CN110059357A - A kind of intelligent electric energy meter failure modes detection method and system based on autoencoder network - Google Patents

Abstract

The invention discloses a method and system for classifying and detecting faults of smart electric energy meters based on a self-encoding network. Encoding the parameters of the network model; select sample data in the training set and input it into the self-encoding network model to obtain signal features and input them into the classifier for classification, and perform iterative training according to the classification results; continuously adjust according to the classification results of the test set The self-encoding network model parameters are used to determine the optimal parameters of the self-encoding network model; and the faults of the smart electric energy meter are classified and detected by using the self-encoding network model corresponding to the optimal parameters. The invention adopts the deep noise reduction self-encoding network to perform unsupervised feature extraction on the collected signals, which can realize fast and accurate classification of fault signals, help to improve the fault identification ability of the smart electric energy meter, and has stronger resistance to the traditional method. noise.

Description

A method and system for fault classification and detection of smart electric energy meters based on self-encoding network

technical field

The invention relates to the technical field of fault analysis of smart electric energy meters, and more particularly, to a method and system for classifying and detecting faults of intelligent electric energy meters based on a self-encoding network.

Background technique

With the use of various nonlinear loads in the power system and the external interference in the power line transmission process, various types of faults in the power grid increase, and the types of faults tend to be more complex. The rapid and accurate detection and classification of various faults in the line can effectively reduce the destructive influence of the power system by the fault.

Extracting features from voltage and current signals helps researchers better understand the nature of faults, feature detection and classification tasks, and facilitates a more consistent and efficient way to accomplish these tasks.

In traditional fault detection methods, converting time-domain signals into frequency-domain signals is the main step in feature extraction. Commonly used methods include discrete wavelet transform (Discrete Fourier transform, DFT), S-transform (S-transform, ST) and other time-frequency analysis methods. In addition, dimensionality reduction analysis can also generate useful features for classification, such as principal component analysis (PCA). Although the above feature extraction techniques have been applied to different types of power system fault identification, certain prior knowledge is the premise of accurate fault classification, and some details in the implementation process often need to be revised and adjusted repeatedly. Not only that, the accuracy of DFT and ST transforms is constrained and affected by other factors. Therefore, implementing these techniques can be very difficult, time-consuming, and lack generality.

Application No. 201810522073.5 Anomaly detection method of electric energy metering device based on long and short-term memory model applies deep learning theory to anomaly detection of digital electric energy meters. It can effectively improve the abnormal detection accuracy of the electric energy metering device; however, this method takes a long time and requires a large amount of calculation. For a large number of redundant electric energy meters to collect signals, timely response to the power grid requires high hardware support.

SUMMARY OF THE INVENTION

The present invention proposes a method and system for classifying and detecting faults of smart electric energy meters based on a self-encoding network, so as to solve the problem of how to efficiently and accurately classify and detect the faults of intelligent electric energy meters.

In order to solve the above problems, according to an aspect of the present invention, a method for classifying and detecting faults of smart electric energy meters based on a self-encoding network is provided, wherein the method includes:

Normalize the acquired historical detection data of the smart energy meter, and divide the normalized historical detection data into a training set and a test set; wherein the historical detection data is a voltage signal sample or a current signal sample , including: normal signal data and fault signal data;

Initially set the parameters of the self-encoding network model; wherein, the parameters include: the number of layers of the self-encoding network model, the number of nodes in each layer, the weight and bias of the encoding structure;

Select the sample data of the number of nodes in the first layer in the training set and input it into the self-encoding network model, so as to obtain the signal feature and input it into the classifier for classification, and perform iterative training according to the classification result;

Calculate the test accuracy according to the classification result obtained by using the test set, and perform an iterative operation according to the test accuracy, and continuously adjust the parameters of the self-encoding network model to determine the optimal parameters of the self-encoding network model;

The faults of the smart electric energy meter are classified and detected by using the self-encoding network model corresponding to the optimal parameters.

Preferably, the normalization of the acquired historical detection data of the smart electric energy meter includes:

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter, E(·) is the mean value of the signal x(n), and Std(·) is the standard deviation of the signal x(n) .

Preferably, wherein the method further comprises:

Before normalizing the acquired historical detection data of the smart electric energy meter, eliminate redundant data and erroneous data in the acquired historical detection data of the smart electric energy meter, and fill in the missing data.

Preferably, wherein the self-encoding network model includes: an encoding structure and a decoding structure,

The coding structure is expressed as the following formula:

h _t =f(Wx'(n)+b) _t ,

The decoding structure is expressed as the following formula:

x _t =f(W'h _t +b') _t ,

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter f(·) is the activation function, W is the weight of the coding structure, b is the bias of the coding structure, t=1, 2,...,N is the number of layers of the self-encoding network model, n is the input dimension, and the dimensions of different layers are gradually reduced to effectively extract signal features; W' is the weight of the decoding structure, b' is the bias of the decoding structure, x _t is the output of the decoding structure, which is equal to the input of the encoding structure.

