. The outputs of GCN graph convolution network model-hopfield neural network classifier 2, TDL beat delay 1, TDL beat delay 2, TDL beat delay 3 and TDL beat delay 4 are used as the corresponding inputs of GCN graph convolution network model-hopfield neural network classifier 1 respectively. The output of GCN graph convolution network model-hopfield neural network classifier 1 is used as the input of TDL beat delay 4. The safety level of the detected configuration device output by GCN graph convolution network model-hopfield neural network classifier 1 is safe, relatively safe, general, dangerous and very dangerous. See the structural diagram of the power device big data intelligent processing subsystem for reference. Figure 2 .

[0008] Furthermore, the sensor output is used as the input of the AANN auto-associative neural network model, the output of the AANN auto-associative neural network model is used as the input of the ensemble empirical mode decomposition model, the input of the ARIMA model and the corresponding input of the fuzzy wavelet neural network model of the TCN time convolutional network model-WOA, the output of the ARIMA model is used as the corresponding input of the fuzzy wavelet neural network model of the TCN time convolutional network model-WOA, the output of the ensemble empirical mode decomposition model is used as the input of the BiLSTM neural network model-ARIMA model and the corresponding input of the GCN graph convolutional neural network model, the output of the BiLSTM neural network model-ARIMA model is used as the corresponding input of the GCN graph convolutional neural network model, the time series difference between the output of the AANN auto-associative neural network model and the output of the BiLSTM neural network model-ARIMA model is used as the input of the BiLSTM neural network model, and the output of the BiLSTM neural network model is used as The corresponding input of the GCN graph convolutional neural network model, the output of the GCN graph convolutional neural network model is used as the corresponding input of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model, the output of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model is used as the input of the TDL beat delay device, the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model and the WOA's fuzzy recursive neural network model, the TDL beat delay device output is used as the corresponding input of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model, the output of the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model is used as the corresponding input of the WOA's fuzzy recursive neural network model, the output of the WOA's fuzzy recursive neural network model is used as the corresponding input of the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model, and the output of the WOA's fuzzy recursive neural network model is used as the output of the parameter detection module. For the structural diagram of the parameter detection module, see Figure 1 .

[0009] Further, the TCN temporal convolutional network model-WOA fuzzy wavelet neural network model is a series connection of the TCN temporal convolutional network model and the WOA fuzzy wavelet neural network model;

[0010] The TCN temporal convolutional network model is a neural network that combines causal convolution and dilated convolution as a convolution layer, fuses it with a 1×1 convolution into a residual module, and then stacks multiple residual modules to form a neural network.

[0011] The fuzzy wavelet neural network model of WOA uses the WOA method to optimize the input layer node weights and hidden layer node thresholds of the fuzzy wavelet neural network model, and uses the WOA optimal fuzzy wavelet neural network model to randomly select the input layer weights w and hidden layer node thresholds h in the fuzzy wavelet neural network model; set the population size s and the maximum number of iterations tmax of WOA, and initialize each individual value; calculate the fitness value of each individual in the initial state, and select the individual with the smallest fitness value; the WOA iterative process adopts the gradual search model, the spiral search model and the random search model respectively; end the iteration, get the optimal solution, output the parameters of the fuzzy wavelet neural network model after WOA optimization, and use the optimized WOA fuzzy wavelet neural network model data set for training.

[0012] Furthermore, the GCN graph convolutional network model-WOA fuzzy wavelet neural network model is a series connection of the GCN graph convolutional network model and the WOA fuzzy wavelet neural network model.

[0013] Furthermore, the fuzzy recursive neural network model of WOA consists of an input layer, a membership function layer, a rule layer, a defuzzification layer and an output layer;

[0014] The input layer has n neurons representing n input variables. The membership function layer uses Gaussian function as the membership function to fuzzify the input variables. The output of each node represents the membership value of the corresponding input quantity. Each neuron in the rule layer represents a fuzzy rule. At the same time, a recursive link is introduced in the rule layer. The output of the rule layer at the previous moment is used as a feedback input of the rule layer at the current moment through the sigmoid function. The number of neurons in the defuzzification layer is the same as that in the rule layer, which is m neurons. The output layer performs linear weighted summation on its input quantity.

[0015] The fuzzy recursive neural network model of WOA is to optimize the parameters of the fuzzy recursive neural network model using WOA.

