CN111060304A

CN111060304A - Circuit breaker fault diagnosis method based on neural network

Info

Publication number: CN111060304A
Application number: CN202010078411.8A
Authority: CN
Inventors: 周福举; 庞吉年; 薛风华
Original assignee: State Grid Jiangsu Electric Power Co Ltd Suqian Power Supply Branch; State Grid Corp of China SGCC; State Grid Jiangsu Electric Power Co Ltd
Current assignee: State Grid Jiangsu Electric Power Co ltd Suqian Power Supply Branch
Priority date: 2020-02-03
Filing date: 2020-02-03
Publication date: 2020-04-24

Abstract

The invention relates to a fault diagnosis method for circuit breakers in a power system, in particular to a fault diagnosis method based on the combination of a deep neural network and a BP neural network. The vibration signals of circuit breakers are different in different states. Firstly, the wavelet packet-energy entropy technology is used to extract the characteristic quantities of vibration signals of vacuum circuit breakers in different states, and then the deep neural network is used to process these eigenvectors. It is classified with the normal state once, and finally the fault state is specifically classified through the BP neural network. Compared with the method of directly using the BP neural network to diagnose the fault of the circuit breaker, the present invention is significantly improved in terms of accuracy and speed.

Description

Circuit breaker fault diagnosis method based on neural network

Technical Field

The invention relates to a fault diagnosis method of a circuit breaker in a power system, in particular to a fault diagnosis method based on combination of a deep neural network and a BP neural network.

Background

The vacuum circuit breaker is used as a switch and a protection device of a power system, is widely applied to a power distribution system, and has the advantages of small volume, low noise, high reliability and the like. When a local fault occurs in a power system, once the fault cannot be timely removed due to the fault of a breaker, large-area power failure in the area is very likely to be caused. Therefore, the vacuum circuit breaker can reliably act, and is particularly important for safe and stable operation of a power system. According to the data, the mechanical faults caused by the problems of the control circuit of the circuit breaker, the operating mechanism and the like account for 70-80% of all the faults.

At present, the fault diagnosis research work aiming at the circuit breaker is successively developed at home and abroad, and in the aspect of the application of an intelligent algorithm, particularly, a neural network is increasingly introduced into the fault diagnosis of the circuit breaker. Such as Deep Neural Networks (DNNs), BP neural networks, radial basis function networks (RBFs), self-organizing map networks (SOMs), and the like. All weights need to be corrected when DNN is input every time, and the DNN cannot meet the real-time requirement when used globally; the BP neural network is used for global influence on convergence speed and is easily limited to a local minimum value; the RBF cannot fully identify the fault type; the SOM employs unsupervised learning rules and lacks classification information.

Disclosure of Invention

The invention further researches the problems and provides a breaker fault diagnosis method based on a neural network. The invention diagnoses various states of the circuit breaker by using a method of combining the deep neural network and the BP neural network, and compared with the method of singly using the BP neural network, the invention obviously improves the accuracy and the speed.

The invention is realized by the following technical scheme, and the method comprises the following steps:

step 1: an acceleration sensor is used for acquiring an acceleration waveform signal generated by the breaker in the switching-on operation process;

step 2: extracting vibration signal characteristic quantities of the vacuum circuit breaker in different states by adopting a wavelet packet-energy entropy technology;

and step 3: classifying the fault state and the normal state for the first time by using a deep neural network;

and 4, step 4: and adopting a BP neural network to specifically classify the fault state.

Further, in step 1, the sampling frequency of the acceleration sensor is 40kHz, sampling points of original signals are set to 8000 points, then a movable rectangular window with 100-150 points is used for searching for an initial point of effective data, a threshold value of the rectangular window is set, when a breaker is switched on and started, a data variance inside the rectangular window changes suddenly, the data variance is used as a judgment condition to intercept an effective vibration data sequence, the length of the intercepted effective signal is 4096 points, and the signal section approximately contains signal content about 100ms after the switching on and the starting.

