Power cable insulation partial discharge defect type and fault level identification method based on multi-feature sequence combination

Abstract

The invention provides a power cable insulation partial discharge defect type and fault level identification method based on multi-feature sequence combination. The method comprises the following steps: carrying out a partial discharge test; establishing a partial discharge feature library, and extracting partial discharge features; according to the partial discharge characteristics, establishing a discharge intensity clustering network and a discharge type identification network; and putting a to-be-detected partial discharge grey-scale map and a related feature sequence into the established discharge intensity clustering network and the discharge type identification network to carry out evaluation so as to obtain the discharge type and the intensity grade of the to-be-detected partial discharge grey-scale map. Compared with a traditional feature extraction method, the method has the advantages that multi-feature combination is fused from multiple angles of the spectrogram and the discharge pulse, the discharge type and the discharge intensity of cable insulation partial discharge can be effectively recognized, the recognition success rate is increased, and a convenient and effective method is provided for the field of cable insulation partial discharge mode recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of power cable fault diagnosis, and in particular relates to a method for identifying power cable insulation partial discharge defect types and fault levels based on a combination of multiple feature sequences. Background Art

[0002] The power cable is one of the core components of the EMU. Its operating stability will directly affect the reliability of the EMU operation. According to the survey, 30% to 40% of power cable failures are related to insulation. In the design, assembly and long-term operation and maintenance of power cables, insulation will be affected by heat, electricity, environment and mechanical factors, and different types of discharge will occur in different parts of the insulation. The types of partial discharge of cable insulation are complex and changeable. Different types of partial discharge represent different insulation degradation mechanisms, and partial discharge also has strength. The type and strength level of partial discharge have certain guiding significance for characterizing the degree of damage to cable insulation.

[0003] At present, the application of partial discharge pattern recognition is more focused on statistical methods based on phase distribution under external working frequency voltage. The signal is distinguished using mathematical statistical relationships. This method has problems such as strong subjectivity, low recognition rate, and large interference. Relevant scholars have done a lot of work on the identification of partial discharge defect types and degrees, but due to the diversity and complexity of discharge types, accurate identification of partial discharge is still a difficult problem that needs to be solved in this field. Summary of the invention

[0004] In order to solve the technical problems in the prior art, the present invention provides a method for identifying the type and fault level of partial discharge defects in power cable insulation based on a combination of multiple feature sequences, comprising the following steps:

[0005] Conduct partial discharge tests;

[0006] Establish a partial discharge feature library and extract partial discharge features, including: time domain and frequency domain features, grayscale image moment features, PRD spectrum distribution features, and grayscale images;

[0007] According to the local discharge characteristics, the discharge intensity clustering network and the discharge type identification network are established;

[0008] The grayscale image of the partial discharge to be detected and the related feature sequence are put into the established discharge intensity clustering network and discharge type recognition network for evaluation to obtain its discharge type and intensity level.

[0009] Furthermore, the partial discharge test includes:

[0010] The typical discharge defects of power cable insulation are simulated, and the discharge pulse waveforms and PRPD spectra of different typical discharge defects and the same discharge defect at different test voltage levels are collected and obtained multiple times.

[0011] Furthermore, establishing a partial discharge feature library includes:

[0012] The time domain and frequency domain features of the discharge pulse waveform and the distribution features of the PRPD spectrum are extracted. At the same time, the PRPD spectrum is preprocessed to obtain a grayscale image, and the moment features and texture features of the grayscale image are extracted to establish a partial discharge feature library.

[0013] Furthermore, according to the local discharge characteristics, a discharge intensity clustering network is established, including:

[0014] Based on the grayscale image, a convolutional neural network is constructed, the features of the multi-pooling layer are fused and the network is trained to obtain the discharge type recognition network 1;

[0015] At the same time, BP neural network is used to construct recognition network 2 based on the pulse waveform time domain and frequency domain feature sequence;

[0016] The BP neural network is used to construct the recognition network three based on the PRPD spectrum distribution characteristics;

[0017] The weight coefficients are obtained according to the recognition rates of the three networks, and the recognition network is constructed according to the weight coefficients to realize the recognition of the discharge type.