Preferably, the weights and biases of the coding structure are initialized using a truncated normal distribution, including:

Among them, f(x) is the density function, u and δ are the mean and standard deviation of the normal distribution, respectively, and the value range of the weight W of the coding structure is (u-2δ, u+2δ).

Preferably, wherein the method further comprises:

The iteration is stopped when the loss value no longer decreases within the consecutive training cycles of the first preset number threshold or the test accuracy begins to decrease after the consecutive training cycles of the second preset number threshold.

Preferably, wherein the method further comprises:

When using the self-encoding network model for iterative training, a preset ratio of noise is added to the input signal, and the optimization objective function of the self-encoding network model is determined as:

Among them, L(W,b) is the loss value, σ is the added noise, W is the weight of the encoding structure, W' is the weight of the decoding structure, the first term is the squared loss function, and the second term is the L2 regularization expression , to prevent the model from overfitting and increase the sparsity of the model.

Preferably, wherein the method further comprises:

Add a Dropout layer to the upper layer of the classifier to randomly discard features with a 1-r probability number to reduce overfitting;

Among them, r conforms to the Bernoulli distribution, which is: r～Bernoulli(p), p is the Bernoulli distribution parameter, the actual number of nodes used for classification is rh _t , and h _t is the output of the auto-encoding network.

Preferably, the grid method is used to adjust the number of layers of the self-encoding network model and the number of nodes in each layer, and p is adjusted according to the test accuracy.

According to another aspect of the present invention, a fault classification and detection system for smart electric energy meters based on a self-encoding network is provided, wherein the system includes:

The normalization processing module is used to normalize the acquired historical detection data of the smart electric energy meter, and divide the normalized historical detection data into a training set and a test set; wherein, the historical detection data It is a voltage signal sample or a current signal sample, including: normal signal data and fault signal data;

A parameter initialization setting module for initializing and setting parameters of the self-encoding network model; wherein, the parameters include: the number of layers of the self-encoding network model, the number of nodes in each layer, and the weight and bias of the encoding structure;

The self-encoding network model iterative training module is used to select the sample data of the number of nodes in the first layer in the training set and input it into the self-encoding network model to obtain signal features and input them into the classifier for classification, and according to the classification results Perform iterative training;

The optimal parameter determination module is used to calculate the test accuracy according to the classification results obtained by using the test set, and perform iterative operations according to the test accuracy, and continuously adjust the parameters of the self-encoding network model to determine the optimal value of the self-encoding network model. optimal parameters;

The fault analysis module is used for classifying and detecting the fault of the smart electric energy meter by using the self-encoding network model corresponding to the optimal parameter.

Preferably, the normalization processing module performs normalization processing on the acquired historical detection data of the smart energy meter, including:

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter, E(·) is the mean value of the signal x(n), and Std(·) is the standard deviation of the signal x(n) .

Preferably, wherein the system further comprises:

The data preprocessing module is used to eliminate redundant data and erroneous data in the acquired historical detection data of the smart energy meter before normalizing the acquired historical detection data of the smart energy meter, and eliminate the missing data. data is populated.

Preferably, wherein the self-encoding network model includes: an encoding structure and a decoding structure,

The coding structure is expressed as the following formula:

h _t =f(Wx'(n)+b) _t ,

The decoding structure is expressed as the following formula:

x _t =f(W'h _t +b') _t ,

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter f(·) is the activation function, W is the weight of the coding structure, b is the bias of the coding structure, t=1, 2,...,N is the number of layers of the self-encoding network model, n is the input dimension, and the dimensions of different layers are gradually reduced to effectively extract signal features; W' is the weight of the decoding structure, b' is the bias of the decoding structure, x _t is the output of the decoding structure, which is equal to the input of the encoding structure.

Preferably, in the parameter initialization setting module, the weights and biases of the coding structure are initialized by using a truncated normal distribution, including:

Among them, f(x) is the density function, u and δ are the mean and standard deviation of the normal distribution, respectively, and the value range of the weight W of the coding structure is (u-2δ, u+2δ).

Preferably, wherein the self-encoding network model iterative training module further includes:

The training is stopped when the loss value no longer decreases within the consecutive training cycles of the first preset number of thresholds or the test accuracy begins to decrease after consecutive training cycles of the second preset number of thresholds.

Preferably, wherein the self-encoding network model iterative training module further includes:

When using the self-encoding network model for iterative training, a preset ratio of noise is added to the input signal, and the optimization objective function of the self-encoding network model is determined as:

Among them, L(W,b) is the loss value, σ is the added noise, W is the weight of the encoding structure, W' is the weight of the decoding structure, the first term is the squared loss function, and the second term is the L2 regularization expression , to prevent the model from overfitting and increase the sparsity of the model.

Preferably, wherein the system further comprises:

Add a Dropout layer to the upper layer of the classifier to randomly discard features with a 1-r probability number to reduce overfitting;

Among them, r conforms to the Bernoulli distribution, which is: r～Bernoulli(p), p is the Bernoulli distribution parameter, the actual number of nodes used for classification is rh _t , and h _t is the output of the auto-encoding network.