Further, the wavelet packet-energy entropy technology in step 2 is to select db4 wavelet to carry out 4-layer decomposition on the collected vibration signal, and then carry out quantitative expression on the decomposed data sequence by using information entropy; the specific process comprises the following steps:

2-1) calculating the energy of each frequency band in an integral mode, wherein the calculation formula is as follows;

wherein

（LTo decompose the number of layers),i=1，2，...，N；

2-2) calculating the energy entropy after normalization processing;

the normalization processing formula is

；

The formula for calculating the energy entropy is

；

2-3) will get a set of vectors containing 16 energy entropiesTI.e. by

。

Further, in step 3, the process of building the deep neural network includes the following steps:

3-1) defining a network structure;

3-2) initializing model parameters;

3-3) calculating in a loop: forward propagation/calculation of current loss/backward propagation/weight update.

Further, in step 4, the BP neural network only adopts 1 hidden layer, the number of neuron nodes of the hidden layer is 25, the input layer includes 16 nodes, the output layer includes 5 nodes, and a gradient descent algorithm with momentum is invoked for learning. When the detection is carried out, only the characteristic vector under a certain condition needs to be input, the learnt BP neural network is used for operation, a 5-element vector is output and is compared with a target vector, and therefore the state of the circuit breaker is judged.

The specific implementation process of the gradient descent algorithm with momentum comprises the following steps:

4-1) inputNA study sample

；

4-2) constructing a BP neural network structure;

4-3) setting an error limit value

Maximum number of iterations

Learning rate

And impulse coefficient

Number of initial iterationst=1, training data sequencek=1；

4-4) taking the followingkA study sample

；

4-5) is prepared from

Carrying out signal forward propagation calculation;

4-6) the input signal is transmitted forward through weight matrix processing, and the error of each node of the BP network output layer is calculated:

；

4-7) if pairNAny data sequence of training datakValue is such that

Or

Then training is finished; if the error does not meet the requirement, the error is reversely propagated according to the network, and the weight matrix is modified;

4-8) calculating the back propagation of errors;

4-9) order

，

And jumping to step 4-4).

Compared with the prior art, the invention has the following remarkable advantages: when various fault states of the circuit breaker are diagnosed, the accuracy and the quick action performance of the algorithm are remarkably improved.

Drawings

Fig. 1 is a simplified model diagram of a vacuum circuit breaker structure under study.

Fig. 2 is a diagnostic flowchart of the present breaker diagnostic apparatus.

Fig. 3 is a flow chart of vibration signal extraction.

FIG. 4 is a schematic representation of the db4 wavelet after 4-layer decomposition.

FIG. 5 is a graph of the error curves obtained after the deep neural network is classified once.

Fig. 6 is a BP neural network learning case that classifies only fault conditions.

FIG. 7 is a graph of the confusion matrix obtained after diagnosis using the present invention.

FIG. 8 shows the learning of the BP network when only the BP neural network is used for diagnosis.

FIG. 9 is a graph of the confusion matrix obtained using only BP neural network diagnostics.

Detailed Description

The following examples are given for the detailed implementation and specific operation of the present invention, and the application scope of the present invention is not limited to the following examples.

The invention is explained by taking a simplified model of a vacuum circuit breaker structure shown in fig. 1 as an example, and five common fault states of the circuit breaker are specifically analyzed: the method comprises the following steps of incapability of keeping the switch-on state, breakage of a switch-off spring, looseness of flexible connection, looseness of an insulating pull rod and abrasion of a single-phase contact.

