CN111123048A - A serial fault arc detection device and method based on convolutional neural network - Google Patents

Abstract

The invention provides a serial fault arc detection device based on a convolutional neural network, comprising a line current signal acquisition unit, a data processing unit, a fault arc model construction unit and a result output unit; the line current signal acquisition unit collects and converts arc current data; The data processing unit organizes the arc current data into a current sample data set, and establishes an arc current database to form a sample training set and a sample test set; the fault arc detection model building unit normalizes the samples, and builds a model based on convolutional neural networks. The series arc fault detection model of the network is trained and tested, and a trained series arc fault detection model is obtained; the result output unit obtains the arc current data to be detected and imports it into the trained series arc fault detection model to determine the arc current to be detected. Whether the data is a series arc fault. The implementation of the invention has the characteristics of strong generalization ability, high detection accuracy, and low misjudgment rate.

Description

A serial fault arc detection device and method based on convolutional neural network

technical field

The invention relates to the technical field of arc detection, in particular to a serial fault arc detection device and method based on a convolutional neural network.

Background technique

Arc, commonly known as "electric spark", has the characteristics of high temperature, small current and short duration. occurs frequently. Since a large amount of heat is generated during arc discharge, which can ignite surrounding flammable and explosive materials, resulting in fire or even explosion, fault arc is identified as an important cause of electrical fire, and it is necessary to conduct in-depth research on fault arc detection.

There are three main types of arc faults, including series arc fault, parallel fault arc and ground fault arc. Among them, the series arc fault is more difficult to detect and relatively dangerous because the current is small when the fault occurs, and the characteristics of the arc fault are related to the type of load connected in series.

At present, in the research of arc fault, the most used research object is the current data when the arc fault occurs, extract the characteristic quantity that can characterize the arc fault in the current, and set the corresponding threshold for detection. For example, the application announcement number is CN103116093A, and the title is an invention patent entitled "Series Arc Fault Early Warning System and Its Detection Method". The series fault arc early warning system of the invention patent includes a power supply circuit, a current sensor, a current sensing circuit, a signal conditioning circuit, a voltage The zero-crossing comparison circuit, the microprocessor and the fault output circuit realize the acquisition of the current signal, process the current signal, and use the adjacent cycle current waveform comparison and threshold for fault judgment, avoiding the complicated calculation in the frequency domain; another example , the application announcement number is CN102981088B, and the invention patent is named "Arc Fault Detection Method". The invention patent is mainly based on the feature extraction of the time domain. By collecting the current data of each cycle to analyze whether the current waveform has lost periodicity, positive and negative Semi-circular asymmetry, flat shoulders and excessive rate of change to determine whether an arc fault has occurred.

Although the above fault arc research can detect the fault arc to a certain extent, it is easy to cause misjudgment when the threshold value changes due to the change of the power consumption environment. Therefore, for security reasons, the stability and generalization ability of the system need to be further improved.

Therefore, in view of the large problem of the series arc fault, there is an urgent need for a series arc fault detection device, which has strong generalization ability, high detection accuracy, and can reduce the misjudgment rate.

SUMMARY OF THE INVENTION

The technical problem to be solved by the embodiments of the present invention is to provide a serial fault arc detection device and method based on a convolutional neural network, which has the characteristics of strong generalization ability, high detection accuracy, and low misjudgment rate.

In order to solve the above technical problems, an embodiment of the present invention provides a series arc fault detection device based on a convolutional neural network, including a line current signal acquisition unit, a data processing unit, a fault arc model construction unit, and a result output unit; wherein,

The line current signal acquisition unit is used to collect and convert the current signal flowing through the arc fault generator when a different load is correspondingly loaded into corresponding arc current data;

The data processing unit is used to organize the collected arc current data into respective corresponding current sample data sets according to the types of different loads, and mark each current sample according to whether it is a series fault arc to establish an arc current database, and further in the arc current database, select corresponding arc current data to form a sample training set and a sample test set respectively;

The arc fault detection model construction unit is used to normalize the sample data in the sample training set and the sample test set, and construct a series arc fault detection model based on a convolutional neural network, and further normalize the normalized arc fault detection model. The unified sample training set and sample test set are imported into the series arc fault detection model for training and testing, and a trained series arc fault detection model is obtained;

The result output unit is used for acquiring the arc current data to be detected, and importing the arc current data to be detected into the obtained trained series arc fault detection model, and according to the obtained trained series arc fault detection model The output tag type result is used to determine whether the arc current data to be detected is a series arc fault.