[0016] Furthermore, the GCN graph convolutional network model-hopfield neural network classifier is a GCN graph convolutional network model connected in series with a hopfield associative memory neural network, and the hopfield associative memory neural network contains n neurons, namely N1, N2, ..., Nn, which are interconnected to form a feedback network;

[0017] Taking the feature vector of the tested power device data as the standard, the GCN graph convolutional network model is introduced as the pre-processing of the Hopfield neural network classifier, so that the attraction domain of the Hopfield neural network classifier is effectively expanded.

[0018] The present invention also discloses an Internet of Things system for electric power devices, which is composed of a data collector, a centralized controller, a remote monitoring terminal, a cloud platform and a mobile phone APP. The data collector and the centralized controller communicate via a LoRa communication module in a self-organizing network manner, and the centralized controller realizes communication between the data collector and the remote monitoring terminal, the cloud platform and the mobile phone APP. The remote monitoring terminal and the cloud platform monitor, store and process big data of electric power devices. The mobile phone APP and the remote monitoring terminal communicate with the centralized controller in a two-way manner via the cloud platform. The management personnel monitor the operating status of the data collector and the centralized controller via the remote monitoring terminal and the mobile phone APP. The Internet of Things system for electric power devices is loaded with computer program steps for realizing the above-mentioned intelligent monitoring method for big data of electric power devices. The structure of the Internet of Things system for electric power devices is shown in FIG. Figure 3 shown.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] 1. The present invention uses the AANN self-associative neural network model to learn the nonlinear relationship between the various variables of the power device. For the data inputted as the normal operation of the power device, the AANN self-associative neural network model solidifies the characteristics and associations between the data during the normal operation of the power device into the AANN self-associative neural network model. The AANN self-associative neural network model is trained with the goal of minimizing the input and output errors, and seeks the optimal network parameters of the AANN self-associative neural network model so that the output of the AANN self-associative neural network model can restore the input data of the power device as much as possible. When the input is degraded or faulty data, the data output by the AANN self-associative neural network model will have a large deviation from the input data of the power device.

[0021] Second, the present invention proposes a method of collective empirical mode decomposition to address the intermittent and volatile characteristics of the power device data sequence in time series. While introducing the modal decomposition of the power device data, it greatly improves the fitting degree between the predicted value of the power device data and the original power device data sequence, avoids the mutual interference of different frequency scales, and reflects the superiority of collective empirical mode decomposition in processing non-stationary sequence data of power devices. Collective empirical mode decomposition eliminates the non-stationary characteristics of the power device data sequence and realizes the stabilization of the power device data sequence.

[0022] 3. The BiLSTM neural network model consists of two forward and backward long short-term memory (LSTM) neural network models, which consists of a forward LSTM hidden layer and a backward LSTM hidden layer, and both hidden layers are connected to the output layer. The forward information and the backward information are simultaneously passed to the BiLSTM neural network model output layer, which improves the accuracy of the BiLSTM neural network model in predicting the output information.

[0023] 4. The input object of the GCN graph convolutional network model is graph data. The GCN graph convolutional neural network model learns the topological structure of the power device data network through the stacking of multiple graph convolutional layers, captures the spatial correlation of the power device data flow, and obtains a deep abstract representation. The GCN graph convolutional network model can better integrate the node information and data structure information in the power device data graph.

[0024] 5. In view of the characteristics of power device data that there are many features and both temporal and spatial features, the TCN temporal convolutional network model uses causal convolution to learn the causal relationship between network traffic features and power device data features while retaining more original information of power device data. By increasing the receptive field of convolution by dilated convolution, one convolution can learn more power device data features; the introduction of the residual module can deepen the number of layers of the TCN temporal convolutional network model to more fully learn the implicit characteristics of power device data flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a parameter detection module of the present invention;

[0026] Figure 2 It is the big data intelligent processing subsystem of the power device of the present invention;

[0027] Figure 3 The power device Internet of Things system of the present invention;

[0028] Figure 4 The data collector of the present invention;

[0029] Figure 5 The centralized controller of the present invention;

[0030] Figure 6 This is the remote monitoring software function of the present invention. DETAILED DESCRIPTION

[0031] In order to better explain the present invention, the following Figure 1-6 The technical solution of the present invention is described in detail. The following embodiments are explanations of the invention, and the present invention is not limited to the following embodiments.

[0032] The present invention discloses a method for intelligent monitoring of big data of electric power equipment, comprising the following steps:

[0033] Step 1: Construct a parameter detection module. The parameter detection module includes AANN autoassociative neural network model, ensemble empirical mode decomposition model, BiLSTM neural network model-ARIMA model, BiLSTM neural network model, ARIMA model, GCN graph convolutional network model, TCN temporal convolutional network model-WOA fuzzy wavelet neural network model, GCN graph convolutional network model-WOA fuzzy wavelet neural network model and WOA fuzzy recursive neural network model.