As shown in fig. 2, the operation steps of this embodiment are as follows:

step 1, firstly, an acceleration sensor is used for extracting a vibration signal of the circuit breaker, and a flow chart is shown in fig. 3. According to relevant documents, the CA-YD-181 customized type of the IEPE acceleration sensor is finally selected, the sensitivity is 1mv/g, the frequency response range is 5-10 kHz, and the measurement range is 5000 g. During experiments, the acceleration sensor is arranged in the center of the top of an operation box of the circuit breaker, and a CM3502 constant current source module is adopted to provide 4mA current to supply power for the IEPE. The back is connected with an AD chip controlled by a DSP, and the DSP development board selected by the invention is research and development of Asahi TMS320F 28335. The voltage signal output range of the IEPE sensor is +/-5V, so that the IEPE sensor needs to be connected with an external AD chip, and the AD7606 is selected here and can bear the input of +/-10V and +/-5V voltage signals.

And 2, performing 4-layer decomposition on the extracted vibration signal by using a Matlab wavelet tool box and selecting a db4 wavelet, wherein a schematic diagram obtained after decomposition is shown in FIG. 4, and then obtaining a feature vector by using an energy entropy. The energy of each frequency band is calculated by means of integration

Wherein

（LTo decompose the number of layers),i=1，2，...，N(ii) a After normalization processing

Then, the energy entropy is calculated:

(ii) a Finally, a group of vectors containing 16 energy entropies is obtainedTI.e. by

。

And then randomly selecting one vibration signal under 6 states of the circuit breaker, splitting a wavelet packet and calculating energy entropy, wherein the corresponding characteristic vectors are shown in a table 1, and the difference of entropy values is utilized to establish a relation with the state of the circuit breaker, so that the fault diagnosis of the circuit breaker is carried out.

TABLE 1 feature vector Table under different states

And 3, carrying out primary classification on the normal state and the fault state of the circuit breaker by using the deep neural network. The present embodiment selects 180 characteristic phasors in various states of the circuit breaker, wherein 120 characteristic phasors are used for training, and the other 60 characteristic phasors are used for testing. The target vector corresponding to each state is a binary vector, the element takes the value of 0 or 1, and the corresponding relationship is as shown in table 2:

TABLE 2 State of Circuit breaker and target vector correspondence

After training and testing, the resulting error plot is shown in fig. 5.

And 4, further diagnosing various fault states of the circuit breaker by using the BP neural network. And selecting 150 groups of feature vectors under the fault state after primary classification of the deep neural network, wherein 100 groups are used for training, and 50 groups are used for testing. The target vector corresponding to each fault state is a five-element vector, the element takes a value of 0 or 1, and the corresponding relationship is shown in table 3.

TABLE 3 Fault State and target vector correspondence for circuit breakers

Constructing a BP neural network, taking the first 20 groups of 30 groups of feature vectors in each fault state, and using 100 groups of feature vectors in total for learning and training of the BP neural network, wherein the specific steps are as follows:

the first step is as follows: reading test data;

the second step is that: preparing data for BP neural network training;

the third step: constructing a BP neural network;

the fourth step: training a network; wherein the display period of the intermediate result is set to be 50, the iteration frequency is 5000, the error target is 0.00001, and the learning rate is 0.05;

the fifth step: simulating the network;

and a sixth step: and storing the trained neural network.

The learning condition of the BP neural network is as shown in fig. 6, and after 61 iterations, the error target is reached. And then, performing pattern recognition on input data by using the trained BP neural network, selecting the last 10 groups of 30 groups of feature vectors in each fault state, and verifying the total of 50 groups, wherein the final result is shown in FIG. 7.

And 5, classifying the 6 states of the circuit breaker under the condition of only using the BP neural network. A total of 180 sets of data were selected for each state of the circuit breaker, 120 for training and 60 for testing. The target vector corresponding to each state is a six-element vector, the elements take the values of 0 or 1 respectively, and the corresponding relationship is shown in table 4.

TABLE 4 State of Circuit breaker and target vector correspondence

And constructing the BP neural network, taking the first 20 groups of 30 groups of feature vectors in each state, and using 120 groups of feature vectors in total for learning and training of the BP neural network. The learning condition of the BP network is as shown in fig. 8, and the error target is reached after 200 iterations. And then, performing pattern recognition on input data by using the trained BP neural network, selecting the last 10 groups of 30 groups of feature vectors in each state, and performing verification on 60 groups in total, wherein the result is shown in fig. 9.