Wherein, the line current signal acquisition unit includes a current transformer, a signal conditioning circuit and a signal conversion processing module; wherein,

The current transformer is used to collect and convert the current signal flowing through the arc fault generator when different loads are loaded each time, and convert it into a voltage signal;

The signal conditioning circuit is used to process the voltage signal, including amplifying the voltage signal, performing low-pass filtering on the voltage signal, and adjusting the level value of the voltage signal through the virtual ground voltage VGND are raised to positive values;

The signal conversion processing module is used for ADC sampling to convert the voltage signal processed by the signal conditioning circuit into corresponding arc current data.

Wherein, the current transformer is a voltage-type current transformer, and the voltage-type current transformer directly converts a current signal into a voltage signal by connecting a sampling resistor in parallel.

Wherein, the signal conditioning circuit includes an operational amplifier IC1A, a resistor R1, a resistor R4 and a capacitor C1; wherein,

The non-inverting input terminal of the operational amplifier IC1A is connected to the virtual ground voltage VGND, the negative phase input terminal is connected to the resistor R4, and the output terminal is connected to the signal conversion processing module;

The resistor R4 and the capacitor C1 are connected in parallel to form a low-pass filter circuit, and both ends of the low-pass filter circuit are connected to the non-inverting input and output of the operational amplifier IC1A;

The ratio of the resistor R4 to the resistor R1 determines the magnification.

Wherein, the arc current database is realized by using MATLAB.

An embodiment of the present invention also provides a method for detecting a series arc fault based on a convolutional neural network, the method comprising the following steps:

Collect and convert the current signal flowing through the fault arc generator when different loads are loaded each time, and convert it into corresponding arc current data;

According to the types of different loads, the collected arc current data is organized into respective corresponding current sample data sets, and each current sample is marked according to whether it is a series fault arc, an arc current database is established, and further in the arc In the current database, select the corresponding arc current data to form a sample training set and a sample test set respectively;

Normalize the sample data in the sample training set and the sample test set, build a series fault arc detection model based on a convolutional neural network, and further normalize the sample training set and sample test after normalization. The set is imported into the series arc fault detection model for training and testing, and a trained series arc fault detection model is obtained;

Obtain the arc current data to be detected, import the arc current data to be detected into the obtained trained series arc fault detection model, and determine according to the result of the label type output by the obtained trained series arc fault detection model Whether the arc current data to be detected is a series fault arc.

Among them, the series arc fault detection model is established by using the deep learning framework Keras to realize the convolutional neural network, and defines a batch function module, a data reading module, a CNN structure module, and a training and testing module; wherein, the convolutional The neural network has four convolutional layers, two pooling layers, one dropout layer, and one fully connected layer.

Implementing the embodiment of the present invention has the following beneficial effects:

The invention can perform self-learning on the characteristics of the current flowing through the arc fault generator when different loads are loaded each time, and obtain the parameters of the series arc fault detection model based on the convolutional neural network through training of a large number of samples, without artificial Feature extraction and threshold setting ensure that the series arc fault detection model based on convolutional neural network can detect series arc faults under different load types. It has good generalization ability and improves the accuracy of detection and false positive rate Low.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention, and for those of ordinary skill in the art, obtaining other drawings according to these drawings still belongs to the scope of the present invention without any creative effort.