Preferably, the grid method is used to adjust the number of layers of the self-encoding network model and the number of nodes in each layer, and p is adjusted according to the test accuracy.

The present invention provides a method and system for classifying and detecting faults of smart electric energy meters based on an auto-encoding network. The parameters of the coding network model; select sample data in the training set and input it into the self-coding network model, to obtain signal features and input them into the classifier for classification, and perform iterative training according to the classification results; According to the classification results of the test set The parameters of the self-encoding network model are continuously adjusted to determine the optimal parameters of the self-encoding network model; the faults of the smart electric energy meter are classified and detected by using the self-encoding network model corresponding to the optimal parameters. The invention firstly sends a large amount of unlabeled data into the self-encoding network for compression, and consists of a multi-layer encoder and a decoder to form a feature extraction network to extract the features of the fault signal and the normal signal; finally, a small number of labeled fault samples are used to classify the classification. device to classify. The invention adopts the deep noise reduction self-encoding network to perform unsupervised feature extraction on the collected signals, which is suitable for the condition of insufficient label information, and combines unsupervised learning and supervised classification to realize fast and accurate classification of fault signals, which is helpful for improving Compared with the traditional method, the fault identification ability of the smart energy meter has stronger noise immunity.

Description of drawings

Exemplary embodiments of the present invention may be more fully understood by reference to the following drawings:

FIG. 1 is a flowchart of a method 100 for classifying and detecting faults of smart electric energy meters based on a self-encoding network according to an embodiment of the present invention;

2 is a schematic diagram of a smart energy meter data transmission system according to an embodiment of the present invention;

3 is a schematic diagram of a self-encoding network model according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of deep noise reduction using an auto-encoding network model according to an embodiment of the present invention; and

FIG. 5 is a schematic structural diagram of a fault classification and detection system 500 for a smart electric energy meter based on a self-encoding network according to an embodiment of the present invention.

Detailed ways

Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings, however, the present invention may be embodied in many different forms and is not limited to the embodiments described herein, which are provided for the purpose of this thorough and complete disclosure invention, and fully convey the scope of the invention to those skilled in the art. The terms used in the exemplary embodiments shown in the drawings are not intended to limit the invention. In the drawings, the same elements/elements are given the same reference numerals.

Unless otherwise defined, terms (including scientific and technical terms) used herein have the commonly understood meanings to those skilled in the art. In addition, it is to be understood that terms defined in commonly used dictionaries should be construed as having meanings consistent with the context in the related art, and should not be construed as idealized or overly formal meanings.

FIG. 1 is a flowchart of a method 100 for classifying and detecting faults of a smart electric energy meter based on an auto-encoding network according to an embodiment of the present invention. As shown in FIG. 1 , the method for classifying and detecting faults of smart electric energy meters based on the self-encoding network provided by the embodiment of the present invention firstly sends a large amount of unlabeled data into the self-encoding network for compression, and is composed of a multi-layer encoder and a decoder. The feature extraction network extracts the features of the fault signal and the normal signal; finally, a small number of labeled fault samples are used to classify the classifier. In this way, the fault signal under the small sample of labeled data can be effectively monitored. The embodiment of the present invention adopts the deep noise reduction auto-encoding network to perform unsupervised feature extraction on the collected signals, which is suitable for the condition of insufficient label information, and combines unsupervised learning and supervised classification to achieve fast and accurate classification of fault signals. It helps to improve the fault identification ability of smart energy meters, and has stronger noise immunity than traditional methods. The method 100 for fault classification and detection of smart electric energy meters based on the self-encoding network provided by the embodiment of the present invention starts from step 101, and in step 101, the acquired historical detection data of the smart electric energy meter is normalized, and the normalized The processed historical detection data is divided into a training set and a test set; wherein, the historical detection data is a voltage signal sample or a current signal sample, including: normal signal data and fault signal data.

Preferably, the normalization of the acquired historical detection data of the smart electric energy meter includes:

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter, E(·) is the mean value of the signal x(n), and Std(·) is the standard deviation of the signal x(n) .

Preferably, wherein the method further comprises:

Before normalizing the acquired historical detection data of the smart electric energy meter, eliminate redundant data and erroneous data in the acquired historical detection data of the smart electric energy meter, and fill in the missing data.

In the embodiment of the present invention, the data transmission system as shown in FIG. 2 is used to obtain the historical data of the smart energy meter in a typical province, including: the signal during normal operation and the fault signal, and the signal can be a voltage signal or a current signal. at least one of. The smart electric energy meter samples the voltage and current signals and sends them to the base station server host through the concentrator. The communication method can be optical fiber or GPRS communication.

After obtaining the historical detection data of the smart electric energy meter, the spline interpolation method is used to interpolate the missing data, and the value, redundant data and error data when the digital signal value is greater than a certain threshold are directly eliminated. Then, the following normalization method is used to normalize all data, and the normalized historical detection data is divided into training set and test set. Among them, the acquired historical detection data of the smart energy meter is normalized, including:

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter, E(·) is the mean value of the signal x(n), and Std(·) is the standard deviation of the signal x(n) .