Step 6, comparing fig. 7 and fig. 9, it can be found that, compared with the case of using the BP neural network alone, the accuracy and the quick-acting performance of the algorithm of the fault diagnosis of the vacuum circuit breaker using the present invention are both significantly improved.

Claims

1. a kind of circuit breaker fault diagnosis method based on neural network, is characterized in that, comprises the following steps:

Step 1, use the acceleration sensor to collect the acceleration waveform signal generated by the circuit breaker during the closing operation;

Step 2, using the wavelet packet-energy entropy technique to extract the vibration signal characteristic quantities of the vacuum circuit breaker in different states;

Step 3, use the deep neural network to classify the fault state and the normal state once;

Step 4, using BP neural network to classify the fault state in detail.

2. The neural network-based circuit breaker fault diagnosis method according to claim 1, wherein in step 1, the sampling frequency of the acceleration sensor is 40 kHz, and the sampling points of the original signal are set to be 8000 points in total, and then pass A 100~150-point movable rectangular window is used to find the starting point of valid data, and the threshold of the rectangular window is set. When the circuit breaker is closed and started, the variance of the data inside the rectangular window will change abruptly, which is used as a judgment condition to intercept valid data. The vibration signal data sequence is 4096 points, and the length of the intercepted effective signal is 4096 points. This signal contains the content about 100ms after the closing start.

3. the circuit breaker fault diagnosis method based on neural network according to claim 1, is characterized in that, the wavelet packet-energy entropy technique described in step 2 is to select db4 wavelet to carry out 4-layer decomposition to the collected vibration signal, and then Use information entropy to quantify the decomposed data sequence;

The specific process of step 2 includes the following steps:

2-1) Calculate the energy of each frequency band by integrating, the calculation formula is as follows;

,in

( L is the number of decomposition layers), i = 1, 2, ..., N ;

2-2) After normalization, calculate the energy entropy;

The normalization processing formula is

;

The formula for calculating energy entropy is

;

2-3) will get a set of vectors T containing 16 energy entropy, namely

.

4. The neural network-based circuit breaker fault diagnosis method according to claim 1, wherein in step 3, the construction process of the deep neural network comprises the following steps:

3-1) Define the network structure;

3-2) Initialize model parameters;

3-3) Circular calculation, forward propagation / calculation of current loss / back propagation / weight update.

5. The neural network-based circuit breaker fault diagnosis method according to claim 1, wherein in step 4, the BP neural network adopts one hidden layer, and the number of neuron nodes in the hidden layer is fixed. It is 25, the input layer contains 16 nodes, and the output layer contains 5 nodes, and the gradient descent algorithm with momentum is called for learning; when testing, only the feature vector in a certain situation is input, and the learned BP neural network is used. The network performs operations and outputs a 5-element vector, which is compared with the target vector to judge the state of the circuit breaker.

6. The neural network-based circuit breaker fault diagnosis method according to claim 5, wherein the specific implementation process of the gradient descent algorithm with momentum comprises the following steps:

4-1) Input N learning samples

;

4-2) Construct BP neural network structure;

4-3) Set the error limit value

, the maximum number of iterations

, the learning rate

and the impulse coefficient

, the initial number of iterations t = 1, the training data sequence k = 1;

4-4) Take the kth learning sample

;

4-5) by

Perform signal forward propagation calculation;

4-6) The input signal is forwarded through the weight matrix processing, and the error of each node in the output layer of the BP network is calculated:

;

4-7) If the value of k for any data sequence of N training data is such that

or

, then the training is over; if the requirements are not met, the error is back-propagated through the network, and the weight matrix is modified;

4-8) Error back propagation calculation;

4-9) Order

,

, skip to step 4-4).