FIG. 1 is a schematic diagram of the system structure of a serial arc fault detection device based on a convolutional neural network provided by an embodiment of the present invention;

FIG. 2 is an application scenario diagram in which a fault arc generator is loaded with different loads in a convolutional neural network-based series arc fault detection device provided by an embodiment of the present invention;

Fig. 3 is the circuit connection diagram of the signal conditioning circuit included in the line current signal acquisition unit in Fig. 1;

Fig. 4 is the structural representation of the arc current database constructed by the data processing unit in Fig. 1;

Fig. 5 is the structural schematic diagram of the convolutional neural network adopted by the arc fault model construction unit in Fig. 1;

FIG. 6 is a flowchart of a method for detecting a series arc fault based on a convolutional neural network provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

As shown in FIG. 1 , in the embodiment of the present invention, a convolutional neural network-based series arc fault detection device is provided, including a line current signal acquisition unit 1, a data processing unit 2, a fault arc model construction unit 3 and a result output unit 4; wherein,

The line current signal acquisition unit 1 is used to collect and convert the current signal flowing through the arc fault generator when a different load (as shown in FIG. 2 ) is correspondingly loaded into corresponding arc current data;

The data processing unit 2 is used to organize the collected arc current data into respective corresponding current sample data sets according to the types of different loads, and mark each current sample according to whether it is a series fault arc, and establish an arc current database. , and further in the arc current database, select the corresponding arc current data to form a sample training set and a sample test set respectively;

The arc fault detection model building unit 3 is used to normalize the sample data in the sample training set and the sample test set, and build a series arc fault detection model based on a convolutional neural network, and further normalize the processed The sample training set and sample test set are imported into the series arc fault detection model for training and testing, and the trained series arc fault detection model is obtained;

The result output unit 4 is used to obtain the arc current data to be detected, and import the arc current data to be detected into the obtained trained series arc fault detection model, and according to the obtained trained series arc fault detection model The output tag type result is used to determine whether the arc current data to be detected is a series arc fault.

In the embodiment of the present invention, the line current signal acquisition unit 1 includes a current transformer, a signal conditioning circuit, and a signal conversion processing module; wherein,

The current transformer is used to collect and convert the current signal that flows through the arc fault generator when different loads are loaded each time, and convert it into a voltage signal; in one embodiment, the current transformer is a voltage type current transformer, The voltage-type current transformer directly converts the current signal into a voltage signal by connecting a sampling resistor in parallel;

A signal conditioning circuit for processing the voltage signal, including amplifying the voltage signal, performing low-pass filtering on the voltage signal, and increasing the level value of the voltage signal through the virtual ground voltage VGND is a positive value; in one embodiment, as shown in Figure 3, the signal conditioning circuit includes an operational amplifier IC1A, a resistor R1, a resistor R4 and a capacitor C1; the non-inverting input terminal of the operational amplifier IC1A is connected to the virtual ground voltage VGND, and the negative phase input terminal The resistor R4 is connected, and the output end is connected to the signal conversion processing module; the resistor R4 and the capacitor C1 are connected in parallel to form a low-pass filter circuit, and both ends of the low-pass filter circuit are connected to the non-inverting input end of the operational amplifier IC1A and the The output terminal is connected; the ratio of the resistance R4 and the resistance R1 determines the magnification;

The signal conversion processing module is used for ADC sampling to convert the voltage signal processed by the signal conditioning circuit into corresponding arc current data.

In this embodiment of the present invention, the data processing unit 2 is implemented by using MATLAB. Set the sampling frequency f of the current to 5KHz, and under the power frequency of 50Hz, it is calculated that the number of sampling points in each cycle is 100 sampling points, and 5 consecutive complete cycles are taken as a current sample (or more than 5 cycles), that is, Each current sample contains 500 sampling points, and each sample is marked with a corresponding label.

As shown in Figure 4, it is a schematic diagram of the division structure of the arc current database; where L represents a load, this example takes a fluorescent lamp as an example, and the collected data samples include the current sample data set A under the arc fault of the fluorescent lamp and the current sample data set A under normal conditions. Current sample dataset M. On the basis of A and M, according to a certain ratio, it can generally be divided into sample training set B and sample test set G at a ratio of 4:1, and sample training set B continues to divide training set E and verification set F according to the ratio. The ratio is generally 4:1, and the arc current database under normal conditions and fault conditions of different load types is established in this way.