[0034] The sensor output is used as the input of the AANN auto-associative neural network model. The output of the AANN auto-associative neural network model is used as the input of the ensemble empirical mode decomposition model, the ARIMA model and the corresponding input of the fuzzy wavelet neural network model of the TCN temporal convolutional network model-WOA. The output of the ARIMA model is used as the corresponding input of the fuzzy wavelet neural network model of the TCN temporal convolutional network model-WOA. The output of the ensemble empirical mode decomposition model is used as the input of the BiLSTM neural network model-ARIMA model and the corresponding input of the GCN graph convolutional neural network model. The output of the BiLSTM neural network model-ARIMA model is used as the corresponding input of the GCN graph convolutional neural network model. The time series difference between the output of the AANN auto-associative neural network model and the output of the BiLSTM neural network model-ARIMA model is used as the input of the BiLSTM neural network model. The output of the BiLSTM neural network model is used as the GCN graph convolutional neural network model. The output of the GCN graph convolutional neural network model is used as the corresponding input of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model. The output of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model is used as the input of the TDL beat delay device, the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model and the fuzzy recursive neural network model of WOA. The output of the TDL beat delay device is used as the corresponding input of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model, the output of the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model is used as the corresponding input of the WOA's fuzzy recursive neural network model, the output of the WOA's fuzzy recursive neural network model is used as the corresponding input of the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model, and the output of the WOA's fuzzy recursive neural network model is used as the output of the parameter detection module. For the structural diagram of the parameter detection module, see Figure 1 .

[0035] Step 2: Construct a power device big data intelligent processing subsystem, which includes 3 parameter detection modules, 4 TDL beat delayers, 2 BiLSTM neural network models and 2 GCN graph convolutional network models-hopfield neural network classifiers;

[0036] The output of the current sensor, the output of the temperature sensor and the output of the vibration sensor are respectively used as the corresponding inputs of the parameter detection module 1, the parameter detection module 2 and the parameter detection module 3 and the corresponding inputs of the GCN graph convolutional network model-hopfield neural network classifier 2. The output of the parameter detection module 1 is respectively used as the input of the TDL beat delay device 1, the corresponding input of the BiLSTM neural network model 1 and the corresponding input of the BiLSTM neural network model 2. The output of the parameter detection module 2 is used as the corresponding input of the BiLSTM neural network model 1, the output of the parameter detection module 3 is used as the corresponding input of the BiLSTM neural network model 2, and the output of the BiLSTM neural network model 1 is used as the input of the TDL beat delay device 2 and the BiLSTM neural network The output of BiLSTM neural network model 2 is used as the input of TDL beat delay 3. The outputs of GCN graph convolution network model-hopfield neural network classifier 2, TDL beat delay 1, TDL beat delay 2, TDL beat delay 3 and TDL beat delay 4 are used as the corresponding inputs of GCN graph convolution network model-hopfield neural network classifier 1 respectively. The output of GCN graph convolution network model-hopfield neural network classifier 1 is used as the input of TDL beat delay 4. The safety level of the detected configuration device output by GCN graph convolution network model-hopfield neural network classifier 1 is safe, relatively safe, general, dangerous and very dangerous. See the structural diagram of the power device big data intelligent processing subsystem for reference. Figure 2 .

[0037] 1. Parameter detection module design

[0038] The parameter detection module includes AANN autoassociative neural network model, ensemble empirical mode decomposition model, BiLSTM neural network model-ARIMA model, BiLSTM neural network model, ARIMA model, GCN graph convolutional network model, TCN temporal convolutional network model-WOA fuzzy wavelet neural network model, GCN graph convolutional network model-WOA fuzzy wavelet neural network model and WOA fuzzy recursive neural network model.

[0039] 1. Design of AANN auto-associative neural network model

[0040] The AANN auto-associative neural network model is a five-layer feedforward transfer network with a symmetrical topological structure. Its characteristics are that the output is equal to the input. It has a symmetrical topological structure and can be used for effective feature extraction in high-dimensional space, system modeling, fault identification and other issues. This patent uses the AANN auto-associative neural network model to extract the main features and correlations of the power device operation data, while solidifying its normal operating mode data into the AANN auto-associative neural network model. The process of the power device input data of the AANN auto-associative neural network model passing through the input layer and the mapping layer to the bottleneck layer belongs to the encoding process; and then passing through the bottleneck layer to the mapping layer and finally to the output layer belongs to the decoding process. First, the high-dimensional current device data is nonlinearly mapped to the bottleneck layer through the input layer to achieve compression of the high-dimensional data and then extract effective feature dimensions.

