Medium-voltage cable defect identification method and system based on multi-source characteristic parameters

Abstract

The application provides a medium-voltage cable defect identification method and system based on multi-source characteristic parameters, and belongs to the technical field of cable defect identification. The method is applied to an identification system including a multi-source data processing module, a model construction module, a defect identification module and a strategy construction module, and specifically comprises the following steps: characteristic extraction is performed on multi-source data of a cable under different operating conditions, and a defect sample set is constructed in combination with corresponding operating condition types and defect data; model parameters of an online sequence extreme learning machine are initialized based on parameter initialization constraints of fault priors, and a cable defect identification model is obtained by training according to the defect sample set; multi-source data of current cable operation are collected, a defect identification result is obtained according to the cable defect identification model, the cable operating state is evaluated through fuzzy evaluation in combination with the current operating condition, and a cable operation and maintenance strategy is established. The application sets the parameter initialization constraints of fault priors, guarantees the performance stability of the cable defect identification model, and guarantees the identification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cable defect identification, in particular to a medium-voltage cable defect identification method and system based on multi-source feature parameters. BACKGROUND

[0002] As a core component of urban distribution network, the operation state of medium-voltage cable is directly related to the safety and stability of the power system. During long-term operation, various defects may occur due to insulation aging, mechanical damage, environmental erosion, etc. If not identified in time, it may cause short circuit, power failure, and even serious accidents such as fire. The online sequence extreme learning machine can accurately adapt to the scene requirements of medium-voltage cable defect identification. It can realize incremental learning on real-time data such as cable continuous partial discharge and temperature without full retraining and with rapid response. It also has strong generalization ability to distinguish multiple types of defects such as insulation aging and mechanical damage and coupled scenes, and can dynamically adjust the model to maintain stable performance as the operating environment changes and defects evolve. It has been widely used in the identification of medium-voltage cable defects.

[0003] In practical application, in order to simplify the training process, the online sequence extreme learning machine randomly initializes the model parameters. This random initialization may cause the hidden layer of the model to lack determinacy in feature mapping of input data. Under the same training data and training process, different initial parameter combinations may cause the model to exhibit significant performance fluctuations, making it difficult to form stable feature extraction logic. This may further lead to inconsistent recognition results for the same batch of test samples after multiple training, resulting in poor model stability. Especially in the case of high multi-source feature dimension, the influence of parameter randomness on model stability is more significant, and the model's ability to capture key defect features is more susceptible to initial parameter interference, making it difficult to ensure the accuracy of defect identification results. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art, in which the initial parameter randomness affects the stability of the recognition model when using an online sequence extreme learning machine for cable defect identification, and the accuracy of the defect identification result is difficult to guarantee. A medium-voltage cable defect identification method and system based on multi-source feature parameters are provided. By setting a parameter initialization constraint based on fault priors, the initial parameters of the model can be close to the distribution pattern of the multi-source features of the medium-voltage cable, ensuring the stability of the cable defect identification model and the accuracy of the defect identification result. In combination with fuzzy evaluation, the defect identification result and the current working condition are fused to further correct the recognition deviation, reduce the risk of misjudgment and omission, and further improve the accuracy of cable defect identification.

[0005] The present application is achieved by the following technical solutions:

[0006] The medium-voltage cable defect identification method based on multi-source feature parameters comprises:

[0007] Collect multi-source data and corresponding defect data of cables under different operating conditions, extract features from the multi-source data, obtain multi-source feature parameters, and construct a defect sample set by combining the corresponding operating condition type and defect data.

[0008] Based on the parameter initialization constraints of fault prior, the model parameters of the online sequence extreme learning machine are initialized, and the online sequence extreme learning machine is trained according to the defect sample set to obtain the cable defect identification model.

[0009] Collect multi-source data on the current cable operation, obtain the cable defect identification results based on the cable defect identification model, and evaluate the cable operation status through fuzzy evaluation based on the defect identification results and the current cable operation conditions.

