High-voltage cable pattern recognition method and device based on time-series topology data analysis

Abstract

The application provides a high-voltage cable mode recognition method and device based on time sequence topology data analysis. The method comprises the following steps: collecting partial discharge signals of cables with known insulation defect types to construct a data set; reconstructing the phase space of each sample in the data set to generate point cloud data of the sample; modeling the point cloud data of the sample in a topological space based on persistent homology to extract topological features; converting the topological features of the sample into a persistent regime curve and / or a betti curve; dividing the data set into a training set and a test set, and taking the persistent regime curve or betti curve of each sample in the training set as the input of a given high-voltage cable mode recognition model to train the model; collecting partial discharge signals of a cable to be detected, and obtaining a persistent regime curve or betti curve of the cable to be detected according to the partial discharge signals; and inputting the persistent regime curve or betti curve into the trained high-voltage cable mode recognition model to obtain a recognition result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of partial discharge detection and pattern recognition of high-voltage cables, and particularly relates to a high-voltage cable pattern recognition method and device based on time series topology data analysis. BACKGROUND

[0002] Partial discharge detection and analysis are widely used in the diagnosis of cable insulation deterioration, and the discharge mode has a strong correlation with the nature of the insulation defect. Fully mining and analyzing power cable operation fault records and various partial discharge test data, timely and accurate identification of fault types has important guiding significance for cable line maintenance, repair strategy formulation, etc. Due to the rise of artificial intelligence in recent years, partial discharge pattern recognition research mainly focuses on two aspects: partial discharge feature extraction method and pattern recognition algorithm iteration update.

[0003] However, the main feature extraction method in high-voltage cable pattern recognition is based on amplitude, frequency, and partial discharge phase analysis (PRPD) spectrum to extract features of different partial discharge modes of the cable. The features extracted by amplitude, frequency, and PRPD spectrum for partial discharge signals corresponding to some samples of the two defects of high-voltage cable insulation surface scratches and insulation main surface contamination have certain similarity, which further increases the difficulty of pattern recognition and has a serious impact on the maintenance and repair strategy formulation of the cable line. SUMMARY

[0004] In view of the problem that the features extracted from the partial discharge mode of the high-voltage cable have low discrimination, which easily causes misjudgment of the pattern recognition model, the present application provides a high-voltage cable pattern recognition method and device based on time series topology data analysis.

[0005] In one aspect, the present application provides a high-voltage cable pattern recognition method based on time series topology data analysis, comprising the following steps:

[0006] Step 1: Collecting partial discharge signals of cables with known insulation defect types, each partial discharge signal as a sample, and the corresponding insulation defect type as a sample label, to construct a data set containing M samples;

[0007] Step 2: Phase space reconstruction is performed on each sample in the data set to generate point cloud data of the sample; topology space modeling is performed on the point cloud data of the sample based on persistent homology to extract topology features; and the topology features of the sample are converted into a persistent state curve and / or a betti curve;

[0008] Step 3: Dividing the data set into a training set and a test set, and taking the persistent state curve or betti curve of each sample in the training set as the input of a given high-voltage cable pattern recognition model to train it;

[0009] Step 4: Collecting the partial discharge signal of the cable to be detected, and obtaining the persistence trend curve or the Betti curve of the cable to be detected according to the partial discharge signal, inputting the persistence trend curve or the Betti curve into the trained high-voltage cable pattern recognition model to obtain a recognition result.

[0010] Further, in step 1, the insulation defect types include at least one of an insulation outer surface scratch, an outer semiconductor creepage, an insulation main surface air gap, and an insulation main surface contamination.

[0011] Further, the high-frequency current coupling method is used to collect the partial discharge signal of the cable.

[0012] Further, in step 2, the phase space reconstruction is performed on each sample in the data set to generate point cloud data of the sample, specifically including:

[0013] Step 2.1: determining the embedding dimension m and the time delay parameter τ of the phase space reconstruction;

[0014] Step 2.2: for a given time series partial discharge signal x(t), t = 1, 2, …, T, obtaining a point cloud data sequence X(n), n = 1, 2, …, N with a sequence length of N according to formula (1); wherein N = T-(m-1)τ;

[0015]

[0016] Further, in step 2, the topological features of the sample are converted into a persistence trend curve or a Betti curve, specifically including:

[0017] Step 2.3: generating a bar code by arranging all the cohomologies on the vertical axis, with the horizontal axis representing the survival time of each cohomology, and the bar code is recorded as the persistence bar code of the point cloud data;

[0018] Step 2.4: generating a scatter plot by arranging all the cohomologies on the vertical axis, with the horizontal axis representing the generation time of each cohomology, and the scatter plot is recorded as the persistence scatter plot of the point cloud data;

[0019] Step 2.5: drawing a persistence trend curve and / or a Betti curve according to the persistence bar code and the persistence scatter plot.