[0018] Furthermore, according to the local discharge characteristics, a discharge type identification network is established including:

[0019] The PCA algorithm is used to reduce the dimension of the feature sequence of the pulse waveform time domain, frequency domain characteristics, PRPD spectrum distribution characteristics, and moment feature combination of the grayscale image to obtain the principal component feature sequence representing the characteristic information of the discharge intensity. K-means clustering analysis is performed on the four discharge intensities of each discharge type to establish a classification network for the discharge intensity.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention classifies and diagnoses the type and intensity of partial discharge of cable insulation through the multi-dimensional combination characteristics of partial discharge phase spectrum and discharge pulse and phase grayscale map. Compared with the traditional feature extraction method, this method integrates multi-feature combination from multiple angles of spectrum and discharge pulse, can effectively identify the discharge type and discharge intensity of partial discharge of cable insulation, improves the recognition success rate, and provides a convenient and effective method for the field of cable insulation partial discharge pattern recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the scheme of the present invention;

[0023] Figure 2 It is a flow chart of the discharge type identification network of the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1 As shown, the present application proposes a method for identifying the type and fault level of partial discharge defects in power cable insulation based on a combination of multiple feature sequences, which specifically includes the following steps:

[0026] 1) Simulate typical discharge defects of power cable insulation, collect and obtain discharge pulse waveforms and PRPD spectra of different typical discharge defects and different test voltage levels under the same discharge defect for multiple times;

[0027] 2) Extract the time domain and frequency domain features of the discharge pulse waveform and the distribution features of the PRPD spectrum. At the same time, preprocess the PRPD spectrum to obtain a grayscale image, extract the moment features and texture features of the grayscale image, and establish a partial discharge feature library;

[0028] 3) Based on the grayscale image, a convolutional neural network is constructed, and the features of the multi-pooling layer are integrated and the network is trained to obtain the discharge type recognition network 1. At the same time, a BP neural network is used to construct a recognition network 2 based on the pulse waveform time domain and frequency domain feature sequences, and a BP neural network is used to construct a recognition network 3 based on the PRPD spectrum distribution characteristics; the weight coefficients are obtained according to the recognition rates of the three networks, and the recognition network is constructed according to the weight coefficients to realize the recognition of the discharge type;

[0029] 4) PCA algorithm is used to reduce the dimension of the characteristic sequence of pulse waveform time domain, frequency domain characteristics, PRPD spectrum distribution characteristics, and moment characteristic combination of grayscale images to obtain the principal component characteristic sequence representing the characteristic information of discharge intensity. K-means cluster analysis is performed on the four discharge intensities of each discharge type to establish a classification network for discharge intensity.

[0030] 5) The grayscale image of the partial discharge to be detected and the related feature sequence are sent to the network in steps 3) and 4) for evaluation to obtain the discharge type and intensity level.

[0031] Four typical insulation defect models of cable insulation were simulated and produced, namely: end discharge, internal discharge, suspended discharge, and surface discharge. First, a partial discharge test platform was built, and the discharge signal of the discharge model was collected based on the pulse current method. The partial discharge pulse waveform and PRPD spectrum were obtained through high-voltage testing. The PRPD spectrum refers to the display of the partial discharge pulse with phase identification according to the phase. The partial discharge sampling frequency is 100MHz, and each PRPD spectrum contains 3000 continuous discharge pulses. For each defect model, two voltage measurement points are evenly taken in the starting voltage and the highest test voltage, and the discharge amplitude and phase of each of the 100 groups of PRPD spectra at the four test voltage measurement points including the starting and highest test voltages are recorded, and the discharge intensity is divided into four levels: initial, attention, abnormal and severe. At the same time, 100 groups of discharge pulse waveforms of each defect type at each test voltage measurement point are recorded, and a total of 1600 groups of discharge PRPD spectra and discharge pulse waveforms are obtained.

[0032] The 12-dimensional features of the time domain and frequency domain of 1600 pulse waveforms are extracted, mainly including the rise time t of the time domain waveform r , Fall time t d , 50% amplitude duration t 50％ , 10% amplitude duration t 10％ , pulse duration t total , equivalent time T i , Maximum voltage amplitude U m , skewness S k1 , steepness K u1 ; Extract the skewness S of the frequency domain waveform k1 , steepness K u2 and equivalent bandwidth F i .

[0033] The 1600 groups of PRPD spectra were transformed to obtain the relationship spectrum q-φ between phase and average discharge amount and the relationship spectrum n-φ between phase and discharge number, and the positive and negative half-cycle skewness S of the q-φ spectrum was extracted. k3 , S k4 and steepness K u3 , K u4 , the positive and negative half-cycle skewness S of the n-φ spectrum k5 , S k6 and steepness K u5 , K u6 , the 11-dimensional eigenvalues ​​such as the mutual correlation coefficient cc, discharge factor Q and modified mutual correlation coefficient mcc of the q-φ spectrum, thus forming an 11-dimensional spectrum distribution feature sequence.

