Method and system for learning-based identification of latent faults in medium voltage distribution cables

Abstract

The application discloses a kind of medium voltage distribution cable latent fault guided learning identification method and system, comprising: obtaining the latent fault current data of cable;Extract the fault feature vector of fault current data under time domain, frequency domain and time-frequency domain working condition, and the dimensionality reduction optimization of fault feature vector is carried out;Based on the optimized fault feature vector, the data-knowledge joint driving fault identification model is utilized to obtain the cable latent fault identification result;Wherein, the data-knowledge joint driving fault identification model is obtained by the pure data driving model based on extreme learning machine and the field experience knowledge fusion;The field experience knowledge of cable latent fault identification is introduced into pure data driving model in the application and is guided to limit, so that the fusion model obtains the ability that strong generalization performance, high stability, strong robustness, serious classification error is not easy to occur, accurate discrimination medium voltage distribution cable latent fault is realized.

Description

Technical Field

[0001] This invention relates to the field of cable latent fault identification technology, and in particular to a guided learning identification method and system for latent faults in medium-voltage distribution cables. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, power cables have been widely used in urban power distribution networks due to their advantages such as high power supply reliability, small footprint, and aesthetic appeal. However, cables buried underground experience a gradual decrease in insulation strength over time due to declines in electrical, thermal, and mechanical strength, eventually leading to permanent faults and jeopardizing the safe and reliable power supply of the urban power distribution network. Operational experience with medium-voltage cables shows that before a permanent fault occurs, a transient, self-recovering arc-to-ground fault may occur at the same location on the cable. Because this type of fault is short-lived and has a small fault current, traditional overcurrent protection devices with inverse-time characteristics cannot activate; this type of fault is called a latent cable fault. Statistics show that 10% to 15% of cable faults have experienced a latent fault stage before their occurrence. Therefore, timely detection and targeted maintenance of latent faults in medium-voltage distribution cables can effectively prevent them from developing into permanent faults and improve the power supply reliability of the distribution network.

[0004] As mentioned earlier, cable latent faults are characterized by short duration and low fault energy, making traditional overcurrent protection devices in power systems ineffective against such faults. Furthermore, cable latent faults share similarities with other overcurrent disturbance waveforms in power systems, such as load surges, capacitor switching, and constant impedance grounding faults, which further complicates the identification of cable latent faults. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a guided learning method and system for identifying latent faults in medium-voltage power distribution cables. By mining experiential knowledge in the field of cable latent fault identification and integrating it with a fault identification model based on Extreme Learning Machine, a data-knowledge jointly driven cable latent fault identification method is established, enabling accurate identification of latent faults in medium-voltage power distribution cables.

[0006] In some implementations, the following technical solutions are adopted:

[0007] A method for learning-guided identification of latent faults in medium-voltage distribution cables includes:

[0008] Obtain fault three-phase current data and zero-sequence current data of the cable;

[0009] Extract the fault feature vectors of the current data under time domain, frequency domain and time-frequency domain conditions respectively, and optimize the dimensionality reduction of the fault feature vectors;

[0010] Based on the optimized fault feature vectors, use the trained data-knowledge joint-driven fault identification model to obtain the identification result of the cable latent fault;

[0011] Among them, the data-knowledge joint-driven fault identification model is obtained by fusing a pure data-driven model based on extreme learning machine and domain experience knowledge;

[0012] Use the cable latent fault arc model to construct domain experience knowledge, and construct a knowledge function based on the loss function of the error between the output of the pure data-driven model and the actual label of the training set samples and the conflict function representing the conflict between the output of the pure data-driven model and the domain experience knowledge; Train and optimize the data-knowledge joint-driven fault identification model with the goal of minimizing the knowledge function value.

[0013] Among them, using the cable latent fault arc model to construct domain experience knowledge specifically includes:

[0014] Knowledge (1): If the duration of the fault or transient disturbance reaches five cycles or more, it must not be a cable latent fault;

[0015] Knowledge (2): If the mutation degree of zero-sequence current before and after the fault < a, it must not be a cable latent fault; where a is a set value.