In step 102, parameters of the self-encoding network model are initially set; wherein, the parameters include: the number of layers of the self-encoding network model, the number of nodes in each layer, and the weight and bias of the encoding structure.

Preferably, wherein the self-encoding network model includes: an encoding structure and a decoding structure,

The coding structure is expressed as the following formula:

h _t =f(Wx'(n)+b) _t ,

The decoding structure is expressed as the following formula:

x _t =f(W'h _t +b') _t ,

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter f(·) is the activation function, W is the weight of the coding structure, b is the bias of the coding structure, t=1, 2,...,N is the number of layers of the self-encoding network model, n is the input dimension, and the dimensions of different layers are gradually reduced to effectively extract signal features; W' is the weight of the decoding structure, b' is the bias of the decoding structure, x _t is the output of the decoding structure, which is equal to the input of the encoding structure.

Preferably, the weights and biases of the coding structure are initialized using a truncated normal distribution, including:

Among them, f(x) is the density function, u and δ are the mean and standard deviation of the normal distribution, respectively, and the value range of the weight W of the coding structure is (u-2δ, u+2δ).

In an embodiment of the present invention, the autoencoder network model is a neural network composed of multiple layers of sparse autoencoders, and the output of each layer is connected to the input of the next layer. The self-encoding network model consists of an encoder and a decoder. For an N-layer deep encoder, the output of the encoding structure is expressed by the following equation: h _t =f(Wx'(n)+b) _t ; where f( ) is the activation function, W represents the weight of the coding structure, b represents the bias of the coding structure, and t=1,2,...,N represents the number of layers; where the input dimension is n, and the dimensions of different layers gradually reduced to effectively extract signal features.

FIG. 3 is a schematic diagram of a self-encoding network model according to an embodiment of the present invention. As shown in Figure 3, it is a 3-layer self-encoding network model.

For the activation function, it can be a linear rectification (RELU) activation function. Simpler and less computationally expensive.

Assuming that the number of output nodes of the encoder is set to m, the decoder structure can be expressed as: x _t =f(W'h _t +b') _t , where W' represents the weight of the decoding structure, and b' represents the bias of the decoding structure set, and the output xt is the decoder output, which is equal to the encoder input.

In the embodiment of the present invention, the weight initialization is performed by using a truncated normal distribution, and the expression of the initialization method is:

Among them, u and δ are the mean and standard deviation of the normal distribution, and the value range of the weight W is (u-2δ, u+2δ).

In step 103, the sample data of the number of nodes in the first layer is selected in the training set and input into the self-encoding network model to obtain signal features and input into the classifier for classification, and iterative training is performed according to the classification result.

Preferably, wherein the method further comprises:

When using the self-encoding network model for iterative training, a preset ratio of noise is added to the input signal, and the optimization objective function of the self-encoding network model is determined as:

Among them, L(W,b) is the loss value, σ is the added noise, W is the weight of the encoding structure, W' is the weight of the decoding structure, the first term is the squared loss function, and the second term is the L2 regularization expression , to prevent the model from overfitting and increase the sparsity of the model.

Preferably, wherein the method further comprises:

The iteration is stopped when the loss value no longer decreases within the consecutive training cycles of the first preset number threshold or the test accuracy begins to decrease after the consecutive training cycles of the second preset number threshold.

Preferably, wherein the method further comprises:

Add a Dropout layer to the upper layer of the classifier to randomly discard features with a 1-r probability number to reduce overfitting;

Among them, r conforms to the Bernoulli distribution, which is: r～Bernoulli(p), p is the Bernoulli distribution parameter, the actual number of nodes used for classification is rh _t , and h _t is the output of the auto-encoding network.

FIG. 4 is a schematic diagram of deep noise reduction using an auto-encoding network model according to an embodiment of the present invention. As shown in FIG. 4 , in order to improve the anti-noise performance of the model, noise reduction auto-encoding is used to train the model in the embodiment of the present invention, that is, a certain proportion of noise is added to the input signal, and the input signal becomes x(n )+σ. In order to solve the parameters of the deep denoising autoencoder model, the optimization objective function of the model is:

Among them, the first term is the square loss function, and the second term is the L2 regularization expression, the purpose is to prevent the model from overfitting and increase the sparsity of the model.

Select the sample data of the number of nodes in the first layer in the training set and input it into the self-encoding network model to obtain signal features, and input the obtained feature signal into the classifier for classification to obtain the classification result. When the loss value within the training period of the first preset number threshold is no longer reduced, the iterative training is stopped. Wherein, the first preset number threshold may be 5, 8, 10 and so on.

In the embodiment of the present invention, the classifier adopts the Softmax classifier. The expression for this classifier is:

Among them, h _l is the output of the deep noise reduction self-encoding coding structure, and K represents the number of fault types collected by the smart energy meter. The purpose of Softmax is to constrain the output value of the model to (0,1], and the position corresponding to the output maximum probability value is the corresponding fault category.