In the embodiment of the present invention, the arc fault detection model building unit 3 adopts the deep learning framework Keras to implement a convolutional neural network to establish a series arc fault detection model based on the convolutional neural network. The series arc fault detection model defines a batch function module , data reading module, CNN structure module, training and testing module; among them, the convolutional neural network has four convolutional layers, two pooling layers, one loss layer and one fully connected layer. Because the data of the current waveform is one-dimensional, the neural network also adopts a one-dimensional convolutional neural network. The width of each layer in the neural network is 1. The default "size" below refers to the length of the size.

First, build the sample normalization module. The sample normalization module performs normalization processing for each sampling point in each sample to eliminate the unit limitation of the data and convert it into a dimensionless pure value. Among them, x represents a sample matrix, the matrix elements are the 500 sampling points of the sample, x represents the average value of the sampling points in the sample matrix, s represents the standard deviation of the sampling points in the sample matrix, and y _i represents the i-th element in the matrix. The conversion value is calculated as follows:

After normalization, the data samples are imported into the convolutional neural network structure as the input layer;

Second, build a convolutional neural network module. The convolutional neural network module adopts the structural schematic diagram of the convolutional neural network shown in FIG. 5 . The convolutional neural network includes four convolution layers H ₁ to H ₄ , two pooling layers P ₁ and P ₂ , a loss layer I ₁ and a fully connected layer J ₁ . The connection sequence between layers is H ₁ -H ₂ -P ₁ -H ₃ -H ₄ -P ₂ -I ₁ -J ₁ . The design of each layer is different, where M ₁ to M ₆ are the number of each layer, Q ₁ to Q ₅ represent the size of each layer, and the size of the samples processed by each layer is the size of _Ni and the size of the convolution kernel. Dimension Q _i-1 is relevant.

Taking H ₁ as an example, the input N ₁ is 500. In this example, the layer contains M ₁ convolution kernels of size Q ₁ , where M ₁ =100, Q ₁ =3, the sample size processed by this layer It becomes N ₂ =N ₁ -Q ₁ +1=498. The calculation of the size of other convolutional layers in this example is analogous. This process realizes the feature extraction of current samples; in particular, after the maximum pooling operation of the pooling layer P ₁ , the calculation formula of N ₄ in the size of the sample is N ₄ =N ₃ /Q ₃ , where Q ₃ is taken as 5, and N ₃ is calculated as 496 after the previous convolutional layer, and the pooled sample size N ₄ =99 can be obtained by calculation. This process further changes the learned features. Extraction; P ₂ is global average pooling, its size N ₆ is the same as the sample size, N ₆ is 95, after pooling, an output of size M ₆ × 1 is obtained, M ₆ is 160, the above related parameters All are set according to the specific experimental results.

Finally, build the training and testing modules. In order to enable the network to train excellent model parameters, the training and testing module adopts the cross entropy as the loss function of the neural network considering that the series fault arc detection is a classification problem, and its expression is as follows:

表示当前输入样本属于k类的真实标签分布向量。

表示模型对当前样本预测为属于k类的预测标签分布向量。采用Adam优化算法，对模型进行优化，Adam优化是一种基于梯度下降的算法，其可以实现自适应的学习率，能更快更好地训练模型。Among them, y represents the true label distribution vector, f(x; θ) represents the learned model, x is the input sample, and θ is the model parameter. m represents the number of samples, K represents the number of categories of sample labels, which is equal to 2 here,

Represents the true label distribution vector that the current input sample belongs to k classes.

Represents the predicted label distribution vector that the model predicts for the current sample to belong to k classes. The Adam optimization algorithm is used to optimize the model. Adam optimization is an algorithm based on gradient descent, which can achieve an adaptive learning rate and train the model faster and better.