[0016] In some other embodiments, the following technical solutions are adopted:

[0017] A method for guiding learning and identifying latent faults of medium-voltage distribution cables includes:

[0018] A data acquisition module for acquiring three-phase and zero-sequence current data of cable faults;

[0019] A feature extraction module for respectively extracting fault feature vectors of current data under time domain, frequency domain and time-frequency domain conditions, and optimizing the dimensionality reduction of the fault feature vectors;

[0020] A fault identification module for obtaining the identification result of the cable latent fault based on the optimized fault feature vectors by using the trained data-knowledge joint-driven fault identification model;

[0021] Among them, the data-knowledge joint-driven fault identification model is obtained by fusing a pure data-driven model based on extreme learning machine and domain experience knowledge;

[0022] Domain experience knowledge is constructed using a cable latent fault arc model. A knowledge function is constructed based on the loss function of the error between the pure data-driven model output and the actual labels of the training set samples, and the conflict function representing the contradiction between the pure data-driven model output and the domain experience knowledge. The data-knowledge joint-driven fault identification model is trained and optimized with the goal of minimizing the value of the knowledge function.

[0023] In other embodiments, the following technical solutions are adopted:

[0024] A terminal device includes a processor and a memory, the processor being used to implement instructions; the memory being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned method for guided learning and identification of latent faults in medium-voltage power distribution cables.

[0025] In other embodiments, the following technical solutions are adopted:

[0026] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described method for guided learning and identification of latent faults in medium-voltage power distribution cables.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] (1) This invention introduces empirical knowledge in the field of cable latent fault identification into the pure data-driven model and guides and restricts it, so that the fusion model retains the powerful pattern mining ability of the pure data-driven algorithm, while also gaining the ability of strong generalization performance, high stability, strong robustness and low risk of serious classification errors brought about by the introduction of empirical knowledge. It can accurately detect and identify latent faults in medium-voltage distribution cables, avoid interference from other overcurrent disturbance waveforms, and finally carry out timely maintenance to avoid the occurrence of permanent faults.

[0029] (2) When introducing domain experience knowledge into a pure data-driven model, this invention incorporates the setting of a random factor, which can be adjusted according to the sample size. When the sample size is insufficient, the value of the random factor is reduced to give fuller play to the guiding role of experience knowledge and reduce the model's excessive dependence on sample data.

[0030] (3) The present invention makes reasonable settings for the calculation rules of the conflict function, which greatly improves the safety performance of the model.

[0031] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of a medium-voltage distribution network model in an embodiment of the present invention;

[0033] Figure 2 This is a waveform diagram of the half-cycle latent fault current of a medium-voltage power distribution cable in an embodiment of the present invention.

[0034] Figure 3 This is a waveform diagram of the multi-cycle latent fault current of a medium-voltage power distribution cable in an embodiment of the present invention.

[0035] Figure 4 This is a graph showing the number of eigenvalues ​​and the cumulative variance contribution rate obtained by principal component analysis in this embodiment of the invention.

[0036] Figure 5 This is a schematic diagram of the Extreme Learning Machine (ELM) network structure in an embodiment of the present invention. Detailed Implementation

[0037] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0038] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0039] Example 1

[0040] Whether it's machine learning algorithms or the increasingly widely used deep learning, they are essentially data-driven algorithms. Data-driven algorithms, especially deep learning algorithms, have high requirements for sample size and quality. Over-reliance on samples, insufficient sample size, or poor sample quality often leads to unsatisfactory results. However, in actual operation, power distribution networks are mostly in normal operating conditions, with relatively few fault samples. Secondly, pure data-driven algorithms are essentially "black box" models, especially deep learning algorithms. Engineers cannot understand their decision-making process, which leads to a lack of security in the output results, posing a serious potential threat to the safe and reliable operation of the power system. Finally, pure data-driven models only extract patterns from the training set data, resulting in limited generalization performance.

