Commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU

Abstract

The invention discloses a commutation failure fault diagnosis method based on fuzzy entropy characteristics and PCNN-GRU. The method comprises the steps that a rectification side direct current fault and an inversion side alternating current fault are included, and rectification side positive pole direct current voltage data after the faults are collected; the method comprises the following steps: firstly, decomposing a DC voltage signal by using a CEEMDAN algorithm to obtain high and low frequency components under a fault voltage subsequence; then, the fuzzy entropy of each feature component is calculated, and an input matrix of the fault diagnosis model is constructed; an input sample is divided into a training set and a test set for training a dual-branch parallel convolution gated cycle unit neural network model, model parameters are finely adjusted during training, the sizes of a convolution kernel and neurons are adjusted, optimal model parameters are obtained, and the model is stored; and finally, testing the data of the test set in the adjusted model, so that different fault types under the condition of causing commutation failure can be effectively judged, and the method has certain anti-noise performance.

Description

Technical Field

[0001] The invention relates to a commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU, and belongs to the technical field of relay protection of high-voltage direct current transmission systems. Background Art

[0002] Commutation failure is the most common fault in DC transmission systems. It will affect the safe and stable operation of the system. In severe cases, it may even cause the system to stop transmitting power and cause the system to crash. In addition to the fault of the converter valve itself, there are many reasons for commutation failure. Especially for the weak receiving end DC transmission system, when the AC system on the inverter side fails, it is very easy to cause commutation failure [3]. The commutation failure process is accompanied by sudden changes in DC current and DC voltage. If it cannot be diagnosed and taken corresponding measures in a timely and accurate manner, it will cause serious harm and great impact to various equipment in the transmission system. At the same time, the changes in DC voltage signals of DC transmission line faults and commutation failures are very similar. It is necessary to diagnose and distinguish between DC line faults and commutation failures.

[0003] The traditional commutation failure fault diagnosis method starts with theoretical analysis and manual extraction of fault features, determines the fault threshold through mathematical mechanisms, compares the collected real-time quantity with the threshold, and detects whether a commutation failure occurs. The detection variable is single and does not cover most commutation failure situations. The occurrence of a fault is often not caused by a single factor. In addition to conventional short-circuit faults, it is also caused by coupling factors such as load changes, harmonics, and excessive resistance. With the continuous development of artificial intelligence, fault diagnosis based on deep learning algorithms has begun to emerge and has gradually been applied in high-voltage direct current (HVDC) systems. Relying on its powerful data feature extraction capabilities and extremely strong nonlinear fitting capabilities, deep learning methods have unique advantages, especially in the feature extraction of time series and non-stationary signals. The application of deep learning methods in commutation failure fault diagnosis not only makes up for the shortcomings of traditional fault identification methods in manually analyzing and extracting fault information, but also improves the fault recognition rate. Summary of the invention

[0004] In order to avoid complicated threshold setting calculations and improve the model's tolerance to transition resistance, the present invention provides a commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU.

[0005] The technical solution of the present invention comprises the following steps:

[0006] S1. According to different fault types on the rectifier and inverter sides, uniformly collect transient voltage data of DC side positive pole fault;

[0007] S2. The CEEMDAN algorithm is used to decompose transient voltage and obtain multiple intrinsic mode subsequences of different scales, which can effectively extract fault features in complex signals.

[0008] S3. Calculate the fuzzy entropy of each decomposed component to obtain fuzzy entropy feature information that can reflect each modal component, which is used for training and testing the diagnosis model.

[0009] S4. Divide the feature samples into training set and test set, use the training set to train the dual-branch parallel convolution gated recurrent unit network, fine-tune the parameters in the training process, and obtain the dual-branch parallel convolution gated recurrent unit network model with the best parameter configuration, which is used for commutation failure fault type diagnosis and output Y i f 1 ~f 6 There are 6 types of failures.

[0010] S5. Input the test set into the dual-branch parallel convolution gated recurrent unit network model for testing to obtain a commutation failure fault type diagnosis result.

[0011] In S1, the sampling frequency is selected as 20kHz, the data window size is 1s, a total of 4000 sampling points, the total system simulation time is 3s, and the fault duration is 100ms.

[0012] The decomposition method is: using the CEEMDAN algorithm to decompose the transient fault current, the transient voltage is the voltage measured by the rectifier side protection device in S1.

[0013] The S2 feature vector is calculated by fuzzy entropy to obtain fuzzy entropy feature information that can reflect each modal component and is used for training and testing the diagnostic model.