[0010] Establish cable operation and maintenance strategies based on cable operating status.

[0011] By incorporating prior knowledge of cable faults, the online sequence learning machine aligns with the distribution patterns of multi-source feature parameters during the initialization phase, avoiding performance fluctuations caused by random parameters and ensuring the operational stability of the cable defect identification model. Furthermore, by fusing defect identification results with the current cable operating conditions, the model's output defect identification results are corrected, avoiding identification biases caused by environmental interference under complex operating conditions, reducing the risk of misjudgments and omissions, and improving the accuracy of cable defect identification.

[0012] Furthermore, the parameter initialization constraints based on fault priors initialize the model parameters of the online sequence extreme learning machine, and train the online sequence extreme learning machine according to the defect sample set to obtain a cable defect identification model, including:

[0013] Screen the multi-source characteristic parameters of cables before defects occur in the defect sample set to obtain fault precursor characteristics;

[0014] Statistical analysis was performed on the corresponding fault precursor characteristics according to the defect type to obtain the distribution pattern and sensitive interval of each defect type;

[0015] The parameter constraints for the input layer weights and hidden layer biases of the online sequence extreme learning machine are set based on the distribution pattern and sensitive intervals.

[0016] The online sequence extreme learning machine (OSM) model parameters are initialized according to parameter constraints, and the corresponding parameter adjustment range boundaries are set. The OSM is then trained using a defect sample set, and the model parameters are iteratively optimized to obtain a cable defect identification model.

[0017] Furthermore, the step of training an online sequence extreme learning machine based on a defect sample set, iteratively optimizing the model parameters, and obtaining a cable defect identification model includes:

[0018] The defect sample set is divided into a training subset and a validation subset. The online sequence extreme learning machine initialized with multi-source feature parameters of the training subset is used to calculate the hidden layer output through forward propagation and solve the output layer weights by the least squares method to obtain the basic cable defect identification model.

[0019] The model performance of the basic cable defect identification model is evaluated based on the validation subset. If the model performance evaluation results do not meet the preset conditions, the output layer weights are iteratively adjusted based on the prediction error of the validation subset until the model performance meets the preset conditions, and the cable defect identification model is obtained.

[0020] Furthermore, the step of evaluating the cable's operating status through fuzzy evaluation based on the defect identification results and the current operating conditions of the cable includes:

[0021] Set up defect indicators and operating condition indicators, and configure the corresponding indicator weights according to the priority of the impact of operating conditions.

[0022] Based on the defect identification results and the current operating conditions of the cable, the values ​​of each indicator are determined, and the indicators are normalized through the deterioration degree model to construct the fuzzy membership matrix of each state corresponding to each indicator.

[0023] The fuzzy measure values ​​of each indicator and indicator combination are calculated based on the preset indicator synergy coefficients, and the comprehensive membership degree of each state is calculated by combining the fuzzy membership degree matrix of each state corresponding to each indicator.

[0024] The overall membership degree of each state is weighted and summed to obtain the cable operating status score, and the cable operating status is identified based on the cable operating status score.

[0025] Furthermore, the establishment of a cable operation and maintenance strategy based on the cable's operating status includes:

[0026] Identify the current cable maintenance objectives based on the cable's operating status, and match corresponding maintenance measures and frequencies accordingly.

[0027] Obtain historical operation and maintenance data of the current cable, and determine the execution parameters of operation and maintenance measures based on the cable's operating status;

[0028] The maintenance frequency is adjusted based on the changing trend of the cable operating status score, and a cable maintenance strategy is established by combining the execution parameters of maintenance measures.

[0029] Furthermore, after assessing the cable's operating status through fuzzy evaluation, the following steps are also performed:

[0030] By comparing the cable operation status assessment results and the defect identification results, a status deviation feedback signal is generated based on the comparison results, and trigger adjustment judgment is made based on the status deviation feedback signal.