[0041] The root cause of the above problems lies in the fact that purely data-driven algorithms adopt the empirical risk minimization criterion, failing to utilize prior knowledge to provide guidance or correction for pattern discovery during algorithm training. To address these issues, this invention mines empirical knowledge in the field of cable latent fault identification and integrates it into a fault identification model based on extreme learning machines, establishing a data-knowledge jointly driven cable latent fault identification model.

[0042] This invention provides guided learning for identifying latent faults in medium-voltage distribution network cables. The first stage involves summarizing domain-specific experience and establishing fault and disturbance condition models. Based on these models, domain-specific experience in cable latent fault identification is summarized. The second stage establishes a cable latent fault identification model based on an extreme learning machine and constructs knowledge functions, forming a data-knowledge jointly driven cable latent fault identification model.

[0043] Based on this, in one or more embodiments, a method for guided learning and identification of latent faults in medium-voltage distribution cables is disclosed, specifically including the following process:

[0044] (1) Acquire experience and knowledge in the field of cable latent fault identification; including the following steps:

[0045] (1-1) Obtain the arc model of latent faults in cables;

[0046] Since latent cable faults are mostly low-energy, intermittent arc faults, this embodiment uses the Kizilcay arc model to simulate latent cable faults. The Kizilcay arc model is a dynamic arc model based on arc gap energy balance and cybernetics, and is widely used in the analysis and research of arc faults. Its expression is:

[0047]

[0048] Where τ is the time constant, i f g(n) is the arc current, u0 is the arc characteristic voltage, r0 is the arc characteristic resistance, and g(n) is the arc conductance.

[0049] (1-2) Build a model of the medium-voltage cable distribution system and obtain the fault current waveform and disturbance current waveform;

[0050] A 10kV medium-voltage cable distribution system was built in PSCAD / EMTDC using the Bergeron cable model, with a sampling frequency of 10kHz. Figure 1 As shown. The waveforms of the latent fault current in half-cycle and multi-cycle cables are obtained, as follows: Figure 2 and Figure 3 As shown.

[0051] Cable latent faults have the characteristics of self - recovery and extremely short duration. It is very easy for relay protection devices to identify them as transient disturbances. At the same time, during the operation of the power system, the normal switching of equipment such as load switching and capacitor switching also belongs to transient disturbances, and the latter has similarities with the transient characteristics of cable latent faults. In addition, since latent faults are mostly single - phase grounding faults, in order to prevent misjudging the two and causing unnecessary tripping or missed tripping, it is necessary to comprehensively consider the influence of over - current disturbances and single - phase grounding constant - impedance faults. Based on this idea, this invention considers five working conditions: half - cycle and multi - cycle cable latent faults, constant - impedance grounding faults, capacitor switching, and load mutation.

[0052] (1 - 3) Summarize domain experience knowledge based on theoretical knowledge and simulation waveforms; where domain experience knowledge refers to the basic knowledge, laws or experiences of the power system. Without deliberately adding assumptions, it guides the sample training during the learning process itself, realizing the transformation from hypothesis - based and statistical reasoning to knowledge - guided statistical reasoning generalization, and integrating knowledge analysis and data analysis and mining.

[0053] This embodiment takes into account the characteristics of domain experience knowledge and the actual requirements of cable latent fault identification, and initially extracts the following two pieces of knowledge:

[0054] Knowledge (1): If the duration of a fault or transient disturbance reaches five cycles or more, it must not be a cable latent fault;

[0055] Knowledge (2): If the mutation degree of zero - sequence current before and after a fault < a, it must not be a cable latent fault.

[0056] The calculation formula for the mutation degree of zero - sequence current before and after a fault is as follows:

[0057]

[0058] In the formula, where, i A , i B , i C are the instantaneous values of the currents of phases A, B, and C respectively, I is the effective value of the current before the occurrence of the fault, and Δi0 is the change in zero - sequence current before and after the fault.