[0014] The dual-branch parallel convolutional gated recurrent unit network has 7 layers: input layer, dual-branch parallel convolutional layer 1, dual-branch parallel convolutional layer 2, Flatten flattening layer, GRU layer, Drorout layer and classification output layer.

[0015] The learner of the dual-branch parallel convolutional gated recurrent unit network adopts the adam gradient descent algorithm. The convolution kernel size and number of the dual-branch parallel residual layer and the pooling kernel size have a total of 8 parameters, plus the number of GRU neurons, the discard rate of the Dropout layer and the learning rate of the model, a total of 11 parameters.

[0016] According to another aspect of the present invention, a commutation failure fault diagnosis model based on fuzzy entropy features and PCNN-GRU is provided, comprising:

[0017] The acquisition module is used to collect transient voltage data of the DC side positive line fault under different fault types;

[0018] The data preprocessing module uses the CEEMDAN decomposition algorithm to decompose the original fault voltage data to obtain preliminary input samples, calculates the fuzzy entropy of the corresponding decomposition components, obtains the characteristic input sample matrix, and divides the input sample matrix into a training set and a test set.

[0019] Model training module, use the training set to train the dual-branch parallel convolution gated recurrent unit network, fine-tune the parameters in the training model, and obtain the dual-branch parallel convolution gated recurrent unit network model with the best parameter configuration for fault area identification, and output Y i f 1 ~f 6 There are 6 types of failures.

[0020] The recognition module, after the model training is completed, is used to diagnose the fault type of commutation failure using the test set, and finally obtain the results of various fault types that cause commutation failure.

[0021] The beneficial effects of the present invention are:

[0022] 1. The CEEMDAN decomposition method is used to extract the deep features of each modal component of the fault signal, and the fuzzy entropy of each component is calculated as the fault feature, which can effectively reflect different fault states. The model effectively suppresses the influence of factors such as transition resistance and noise, and has strong anti-interference ability.

[0023] 2. The fuzzy entropy feature vector reflecting commutation failure is used as the model input. The diverse feature extraction capabilities of the PCNN-GRU model can be used to accurately diagnose the causes of normal states, DC line faults, and commutation failures.

[0024] 3. Compared with the simple deep learning models of BP, GRU and CNN models, the PCNN-GRU model in this paper has higher fault diagnosis accuracy. And this method has strong noise resistance. Even under 20dB and 30dB noise, the model still has high diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the present invention;

[0026] Figure 2 A topological diagram of the Yongfu high-voltage direct current transmission system used in the embodiment;

[0027] Figure 3 It is the equivalent circuit diagram of 6-pulse inverter;

[0028] Figure 4 This is a network structure diagram of a dual-branch parallel convolutional gated recurrent unit;

[0029] Figure 5 When commutation fails, U dCEEMDAN decomposition result diagram under various commutation failures;

[0030] Figure 6 It is the fuzzy entropy diagram of each mode calculated after CEEMDAN decomposition under commutation failure;

[0031] Figure 7 It is the comparison diagram of fuzzy entropy under different fault states;

[0032] Figure 8 This is the confusion matrix diagram under 20dB and 30dB.

[0033] Note: Figure 2 In the topology diagram shown, the data acquisition zone is set at the fault location of the positive line on the DC side. DETAILED DESCRIPTION

[0034] The invention is further described below in conjunction with the accompanying drawings and embodiments, but the content of the invention is not limited to the scope of the embodiments.

[0035] Example 1: Figure 1-7 As shown, a commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU includes:

[0036] S1. According to different fault types on the rectifier and inverter sides, uniformly collect transient voltage data of DC side positive pole fault;

[0037] S2. The CEEMDAN algorithm is used to decompose transient voltage and obtain multiple intrinsic mode subsequences of different scales, which can effectively extract fault features in complex signals.

[0038] S3. Calculate the fuzzy entropy of each decomposed component to obtain fuzzy entropy feature information that can reflect each modal component, which is used for training and testing the diagnosis model.

[0039] S4. Divide the feature samples into training set and test set, use the training set to train the dual-branch parallel convolution gated recurrent unit network, fine-tune the parameters in the training process, and obtain the dual-branch parallel convolution gated recurrent unit network model with the best parameter configuration, which is used for commutation failure fault type diagnosis and output Y i f 1 ~f 6 There are 6 types of failures.

[0040] S5. Input the test set into the dual-branch parallel convolution gated recurrent unit network model for testing to obtain a commutation failure fault type diagnosis result.