[0031] When it is judged that the state deviation frequency ratio exceeds the preset threshold, the model optimization is triggered, and the cable defect identification model is optimized by incremental updating.

[0032] Further, the cable defect identification model optimized by incremental updating comprises:

[0033] The newly added multi-source feature samples are continuously received according to the preset data block size, and the newly added multi-source feature samples are input into the cable defect identification model;

[0034] The corresponding newly added hidden layer output is obtained by forward propagation, the output layer weight of the cable defect identification model is adjusted in combination with the defect label corresponding to the newly added multi-source feature sample, and the cable defect identification model is optimized.

[0035] Further, the feature extraction of the multi-source data and the acquisition of the multi-source feature parameters comprise:

[0036] The data types of the multi-source data are identified, and corresponding feature dimensions are selected according to the relevance of each data type to the cable fault;

[0037] According to the corresponding feature dimension matching extraction algorithm, the multi-source data of each data type is extracted to obtain the multi-source feature parameters.

[0038] The medium-voltage cable defect identification system based on the multi-source feature parameters is used to execute any one of the identification methods, and comprises:

[0039] The multi-source data processing module is used to collect the multi-source data and corresponding defect data of the cable under different operating conditions, extract features from the multi-source data, acquire multi-source feature parameters, and construct a defect sample set in combination with the corresponding operating condition type and defect data;

[0040] The model construction module is connected with the multi-source data processing module, and is used to initialize the model parameters of the online sequence extreme learning machine based on the parameter initialization constraint of the fault priori, and train the online sequence extreme learning machine according to the defect sample set to acquire the cable defect identification model;

[0041] The defect identification module is connected with the model construction module, collects the multi-source data of the current cable operation, acquires the defect identification result of the cable according to the cable defect identification model, and evaluates the cable operation state through fuzzy evaluation according to the defect identification result and the operating condition of the current cable;

[0042] The strategy construction module is connected with the defect identification module, and is used to establish a cable operation and maintenance strategy according to the cable operation state.

[0043] Further, the identification system further comprises:

[0044] The model optimization module is connected with the model construction module and the defect identification module respectively, and is used for comparing the cable operation state evaluation result and the defect identification result to generate a state deviation feedback signal, and when the model optimization is triggered by the feedback signal, the cable defect identification model constructed is optimized through incremental updating.

[0045] The beneficial effects of the present application are:

[0046] (1) By integrating prior knowledge of cable faults, the online sequence learning machine is fitted to the distribution law of multi-source feature parameters in the initialization stage, avoiding model performance fluctuations caused by random parameters and ensuring the operation stability of the cable defect identification model. On this basis, the defect identification result is fused with the current cable operation condition to correct the defect identification result of the model output, avoid the recognition deviation caused by environmental interference under complex working conditions, reduce the risk of misjudgment and omission, and improve the accuracy of cable defect identification.

[0047] (2) By screening fault precursor features, statistics of the distribution law and sensitive interval of various defects, parameter constraints are established for the input layer weight and hidden layer bias of the online sequence extreme learning machine, and the corresponding parameter adjustment amplitude boundary is clear, avoiding performance fluctuations caused by random initial parameters, so that the initial state of the model can be fitted to the actual distribution of cable defect features. And in the training process, by dividing the training subset and the verification subset, the output layer weight is solved by the least square method to construct the basic model, and then the weight is iteratively optimized based on the prediction error of the verification subset, which not only ensures sufficient learning of the sample features by the model, but also effectively improves the generalization ability of the model, avoids overfitting or underfitting problems, and further improves the accuracy of defect identification.

[0048] (3) By setting state deviation feedback and incremental updating, the cable defect identification model can adapt to new working conditions and new defect features in the cable operation process in real time, and maintain high recognition accuracy and stability. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flowchart of the present application;

[0050] Figure 2 is a structural schematic diagram of an embodiment of the present application.

[0051] Among them: 1, multi-source data processing module, 2, model construction module, 3, defect identification module, 4, strategy construction module, 5, model optimization module. DETAILED DESCRIPTION

[0052] The present application will be further described below in conjunction with the drawings and embodiments.