[0034] The 1600 sets of PRPD spectra obtained were preprocessed. The pulse signal amplitude q was divided into 180 intervals, the power frequency phase φ was divided into 180 intervals, and then the q-φ plane was divided into 180×180 intervals; the number of partial discharge pulses n in each sub-interval was counted. a,b (a∈[0,180],b∈[0,180]), we get a three-dimensional q-φ-n spectrum; then we calculate the pulse number n a,b Normalization is performed: n a,b '=n a,b n max ; where n a,b ' is the normalized pulse number, n max is the maximum number of pulses in all intervals of the q-φ-n spectrum; set n a,b 'The gray value of each pixel is K a,b :K a,b =(1-n a,b ')×255, and then a grayscale image that can characterize the discharge characteristics is obtained.

[0035] For 1600 groups of grayscale images, the horizontal and vertical coordinates of the grayscale center of gravity of the positive half cycle of the power frequency are f x + 、f y + , the horizontal and vertical coordinates of the grayscale center of gravity of the negative half cycle of the power frequency are f x - 、f y - , the horizontal and vertical coordinates of the power frequency grayscale center of gravity f x 、f y , characteristic parameter f of the main axis direction in the positive half cycle of the power frequency m + , characteristic parameter f of the main axis direction in the negative half cycle of the power frequency m - , characteristic parameter f of the power frequency spindle direction m The 9-dimensional feature parameters are extracted.

[0036] The grayscale images of each discharge type and the corresponding four discharge type labels are used as input and output respectively, and a convolutional neural network is built and trained as the discharge type recognition network 1. The convolutional neural network model adopts the pooling layer fusion method, and its process is as follows: Figure 2 As shown in the figure, the model consists of an input layer, three convolutional layers, three pooling layers, two fully connected layers and a final output layer. The output feature sequences of the three pooling layers are fused and input into the fully connected layer.

[0037] Then output the discharge type, compare the output result with the original grayscale image label, perform back propagation, and modify the weights and bias of each layer according to the error between the two to obtain a new network for the next training. After feature fusion, use the Droput method to reduce the amount of calculation. Average pooling is used for pooling, Softmax classifier is used as the classifier, cross entropy cost function is used as the classification algorithm, ReLU is used as the activation function, and SGD method is used to train and update model parameters. 800 groups of samples are selected for the training set, and the remaining 800 groups of samples are used for the test set to evaluate the classification performance of the algorithm and output the classification recognition rate. After training, the grayscale image to be detected is sent to the network, and the grayscale image is feature fused according to the network requirements, and then the probability value sequence L corresponding to each discharge type of the image is obtained in the output layer. 1 . Set L 1 :[X 11 , X 12 , X 13 , X 14 ].

[0038] Based on 1600 sets of 12-dimensional feature sequences of discharge pulse time domain and frequency domain waveforms, in order to reduce the difference in individual data levels, the data was normalized, and then the discharge pulse feature sequences of each discharge type and the corresponding four discharge type labels were used as input and output respectively to train and establish a BP neural network, of which 800 sets were used for training and 800 sets were used for testing, and then the classification performance of the algorithm was evaluated, and the output recognition was successful as the discharge type recognition network 2. The trained BP neural network classifier was used to classify the time domain and frequency domain feature sequences of the discharge pulse to be detected, and the probability value sequence L corresponding to each discharge type of the discharge pulse was obtained at the output layer. 2 . Set L 2 :[X 21 , X 22 , X 23 , X 24 ].

[0039] Based on the 11-dimensional feature sequence of 1600 groups of PRPD spectra, the PRPD spectra feature sequence of each discharge type and the corresponding four discharge type labels are used as input and output respectively, and the BP neural network is trained and established, of which 800 groups are used for training and 800 groups are used for testing, so as to evaluate the classification performance of the algorithm and output the recognition success rate as the discharge type recognition network 3. The trained BP neural network classifier is used to classify the PRPD spectra feature sequence to be detected, and the probability value sequence L corresponding to each discharge type of the PRPD spectra is obtained at the output layer. 3 . Set L 3 :[X 31 , X 32 , X 33 , X 34 ].

[0040] The discharge type comprehensive evaluation criteria are set according to the success rate of the three discharge type recognition networks on the test set. L is the probability value sequence corresponding to each discharge type after comprehensive evaluation. Where L:[X 1 , X 2 , X 3 , X 4 ].