[0020] Further, in step 2.1, the determination of the time delay parameter τ of the phase space reconstruction specifically includes:

[0021] After determining the embedding dimension m of the phase space reconstruction, an m-dimensional vector is initialized to represent a point cloud data;

[0022] A partitioning indicator function is defined, and the vectorized point cloud data is partitioned using the partitioning indicator function. The partitioning probability p when the point cloud data belongs to a certain partition A is calculated. A ;

[0023] Calculate the sequence symbol entropy h(m,τ) of each point cloud data according to formula (2):

[0024]

[0025] Among them, Γ m Represents a set of partition labels;

[0026] The fitness function of the particle swarm algorithm is set as shown in formula (3), and the optimal partition indicator value ε is determined by the particle swarm algorithm.

[0027]

[0028] Among them, FP + (k,τ) and FP - (k,τ) represents the maximum and minimum values ​​of h(m,τ) in the interval (τ-k,τ+k), where k is the given time delay parameter and peak search parameter.

[0029] The optimal time delay parameter τ is obtained by setting the sequence symbol entropy of the partition to the optimal partition indicator value ε and obtaining the time delay parameter corresponding to the maximum value. * As shown in formula (4):

[0030] τ * =argmax{h(m,τ)} (4).

[0031] Furthermore, the partition indicator function is as shown in formula (5):

[0032]

[0033] Among them, if|v i |,|v j |<εor|v i |,|v j |≥ε means when|v i |and|v j |Simultaneously satisfying either less than or greater than or equal to the partition marker value ε;

[0034] For any vectorized point cloud data, the partitioning mapping result is shown in formula (6):

[0035] f ε (v=(v1,v2,...,v m ))=(δ 12 (v),δ 13 (v),...,δ1m (v)) (6)

[0036] The partition probability p when the point cloud data belongs to a certain A-type partition is calculated according to formula (7). A :

[0037]

[0038] Where #{X(t)of A-type} indicates that X(t) in the point cloud data sequence belongs to the A-type partition.

[0039] Furthermore, the persistent situation curve Λ is expressed using formula (8). p (t):

[0040]

[0041] Where d is the extinction time of a homology, b is the generation time of a homology, α represents a certain homology, and the horizontal axis in the persistence curve is...

[0042] Furthermore, the high-voltage cable pattern recognition model employs any one of the following: random forest, support vector machine, backpropagation neural network, and one-dimensional convolutional neural network.

[0043] On the other hand, the present invention provides a high-voltage cable pattern recognition device based on time-series topology data analysis, comprising:

[0044] Partial discharge signal acquisition module, used to acquire partial discharge signals of cables;

[0045] The dataset construction module is used to construct a dataset containing M samples by taking the partial discharge signal of each known insulation defect type of cable as a sample and the corresponding insulation defect type as the sample label.

[0046] The phase space reconstruction module is used to reconstruct the phase space of each partial discharge signal and generate the corresponding point cloud data;

[0047] The topology data analysis module is used to perform topological space modeling on each point cloud data based on continuous coherence to extract topological features, and to convert the topological features into persistent situation curves and / or Betty curves.

[0048] The high-voltage cable pattern recognition module is used to divide the dataset into a training set and a test set, use the persistent state curve or Betty curve of each sample in the training set as input to train a given high-voltage cable pattern recognition model, and use the trained high-voltage cable pattern recognition model to identify the partial discharge signal of the cable to be detected.

[0049] The beneficial effects of this invention are:

[0050] This invention provides a high-voltage cable pattern recognition method and apparatus based on temporal topology data analysis. It uses phase space reconstruction to extract dynamic change information of partial discharge signals. The point cloud geometric distribution corresponding to partial discharge signals under different defect types varies. The topology data analysis method can describe the geometric features of the point cloud distribution through the duration of coherence, encompassing the positional distribution information of each point within continuous coherence. This means that even subtle changes in the partial discharge signal are reflected in the topological features. Furthermore, for partial discharge signals of high-voltage cables with two defects—scratches on the outer surface of the insulation and contamination on the main surface of the insulation—although there are similarities in amplitude, frequency, and PRPD spectra, there are still some differences in phase distribution. Therefore, this invention performs phase space reconstruction on the partial discharge signal, extracts topological features, and performs subsequent model training and testing. This avoids misjudgments in the pattern recognition model caused by phase shifts in the test voltage and similar phase distributions of different pattern samples when there is no reference voltage. It achieves a high high-voltage cable pattern recognition rate and has significant guiding significance for cable operation, maintenance, and fault analysis and repair. Attached Figure Description