[0053] Embodiment:

[0054] A medium-voltage cable defect identification method based on multi-source characteristic parameters, as shown in Figure 1

[0055] Collecting multi-source data and corresponding defect data of the cable under different operating conditions, performing feature extraction on the multi-source data, obtaining multi-source characteristic parameters, and combining the corresponding operating condition types and defect data to construct a defect sample set;

[0056] Initializing the model parameters of the online sequence extreme learning machine based on parameter initialization constraints of the fault prior, and training the online sequence extreme learning machine according to the defect sample set to obtain a cable defect identification model;

[0057] Collecting multi-source data of the current cable operation, obtaining a defect identification result of the cable according to the cable defect identification model, and evaluating the cable operation state through fuzzy evaluation according to the defect identification result and the operating condition of the current cable;

[0058] Establishing a cable operation and maintenance strategy according to the cable operation state.

[0059] Considering that when using an online sequence extreme learning machine for defect identification, if the initial parameters of the model lack effective constraints, the final model obtained may lack stability due to the actual feature distribution of the cable. Therefore, before model training, multi-source data of the cable under different operating conditions and corresponding defect data are collected, and feature extraction is performed on the operating conditions and defect labels to present the distribution law and sensitive interval of different defects, providing basic data for setting the parameter initialization constraints of the fault prior, so that the initialization parameters of the subsequent model can be close to the real operating features of the cable, ensuring the stability and identification accuracy of the model.

[0060] The multi-source data includes temperature, partial discharge, acoustic signal, insulation resistance, etc., and the defect data includes corresponding defect types, defect severity, defect location, discovery time, etc. The multi-source data and defect data can be matched and bound according to the specific collection source.

[0061] The collected original multi-source data mostly contains a large amount of redundant information and environmental noise, and the dimensions and formats of different types of data differ greatly. Therefore, the obtained multi-source data needs to be feature extracted to filter irrelevant noise and redundant information, and to realize the standardization and integration of multi-source data.

[0062] The feature extraction on the multi-source data to obtain multi-source characteristic parameters includes:

[0063] Identifying the data types of the multi-source data, and selecting corresponding feature dimensions according to the relevance of each data type to cable faults;

[0064] ​According to the corresponding feature dimension matching extraction algorithm, the multi-source data of each data type is subjected to feature extraction to obtain multi-source feature parameters.

[0065] Firstly, the multi-source data is classified based on the collection source and signal type, and is divided into electrical parameter data, environmental working condition data, physical state data, acoustic and optical data, etc., so as to avoid interference between data features of different properties.

[0066] Secondly, the correlation between various data types and cable faults is obtained through a correlation algorithm such as grey correlation degree, so as to filter out the feature dimensions that have the identification of the corresponding cable faults. Specifically, the feature dimensions that are strongly correlated with cable defects, such as the waveform features of partial discharge data and the spectrum features of acoustic signals, can be selected as the feature dimensions for feature extraction, so as to eliminate redundant data to reduce the complexity of the subsequent model, and to retain key features to enhance the correlation between data and faults.

[0067] The attributes of different feature dimensions also differ. If the algorithm does not match the feature dimension, the feature will be distorted, which will affect the recognition accuracy of the subsequently constructed model for cable defects. Therefore, after determining the feature dimension, a matching algorithm should be used to accurately extract the feature. For example, the spectrum feature of acoustic signals belongs to the frequency domain feature, so the time domain signal can be converted into a frequency spectrum graph through short-time Fourier transform for extraction.

[0068] After extracting the multi-source feature parameters, the multi-source feature parameters, working condition information and defect data are often independent of each other, and the subsequent online sequence extreme learning machine cannot capture the defect rules in different scenarios. Therefore, based on the determination of the multi-source feature parameters, the matching working condition information and defect data are obtained in combination with the corresponding data sources, and then the working condition labels and defect labels of the multi-source feature parameters are established, taking one of the multi-source feature parameters and the corresponding working condition labels and defect labels as a sample to establish a defect sample set.