[0041] Optionally, the different fault types include commutation failure of the AC / DC system under different regional faults, different transition resistances, and different fault distances; the different regions are DC line faults on the rectifier side and AC faults on the inverter side (single-phase grounding, two-phase short circuit, two-phase grounding short circuit, and three-phase short circuit); the different transition resistances are 0 to 300Ω; the different fault distances are 50km, 100km, 200km, 300km, 400km, and 500km of the transmission line.

[0042] Optionally, in S1, for the convenience of hardware implementation, the sampling frequency is selected to be 20 kHz, the data window size is 1 s, a total of 4000 sampling points, the total system simulation time is 3 s, and the fault duration is 100 ms.

[0043] Alternatively, if Figure 4 As shown in the figure, there are 7 layers in total, namely input layer, dual-branch parallel convolution layer 1, dual-branch parallel convolution layer 2, Flatten layer, GRU layer, Dropout layer and classification output layer; GRU and Dropout layers 1-n represent n neurons respectively, and classification output layers 1-6 represent 6 types of fault categories, 1 is the normal state, and the others are AC and DC faults respectively;

[0044] Alternatively, if Figure 5 As shown in FIG. 1 , the original fault voltage matrix is ​​decomposed using the CEEMDAN decomposition algorithm to obtain the intrinsic mode subsequences under each high and low frequency component.

[0045] Alternatively, if Figure 6 As shown in the figure, the fuzzy entropy of each decomposition component is calculated to obtain the fuzzy entropy feature information that can reflect each modal component, which is used for the training and testing of the diagnosis model. 1 ~R 5 is the transition resistance that increases successively, R 1 =11Ω、R 2 =12Ω、R 3 =13Ω、R 4 =14Ω、R 5 =15Ω.

[0046] Alternatively, if Figure 7 As shown in the figure, the fuzzy entropy curves under different fault conditions show an overall trend of first rising and then falling. The greater the difference in the fuzzy entropy distribution curves, the more separable the corresponding fuzzy entropy feature vectors are.

[0047] The dual-branch parallel convolutional gated recurrent unit network has 7 layers: input layer, dual-branch parallel convolutional layer 1, dual-branch parallel convolutional layer 2, Flatten flattening layer, GRU layer, Drorout layer and classification output layer.

[0048] The learner of the dual-branch parallel convolutional gated recurrent unit network adopts the adam gradient descent algorithm. The convolution kernel size and number of the dual-branch parallel residual layer and the pooling kernel size have a total of 8 parameters, plus the number of GRU neurons, the discard rate of the Dropout layer and the learning rate of the model, a total of 11 parameters.

[0049] An optional implementation manner of the present invention is described in detail below.

[0050] The present invention constructs a simulation model of Yongfu high-voltage direct current transmission system in PSCAD / EMTDC simulation software, and builds a dual-branch parallel convolutional gated recurrent unit neural network on the MATLAB software platform. Through theoretical and experimental analysis, it is found that:

[0051] The equivalent circuit of 6-pulse inverter is as follows Figure 3 As shown, when the inverter is commutating normally, the thyristors V1V2V3V4V5V6 are turned on in the order of V1V2→V2V3→V3V4→V4V5→V5V6→V6V1→V1V2 in turn to realize the conversion of DC to AC.

[0052] Normal commutation process: Assume that before a certain moment, the conduction valve of the inverter side converter is V5V6, and the DC current previously flows from phase a through valve V5. At a certain moment, the converter receives a trigger command and starts commutation. Valve V1 is turned on and valve V5 is turned off. The DC current flows from phase a through valve V5 to phase c through valve V1, that is, from phase a to phase c. At this time, the converter conduction valve is V6V1, and this process is called commutation. During the entire commutation process, the working status of each valve and the commutation situation on the AC side are shown in Table 1.

[0053] The specific process of data preprocessing in the dual-branch parallel convolutional gated recurrent unit network model is as follows:

[0054] S2. The CEEMDAN algorithm is used to decompose transient voltage and obtain multiple intrinsic mode subsequences of different scales, which can effectively extract fault features in complex signals.

[0055] The CEEMDAN method is used to analyze the DC voltage signal U d Then the original signal U d is broken down into:

[0056]

[0057] Where: i = 1, 2, ..., N, N is the number of signal data samples; k = 1, 2, ..., K, K is the number of eigenvalue sequences; is the kth modal component of the ith sample; Ri(t) is the residual component of the ith sample.