In addition, a Dropout layer is added to the upper layer of the classifier to randomly discard the features of the 1-r probability number to reduce overfitting; among them, r conforms to the Bernoulli distribution, which is: r～Bernoulli(p), p is the Bernoulli distribution parameter, the actual number of nodes used for classification is rh _t , and h _t is the output of the autoencoder network.

In step 104, the test accuracy is calculated according to the classification result obtained by using the test set, and an iterative operation is performed according to the test accuracy, and the parameters of the self-encoding network model are continuously adjusted to determine the optimal parameters of the self-encoding network model.

Preferably, the grid method is used to adjust the number of layers of the self-encoding network model and the number of nodes in each layer, and p is adjusted according to the test accuracy.

When the test set is used for testing, the test accuracy is calculated, and when the test accuracy begins to decrease after the second consecutive training period of the preset number threshold, the iteration is stopped; otherwise, the grid method is used to adjust the self-encoding network model. The number of layers and the number of nodes in each layer, and p is adjusted according to the test accuracy until the test accuracy meets the set requirements.

In step 105, the fault of the smart electric energy meter is classified and detected by using the self-encoding network model corresponding to the optimal parameter.

The trained deep noise reduction self-encoding model is placed in the host of the base station server, and the type of grid fault is analyzed in real time through the data collected by the smart energy meter.

The following specific examples illustrate the embodiments of the present invention

Taking the data of smart energy meters in a typical province as an example, the fault classification and detection method of smart energy meters based on self-coding network includes the following steps:

S1: Use the data transmission system as shown in Figure 2 to obtain the historical data of smart energy meters in a typical province, including voltage and current signals. The smart energy meter samples the voltage and current signals and sends them to the base station server host through the concentrator. The communication method can be optical fiber or GPRS communication.

S2: After removing and filling the acquired historical data of the smart electric energy meter, normalization is performed to reduce the influence of the dimension on the result. Specifically include:

Among them, x'(n) represents the voltage and current signals collected by the smart energy meter, E( ) represents the mean value of the signal x(n), and Std represents the standard deviation of the signal x(n).

In this embodiment, for the normalized signal, 70% is selected as the training set, and 30% is selected as the test set.

In this embodiment, 6 fault types are classified, and 1000 samples are taken for each fault type, that is, a total of 6000 fault samples. In addition, 1000 groups of normal samples are taken for analysis, that is, the total number of training samples is 7000. In addition, for the different lengths of different samples, the signals of all samples are uniformly down-sampled to 600 data points. That is, the length of each sample is 600, which ensures the unity of the data and the consistency of the model structure.

S3: Initialize the parameters of the auto-encoder network model, and use the auto-encoder to train all the data.

In order to improve the anti-noise performance of the model, the model is trained with noise reduction auto-encoding, that is, a certain proportion of noise is added to the input signal, and the input signal becomes x(n)+σ at this time.

In this embodiment, white Gaussian noise with signal-to-noise ratios of 20dB and 40dB is added to the input signal. In order to optimize the parameters of the deep denoising autoencoder model, the optimization function of the model is:

Among them, the first term is the square loss function, and the second term is the L2 regularization expression, the purpose is to prevent the model from overfitting and increase the sparsity of the model.

In this embodiment, the selected activation function is a linear rectification (RELU) activation function, and the expression is: f(x)=max(0,x); RELU can effectively ensure gradient descent solution and direction propagation, and avoid gradient disappearance The problem is that the calculation is simpler and the calculation amount of the network is reduced.

In this embodiment, a truncated normal distribution is used to initialize the weight and bias parameters, and the expression of the initialization method is:

Among them, u and δ are the mean and standard deviation of the normal distribution, then the value range of the corresponding weight W is (u-2δ, u+2δ).

When the model is trained using the self-encoding structure shown in Figure 4, the network structure includes a 5-layer deep denoising self-encoding network and a classification layer. In this embodiment, appropriate network hyperparameters are set according to the amount of data. The size of the input layer is set to 600 nodes, that is, the size of the first layer is 600, the number of nodes in the second layer is 300, and the number of nodes in the feature output layer of the third layer is set to 100, that is, m=100. The number of nodes in the fourth layer of the decoder structure is 300, and the number of nodes in the fifth layer is 600.

During training, when the loss value does not decrease for 10 consecutive training epochs, or when the test accuracy starts to decrease after 10 training epochs, the training is stopped.

Among them, the classifier adopts the Softmax classifier, and the expression of the classifier is:

Among them, hl is the output of the deep noise reduction self-encoding coding structure, and K represents the number of fault types collected by the smart energy meter. The purpose of Softmax is to constrain the objective function value to (0,1]. The position corresponding to the output maximum probability value is the corresponding fault category.

In order to prevent the model from overfitting, a layer of Dropout is added to the Softmax classifier to reduce model overfitting. After the network is added to this layer, the number of features with a certain probability of 1-r is randomly discarded to reduce overfitting, and the actual number of nodes used for classification is rht.

In this embodiment, the number of nodes used for the Softmax classifier is 100, and the parameter r of Dropout is set to 0.4.