[0069] The established defect sample set has integrated the complete feature evolution process of different defects from the beginning to the appearance, as well as the performance rules of defect features under different working conditions, etc. Through the classification and statistics of the defect sample set, the feature distribution rules and sensitive intervals of each defect can be directly output, so as to quantitatively obtain the corresponding parameter constraint rules, optimize the rationality of the initialization parameters of the online sequence extreme learning machine, and then ensure the stability of the subsequently obtained cable defect recognition model.

[0070] The parameter initialization constraint based on the fault prior is used to initialize the model parameters of the online sequence extreme learning machine, and the online sequence extreme learning machine is trained according to the defect sample set to obtain a cable defect recognition model, which includes:

[0071] Screening multi-source characteristic parameters before cable defect occurs in the defect sample set to obtain fault precursor characteristics;

[0072] Statistically analyzing the corresponding fault precursor characteristics according to the defect types to obtain the distribution law and sensitive interval of each defect type;

[0073] Setting the parameter constraints of the input layer weight and the hidden layer bias of the online sequence extreme learning machine based on the distribution law and the sensitive interval;

[0074] Initializing the model parameters of the online sequence extreme learning machine according to the parameter constraints, setting the corresponding parameter adjustment amplitude boundary, training the online sequence extreme learning machine according to the defect sample set, iteratively optimizing the model parameters, and obtaining the cable defect recognition model.

[0075] Screening the multi-source characteristic parameters within a preset time window before the defect diagnosis time in the defect sample set as the fault precursor characteristics. In the screening process, the jump values caused by sensor failure, data transmission error, etc. need to be removed, and the data is smoothed by the moving average method to ensure that the obtained fault precursor characteristics can truly reflect the defect evolution trend and avoid noise interference.

[0076] Classifying the screened fault precursor characteristics according to the defect types, and quantitatively analyzing the corresponding fault precursor characteristics of each defect to obtain the corresponding distribution law and sensitive interval. The distribution law includes the distribution interval and change law of the fault precursor characteristics, and the sensitive interval is the range of the fault precursor characteristic values with a significantly increased defect occurrence probability.

[0077] According to the feature importance and defect relevance, the constraint range of the input layer weight and the hidden layer bias of the online sequence extreme learning machine is matched. For the precursor characteristics that are strongly associated with defects and in the sensitive interval, the corresponding input layer weight is constrained in the high response interval to ensure that the model pays more attention to the core characteristics. For the characteristics that are weakly associated with defects or in the normal distribution interval, the corresponding weight is constrained in the low response interval to avoid redundant feature interference. The hidden layer bias is set according to the feature sensitive interval, so that when the feature value enters the sensitive interval, the hidden layer output signal is enhanced, and the sensitivity of the model to the defect risk is improved.

[0078] Within the parameter initialization constraint interval set based on the fault prior, the model parameters of the online sequence extreme learning machine are initialized by uniform random sampling to ensure that the initial parameters are always within a reasonable range. At the same time, the parameter adjustment amplitude boundary is set to reserve optimization space for model adaptation to specific sample data, and to prevent the parameters from deviating from the prior law during training, affecting the stability and recognition accuracy of the model.

[0079] After the initialization parameters are determined, the online sequence extreme learning machine is trained according to the defect sample set, the model parameters are iteratively optimized, and the cable defect recognition model is obtained. Specifically:

[0080] The defect sample set is divided into a training subset and a validation subset. The multi-source feature parameters of the training subset are input into the initialized online sequence extreme learning machine, the hidden layer output is calculated through forward propagation, and the output layer weight is solved through the least square method to obtain the basic cable defect recognition model.