[0059] Of the two points above, knowledge (1) is based on the basic definition of cable latent faults. Cable latent faults generally last for 1 / 4 to 4 cycles. By clearly defining the physical meaning, constant impedance grounding faults and partial load sudden disturbances can be distinguished from the other three types of faults and disturbances, effectively avoiding misjudging constant impedance grounding faults as cable latent faults. Knowledge (2) is based on the instantaneous overcurrent characteristics of cable latent faults. Cable latent faults are generally single-phase faults. During the duration of the fault, there is a large zero-sequence current. Except for constant impedance grounding faults, the zero-sequence current values ​​of the other two types of transient overcurrent disturbances are small. They can be distinguished by constructing a criterion through an empirical threshold a. The setting of the empirical threshold a can be set according to the actual situation. It is sufficient to provide directional guidance for the mining of the pattern of the training set samples. This ensures the requirements of knowledge universality and scalability in the fusion-driven model.

[0060] (2) Obtain fault three-phase current data and zero-sequence current data of the cable;

[0061] (3) Extract the fault feature vectors of the fault current data in the time domain, frequency domain and time-frequency domain respectively, and perform dimensionality reduction optimization on the fault feature vectors;

[0062] In this embodiment, fault features for each operating condition, as shown in Table 1, including time domain, frequency domain, and time-frequency domain features, are extracted. The time-frequency domain features are decomposed into three levels using stationary wavelet transform. Time domain extraction includes peak-to-peak value (C). t1 ), rectified average value (C) t2 ), variance (C) t3 Standard deviation (C) t4 ), effective value (C) t5 ), skewness (C) t6 ), kurtosis (C) t7 Peak factor (C) t8 Waveform factor (C) t9 ), extract the centroid frequency (C) in the frequency domain F1 ), average frequency (C F2 ), root mean square frequency (C F3 ), frequency standard deviation (C F4 Energy extraction in the time and frequency domain (C) w1 Energy entropy (C) w2 Information entropy (C) w3 ).

[0063] Taking the A-phase current of various faults as an example, calculate its 9-dimensional time-domain eigenvector C. ti Perform spectral analysis on it and calculate the 4-dimensional frequency domain eigenvector C. FiA dB4 mother wavelet was used to perform a three-level stationary wavelet transform on the sub-band. Each sub-band was divided into five time periods with an equal interval of 0.02s (power frequency period). Three feature quantities were calculated for each time period of each sub-band to extract a 60-dimensional time-frequency domain feature vector C. wi By combining the three-domain feature vectors, a 73-dimensional feature vector can be constructed for each fault sample [C]. ti C Fi C wi ].

[0064] Table 1 shows the extracted time-domain, frequency-domain, and time-frequency-domain fault feature vectors.

[0065]

[0066]

[0067] The meanings of the parameters involved in Table 1 are explained as follows:

[0068] x i Here, N represents the number of labeled samples, which is the sample feature data. s(k) is the average of all samples; s(k) is the spectrum of the original signal, where k represents the number of spectral lines, K is the total number of spectral lines, and f(k) represents the frequency amplitude corresponding to the k-th spectral line; ij The wavelet coefficients s of the j-th level obtained by performing stationary wavelet transform decomposition on the current signal. j =[s 1j ,s 2j ,…s Nj ].

[0069] Multivariate analysis of variance (MANOVA) was used to study the relationship between independent and dependent variables, and 21 effective features were obtained. Their p-values ​​and F-values ​​are shown in Table 2. Among them, C... w1(12) This represents the energy characteristics corresponding to the second time period of the first sub-band. Similarly, C w2(31) This represents the energy entropy characteristic corresponding to the first time period of the third sub-band. Multivariate analysis of variance (MAO) selected 21 effective features from 73 features, ensuring the classification performance of individual features while reducing feature dimensionality. Principal component analysis (PCA) was then used to further reduce the dimensionality of the feature vectors. Figure 4 As shown, when the cumulative variance contribution rate of the features reaches 95%, the principal component analysis method reduces the feature vector to five dimensions. The feature vector composed of the first five principal components replaces the original 21-dimensional feature vector. The dimensionality is greatly reduced while retaining the effective information of the original features to the greatest extent.

[0070] Table 2. Table of 21-dimensional effective characteristic parameters obtained after multivariate analysis of variance.