S4: According to the performance of the model classifier, continuously adjust the hyperparameters of the deep noise reduction auto-encoding network. The hyperparameters that need to be adjusted mainly include the number of model layers, the number of nodes in each layer of the model, and the probability p of Dropout. Repeat the above steps until satisfied Preset test accuracy threshold.

S5: Place the trained deep noise reduction self-encoding model in the host of the base station server, and analyze the power grid fault type in real time through the data collected by the smart energy meter.

In this embodiment, the three-phase or single-phase voltage and current signals are respectively input into a plurality of self-encoding models according to the sample sequence. In this embodiment, single-phase voltage and current signals are used, that is, two deep noise reduction auto-encoding models are required for training.

In this embodiment, the grid method is used to search for the optimal hyperparameters of the model. Taking the number of network layers as an example, the number of layers is increased from 3 to 10 in order. According to the results of network fault diagnosis, after determining the optimal number of layers, Then perform a similar search on the number of model nodes.

In this embodiment, the number of layers is taken as an example, and the output result of the maximum probability of the Softmax classifier is used as the final fault classification result. The accuracy of the deep denoising autoencoder network is tested, and Table 1 shows the relationship between the number of layers and the accuracy. In order to prevent the interference of the number of nodes, the number of output nodes of the feature layer of the middle layer is set to 100.

Table 1. Results of fault signal diagnosis and classification of smart energy meters based on deep noise reduction self-encoding

It can be seen from Table 1 that with the increase of the number of layers of the deep noise reduction self-encoding network, the accuracy increases greatly, but when the number of layers is greater than a certain number, the accuracy begins to decline, indicating that the number of layers is the main factor affecting the fault diagnosis accuracy of smart energy meters one. When the number of layers increases, the model parameters increase, and the training time also increases. It shows that more model parameters have a certain influence on the fault diagnosis performance. In addition, for 6 kinds of fault signals, the comprehensive accuracy and training time can show that 5 layers is a better choice for the number of layers.

FIG. 5 is a schematic structural diagram of a fault classification and detection system 500 for a smart electric energy meter based on a self-encoding network according to an embodiment of the present invention. As shown in FIG. 5 , a fault classification and detection system 500 for smart electric energy meters based on an auto-encoding network provided by an embodiment of the present invention includes: a normalization processing module 501 , a parameter initialization setting module 502 , and an iterative training module 503 for an auto-encoding network model , an optimal parameter determination module 504 and a failure analysis module 505 .

Preferably, the normalization processing module 501 is used to normalize the acquired historical detection data of the smart electric energy meter, and divide the normalized historical detection data into a training set and a test set; wherein , the historical detection data is a voltage signal sample or a current signal sample, including: normal signal data and fault signal data.

Preferably, the normalization processing module 501 performs normalization processing on the acquired historical detection data of the smart energy meter, including:

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter, E(·) is the mean value of the signal x(n), and Std(·) is the standard deviation of the signal x(n) .

Preferably, the system further comprises: a data preprocessing module for eliminating redundancy in the acquired historical detection data of the smart energy meter before normalizing the acquired historical detection data of the smart energy meter Data and erroneous data were eliminated, and missing data were filled.

Preferably, the parameter initialization setting module 502 is used to initialize and set the parameters of the self-encoding network model; wherein, the parameters include: the number of layers of the self-encoding network model, the number of nodes in each layer, the weight and bias of the encoding structure .

Preferably, wherein the self-encoding network model includes: an encoding structure and a decoding structure, and the encoding structure is expressed as the following formula:

h _t =f(Wx'(n)+b) _t ,

The decoding structure is expressed as the following formula:

x _t =f(W'h _t +b') _t ,

Among them, x'(n) is the voltage signal or current signal of the normalized smart energy meter f(·) is the activation function, W is the weight of the coding structure, b is the bias of the coding structure, t=1, 2,...,N is the number of layers of the self-encoding network model, n is the input dimension, and the dimensions of different layers are gradually reduced to effectively extract signal features; W' is the weight of the decoding structure, b' is the bias of the decoding structure, x _t is the output of the decoding structure, which is equal to the input of the encoding structure.

Preferably, in the parameter initialization setting module 502, the weights and deviations of the coding structure are initialized by using a truncated normal distribution, including:

Among them, f(x) is the density function, u and δ are the mean and standard deviation of the normal distribution, respectively, and the value range of the weight W of the coding structure is (u-2δ, u+2δ).

Preferably, the self-encoding network model iterative training module 503 is used for selecting the sample data of the number of nodes in the first layer in the training set and inputting it into the self-encoding network model, so as to obtain signal features and input them into the classifier. Classification, and iterative training is performed according to the classification results.

Preferably, the self-encoding network model iterative training module 503 further includes: when the loss value is no longer reduced in the training period of the consecutive first preset number threshold or the test accuracy is in the second consecutive preset number When the number threshold starts to decrease after the training period, the training is stopped.