[0051] Figure 1 A flowchart illustrating a high-voltage cable pattern recognition method based on temporal topology data analysis provided in an embodiment of the present invention;

[0052] Figure 2 Schematic diagrams of four types of cable insulation defects provided in embodiments of the present invention;

[0053] Figure 3 A schematic diagram of a partial discharge detection system for cables provided in an embodiment of the present invention;

[0054] Figure 4 The embodiments of the present invention provide a basis for... Figure 2 The point cloud distribution corresponding to the partial discharge signals of the four types of cable defects shown is reconstructed from the phase space;

[0055] Figure 5 A schematic diagram illustrating topological space modeling and topological feature extraction provided in an embodiment of the present invention;

[0056] Figure 6 A diagram showing the correspondence between continuous barcodes and continuous scatter plots provided in embodiments of the present invention;

[0057] Figure 7 A diagram showing the correspondence between a persistent scatter plot and a persistent situation curve provided in an embodiment of the present invention;

[0058] Figure 8A diagram showing the correspondence between continuous barcodes and Betty curves provided in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment of the invention provides a high-voltage cable pattern recognition method based on time-series topology data analysis, including the following steps:

[0062] S101: Collect partial discharge signals of cables with known insulation defect types. Each partial discharge signal is a sample, and the corresponding insulation defect type is used as the sample label. Construct a dataset containing M samples.

[0063] For example, Figure 2 The document showcases four common insulation defect types in cross-linked polyethylene cables. Defects 1 to 4 represent scratches on the outer surface of the insulation, creepage on the outer semiconductor, air gaps on the main surface of the insulation, and contamination on the main surface of the insulation, respectively.

[0064] To construct the dataset, this embodiment provides the fabrication process for the above four types of insulation defects, as detailed below:

[0065] Defect 1: During the manufacturing process of cable terminals, it is necessary to peel off the outer semiconductive layer of the insulation layer. The peeling process can easily damage the main insulation, thus forming this type of defect. In this embodiment, a scratch with a length of 3cm and a depth of 1mm is processed on the outer surface of the insulation layer of the sample cable.

[0066] Defect 2: If there is any residue at the break point of the outer semiconductive layer during the cable joint manufacturing process, creepage is likely to occur. In this embodiment, during the process of stripping the outer semiconductive layer of the sample cable, a 2cm wide and 10cm long portion of the outer semiconductive layer was left at one end that was not completely removed.

[0067] Defect 3: Cross-linked polyethylene insulation materials are prone to uneven foaming reaction and concentrated gas formation during extrusion, thus forming this type of defect. In this embodiment, multiple epoxy resin sheets are tightly stacked and the inner sheets are perforated to simulate this type of defect.

[0068] Defect 4: During the cable terminal manufacturing process, metal debris can easily adhere to the surface of the main insulation, thus forming this type of defect. In this embodiment, a small amount of metal debris is attached to the main insulation surface of the sample cable to simulate this type of defect.

[0069] It should be noted that the above-mentioned insulation defect types are only illustrative explanations of the identification method of the present invention and do not constitute an improper limitation on the technical solution of the present invention. Those skilled in the art can select several types of insulation defects according to the specific circumstances of the implementation and identify such types of insulation defects, and are not limited to the four types of insulation defects mentioned above.

[0070] In this embodiment, a high-frequency current coupling method is used to acquire the partial discharge signal of the cable. Specifically, Figure 3 A schematic diagram of a partial discharge detection system for cables is shown. (For example...) Figure 3 As shown. During the test, the 380V power frequency voltage was converted to the required high voltage after passing through voltage regulator T1 and step-up transformer T2. The high-voltage power frequency current flowed into the cable through protective resistor R1. Capacitors C1 / C2 acted as voltage dividers, C3 was a coupling capacitor used to couple the high-frequency pulse current, and R2 was the sensing impedance. The HFCT-49 high-frequency current sensor was installed on the grounding wire of the cable's copper shielding layer. The signal waveform was observed in a Tektronix 7000 oscilloscope with a sampling rate set to 100MS / s. For each defect model, the test was repeated 10 times at each corresponding voltage, and the partial discharge signal data was recorded.