[0071]

[0072] (4) Based on the optimized fault feature vector, the trained data-knowledge joint-driven fault identification model is used to obtain the identification result of whether the cable fault is a latent fault.

[0073] Among them, the data-knowledge joint-driven fault identification model is obtained by fusing a pure data-driven model based on extreme learning machine (hereinafter referred to as ELM) and domain experience knowledge;

[0074] Domain experience knowledge is constructed using a cable latent fault arc model. A knowledge function is constructed based on the loss function of the error between the pure data-driven model output and the actual labels of the training set samples, and the conflict function representing the contradiction between the pure data-driven model output and the domain experience knowledge. The data-knowledge joint-driven fault identification model is trained and optimized with the goal of minimizing the value of the knowledge function.

[0075] Specifically, Extreme Learning Machine (ELM) is used to train a single hidden layer feedforward neural network (SLFN). ELM can randomly select input layer weights and hidden layer biases. The output layer weights are calculated analytically based on Moore-Penrose (MP) generalized inverse matrix theory by minimizing a loss function consisting of the training error term and the regularization term of the output layer weight norm.

[0076] Given a training set {x} i ,t i |x i ∈R D For a network with i = 1, 2, ..., N, the Extreme Learning Machine has L hidden layer nodes, and its network structure is as follows: Figure 5 As shown. Its hidden layer output is H(x) = [h1(x), ..., h...]. L (x)], where h i (x) is the output of the i-th hidden layer node, usually represented as:

[0077] h i (x)=g(w i ,b i ,x)=g(w i x+b i ),w i ∈R D ,b i ∈R (3)

[0078] Among them, g(w i ,b i ,x) is the activation function, which is a nonlinear piecewise continuous function that satisfies the general approximation theorem of ELM. Commonly used functions include the Sigmoid function and the Gaussian function.

[0079] After passing through the hidden layer, it enters the output layer, and its output is:

[0080]

[0081] where β = [β1, …, β L is the output weight between the hidden layer and the output layer.

[0082] The knowledge function of the data-knowledge jointly driven fault identification model is composed of two parts: one is the loss function that characterizes the error between the output of the pure data-driven model and the actual labels of the training set samples, and the other is the conflict function that characterizes the conflict between the output of the pure data-driven model and the domain experience knowledge. The following will be introduced in detail respectively.

[0083] For the loss function, cross-entropy stems from information theory,刻画 the difference degree between two different probability distributions in the same random variable, and is often used in machine learning algorithms to characterize the distance between the actual output and the true label. The smaller the cross-entropy, the closer the probability distributions of the two are. Its calculation formula is shown in Equation (5).

[0084]

[0085] In the formula, i is the dimension of the output vector, and p(i) and q(i) are the values of the i-th element in the actual label vector and the predicted output vector respectively.

[0086] For the conflict function, determine the number of cycles of the transient process of each fault sample and the mutation degree of the zero-sequence current before and after the fault, and thus determine the corresponding calculation rules for the conflict function:

[0087] Rule 1: If M = 1 and N = 1, then the conflict function value is increased by 1;

[0088] Rule 2: If M = 1 and Y = 1, then the conflict function value is increased by 1;

[0089] where M, N, and Y are data sets. For the output vector of the ELM model, if the column number where the maximum value is located is 1 or 2, then M is set to 1, otherwise 0. If the transient duration of the fault sample is greater than or equal to 5 cycles, then N is set to 1, otherwise 0; if the mutation degree of the zero-sequence current before and after a certain fault sample is calculated to be <a, then Y is set to 1, otherwise 0. M = 1 means that for a group of fault input samples, the pure data-driven model believes that the probability of its fault type being a latent fault is the largest; N = 1 and Y = 1 mean that judged by the domain experience knowledge, it is considered that the fault input sample must not be a cable latent fault.