Preferably, the self-encoding network model iterative training module 503 further includes: when using the self-encoding network model for iterative training, adding a preset proportion of noise to the input signal, and determining the self-encoding network model The optimization objective function is:

Among them, L(W,b) is the loss value, σ is the added noise, W is the weight of the encoding structure, W' is the weight of the decoding structure, the first term is the squared loss function, and the second term is the L2 regularization expression , to prevent the model from overfitting and increase the sparsity of the model.

Preferably, the system further comprises: adding a Dropout layer to the upper layer of the classifier, for randomly discarding the features of the 1-r probability number, so as to reduce overfitting; wherein, r conforms to the Bernoulli distribution, which is: r～Bernoulli(p), p is the Bernoulli distribution parameter, the actual number of nodes used for classification is rh _t , h _t is the output of the auto-encoding network.

Preferably, the optimal parameter determination module 504 is configured to calculate the test accuracy according to the classification result obtained by using the test set, and perform an iterative operation according to the test accuracy, and continuously adjust the parameters of the self-encoding network model to determine the Optimal parameters of the autoencoder network model.

Preferably, the grid method is used to adjust the number of layers of the self-encoding network model and the number of nodes in each layer, and p is adjusted according to the test accuracy.

Preferably, the fault analysis module 505 is configured to use the self-encoding network model corresponding to the optimal parameter to classify and detect the fault of the smart electric energy meter.

The fault classification and detection system 500 based on the self-coding network of the smart electric energy meter of the embodiment of the present invention corresponds to the fault classification and detection method 100 of the smart electric energy meter based on the self-coding network of another embodiment of the present invention, and will not be repeated here.

The present invention has been described with reference to a few embodiments. However, as is known to those skilled in the art, other embodiments than the above disclosed invention are equally within the scope of the invention, as defined by the appended patent claims.

Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a/the/the [means, component, etc.]" are open to interpretation as at least one instance of said means, component, etc., unless expressly stated otherwise. The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.

Claims (18)

1. a kind of intelligent electric energy meter failure modes detection method based on autoencoder network, which is characterized in that the described method includes:

The history detection data of the intelligent electric energy meter of acquisition is normalized, and the history Jing Guo normalized is examined Measured data is divided into training set and test set；Wherein, the history detection data is voltage signal sample or current signal sample, packet It includes: normal signal data and fault-signal data；

The parameter of Initialize installation autoencoder network model；Wherein, the parameter includes: the number of plies, every of autoencoder network model Layer number of nodes, coding structure weight and biasing；

The sample data that the first-level nodes number is chosen in the training set is input in the autoencoder network model, to obtain Signal characteristic, which is input in classifier, classifies, and is iterated training according to classification results；

Measuring accuracy is calculated according to the classification results obtained using test set, and operation is iterated according to the measuring accuracy, The autoencoder network model parameter is adjusted, constantly with the optimized parameter of the determination autoencoder network model；

Classification and Detection is carried out using failure of the corresponding autoencoder network model of the optimized parameter to intelligent electric energy meter.

2. the method according to claim 1, wherein the history detection data of the intelligent electric energy meter of described pair of acquisition It is normalized, comprising:

Wherein, x'(n) be intelligent electric energy meter by normalized voltage signal or current signal, E () is signal x (n) mean value, Std () are the standard deviation of signal x (n).

3. the method according to claim 1, wherein the method also includes:

Before the history detection data to the intelligent electric energy meter of acquisition is normalized, to the intelligent electric energy meter of acquisition Rejecting redundant data and wrong data in history detection data are rejected, and are filled to missing data.

4. the method according to claim 1, wherein the autoencoder network model includes: coding structure reconciliation Code structure,

The coding structure is expressed as following formula:

h_t=f (Wx'(n)+b)_t,

The decoding representation is following formula:

x_t=f (W'h_t+b')_t,

Wherein, x'(n) it is the voltage signal of intelligent electric energy meter by normalized or current signal f () be to activate letter Number, W are the weight of coding structure, and b is the biasing of coding structure, and t=1,2 ..., N are the number of plies of autoencoder network model, and n is Dimension is inputted, the dimension of different layers gradually decreases, effectively to extract signal characteristic；W ' is the weight for decoding structure, and b ' is decoding The biasing of structure, x_tIt is equal with the input of coding structure for the output for decoding structure.

5. according to the method described in claim 4, it is characterized in that, using transversal normal distribution to the weight of coding structure and partially Difference is initialized, comprising:

Wherein, f (x) is density function, and u and δ are respectively the mean value and standard deviation of normal distribution, and the weight W's of coding structure takes Being worth range is (u-2 δ, u+2 δ).

6. the method according to claim 1, wherein the method also includes:

When the penalty values in the cycle of training in continuous first predetermined number threshold value no longer reduce or measuring accuracy is continuous When starting to reduce after the cycle of training of the second predetermined number threshold value, stop iteration.

7. the method according to claim 1, wherein the method also includes:

When being iterated trained using the autoencoder network model, input signal is added the noise of preset ratio, and really The optimization object function of the fixed autoencoder network model are as follows:

Wherein, L (W, b) is penalty values, and σ is the noise being added, and W is the weight of coding structure, and W ' is the weight for decoding structure, the One is quadratic loss function, and Section 2 increases the sparsity of model for L2 regularization expression formula to prevent model over-fitting.