[0041] Let g(x) denote the encoding function, x i As input, i is the i-th dimension of x. The data of the power device is encoded to obtain y i , whose expression is:

[0042] y i =g θ (x i )=s(Wx i +b) (1)

[0043] The hidden layer y i The bottleneck layer is mapped to the output layer of the same dimension as the input layer through a nonlinear function to realize the decoding of the input information, that is, the reconstruction of the input data of the power device. Through the nonlinear compression and reconstruction of the power device data, the AANN self-associative neural network model will learn the nonlinear relationship between the various variables of the power device. For the data input as the normal operation of the power device, the AANN self-associative neural network model solidifies the characteristics and associations between the data during the normal operation of the power device into the AANN self-associative neural network model. When training the AANN self-associative neural network model, the goal is to minimize the input and output errors, and seek the optimal network parameters of the AANN self-associative neural network model so that the output of the AANN self-associative neural network model can restore the input data of the power device as much as possible. When the input is degraded or faulty data, the data output by the AANN self-associative neural network model will have a large deviation from the input data of the power device.

[0044] 2. Design of ensemble empirical mode decomposition model

[0045] The core of empirical mode decomposition is to screen intrinsic mode functions. However, for non-stationary data, EMD decomposition also has defects, mainly because modal aliasing is prone to occur. Modal aliasing refers to the fact that an IMF component contains extremely different characteristic time scales. The ensemble empirical mode decomposition model is an ensemble empirical mode decomposition based on noise-assisted analysis, which can eliminate modal aliasing. The steps of ensemble empirical mode decomposition are as follows:

[0046] 1) Add the standard normally distributed white noise n(t) to the original signal of the power device data to obtain the signal X(t):

[0047] X(t)=x(t)+n(t) (2)

[0048] 2) Add white noise to the power device data signal X i (t) Perform empirical mode decomposition to obtain j IMF components:

[0049]

[0050] Where: X i (t) represents the signal obtained by adding white noise for the ith time; c i,j (t) is the jth IMF of the ith trial; r i,j (t) is the remainder.

[0051] 3) Repeat steps 1) and 2) M times, adding white noise n with different normal distributions i (t) Obtain the IMF set.

[0052] 4) Perform integrated average calculation on the IMF obtained above to obtain the final result:

[0053]

[0054] Aiming at the intermittent and volatile characteristics of the power device data series in time series, a method of ensemble empirical mode decomposition is proposed. The introduction of power device data mode decomposition greatly improves the fitting degree between the power device data prediction value and the original power device data series, avoids the mutual interference of different frequency scales, and reflects the superiority of ensemble empirical mode decomposition in processing non-stationary series data of power devices. Ensemble empirical mode decomposition eliminates the non-stationary characteristics of the power device data series and realizes the stabilization of the power device data series. The combined reconstruction of the prediction results of the BiLSTM neural network model and the ARIMA model enhances the model's fitting ability and prediction accuracy.

[0055] 3. BiLSTM neural network model-ARIMA model design

[0056] The BiLSTM neural network model-ARIMA model is a BiLSTM neural network model connected in series with the ARIMA model. The BiLSTM neural network model consists of two forward and backward long short-term memory (LSTM) neural network models, which consist of a forward LSTM hidden layer and a backward LSTM hidden layer, and both hidden layers are connected to the output layer, and the forward information and backward information are simultaneously transmitted to the output layer of the BiLSTM neural network model. The LSTM neural network model is an improved model of the recurrent neural network, which uses a structure composed of three "gates" of input gate, output gate, and forget gate and cell states to replace the hidden layer in the traditional neural network. The LSTM neural network model algorithm can more accurately predict the long-term series of power device data, and can also overcome the problems of gradient disappearance and explosion in the calculation process.

[0057] The ARIMA model is a differential autoregressive moving average prediction model, which mainly includes two parts: the autoregressive model and the moving average model. The order of the autoregressive model is p, the order of the signal difference is d, and the order of the moving average model is q. The ARIMA model is usually expressed as ARIMA (p, d, q). The specific implementation process of ARIMA is as follows:

[0058] (1) The stationarity of the time series data of the power device data prediction value output by the BiLSTM neural network model is verified. If it does not meet the requirements, it is converted into a stationary time series by performing d-order difference.