Taking the load L (fluorescent lamp) in Fig. 4 as an example, during the training process, the training set E is input, and after the training of each sample x is completed, the calculation result of the above loss function is calculated. At the same time, the verification set F is input to verify the performance of the model. When the value of the loss function is close to 0 and remains stable, the training will end when the training standard is met. Otherwise, the model will continue to be trained based on the Adam algorithm until the optimal model parameters are obtained.

After the network training, the loss function is below 0.0007, which is close to 0; in the testing process, the test set G data is imported into the trained model, and the network model is evaluated according to the evaluation criteria. The selected evaluation indicators include accuracy and recall. The calculation method of the accuracy rate is the ratio of the number of predicted accurate samples to the number of all test samples, and the calculation method of the recall rate is the ratio of the number of samples with arc faults to all the arc faults in the test samples. When the precision and recall rates approach 1, it means that the network model can detect series arc faults.

After the test is successful, the new current samples can be directly input for classification without further training. Classification is achieved by a softmax classifier in the last layer of the network, as shown below

Where w _c is the connection weight vector between the output of the c-th class and the previous layer, first calculate the probability P(y=c|x) that the sample x belongs to the c-th class, c∈[1,C], and finally take the one with the largest probability category as the category of the current sample. Here there are only two categories, so C is 2. In the end, the fault arc is judged by a probability model, rather than by setting a fixed threshold. Among them, C is 1 for normal arc, C is 2 for series fault arc.

In this specific embodiment, the accuracy rate on the test set data reaches more than 99.97%, and the recall rate is 100%, indicating that the accuracy and reliability of the network judgment are high.

In the embodiment of the present invention, the result output unit 4 re-acquires the arc current data to be detected and imports it into the above-mentioned trained series arc fault detection model, and selects the category label with the highest probability to determine whether the arc current data to be detected is a series arc fault. .

As shown in FIG. 6 , in an embodiment of the present invention, a method for detecting a series arc fault based on a convolutional neural network is provided, and the method includes the following steps:

Step S1, collecting and converting the current signal that flows through the arc fault generator when different loads are correspondingly loaded each time into corresponding arc current data;

Step S2, according to the types of different loads, organize the collected arc current data into respective corresponding current sample data sets, and mark each current sample according to whether it is a series fault arc, establish an arc current database, and further In the arc current database, select corresponding arc current data to form a sample training set and a sample test set respectively;

Step S3, normalize the sample data in the sample training set and the sample test set, and construct a series arc fault detection model based on a convolutional neural network, and further normalize the sample training set after processing. Import the series arc fault detection model with the sample test set for training and testing, and obtain a trained series arc fault detection model;

Step S4, obtaining the arc current data to be detected, and importing the arc current data to be detected into the obtained trained series arc fault detection model, according to the label type output by the obtained trained series arc fault detection model As a result, it is determined whether the arc current data to be detected is a series arc fault.

Among them, the series arc fault detection model is established by using the deep learning framework Keras to realize the convolutional neural network, and defines a batch function module, a data reading module, a CNN structure module, and a training and testing module; wherein, the convolutional The neural network has four convolutional layers, two pooling layers, one dropout layer, and one fully connected layer.

Implementing the embodiment of the present invention has the following beneficial effects:

The invention can perform self-learning on the characteristics of the current flowing through the arc fault generator when different loads are loaded each time, and obtain the parameters of the series arc fault detection model based on the convolutional neural network through training of a large number of samples, without artificial Feature extraction and threshold setting ensure that the series arc fault detection model based on convolutional neural network can detect series arc faults under different load types. It has good generalization ability and improves the accuracy of detection and false positive rate Low.

It is worth noting that, in the above device embodiments, the modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be implemented; in addition, the specific names of the functional units It is only for the convenience of distinguishing from each other, and is not used to limit the protection scope of the present invention.

Those of ordinary skill in the art can understand that all or part of the steps in the methods of the above embodiments can be implemented by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium, such as ROM/RAM, magnetic disk, optical disk, etc.