[0071] It should be noted that the partial discharge signals acquired using the high-frequency current coupling method generally contain fundamental noise. To remove this noise, wavelet packet thresholding can be used to further extract the effective partial discharge signal from the original signal. Furthermore, to reduce the computational load of subsequent phase space reconstruction and topological feature extraction processes, the denoised partial discharge signal is divided into 3600 phase partitions. The amplitude of each partition is taken as the maximum discharge magnitude within that partition, and the 3600 maximum discharge magnitudes are combined to form the final partial discharge signal.

[0072] S102: Reconstruct the phase space for each sample in the dataset to generate point cloud data for that sample; perform topological space modeling on the point cloud data of that sample based on persistent homology to extract topological features; convert the topological features of that sample into persistent situation curves and / or Betty curves.

[0073] Specifically, in the process of modeling point cloud data using the continuous cohomology method, the extracted topological features are cohomology, the duration of each cohomology is recorded, and the cohomology is classified according to its type.

[0074] For example, if the point cloud data has a dimension of 3, the topological space modeling process involves generating spheres with radius ε, centered on each point cloud data point. Connectivity is generated when the spheres intersect to reach a critical value. If a "hole" is formed in the connectivity, it is considered homology, or homology is "generated." If the hole structure changes as the radius increases, homology "disappears." The homology recording method for higher dimensions follows the same logic and will not be elaborated here.

[0075] S103: Divide the dataset into a training set and a test set (e.g., randomly select 80% of the samples as the training set and the remaining 20% ​​as the test set), and use the persistent state curve or Betty curve of each sample in the training set as input to train the given high-voltage cable pattern recognition model.

[0076] Specifically, both persistent state curves and Betty curves are one-dimensional sequence data, thus offering strong compatibility with machine learning and deep learning. The Betty curve's advantage over the persistent state curve lies in its inclusion of homogeneous order information, resulting in richer information content. Conversely, the persistent state curve's advantage over the Betty curve lies in its smaller data volume, faster pattern recognition convergence speed, and ability to adapt to a wider range of pattern recognition models. Therefore, one type of curve or a combination of both can be chosen as the model input, depending on the requirements.

[0077] It is understandable that, for persistent situation curves, zero-persistence persistent situation curves are close to proportional function curves in shape. Two-dimensional persistent situation curves have relatively few two-dimensional cohomology and strong randomness. Therefore, taking all factors into consideration, in this embodiment, the persistent situation curve corresponding to one-dimensional cohomology is selected as the input for pattern recognition.

[0078] S104: Acquire the partial discharge signal of the cable under test, and obtain the persistent state curve or Betty curve of the cable under test based on the partial discharge signal. Input the persistent state curve or Betty curve into the trained high-voltage cable pattern recognition model to obtain the recognition result.

[0079] Specifically, the high-voltage cable pattern recognition model employs any one of the following: Random Forest (RF), Support Vector Machine (SVM), Backpropagation Neural Network (BPNN), and One-Dimensional Convolutional Neural Network (1D-CNN). Random Forest is an ensemble learning machine learning method, evolved from the decision tree algorithm. It simultaneously creates multiple tree models, each independent of the others, using the same parameters. It requires random sampling of the dataset, including random sample sampling and feature sampling, and finally combines all the tree models, taking the mode as the final classification result. Support Vector Machine is a non-linear classifier. Many samples are linearly inseparable in low-dimensional space; by selecting a kernel function, the data can be mapped to a high-dimensional space, thus resolving the linear inseparability problem in the original space. Backpropagation Neural Network is the foundation of neural networks, essentially constructing a regression model simultaneously with the training samples. Since the gradient describes the direction of the fastest growth of the function value, to find the point that minimizes the cost function, it is necessary to move backward along the gradient direction, iteratively improving the weights and offsets of each neuron so that the neural network ultimately outputs the desired result. Convolutional neural networks (CNNs) are models that perform convolution operations on input data to complete classification tasks. They exhibit excellent performance in image recognition and classification problems involving input samples with relatively large datasets and difficult feature extraction. Unlike ordinary CNNs, which use two-dimensional images or data as input, one-dimensional CNNs use one-dimensional data as input, making them compatible with persistent state curves and Betty curves.