[0059] (2) Determine the orders p and q through the autocorrelation coefficient graph, partial autocorrelation coefficient graph and Bayesian information criterion.

[0060] (3) ARIMA (p, d, q) of a determined order is used to fit the time series of the power device data output by the BiLSTM neural network model, and the results are statistically analyzed and the prediction accuracy is analyzed based on the data predicted by ARIMA (p, d, q) and the original BiLSTM neural network model time series.

[0061] 4. GCN graph convolutional network model design

[0062] The GCN graph convolutional network model learns the topological structure of the power device data network to capture the spatial correlation of the power device data flow. The GCN graph convolutional network model is a feature extractor like the CNN convolutional neural network model. The object of the GCN graph convolutional network model is graph data. Assume that there are n nodes in the topological graph of the power device data network, and each node has a T-dimensional power device data time series. Let the feature composition matrix of these nodes be X∈R n The topological relationship between each node in ×T forms an adjacency matrix A∈R n×n, X and A are used as the input of the GCN graph convolutional network model. The GCN graph convolutional network model is the output for predicting the future t time series. Different from the CNN convolutional neural network model, the GCN graph convolutional neural network model acts on the nodes in the topological graph of the power device data network. Through the stacking of multiple graph convolutional layers, the GCN graph convolutional network model obtains a deep abstract representation. The GCN graph convolutional network model can better integrate the node information and data structure information in the power device data graph.

[0063] 5. TCN temporal convolutional network model-WOA fuzzy wavelet neural network model design

[0064] The TCN temporal convolutional network model - the fuzzy wavelet neural network model of WOA is a series connection of the TCN temporal convolutional network model and the fuzzy wavelet neural network model of WOA. The TCN temporal convolutional network model uses the temporal convolution combining causal convolution and dilated convolution as the convolution layer, and fuses it with a 1×1 convolution into a residual module, which is then stacked into a neural network composed of multiple residual modules, and also has weight normalization and Dropout.

[0065] The convolution in the TCN temporal convolutional network model includes causal convolution, which can better handle the time series problem of power device data. The use of dilated convolution expands the receptive field of convolution, and the addition of residual module can well preserve the historical information of power devices. The TCN temporal convolutional network model uses more powerful causal convolution to prevent the omission of power device information. Causal convolution can only see future information. However, causal convolution also has the problem of traditional convolution. When a lot of historical information is needed, it is necessary to stack layers or expand the convolution kernel, which will bring high computational cost, so dilated convolution is needed to solve this problem. Applying dilated convolution to causal convolution can obtain more historical information without increasing the amount of calculation. Dilated convolution mainly expands its receptive field through "expansion rate". The TCN temporal convolutional network model can obtain a larger receptive field with relatively few dilated convolution layers. Through the fusion of dilated convolution and causal convolution, the TCN temporal convolutional network model can learn the data flow characteristics of power devices with a larger receptive field, ensuring that more features of the data flow of power devices can be more fully learned on a single convolution kernel. The TCN temporal convolutional network model uses causal convolution to learn the causal relationship between network traffic characteristics and power device data characteristics while retaining more original information of power device data. By expanding the convolution, the receptive field of the convolution is increased, and one convolution can learn more power device data characteristics; the introduction of the residual module can deepen the number of layers of the TCN temporal convolutional network model to more fully learn the implicit characteristics of power device data flow, and it is not easy to have gradient problems, and the model performance can be better improved. In view of the characteristics of power device data with many features and the coexistence of temporal and spatial features, the TCN temporal convolutional network is very suitable for feature learning of power device data, and at the same time, in order to strengthen the temporal feature learning of power device data.

[0066] The wavelet neural network is embedded in the fuzzy model to form a fuzzy wavelet neural network model. The fuzzy wavelet neural network model consists of 5 layers. The first layer is the input layer, and the input directly leads to the second layer of membership function layer. The third layer is the fuzzy rule layer, and the fourth layer is the wavelet function layer. The first-order partial derivative of the Gaussian function is selected as the mother wavelet function. According to the selected mother wavelet function, it is put into the neurons of the second layer as the activation function after scaling and translation transformation. In the fifth layer, the output of the fourth wavelet layer × the node output of the third fuzzy rule layer. The fuzzy wavelet neural network model of WOA uses the WOA method to optimize the input layer node weights and hidden layer node thresholds of the fuzzy wavelet neural network model. The WOA optimal fuzzy wavelet neural network model is used to randomly select the input layer weights w and hidden layer node thresholds h in the fuzzy wavelet neural network model; the population size s and the maximum number of iterations tmax of WOA are set, and each individual value is initialized; the fitness value of each individual in the initial state is calculated, and the individual with the smallest fitness value is selected; the WOA iteration process uses a gradual search model, a spiral search model, and a random search model; the iteration is terminated to obtain the optimal solution. The parameters of the fuzzy wavelet neural network model after WOA optimization are output, and the fuzzy wavelet neural network model data set of the optimized WOA is used for training.