[0058] The CEEMDAN decomposition results for a DC line short circuit fault are shown in Figure 2. Figure 5 As shown in (a), the DC voltage U d Contains rich high-frequency components, and the high-frequency components C 5 , C 6 and C 7 The decrease is obvious during the fault period. The CEEMDAN decomposition results of different causes of commutation failure are as follows: Figure 5 As shown in (b), both contain rich low-frequency components C 1 , C 2 and residual component R, and the high-frequency component decays slowly during the fault. The components of different commutation failure causes have different volatility, and the different characteristic information of these modal components provides a basis for constructing the fuzzy entropy feature vector.

[0059] S3. Calculate the fuzzy entropy of each decomposed component to obtain fuzzy entropy feature information that can reflect each modal component, which is used for training and testing the diagnosis model.

[0060] If the DC voltage U under different fault conditions is d The CEEMDAN components of the model are used as features for model input to diagnose faults, but there are problems such as unclear fault features, redundant data and long training time. In this regard, the fuzzy entropy is used to quantify the values ​​of each component and extract the fault feature information in the component, taking advantage of its small data volume, good consistency and strong anti-noise ability.

[0061] CEEMDAN components C under different fault conditions 1 -C 7 and R, calculate their fuzzy entropy and use them as the feature vector reflecting the fault characteristics. Calculate the fault signal U d The fuzzy entropy of each component is:

[0062]

[0063] Where: The components C of the CEEMDAN of the i-th sample are 1 -C 7 and the fuzzy entropy value of R; F i is the fuzzy entropy feature vector of the i-th sample.

[0064] like Figure 6 As shown in the figure, the fuzzy entropy value distributions of different transition resistances in the same fault state are basically the same, while the fuzzy entropy values ​​of each modal component in different fault states are different and their changing trends are different.

[0065] like Figure 7As shown in the figure, the fuzzy entropy curves under different fault conditions show an overall trend of first rising and then falling. The greater the difference in the fuzzy entropy distribution curves, the more separable the corresponding fuzzy entropy feature vectors are. 1 The fuzzy entropy value of is almost 0, indicating that the complexity of these two components is low and the fuzzy entropy feature information is small. The fuzzy entropy feature information is mainly contained in component C 2 -C 7 Among them, the various AC system faults that cause commutation failure are particularly prominent. The modal component C 6 The maximum values ​​are 0.045, 0.074, 0.062 and 0.051 respectively.

[0066] In normal state, the curve fluctuates most gently; when the DC line is short-circuited, the curve 3 -C 7 The curve increases slightly and gently; the fuzzy entropy curves of various faults that cause commutation failure, from C 2 The upward trend begins, C 6 The maximum value is taken when the modal component is selected, and the fuzzy entropy curves of the four types of fault states have different change trends and different values. Different fuzzy entropy feature information of each modal component is extracted for training and testing of the diagnosis model, thereby realizing the fault diagnosis of commutation failure.

[0067] The specific application process is as follows:

[0068] A: According to the different fault types on the rectifier and inverter sides, the transient voltage data of the DC side positive pole fault is collected uniformly;

[0069] B. Using the CEEMDAN algorithm to decompose transient voltage, multiple intrinsic mode subsequences of different scales are obtained, which can effectively extract fault features in complex signals.

[0070] C. Calculate the fuzzy entropy of each decomposition component to obtain the fuzzy entropy feature information that can reflect each modal component, which is used for training and testing the diagnostic model.

[0071] D. Divide the sample input obtained in step C into a training set and a test set, as shown in Table 1. Use the training set to train the dual-branch parallel convolutional gated recurrent unit network, fine-tune the parameters during the training process, and obtain the dual-branch parallel convolutional gated recurrent unit network model with the best parameter configuration for commutation failure fault type diagnosis, and output Y i f 1 ~f 6 There are 6 types of failures.

[0072] E. Input the test set into the multi-branch parallel residual network model for testing, and obtain the fault area recognition result. In order to verify the accuracy under different parameters, this paper adjusts the convolution kernel size and the number of neurons in CNN, and the model recognition results are shown in Table 2.

[0073] Table 1 Fault data division

[0074]

[0075] Table 2 Accuracy under different parameters

[0076]

[0077]

[0078] In order to verify the effectiveness of the PCNN-GRU diagnosis model, different comparison models were used to conduct diagnosis experiments on fault data. Each model was repeated 7 times, and the accuracy was taken as the average of 7 times. The results are shown in Table 3.

[0079] Table 3 Accuracy of different diagnostic models

[0080]

[0081] When collecting various fault signals in an actual HVDC transmission system, the original fault data will inevitably be interfered by noise signals, which will affect the accuracy of the diagnosis method. Figure 8 As shown in the figure, Gaussian white noise with signal-to-noise ratio of 20dB and 30dB is added to the collected various fault data. The smaller the signal-to-noise ratio, the greater the noise. The input feature vector of the model is also obtained. The trained PCNN-GRU model is used to perform noise resistance test on 240 test sets to further verify the effectiveness and generalization ability of the model in this paper.