[0080] It is understandable that phase space reconstruction describes the dynamic changes of partial discharge signals. The point cloud geometric distributions corresponding to the partial discharge signals under the four defect types have certain differences. Topological data analysis can describe the geometric features of the point cloud distribution through the duration of coherence. Topological data analysis can include the positional distribution information of each point in the continuous coherence, and is a feature extraction method that can automatically encompass rich information. Subtle changes in the partial discharge signal will be reflected in the topological features. Furthermore, for the partial discharge signals of cables with two defects—scratches on the outer surface of the high-voltage cable insulation and contamination on the main surface of the insulation—although there are certain similarities in amplitude, frequency, and PRPD spectrum, there are still certain differences in phase distribution. Therefore, this embodiment performs phase space reconstruction on the partial discharge signal, and then extracts topological features for subsequent model training and testing. This can avoid the phenomenon of misjudgment by the pattern recognition model due to phase shift of the test voltage and the similarity of phase distributions of different mode samples when there is no reference voltage.

[0081] Example 2

[0082] Based on the above embodiments, this embodiment mainly provides one possible implementation method for step S102. Regarding phase space reconstruction, it specifically includes the following steps:

[0083] Step S1021: Determine the embedding dimension m and time delay parameter τ for phase space reconstruction;

[0084] Specifically, since a higher embedding dimension does not improve the accuracy of the pattern recognition model, the embedding dimension of the phase space reconstruction should not exceed 3 in order to balance the difficulty of topological feature extraction (or the computational cost of feature extraction) and the visualization effect of topological features; in this embodiment, the embedding dimension is chosen to be m = 3. After determining the embedding dimension m of the phase space reconstruction, the time delay parameter τ needs to be determined, which will be exemplified below and will not be repeated here.

[0085] Step S1022: Based on the determined embedding dimension m and time delay parameter τ, for a given time-series partial discharge signal x(t), t=1,2,…,T, obtain a point cloud data sequence X(n), n=1,2,…,N with a sequence length of N according to formula (1); where N=T-(m-1)τ;

[0086]

[0087] Specifically, after determining the embedding dimension m of the phase space reconstruction, determining the time delay parameter τ of the phase space reconstruction includes the following sub-steps:

[0088] Step A1: Initialize m-dimensional vector To represent a point cloud data;

[0089] Step A2: Set a partitioning indicator function, use the partitioning indicator function to perform partition mapping on each vectorized point cloud data, and calculate the partition probability p when the point cloud data belongs to a certain partition A. A ;

[0090] Step A3: Calculate the sequence symbol entropy h(m,τ) of each point cloud data according to formula (2):

[0091]

[0092] Among them, Γ m This represents the set of partition labels.

[0093] Step A4: Set the fitness function of the particle swarm algorithm as shown in formula (3), and use the particle swarm algorithm to determine the optimal partition indicator value ε;

[0094]

[0095] Among them, FP + (k,τ) and FP- (k,τ) represents the maximum and minimum values ​​of h(m,τ) in the interval (τ-k,τ+k), where k is a given time delay parameter and a peak search parameter.

[0096] Specifically, the peak values ​​are first detected using a peak detection function, and then the maxima and minima of h(m,τ) are calculated based on these peak values. In this embodiment, the peak detection function is set as follows:

[0097]

[0098]

[0099]

[0100]

[0101] Where, when l=0,1,…,k, h τ+l =h(m,τ+l). FP + (k,τ),FP - (k,τ) are the maximum and minimum values ​​of h(m,τ) in the interval (τ-k,τ+k), respectively. The range of k is usually [3,5], and k is an integer.

[0102] Furthermore, FP + (k,τ) and FP - The calculation method for (k,τ) is as follows:

[0103]

[0104]

[0105] At the same time, the optimal ε should satisfy: "*" indicates the optimal value, 0≤ε≤max{x(t)}. To calculate the optimal ε, an exhaustive search method can be used, but this method involves a large amount of computation. Preferably, in this embodiment, the particle swarm optimization algorithm is chosen to find the optimal ε to reduce the amount of computation and speed up the convergence. Combining the conditions that the optimal ε should satisfy, the fitness function shown in formula (3) is set to evaluate the quality of the particle position. At the same time, for the particle swarm optimization algorithm, each particle (i.e., the value of parameter ε) updates its velocity and position through the following two formulas:

[0106]

[0107] x t+1 (i,k)=x t (i,k)+v t (i,k)

[0108] In the formula, t represents the number of iterations; w is the inertia weight; c1 and c2 are learning factors; rand is a random number function with a range of [0,1]; p best Indicates the optimal position the particle has experienced; g best This indicates the optimal position that the population has experienced.

[0109] Step A5: Using the optimal partition indicator value ε, establish the sequence symbol entropy of the partition and obtain the delay parameter corresponding to the maximum value as the optimal delay parameter τ. * As shown in formula (4):

[0110] τ * =argmax{h(m,τ)} (4).