[0067] 6. GCN graph convolutional network model-WOA fuzzy wavelet neural network model design

[0068] The GCN graph convolutional network model-WOA fuzzy wavelet neural network model is a series connection of the GCN graph convolutional network model and the WOA fuzzy wavelet neural network model. The design process of the GCN graph convolutional network model and the WOA fuzzy wavelet neural network model refers to the relevant design process of this patent.

[0069] 7. Design of Fuzzy Recursive Neural Network Model for WOA

[0070] The fuzzy recursive neural network model of WOA consists of an input layer, a membership function layer, a rule layer, a defuzzification layer and an output layer; the input layer has n neurons representing n input variables, and the output of this layer is equal to the input; the membership function layer uses a Gaussian function as a membership function to fuzzify the input variables, and the output of each node represents the membership value of the corresponding input quantity: each neuron in the rule layer represents a fuzzy rule, and a recursive link is introduced in this layer, and the output of the rule layer at the previous moment is used as a feedback input of the rule layer at the current moment through the sigmoid function; the number of neurons in the defuzzification layer is the same as that in the rule layer, both of which are m neurons; the output layer is a linear weighted summation of its input quantity. The fuzzy recursive neural network model of WOA uses WOA to optimize the parameters of the fuzzy recursive neural network model. The design process of the fuzzy recursive neural network model of WOA refers to the design process of the fuzzy wavelet neural network model of WOA in this patent.

[0071] 2. Design of the intelligent processing subsystem for big data of power equipment

[0072] The power device big data intelligent processing subsystem includes a parameter detection module, a TDL beat delay device, a BiLSTM neural network model, and a GCN graph convolutional network model-hopfield neural network classifier. The GCN graph convolutional network model-hopfield neural network classifier is a GCN graph convolutional network model connected in series with a hopfield associative memory neural network. The hopfield associative memory neural network contains n neurons, namely N1, N2, ..., Nn, which are interconnected to form a feedback network. Since there are only two input forms of the Hopfield associative memory neural network, namely {-1, 1} or {0, 1}, when it is used for fault safety mode recognition, the feature vector extracted from the signal must be converted into a coded input form that can be recognized by the Hopfield associative memory neural network. The feature vector of the tested power device data is used as the standard. By introducing the GCN graph convolutional network model as the pre-processing of the Hopfield neural network classifier, its powerful nonlinear mapping ability is used to effectively expand the attraction domain of the Hopfield neural network classifier, making its associative recall ability more powerful. Therefore, the associative memory neural network with structural improvement, namely the GCN graph convolutional network model-Hopfield neural network classifier, not only solves the coding problem of the Hopfield neural network classifier network, but also improves the network storage fault tolerance performance.

[0073] The relationship between the GCN graph convolutional network model-hopfield neural network classifier output and the safety mode of the power device is shown below:

[0074] Serial number Safety level of electrical installations Classifier output value 1 Very safe [1，1，1,1,1] 2 Safer [-1，1，-1,-1,-1] 3 Safety [-1，-1，1,-1,-1] 4 Unsafe [-1，-1，-1,1,-1] 5 Very unsafe [-1，-1，-1,-1,1]

[0075] 3. Design of an Internet of Things System for Power Installations

[0076] The Internet of Things system for power devices consists of a data collector, a centralized controller, a remote monitoring terminal, a cloud platform and a mobile phone APP. The data collector and the centralized controller communicate through a self-organizing network through the LoRa communication module. The centralized controller realizes communication between the data collector and the remote monitoring terminal, the cloud platform and the mobile phone APP. The remote monitoring terminal and the cloud platform monitor, store and process the big data of the power device. The mobile phone APP and the remote monitoring terminal communicate with the centralized controller in a two-way manner through the cloud platform. The management personnel monitor the operating status of the data collector and the centralized controller through the remote monitoring terminal and the mobile phone APP. A structure of the Internet of Things system for power devices is shown in FIG. Figure 3 shown.