[0082] Attached to the instruction manual Figure 8 From the confusion matrix shown, it can be seen that when the signal-to-noise ratio is 30dB, the model can still make accurate diagnosis of different fault types, with a diagnostic accuracy of 100%, and the model has a certain anti-noise ability. When the signal-to-noise ratio is 20dB, the model diagnoses two groups of two-phase short-circuit grounding faults as single-phase grounding faults, and the diagnostic accuracy drops to 99.1667%, indicating that increasing the noise level weakens the representation of different fault characteristics and increases the difficulty of model classification.

[0083] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU, characterized in that: The steps include: S1. According to different fault types on the rectifier and inverter sides, uniformly collect transient voltage data of DC side positive pole fault; S2. Using CEEMDAN algorithm to decompose transient voltage, multiple intrinsic mode subsequences of different scales are obtained, which can effectively extract fault features in complex signals; S3, calculating the fuzzy entropy of each decomposed component to obtain fuzzy entropy feature information that can reflect each modal component for training and testing of the diagnostic model; S4. Divide the feature samples into training set and test set, use the training set to train the dual-branch parallel convolution gated recurrent unit network, fine-tune the parameters in the training process, and obtain the dual-branch parallel convolution gated recurrent unit network model with the best parameter configuration, which is used for commutation failure fault type diagnosis and output Y i There are 6 fault types from f1 to f6; S5. Input the test set into the dual-branch parallel convolution gated recurrent unit network model for testing to obtain a commutation failure fault type diagnosis result.

2. According to claim 1, a commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU is characterized in that: The different fault types include AC and DC system faults under different fault areas, different transition resistances, and different fault distances; the different areas are DC line faults on the rectifier side and AC faults on the inverter side (single-phase grounding, two-phase short circuit, two-phase grounding short circuit, and three-phase short circuit); the different transition resistances are 0 to 300Ω; the different fault distances are 50km, 100km, 200km, 300km, 400km, and 500km of the transmission line.

3. The commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU according to claim 1 is characterized in that: In S1, the sampling frequency is selected as 20kHz, the data window size is 1s, a total of 4000 sampling points, the total system simulation time is 3s, and the fault duration is 100ms.

4. The commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU according to claim 1 is characterized in that: The CEEMDAN decomposition algorithm is used to decompose the original fault voltage matrix and obtain the intrinsic mode subsequences of each high and low frequency component.

5. The commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU according to claim 1 is characterized in that: The fuzzy entropy is used to calculate the components under each mode, and the fuzzy entropy feature information that can reflect the components of each mode is obtained.

6. The commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU according to claim 1 is characterized in that: The dual-branch parallel convolutional gated recurrent unit network is mainly divided into an input layer, a parallel convolutional layer (PCNN), a GRU layer, a fully connected layer, a Softmax layer and an output layer.

7. The commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU according to claim 1 is characterized in that: The dual-branch parallel convolution gated recurrent unit network learner adopts the adam gradient descent algorithm. The dual-branch parallel convolution kernel size and number, pooling kernel size, a total of 8 parameters, plus the number of GRU neurons, the discard rate of the Dropout layer and the learning rate of the model, a total of 11 parameters. Continuous fine-tuning is required during the model training process to achieve the best output effect.

8. The commutation failure fault diagnosis method based on fuzzy entropy features and PCNN-GRU according to claim 1 is characterized in that: The constructed diagnostic model includes: an acquisition module, a data preprocessing module, a model training module and an identification module; the acquisition module is used to collect transient voltage data of the DC positive line fault under different fault types; the data preprocessing module uses the CEEMDAN decomposition algorithm to decompose the original fault voltage data to obtain preliminary input samples, calculates the fuzzy entropy of the corresponding decomposition components, obtains the feature input sample matrix, and divides the input sample matrix into a training set and a test set; the model training module uses the training set to train multiple dual-branch parallel convolution gated recurrent unit networks, fine-tunes the parameters in the training model, and obtains the dual-branch parallel convolution gated recurrent unit network model with the best parameter configuration, which is used for fault area identification and outputs Y i There are 6 fault types f1 to f6 respectively; the identification module refers to the test set used for diagnosing the commutation failure fault type after the model training is completed, and finally obtaining the results of each fault type that causes the commutation failure.

Images