8. the method according to claim 1, wherein the method also includes:

In upper one layer of Dropout layers of addition of classifier, for giving up the feature of 1-r probability quantity at random, to mitigate over-fitting；

Wherein, r meets Bernoulli Jacob's distribution, are as follows: r~Bernoulli (p), p are Bernoulli Jacob's distribution parameter, the reality for classification Number of nodes is rh_t, h_tFor the output of autoencoder network.

9. according to the method described in claim 8, it is characterized in that, adjusting the layer of the autoencoder network model using gridding method Several and every layer of number of nodes adjusts p according to measuring accuracy.

10. a kind of intelligent electric energy meter failure modes detection system based on autoencoder network, which is characterized in that the system packet It includes:

Normalized module, the history detection data for the intelligent electric energy meter to acquisition are normalized, and will be through The history detection data for crossing normalized is divided into training set and test set；Wherein, the history detection data is voltage signal Sample or current signal sample, comprising: normal signal data and fault-signal data；

Parameter initialization setup module, the parameter for Initialize installation autoencoder network model；Wherein, the parameter includes: The weight and biasing of the number of plies of autoencoder network model, every layer of number of nodes, coding structure；

Autoencoder network model repetitive exercise module, the sample data for choosing the first-level nodes number in the training set are defeated Enter into the autoencoder network model, is classified with obtaining signal characteristic and being input in classifier, and according to classification results It is iterated training；

Optimized parameter determining module, for calculating measuring accuracy according to the classification results obtained using test set, and according to described Measuring accuracy is iterated operation, constantly adjusts the autoencoder network model parameter, with the determination autoencoder network model Optimized parameter；

Failure analysis module, for using the corresponding autoencoder network model of the optimized parameter to the failure of intelligent electric energy meter into Row classification and Detection.

11. system according to claim 10, which is characterized in that the normalized module, to the intelligence electricity of acquisition The history detection data of energy table is normalized, comprising:

Wherein, x'(n) be intelligent electric energy meter by normalized voltage signal or current signal, E () is signal x (n) mean value, Std () are the standard deviation of signal x (n).

12. system according to claim 10, which is characterized in that the system also includes:

Data preprocessing module, for before the history detection data to the intelligent electric energy meter of acquisition is normalized, To in the history detection data of the intelligent electric energy meter of acquisition rejecting redundant data and wrong data reject, and to missing number According to being filled.

13. system according to claim 10, which is characterized in that the autoencoder network model include: coding structure and Structure is decoded,

The coding structure is expressed as following formula:

h_t=f (Wx'(n)+b)_t,

The decoding representation is following formula:

x_t=f (W'h_t+b')_t,

Wherein, x'(n) it is the voltage signal of intelligent electric energy meter by normalized or current signal f () be to activate letter Number, W are the weight of coding structure, and b is the biasing of coding structure, and t=1,2 ..., N are the number of plies of autoencoder network model, and n is Dimension is inputted, the dimension of different layers gradually decreases, effectively to extract signal characteristic；W ' is the weight for decoding structure, and b ' is decoding The biasing of structure, x_tIt is equal with the input of coding structure for the output for decoding structure.

14. system according to claim 13, which is characterized in that in the parameter initialization setup module, using truncation Normal distribution initializes the weight and deviation of coding structure, comprising:

Wherein, f (x) is density function, and u and δ are respectively the mean value and standard deviation of normal distribution, and the weight W's of coding structure takes Being worth range is (u-2 δ, u+2 δ).

15. system according to claim 10, which is characterized in that the autoencoder network model repetitive exercise module, also Include:

When the penalty values in the cycle of training in continuous first predetermined number threshold value no longer reduce or measuring accuracy is continuous When starting to reduce after the cycle of training of the second predetermined number threshold value, deconditioning.

16. system according to claim 10, which is characterized in that the autoencoder network model repetitive exercise module, also Include:

When being iterated trained using the autoencoder network model, input signal is added the noise of preset ratio, and really The optimization object function of the fixed autoencoder network model are as follows:

Wherein, L (W, b) is penalty values, and σ is the noise being added, and W is the weight of coding structure, and W ' is the weight for decoding structure, the One is quadratic loss function, and Section 2 increases the sparsity of model for L2 regularization expression formula to prevent model over-fitting.

17. system according to claim 10, which is characterized in that the system also includes:

In upper one layer of Dropout layers of addition of classifier, for giving up the feature of 1-r probability quantity at random, to mitigate over-fitting；

Wherein, r meets Bernoulli Jacob's distribution, are as follows: r~Bernoulli (p), p are Bernoulli Jacob's distribution parameter, the reality for classification Number of nodes is rh_t, h_tFor the output of autoencoder network.

18. system according to claim 17, which is characterized in that adjust the autoencoder network model using gridding method The number of plies and every layer of number of nodes adjust p according to measuring accuracy.