[0081] The model performance of the basic cable defect recognition model is evaluated according to the validation subset. If the model performance evaluation result does not reach the preset condition, the output layer weight is iteratively adjusted based on the prediction error of the validation subset until the model performance reaches the preset condition, and the cable defect recognition model is obtained.

[0082] The network architecture of the online sequence extreme learning machine is composed of an input layer, a single hidden layer and an output layer, supports batch sample progressive training, and establishes a parameter dynamic updating link between new and old data. When new samples arrive, only incremental correction of parameters based on new data is needed, avoiding repeated calculation of historical samples, which can effectively improve the model training efficiency.

[0083] Therefore, the training process of the online sequence extreme learning machine is divided into two stages. In the initial stage, a certain amount of initial samples are selected as input, and the initial output layer weight is obtained through output, thereby obtaining the initial basic cable defect recognition model. In the second stage, the output layer weight is updated through online sequence learning.

[0084] Specifically, the defect sample set is divided into a training subset and a validation subset according to the time sequence, and during the division, the sample features of all types of defects and working conditions in the training subset and the validation subset are completely covered.

[0085] Then in the initial stage, the multi-source feature parameters of the training subset are input into the online sequence extreme learning machine initialized through the fault prior constraint. The online sequence extreme learning machine first calculates the hidden layer output through forward propagation, and then solves the output layer weight by the least square method to quickly establish the mapping relationship between the hidden layer output and the defect label, and obtain the basic cable defect recognition model.

[0086] The expression of the hidden layer output is:

[0087] ;

[0088] wherein, is the output weight matrix of the output layer, is an activation function, and in this embodiment, a Sigmoid function is specifically used, is the first an input layer weight corresponding to an input sample, is the number of samples in the training subset, is the bias of the hidden layer corresponding to the i-th input sample, , , is the number of samples in the training subset, is the number of hidden layer neurons.

[0089] Then, in the second stage, the performance of the basic cable defect recognition model is evaluated by the validation subset. The multi-source feature parameters of the validation subset are input into the basic cable defect recognition model to obtain the predicted defect type and severity, and the results are compared with the true defect labels of the validation subset to obtain performance index values including recognition accuracy, recall rate, precision rate, and confusion matrix.

[0090] Then, according to the performance index values, it is verified whether the model has overfitting or recognition deviation. If the validation evaluation result meets the preset condition, such as no overfitting and the recognition deviation is within a certain deviation threshold, the final cable defect recognition model can be output.

[0091] If the preset condition is not met, a part of the samples in the validation subset can be selected as new samples according to the time sequence, the output layer weight is updated, and the updated model is verified by the remaining samples in the validation subset until the performance index of the basic cable defect recognition model meets the preset condition, so as to output the final cable defect recognition model.

[0092] When the basic cable defect recognition model is optimized by the validation subset, it is assumed that wherein, is the total number of old data samples that have been learned, is the optimization frequency, is the total optimization frequency, and the number of newly added data samples is , and the newly added new sample data .

[0093] On this basis, the output weight matrix of the output layer under the current optimization frequency is is:

[0094] .

[0095] By solving the output weight matrix, the output layer weight of the optimized basic cable defect recognition model is updated.

[0096] After obtaining the final cable defect identification model, considering that the defect identification results can only output a single conclusion—whether a defect exists, its type, and its severity—without taking into account the amplification or mitigation effect of operating conditions on defect risk, judging the operating status solely based on the defect identification results would ignore the dynamic influence of operating conditions, leading to a disconnect between the defect identification results and the actual situation. Furthermore, although the model has undergone prior constraints and iterative optimization, slight misjudgments or omissions may still occur in actual operation. To avoid such situations, the cable's operating status is further evaluated in conjunction with the current cable operating conditions, ensuring the comprehensiveness of the identification results while correcting the defect identification results.

[0097] Specifically, the step of evaluating the cable's operating status through fuzzy evaluation based on the defect identification results and the current operating conditions of the cable includes:

[0098] Set up defect indicators and operating condition indicators, and configure the corresponding indicator weights according to the priority of the impact of operating conditions.