What is disclosed above is only a preferred embodiment of the present invention, and of course it cannot limit the scope of the rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims (6)

1. A series fault arc detection device based on a convolutional neural network is characterized by comprising a line current signal acquisition unit, a data processing unit, a fault arc model construction unit and a result output unit; wherein,

the line current signal acquisition unit is used for acquiring current signals flowing through the fault arc generator when different loads are correspondingly loaded each time and converting the current signals into corresponding arc current data;

the data processing unit is used for sorting the acquired arc current data into respective corresponding current sample data sets according to the types of different loads, marking each current sample according to whether the current sample is a series fault arc, establishing an arc current database, and further selecting corresponding arc current data from the arc current database to form a sample training set and a sample testing set respectively;

the fault arc detection model building unit is used for carrying out normalization processing on the sample data in the sample training set and the sample testing set, building a series fault arc detection model based on a convolutional neural network, and further importing the sample training set and the sample testing set after the normalization processing into the series fault arc detection model for training and testing to obtain a trained series fault arc detection model;

and the result output unit is used for acquiring arc current data to be detected, importing the arc current data to be detected into the obtained trained series fault arc detection model, and determining whether the arc current data to be detected is the series fault arc according to the label type result output by the obtained trained series fault arc detection model.

2. The convolutional neural network-based series fault arc detection device as claimed in claim 1, wherein the line current signal acquisition unit comprises a current transformer, a signal conditioning circuit and a signal conversion processing module; wherein,

the current transformer is used for collecting current signals flowing through the fault arc generator when different loads are correspondingly loaded each time and converting the current signals into voltage signals;

the signal conditioning circuit is used for processing the voltage signal, and comprises the steps of amplifying the voltage signal, carrying out low-pass filtering on the voltage signal and raising the level value of the voltage signal to a positive value through virtual ground voltage VGND;

and the signal conversion processing module is used for carrying out ADC sampling on the voltage signal processed by the signal conditioning circuit and converting the voltage signal into corresponding arc current data.

3. The convolutional neural network-based series fault arc detection device as claimed in claim 2, wherein the current transformer is a voltage-type current transformer which directly converts a current signal into a voltage signal by connecting sampling resistors in parallel.

4. The convolutional neural network-based series fault arc detection device of claim 2, wherein the signal conditioning circuit comprises an operational amplifier IC1A, a resistor R1, a resistor R4, and a capacitor C1; wherein,

the non-inverting input end of the operational amplifier IC1A is connected with a virtual ground voltage VGND, the negative phase input end is connected with a resistor R4, and the output end is connected with the signal conversion processing module;

the resistor R4 and the capacitor C1 are connected in parallel to form a low-pass filter circuit, and two ends of the low-pass filter circuit are connected with the non-inverting input end and the output end of the operational amplifier IC 1A;

the ratio of the resistance R4 to the resistance R1 determines the amplification.

5. A series fault arc detection method based on a convolutional neural network is characterized by comprising the following steps:

collecting current signals flowing through the fault arc generator when different loads are correspondingly loaded each time, and converting the current signals into corresponding arc current data;

the method comprises the steps that collected arc current data are arranged into corresponding current sample data sets according to the types of different loads, an arc current database is established according to whether series fault arcs are marked for each current sample, and corresponding arc current data are further selected from the arc current database to form a sample training set and a sample testing set respectively;

normalizing the sample data in the sample training set and the sample testing set, constructing a series fault arc detection model based on a convolutional neural network, and further importing the sample training set and the sample testing set after the normalization processing into the series fault arc detection model for training and testing to obtain a trained series fault arc detection model;

and acquiring arc current data to be detected, importing the arc current data to be detected into the obtained trained series fault arc detection model, and determining whether the arc current data to be detected is the series fault arc according to a label type result output by the obtained trained series fault arc detection model.

6. The convolutional neural network-based series fault arc detection method as claimed in claim 5, wherein the series fault arc detection model is established by implementing a convolutional neural network by using a deep learning framework Keras, and is defined by a batch function module, a data reading module, a CNN structure module, and a training and testing module; wherein the convolutional neural network has four convolutional layers, two pooling layers, a loss layer and a full-link layer.

Images