[0090] It should be noted that the Chinese character "刻画" in the original text has been translated as "刻画" first and then left in Chinese in the translation as it seems to be a specific term that might not have a proper English equivalent in this context. If there is a more accurate English expression for it, it can be adjusted accordingly.The above two rules mean that, regarding the diagnosis of whether a fault belongs to a latent fault, if the output result of the ELM model is exactly the opposite of the diagnosis result obtained from empirical knowledge, the conflict function value is increased. This guides and restricts the training of the pure data-driven model, reduces the probability of low-level diagnostic errors, and reduces the probability of diagnosing a serious fault such as constant impedance grounding fault or other transient disturbances as a latent cable fault, thereby greatly improving the safety and reliability of the output result of the fusion model.

[0091] The knowledge function is a comprehensive indicator for evaluating the data-knowledge joint-driven fault identification model. The training and optimization of the data-knowledge joint-driven fault identification model involves adjusting and combining the parameters in the model, with the objective function being to achieve the optimal (minimum) value of the knowledge function.

[0092] The knowledge function of the data-knowledge joint-driven fault identification model can be obtained by adding the calculated cross-entropy loss function to the conflict function.

[0093] Suppose we have N labeled samples {(x i ,t i ), i = 1, 2, ..., N},x i t represents sample feature data. i If the true label of the sample is denoted as , then the knowledge function of the data-knowledge joint-driven fault identification model can be expressed as:

[0094]

[0095] In the formula, h(θ,x) i H[h(θ,x)] represents the output of the ELM model, and θ represents the model parameters. i ),t i ] represents the cross-entropy loss function between the calculated ELM model output and the actual labels; β i S(x) represents the degree of trust that the data-knowledge jointly driven fault identification model places in domain experience knowledge; i If P represents the classification result derived from domain experience knowledge, then P... i (x) represents the conflict value between the ELM model output calculated according to the rules and the empirical knowledge, and its specific value is shown in equation (7).

[0096]

[0097] Considering that the threshold of knowledge (2) in domain experience knowledge is an empirical value, its diagnostic results will inevitably have a small error. Therefore, this embodiment refers to the greedy algorithm in reinforcement learning to handle this problem. The concept of a random factor is set, and the trust level β is introduced. i This parameter has the following specific values:

[0098]

[0099] In the formula, ε is the random factor, reflecting the degree of involvement of empirical knowledge in the training process of the data-knowledge jointly driven fault identification model. Its value depends on the size of the training set. If the training set has sufficient samples, the classification effect of the purely data-driven model can be guaranteed, and ε should be taken to reduce the involvement of empirical knowledge; conversely, if the training set has insufficient samples, a smaller value can be taken to give full play to the guiding role of empirical knowledge and reduce the dependence on sample data. The value of c represents the specific trust level of empirical knowledge. When the domain empirical knowledge is deterministic knowledge with clear physical meaning, c can be taken to give the diagnostic results based on empirical knowledge a greater weight; conversely, the value of c can be adjusted appropriately.

[0100] The ELM model enhances training speed and generalization ability through random weights and biases, but it also causes the hidden layer output matrix to be ill-conditioned in some cases, thus affecting the stability of the model classification. Therefore, it is generally necessary to use an optimization algorithm to optimize the random initial parameters of the model. In this embodiment, the particle swarm optimization algorithm is used to optimize its parameters.

[0101] Particle Swarm Optimization (PSO) is a swarm intelligence evolutionary computation method that simulates the foraging behavior of bird flocks, achieving simplicity, high accuracy, and fast convergence. It first initializes the particle swarm and then tracks the individual extreme values ​​P. best and global extremum G best This is done to update the particle's position. Each time the particle's position is updated, the fitness value needs to be recalculated, and the particle's new fitness value is compared with P. best and G best Compare the results to find the optimal fitness value.

[0102] The training objective of the data-knowledge joint-driven fault identification model is to minimize the knowledge function. Therefore, the fitness function of the PSO algorithm is the knowledge function shown in equation (6), and the optimization objective is the random initial weights and biases of the ELM model. The joint-driven model is trained in this way.