[0041] (where a 1 、a 2 、a 3 The three recognition networks account for the total network weights respectively)

[0042] The algorithm recognition success rate was tested. The average recognition rate of the test set of network 1 was the highest, reaching more than 95%, while the average recognition rates of network 2 and 3 were about 82% and 91% respectively. 1 is 0.4, a 2 is 0.35, a 3 is 0.25.

[0043] Compare X 1 , X 2 , X 3 , X 4 The size of the probability value is determined by taking the discharge type corresponding to the maximum probability value as the output discharge type of the partial discharge data.

[0044] 1600 groups of 12-dimensional pulse features, 11-dimensional PRPD spectrum features, and 9-dimensional grayscale image features are combined into 32-dimensional combined features. The 1600 groups of feature sequences are divided into 4 parts, representing the classification of the four discharge intensity levels of each discharge type. The principal component analysis (PCA) method is used to transform the data from high-dimensional space to low-dimensional space. The principal component analysis method is used to analyze the 4 groups of 32 feature parameters respectively, retaining the principal components with cumulative contribution rates exceeding 0.95, and then obtain 7, 8, 8, and 9-dimensional new feature factors to characterize the discharge intensity of the four discharge types.

[0045] Based on four sets of new multidimensional feature factors, each set of multidimensional feature factors was divided into 200 training sets and 200 test sets. The K-means clustering algorithm was used to classify the discharge intensity of the four types of discharges, and four K-means discharge intensity clustering networks were obtained. The recognition rate of the network for the four discharge intensities of each type of discharge was higher than 95%.

[0046] The multi-dimensional feature quantity and grayscale image of the cable insulation partial discharge test data to be tested are extracted and sent to the discharge type recognition network one, two, and three, and then the partial discharge type of the test cable insulation is output according to the above discharge type judgment standard. Then the 32-dimensional feature vector of the test data is combined with the 400×32-dimensional feature vector of the original feature library corresponding to the discharge type into a 401×32-dimensional feature vector, and the combined feature vector is subjected to PCA dimensionality reduction, and the K-means algorithm is used to cluster the reduced-dimensional feature vector to obtain the discharge intensity of the test data.

[0047] In summary, the above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereto. All equivalent changes and modifications made according to the patent scope of the present invention and the contents of the specification are within the scope covered by the patent of the present invention.

Claims

1. A method for identifying the type and fault level of partial discharge defects in power cable insulation based on a combination of multiple feature sequences, characterized in that: The following steps are involved: Conduct partial discharge tests; Establish a partial discharge feature library and extract partial discharge features, including: time domain and frequency domain features, grayscale image moment features, PRD spectrum distribution features, and grayscale images; According to the local discharge characteristics, the discharge intensity clustering network and the discharge type identification network are established; The grayscale image of the partial discharge to be detected and the related feature sequence are put into the established discharge intensity clustering network and discharge type recognition network for evaluation to obtain its discharge type and intensity level.

2. The method according to claim 1, characterized in that Partial discharge testing includes: The typical discharge defects of power cable insulation are simulated, and the discharge pulse waveforms and PRPD spectra of different typical discharge defects and the same discharge defect at different test voltage levels are collected and obtained multiple times.

3. The method according to claim 2, characterized in that Building a partial discharge feature library includes: The time domain and frequency domain features of the discharge pulse waveform and the distribution features of the PRPD spectrum are extracted. At the same time, the PRPD spectrum is preprocessed to obtain a grayscale image, and the moment features and texture features of the grayscale image are extracted to establish a partial discharge feature library.

4. The method according to claim 3, characterized in that According to the local discharge characteristics, the discharge intensity clustering network is established, including: Based on the grayscale image, a convolutional neural network is constructed, the features of the multi-pooling layer are fused and the network is trained to obtain the discharge type recognition network 1; At the same time, BP neural network is used to construct recognition network 2 based on the pulse waveform time domain and frequency domain feature sequence; The BP neural network is used to construct the recognition network three based on the PRPD spectrum distribution characteristics; The weight coefficients are obtained according to the recognition rates of the three networks, and the recognition network is constructed according to the weight coefficients to realize the recognition of the discharge type.

5. The method according to claim 4, characterized in that According to the local discharge characteristics, the discharge type identification network is established including: The PCA algorithm is used to reduce the dimension of the feature sequence of the pulse waveform time domain, frequency domain characteristics, PRPD spectrum distribution characteristics, and moment feature combination of the grayscale image to obtain the principal component feature sequence representing the characteristic information of the discharge intensity. K-means clustering analysis is performed on the four discharge intensities of each discharge type to establish a classification network for the discharge intensity.

Images