[0111] Specifically, regarding step A2, the partitioning indicator function used in this embodiment is shown in formula (5); for any vectorized point cloud data, the partitioning mapping result is shown in formula (6); the partitioning probability p when the point cloud data belongs to a certain A-type A partition is calculated according to formula (7). A .

[0112]

[0113] f ε (v=(v1,v2,…,v m ))=(δ 12 (v),δ 13 (v),…,δ 1m (v))(6)

[0114]

[0115] Among them, if|v i |,|v j |<εor|v i |,|v j |≥ε means when|v i |and|v j |Simultaneously satisfying either less than or greater than or equal to the partition marker value ε; The function representing the mapping from point cloud data vectors to partitions, due to Γ m The elements in f are symbols, therefore f ε Also known as symbolic mapping; #{X(t)of A-type} indicates that X(t) in the point cloud data sequence belongs to type A partition, and the partition label set Γ is as follows. m The elements consist of {0,1}, have a length of m-1, and a base of 2. m-1 Point cloud data can be divided into 2 using formula (6). m-1 Disjoint partitions. For example: if (0,1)∈Γ m and Then there exists a mapping f ε If (v) = (0,1), then we call v a (0,1)-type.

[0116] Based on this, the range of symbol entropy h(m,τ) is [0, ln(1 / I)]. When the number of partitions of this sequence is I = 1, the lower bound of symbol entropy h(m,τ) is 0. When the probability of the sequence belonging to each partition is equal, the upper bound of symbol entropy h(m,τ) is ln(1 / I).

[0117] Furthermore, regarding the conversion of topological features into persistent situation curves or Betty curves, the specific details include:

[0118] S1023: The horizontal axis represents the duration of each homology, and all homology are arranged on the vertical axis to generate a barcode as the continuous barcode of the point cloud data.

[0119] S1024: The horizontal axis represents the generation time of each homology, and the vertical axis represents the extinction time of each homology. A scatter plot is generated and recorded as the persistent scatter plot of the point cloud data.

[0120] S1025: Draw a persistent situation curve and / or a Betty curve based on the persistent barcode and the persistent scatter plot.

[0121] Specifically, persistent scatter plots can intuitively describe the duration of each homology and have a strong correspondence with persistent state curves; persistent barcodes not only include the duration of homology but also the order of homology, and have a strong correspondence with Betty curves.

[0122] As one possible implementation method, the persistent situation curve Λ is represented by formula (8). p (t):

[0123]

[0124] Where d is the extinction time of a homology, b is the generation time of a homology, α represents a certain homology, and the horizontal axis in the persistence curve is...

[0125] The Betty curve describes the Betty number of a point cloud dataset at different radius values ​​ε, while the m-order Betty number β... m That is, the homology group H m The algebraic rank of β, i.e.: m =rankH m At this point, the homology group is set as a homology set with the same dimension for each sample.

[0126] Example 3

[0127] Corresponding to the above method, this embodiment of the invention provides a high-voltage cable pattern recognition device based on time-series topology data analysis, including: a partial discharge signal acquisition module, a dataset construction module, a phase space reconstruction module, a topology data analysis module, and a high-voltage cable pattern recognition module;

[0128] The system comprises several modules: a partial discharge signal acquisition module for acquiring partial discharge signals from the cable; a dataset construction module for constructing a dataset containing M samples, using the partial discharge signal of each known insulation defect type as a sample and the corresponding insulation defect type as the sample label; a phase space reconstruction module for reconstructing the phase space of each partial discharge signal to generate corresponding point cloud data; a topology data analysis module for performing topology space modeling on each point cloud data based on persistent homology to extract topology features, and converting the topology features into persistent state curves and / or Betty curves; and a high-voltage cable pattern recognition module for dividing the dataset into training and testing sets, using the persistent state curve or Betty curve of each sample in the training set as input to a given high-voltage cable pattern recognition model for training, and using the trained high-voltage cable pattern recognition model to identify the partial discharge signals of the cable under test.

[0129] It should be noted that the apparatus provided in the embodiments of the present invention is for implementing the above method embodiments, and its specific functions can be referred to the above method embodiments, which will not be repeated here.

[0130] To verify the effectiveness of the present invention, based on the above embodiments, the present invention also provides the following experimental data.

[0131] The phase space reconstruction effects of partial discharge signals for four typical defects after selecting the optimal time delay parameters are as follows: Figure 4 As shown, observations reveal that the point cloud geometric distributions of the partial discharge signals corresponding to the four typical defects exhibit good discriminative power.