[0103] This embodiment introduces experiential knowledge from the field of cable latent fault identification into a pure data-driven model and guides and restricts it. This allows the fusion model to retain the powerful pattern-mining ability of the pure data-driven algorithm while also gaining the ability to generalize strongly, stably, robustly, and less prone to serious classification errors brought about by the introduction of experiential knowledge. It can accurately detect and identify latent faults in medium-voltage distribution cables, avoid interference from other overcurrent disturbance waveforms, and ultimately carry out timely maintenance to avoid the occurrence of permanent faults.

[0104] Example 2

[0105] In one or more embodiments, a method for guided learning identification of latent faults in medium-voltage distribution cables is disclosed, including:

[0106] The data acquisition module is used to acquire the three-phase current and zero-sequence current data of cable faults;

[0107] The feature extraction module is used to extract fault feature vectors of current data in the time domain, frequency domain, and time-frequency domain, respectively, and to perform dimensionality reduction optimization on the fault feature vectors.

[0108] The fault identification module is used to obtain the identification results of cable latent faults based on the optimized fault feature vector and the trained data-knowledge joint driving fault identification model.

[0109] The data-knowledge joint-driven fault identification model is obtained by fusing a pure data-driven model based on extreme learning machine and domain experience knowledge.

[0110] Domain experience knowledge is constructed using a cable latent fault arc model. A knowledge function is constructed based on the loss function of the error between the pure data-driven model output and the actual labels of the training set samples, and the conflict function representing the contradiction between the pure data-driven model output and the domain experience knowledge. The data-knowledge joint-driven fault identification model is trained and optimized with the goal of minimizing the value of the knowledge function.

[0111] It should be noted that the specific implementation methods of the above modules have been described in Example 1, and will not be detailed here.

[0112] Example 3

[0113] In one or more embodiments, a terminal device is disclosed, including a server. The server includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method for guiding learning and identifying latent faults in medium-voltage power distribution cables as described in Embodiment 1. For the sake of brevity, further details are omitted here.

[0114] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0115] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0116] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0117] Example 4

[0118] In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed by the guided learning and identification method for latent faults in medium-voltage distribution cables described in Embodiment 1.

[0119] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for guided learning and identification of latent faults in medium-voltage power distribution cables, characterized in that, include: Obtain three-phase current and zero-sequence current data for cable faults; Fault feature vectors of current data under time, frequency, and time-frequency conditions are extracted respectively, and the fault feature vectors are optimized by dimensionality reduction. Based on the optimized fault feature vector, the trained data-knowledge joint-driven fault identification model is used to obtain the cable latent fault identification results. The data-knowledge joint-driven fault identification model is obtained by fusing a pure data-driven model based on extreme learning machine and domain experience knowledge. Domain experience knowledge is constructed using a cable latent fault arc model. A knowledge function is constructed based on the loss function of the error between the pure data-driven model output and the actual labels of the training set samples, and the conflict function representing the contradiction between the pure data-driven model output and the domain experience knowledge. The data-knowledge joint-driven fault identification model is trained and optimized with the goal of minimizing the knowledge function value. Specifically, the use of domain experience knowledge from the cable latent fault arc model is as follows: Knowledge (1): If the duration of a fault or transient disturbance reaches five cycles or more, it is definitely not a latent fault in the cable. Knowledge (2): If the change in zero-sequence current before and after the fault is < a If so, it is definitely not a latent cable fault; among them, a Set value; The knowledge function is specifically as follows: ； Where N is the number of labeled samples, {( x i , t i ), i =1,2,…,N}, x i Represents sample feature data, t i For the true label of the sample, h ( θ , x i () represents the output of a purely data-driven model based on Extreme Learning Machine. θ Indicates model parameters; H [ h ( θ , x i ), t i [] represents the calculated cross-entropy loss function between the output of the pure data-driven model based on Extreme Learning Machine and the actual label; β i This indicates the degree of trust the fusion model has in domain experience and knowledge; S ( x i This represents the classification result derived from domain experience knowledge. This represents the conflict value between the output of the pure data-driven model based on extreme learning machine calculated according to the rules and empirical knowledge. If the two classification results are consistent, it is 1; otherwise, it is 0. The degree of trust that the fusion model places in domain experience knowledge is specifically as follows: ； in, ε That is, the random factor. c The value represents the degree of trust in the empirical knowledge.