[0077] 1. Data collector design

[0078] The present invention adopts a data collector as a parameter sensing terminal for power devices. The data collector and the centralized controller realize two-way communication through the LoRa communication module. The data collector includes sensors for collecting temperature, current, voltage, vibration and humidity, and corresponding signal conditioning circuits and STM32 microprocessors. The software of the data collector mainly realizes LoRa module communication and the collection and preprocessing of power device parameters. The software is designed with C language programming, which has a high degree of compatibility, greatly improves the work efficiency of software design and development, and enhances the reliability, readability and portability of the program code. See the data collector for details. Figure 4 shown.

[0079] 2. Centralized controller design

[0080] The centralized controller includes an STM32 microprocessor, a LoRa communication module and a 4G module, 10 relay output circuits, 10 switch input circuits, 2 D / A conversion circuits and a touch screen. A self-organizing communication network is constructed to realize data interaction between the data collector and the centralized controller. The centralized controller realizes two-way data interaction with the cloud platform through the 4G module, collects the power grid parameters of the data collector, uploads the power grid parameters to the cloud platform server, and realizes monitoring of the power grid power devices. The centralized controller can be seen Figure 5 .

[0081] 3. Remote monitoring software design

[0082] The remote monitoring terminal is an industrial control computer. It realizes information exchange between the centralized controller and the data collector through the cloud platform server, and realizes the monitoring of power grid parameters and the intelligent processing subsystem of power device big data. The main functions of the remote monitoring terminal are communication parameter setting, data analysis and the intelligent processing subsystem of power device big data. Figure 6 shown.

[0083] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intelligent monitoring of power equipment based on big data. It is characterized in that The following steps are involved: Step 1: Construct parameter detection module; Step 2: Construct a power device big data intelligent processing subsystem, which includes 3 parameter detection modules, 4 TDL beat delayers, 2 BiLSTM neural network models and 2 GCN graph convolutional network models-hopfield neural network classifiers; The output of the current sensor, the output of the temperature sensor and the output of the vibration sensor are respectively used as the corresponding inputs of the parameter detection module 1, the parameter detection module 2 and the parameter detection module 3 and the corresponding inputs of the GCN graph convolutional network model-hopfield neural network classifier 2. The output of the parameter detection module 1 is respectively used as the input of the TDL beat delay device 1, the corresponding input of the BiLSTM neural network model 1 and the corresponding input of the BiLSTM neural network model 2. The output of the parameter detection module 2 is used as the corresponding input of the BiLSTM neural network model 1, the output of the parameter detection module 3 is used as the corresponding input of the BiLSTM neural network model 2, and the output of the BiLSTM neural network model 1 is used as the input of the TDL beat delay device 2 and the BiLSTM neural network The output of BiLSTM neural network model 2 is used as the input of TDL beat delay 3, the outputs of GCN graph convolution network model-hopfield neural network classifier 2, TDL beat delay 1, TDL beat delay 2, TDL beat delay 3 and TDL beat delay 4 are respectively used as the corresponding inputs of GCN graph convolution network model-hopfield neural network classifier 1, the output of GCN graph convolution network model-hopfield neural network classifier 1 is used as the input of TDL beat delay 4, and the safety level of the detected configuration device output by GCN graph convolution network model-hopfield neural network classifier 1 is safe, relatively safe, general, dangerous and very dangerous.

2. The method for intelligent monitoring of power devices based on big data according to claim 1, It is characterized in that The parameter detection module includes AANN autoassociative neural network model, ensemble empirical mode decomposition model, BiLSTM neural network model-ARIMA model, BiLSTM neural network model, ARIMA model, GCN graph convolutional network model, TCN temporal convolutional network model-WOA fuzzy wavelet neural network model, GCN graph convolutional network model-WOA fuzzy wavelet neural network model and WOA fuzzy recursive neural network model; The sensor output is used as the input of the AANN auto-associative neural network model. The output of the AANN auto-associative neural network model is used as the input of the ensemble empirical mode decomposition model, the ARIMA model and the corresponding input of the fuzzy wavelet neural network model of the TCN temporal convolutional network model-WOA. The output of the ARIMA model is used as the corresponding input of the fuzzy wavelet neural network model of the TCN temporal convolutional network model-WOA. The output of the ensemble empirical mode decomposition model is used as the input of the BiLSTM neural network model-ARIMA model and the corresponding input of the GCN graph convolutional neural network model. The output of the BiLSTM neural network model-ARIMA model is used as the corresponding input of the GCN graph convolutional neural network model. The time series difference between the output of the AANN auto-associative neural network model and the output of the BiLSTM neural network model-ARIMA model is used as the input of the BiLSTM neural network model. The output of the BiLSTM neural network model is used as the GCN graph convolutional neural network model. The output of the GCN graph convolutional neural network model is used as the corresponding input of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model, the output of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model is used as the input of the TDL beat delay device, the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model and the fuzzy recursive neural network model of WOA, the output of the TDL beat delay device is used as the corresponding input of the TCN temporal convolutional network model-WOA's fuzzy wavelet neural network model, the output of the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model is used as the corresponding input of the fuzzy recursive neural network model of WOA, the output of the fuzzy recursive neural network model of WOA is used as the corresponding input of the GCN graph convolutional network model-WOA's fuzzy wavelet neural network model, and the output of the fuzzy recursive neural network model of WOA is used as the output of the parameter detection module.