[0132] The process of topological space modeling is as follows Figure 5 As shown. During the modeling process, each independent point cloud data point is considered an independent sphere with a radius of ε = 0, and is called a 0-simplex in the complex construction process. As the radius ε increases, two spheres will merge. When the radius of the sphere equals the distance between two data points, a connecting line appears at the center of the sphere, and the data points "disappear." This connected body is the 1-simplex. As the radius increases, one-dimensional or even high-dimensional holes appear. In topology, holes are called homology. For example, the first hole appears when ε = ε2, which is a one-dimensional homology. When ε = ε3, four data points generate a 3-simplex, at which point the one-dimensional homology "disappears." The duration of each homology structure is called persistent homology, and persistent homology is a topological feature.

[0133] To visualize topological features and convert them into a format more compatible with machine learning, persistent coherence is transformed into persistent barcodes and persistent scatter plots. For example... Figure 6 The horizontal axis on the left represents the radius of the sphere during the modeling process, and each barcode corresponds to the time of homology generation and destruction. The vertical axis represents the arrangement of each homology, which has no practical meaning. Figure 6 As shown on the right, a persistent scatter plot is another visualization method for persistent homology. Its horizontal axis represents the generation time of homology, and its vertical axis represents the extinction time of homology. The generation and extinction time of each homology is intuitively displayed in the form of a scatter plot. However, both persistent barcodes and persistent scatter plots are two-dimensional data formats, and their compatibility with machine learning needs to be improved.

[0134] To improve the compatibility of topological features with machine learning, the formats of persistent barcodes and persistent scatter plots are converted. For example... Figure 7 As shown, the correspondence between the persistent scatter plot and the persistent situation curve can be geometrically viewed as rotating the persistent scatter plot clockwise and drawing a piecewise function according to the vertices of the persistent situation.

[0135] like Figure 8 As shown, the Betty curve is drawn based on the distribution of continuous barcodes. The Betty curve represents the overall homology during the modeling process and is arranged according to the corresponding radius relationship during the modeling process, which is the time sequence.

[0136] To verify the effectiveness of topological features in partial discharge signal pattern recognition, random forest (RF), support vector machine (SVM), backpropagation neural network (BPNN), and one-dimensional convolutional neural network (1D-CNN) pattern recognition models were selected for training, and the pattern recognition results were validated. The pattern recognition results are shown in Table 1.

[0137] Table 1

[0138]

[0139] Analysis of the data in Table 1 shows that the pattern recognition model using persistent situation curves as input achieves a maximum recognition rate of 98.75% and a minimum of 94.45%. The pattern recognition model using Betty curves as input achieves a maximum recognition rate of 99.00% and a minimum of 92.50%. Because Betty curves contain homology order information and have a relatively large data volume, they pose a certain obstacle to the training process of common machine learning models such as RF and SVM, ultimately affecting the pattern recognition rate. However, they show relatively high pattern recognition rates and greater adaptability for deep learning models such as BPNN and 1D-CNN. Persistent situation curves demonstrate good pattern recognition rates for all four models, indicating that they are a more adaptable form of topological feature. Therefore, in practical pattern recognition processes, it is necessary to select an appropriate pattern recognition model based on the corresponding topological features.

[0140] The experimental data above show that the present invention can effectively extract the relevant features of partial discharge signals and achieve good performance in pattern recognition rate. It has important guiding significance for the operation, maintenance and fault analysis and repair of high-voltage cables, and has positive significance for the development of intelligent detection systems.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-voltage cable pattern recognition method based on time-series topology data analysis, characterized in that, include: Step 1: Collect partial discharge signals of cables with known insulation defect types. Each partial discharge signal is used as a sample, and the corresponding insulation defect type is used as the sample label to construct a dataset containing M samples. Step 2: Reconstruct the phase space for each sample in the dataset to generate the point cloud data of that sample; perform topological space modeling on the point cloud data of that sample based on continuous homology to extract topological features; The topological features of the sample are converted into persistent situation curves and / or Betty curves; Step 3: Divide the dataset into a training set and a test set, and use the persistent state curve or Betty curve of each sample in the training set as input to train the given high-voltage cable pattern recognition model. Step 4: Collect the partial discharge signal of the cable under test, and obtain the persistent state curve or Betty curve of the cable under test based on the partial discharge signal. Input the persistent state curve or Betty curve into the trained high-voltage cable pattern recognition model to obtain the recognition result.

2. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 1, characterized in that, In step 1, the insulation defect type includes at least one of the following: insulation outer surface scratches, external semiconductor creepage, insulation main surface air gaps, and insulation main surface contamination.

3. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 1, characterized in that, The partial discharge signal of the cable was acquired using a high-frequency current coupling method.

4. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 1, characterized in that, Step 2, which involves reconstructing the phase space for each sample in the dataset to generate the point cloud data for that sample, specifically includes: Step 2.1: Determine the embedding dimension m and time delay parameter τ for phase space reconstruction; Step 2.2: For a given time-series partial discharge signal x(t), t=1,2,…,T, obtain a point cloud data sequence X(n), n=1,2,…,N with a sequence length of N according to formula (1); where N=T-(m-1)τ; 5. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 1, characterized in that, In step 2, the topological features of the sample are converted into a persistent situation curve or a Betty curve, specifically including: Step 2.3: Use the horizontal axis to represent the duration of each homology, arrange all homology on the vertical axis, and generate a barcode to record the persistence barcode of the point cloud data; Step 2.4: Use the horizontal axis to represent the generation time of each homology and the vertical axis to represent the extinction time of each homology, and generate a scatter plot as the persistent scatter plot of the point cloud data. Step 2.5: Draw the persistent situation curve and / or Betty curve based on the persistent barcode and the persistent scatter plot.

6. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 4, characterized in that, In step 2.1, determining the time delay parameter τ for phase space reconstruction specifically includes: After determining the embedding dimension m of the phase space reconstruction, initialize the m-dimensional vector. To represent a point cloud data; A partitioning indicator function is defined, and the vectorized point cloud data is partitioned using the partitioning indicator function. The partitioning probability p when the point cloud data belongs to a certain partition A is calculated. A ; Calculate the sequence symbol entropy h(m,τ) of each point cloud data according to formula (2): Among them, Γ m Represents a set of partition labels; The fitness function of the particle swarm algorithm is set as shown in formula (3), and the optimal partition indicator value ε is determined by the particle swarm algorithm. Among them, FP + (k,τ) and FP - (k,τ) represents the maximum and minimum values ​​of h(m,τ) in the interval (τ-k,τ+k), where k is the given time delay parameter and peak search parameter. The optimal time delay parameter τ is obtained by setting the sequence symbol entropy of the partition to the optimal partition indicator value ε and obtaining the time delay parameter corresponding to the maximum value. * As shown in formula (4): t * = arg max{h(m,τ)} (4).

7. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 6, characterized in that, The partition indicator function is shown in formula (5): Among them, if|v i |,|v j |<εor|v i |,|v j |≥ε means when|v i |and|v j |Simultaneously satisfying either less than or greater than or equal to the partition marker value ε; For any vectorized point cloud data, the partitioning mapping result is shown in formula (6): f ε (v=(v1,v2,...,v m ))=(δ 12 (v),δ 13 (v),...,δ 1m (v)) (6) The partition probability p when the point cloud data belongs to a certain A-type partition is calculated according to formula (7). A : Here, #{X(t) of A-type} indicates that X(t) in the point cloud data sequence belongs to the A-type partition.

8. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 5, characterized in that, The persistent situation curve Λ is represented by formula (8). p (t): Where d is the extinction time of a homology, b is the generation time of a homology, α represents a certain homology, and the horizontal axis in the persistence curve is...

9. The high-voltage cable pattern recognition method based on time-series topology data analysis according to claim 6, characterized in that, The high-voltage cable pattern recognition model adopts any one of random forest, support vector machine, backpropagation neural network and one-dimensional convolutional neural network.

10. A high-voltage cable pattern recognition device based on time-series topology data analysis, characterized in that, include: Partial discharge signal acquisition module, used to acquire partial discharge signals of cables; The dataset construction module is used to construct a dataset containing M samples by taking the partial discharge signal of each known insulation defect type of cable as a sample and the corresponding insulation defect type as the sample label. The phase space reconstruction module is used to reconstruct the phase space of each partial discharge signal and generate the corresponding point cloud data; The topology data analysis module is used to perform topological space modeling on each point cloud data based on continuous coherence to extract topological features, and to convert the topological features into persistent situation curves and / or Betty curves. The high-voltage cable pattern recognition module is used to divide the dataset into a training set and a test set, use the persistent state curve or Betty curve of each sample in the training set as input to train a given high-voltage cable pattern recognition model, and use the trained high-voltage cable pattern recognition model to identify the partial discharge signal of the cable to be detected.

Images