2. The method for guided learning and identification of latent faults in medium-voltage distribution cables as described in claim 1, characterized in that, The specific abrupt change in zero-sequence current before and after the fault is as follows: ； in, , , These are the instantaneous values ​​of the three-phase currents A, B, and C, respectively, and I is the effective value of the current before the fault occurred. This represents the change in zero-sequence current before and after the fault.

3. The method for guided learning and identification of latent faults in medium-voltage distribution cables as described in claim 1, characterized in that, The loss function is specifically as follows: ； in, i The dimension of the output vector. p ( i ), q ( i ) are the th elements in the actual label vector and the predicted output vector, respectively. i The value of each element.

4. The method for guided learning and identification of latent faults in medium-voltage distribution cables as described in claim 1, characterized in that, The calculation rule for the conflict function is as follows: Rule 1: If M=1 and N=1, then the conflict function value is incremented by 1; Rule 2: If M=1 and Y=1, then the conflict function value is incremented by 1; Where M, N, and Y are all datasets. For the output vector of a pure data-driven model based on Extreme Learning Machine, if the column number containing the maximum value is 1 or 2, then M is set to 1; otherwise, M is set to 0. If the transient duration of a fault sample is greater than or equal to 5 cycles, then N is set to 1; otherwise, N is set to 0. If the calculated change in zero-sequence current before and after a fault sample is < a If the value is positive, then Y is set to 1; otherwise, it is set to 0. a This is the set value.

5. A latent fault-guided learning and identification system for medium-voltage power distribution cables, characterized in that, include: The data acquisition module is used to acquire the three-phase current and zero-sequence current data of cable faults; The feature extraction module is used to extract fault feature vectors of current data in the time domain, frequency domain, and time-frequency domain, respectively, and to perform dimensionality reduction optimization on the fault feature vectors. The fault identification module is used to obtain the identification results of cable latent faults based on the optimized fault feature vector and the trained data-knowledge joint driving fault identification model. The data-knowledge joint-driven fault identification model is obtained by fusing a pure data-driven model based on extreme learning machine and domain experience knowledge. Domain experience knowledge is constructed using a cable latent fault arc model. A knowledge function is constructed based on the loss function of the error between the pure data-driven model output and the actual labels of the training set samples, and the conflict function representing the contradiction between the pure data-driven model output and the domain experience knowledge. The data-knowledge joint-driven fault identification model is trained and optimized with the goal of minimizing the knowledge function value. The domain experience and knowledge are built using a cable latent fault arc model, specifically as follows: Knowledge (1): If the duration of a fault or transient disturbance reaches five cycles or more, it is definitely not a latent fault in the cable. Knowledge (2): If the change in zero-sequence current before and after the fault is < a If so, it is definitely not a latent cable fault; among them, a Set value; The knowledge function is specifically as follows: ； Where N is the number of labeled samples, {( x i , t i ), i =1,2,…,N}, x i Represents sample feature data, t i For the true label of the sample, h ( θ , x i () represents the output of a purely data-driven model based on Extreme Learning Machine. θ Indicates model parameters; H [ h ( θ , x i ), t i [] represents the calculated cross-entropy loss function between the output of the pure data-driven model based on Extreme Learning Machine and the actual label; β i This indicates the degree of trust the fusion model has in domain experience and knowledge; S ( x i This represents the classification result derived from domain experience knowledge. This represents the conflict value between the output of the pure data-driven model based on extreme learning machine calculated according to the rules and empirical knowledge. If the two classification results are consistent, it is 1; otherwise, it is 0. The degree of trust that the fusion model places in domain experience knowledge is specifically as follows: in, ε That is, the random factor. c The value represents the degree of trust in the empirical knowledge.

6. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-4, a method for guided learning and identification of latent faults in medium-voltage power distribution cables.

7. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded and executed by the processor of the terminal device, and to provide a method for guided learning and identification of latent faults in medium-voltage power distribution cables as described in any one of claims 1-4.

Images