3. The method for intelligent monitoring of power devices based on big data according to claim 1, It is characterized in that The TCN temporal convolutional network model-WOA fuzzy wavelet neural network model is a series connection of the TCN temporal convolutional network model and the WOA fuzzy wavelet neural network model; The TCN temporal convolutional network model is a neural network that combines causal convolution and dilated convolution as a convolution layer, fuses it with a 1×1 convolution into a residual module, and then stacks multiple residual modules to form a neural network. The fuzzy wavelet neural network model of WOA uses the WOA method to optimize the input layer node weights and hidden layer node thresholds of the fuzzy wavelet neural network model, and uses the WOA optimal fuzzy wavelet neural network model to randomly select the input layer weights w and hidden layer node thresholds h in the fuzzy wavelet neural network model; set the population size s and the maximum number of iterations tmax of WOA, and initialize each individual value; calculate the fitness value of each individual in the initial state, and select the individual with the smallest fitness value; the WOA iterative process uses a gradual search model, a spiral search model, and a random search model respectively; end the iteration, get the optimal solution, output the parameters of the fuzzy wavelet neural network model after WOA optimization, and use the optimized WOA fuzzy wavelet neural network model data set for training.

4. The method for intelligent monitoring of power equipment based on big data according to claim 3, It is characterized in that The GCN graph convolutional network model-WOA fuzzy wavelet neural network model is a series connection of the GCN graph convolutional network model and the WOA fuzzy wavelet neural network model.

5. The method for intelligent monitoring of power devices based on big data according to claim 3, It is characterized in that The fuzzy recursive neural network model of WOA consists of an input layer, a membership function layer, a rule layer, a defuzzification layer and an output layer; The input layer has n neurons representing n input variables. The membership function layer uses Gaussian function as the membership function to fuzzify the input variables. The output of each node represents the membership value of the corresponding input quantity. Each neuron in the rule layer represents a fuzzy rule. At the same time, a recursive link is introduced in the rule layer. The output of the rule layer at the previous moment is used as a feedback input of the rule layer at the current moment through the sigmoid function. The number of neurons in the defuzzification layer is the same as that in the rule layer, which is m neurons. The output layer performs a linear weighted sum of its inputs; The fuzzy recursive neural network model of WOA is to optimize the parameters of the fuzzy recursive neural network model using WOA.

6. The method for intelligent monitoring of power devices using big data according to claim 1, It is characterized in that The GCN graph convolutional network model-Hopfield neural network classifier is a GCN graph convolutional network model connected in series with a Hopfield associative memory neural network. The Hopfield associative memory neural network contains n neurons, namely N1, N2, ..., Nn, which are interconnected to form a feedback network.

7. An Internet of Things system for power devices, It is characterized in that The invention is composed of a data collector, a centralized controller, a remote monitoring terminal, a cloud platform and a mobile phone APP. The data collector and the centralized controller communicate with each other through a self-organizing network via a LoRa communication module. The communication between the data collector and the remote monitoring terminal, the cloud platform and the mobile phone APP is realized through the centralized controller. The remote monitoring terminal and the cloud platform monitor, store and process the big data of the power device. The mobile phone APP and the remote monitoring terminal communicate with the centralized controller in two directions via the cloud platform. The management personnel monitor the operating status of the data collector and the centralized controller via the remote monitoring terminal and the mobile phone APP. The power device Internet of Things system is loaded with computer program steps for realizing the intelligent monitoring method of big data of power devices as described in any one of claims 1 to 6.

Images