[0099] Based on the defect identification results and the current operating conditions of the cable, the values ​​of each indicator are determined, and the indicators are normalized through the deterioration degree model to construct the fuzzy membership matrix of each state corresponding to each indicator.

[0100] The fuzzy measure values ​​of each indicator and indicator combination are calculated based on the preset indicator synergy coefficients, and the comprehensive membership degree of each state is calculated by combining the fuzzy membership degree matrix of each state corresponding to each indicator.

[0101] The overall membership degree of each state is weighted and summed to obtain the cable operating status score, and the cable operating status is identified based on the cable operating status score.

[0102] In this embodiment, the severity level of defects, the rate of defect development, and the extent to which characteristic parameters deviate from the normal range are set as defect indicators, and the duration of load rate exceeding the standard, the frequency of environmental temperature and humidity exceeding the limit, and the voltage fluctuation coefficient are set as operating condition indicators.

[0103] Considering the significant differences in the dimensions and value ranges of different indicators, we first normalize them using a degradation degree model, mapping the values ​​of each indicator to the range of [0, 1]. Here, 0 represents no degradation, i.e. the best state, while 1 represents severe degradation, i.e. the worst state.

[0104] Furthermore, for each indicator, typical operating states of the cable are defined, including normal, warning, abnormal, and verification. The degree to which the indicator belongs to each state is then calculated using membership functions such as trapezoidal functions.

[0105] The membership degree of a single index cannot reflect the actual influence of multi-factor coupling, therefore, the importance of the index and the index combination is quantified by the fuzzy measurement value. The fuzzy measurement value is determined by the preset index synergy coefficient, and the index synergy coefficient can be set according to the correlation strength between the indexes.

[0106] In the calculation of the fuzzy measurement, a basic measurement can be assigned to each index according to actual needs, and then a combined measurement is superimposed according to the index synergy coefficient, so as to obtain the fuzzy measurement values of all indexes and index combinations.

[0107] For each state, the membership degrees of all indexes belonging to the state are weighted and summed according to the fuzzy measurement values of the corresponding indexes or index combinations, to obtain the comprehensive membership degrees of each state.

[0108] The comprehensive membership degrees of each state only reflect the possibility of each state, therefore, the preset scores corresponding to each state are multiplied by the comprehensive membership degrees of each state as weights, and then summed, to determine the cable operation state score, and the cable operation state score is mapped to a specific operation state level according to a preset threshold, to determine the current cable operation state.

[0109] In the long-term operation of the medium-voltage cable, the working conditions and defect modes can change, and the cable defect model obtained by the initial training can have recognition deviation due to data distribution drift, and only relying on the fixed model cannot adapt to the dynamic scene, therefore, after the fuzzy evaluation is used to evaluate the cable operation state, the following is further performed:

[0110] The cable operation state evaluation result is compared with the defect recognition result, a feedback signal of state deviation is generated according to the comparison result, and the feedback signal of state deviation is used for triggering adjustment and judgment;

[0111] When it is judged that the frequency ratio of state deviation exceeds a preset threshold, the model optimization is triggered, and the cable defect recognition model is optimized by incremental updating.

[0112] The cable operation state evaluation result obtained by the fuzzy evaluation is compared with the original defect recognition result output by the defect recognition model one by one, the deviation types of the two are analyzed, and a standardized state deviation feedback signal is generated according to the degree and type of the deviation, and the time, corresponding working condition and characteristic parameter of the deviation are recorded synchronously.

[0113] The frequency ratio of the state deviation sample in a preset statistical period is counted, it is judged whether the ratio exceeds a preset threshold, if the ratio does not exceed the threshold, it is indicated that the deviation is caused by accidental factors such as short-term extreme working condition interference, and the model optimization is not triggered, if the ratio exceeds the threshold, it is judged that the current cable defect prediction model has appeared adaptive attenuation, and the model optimization is needed.

[0114] The cable defect recognition model in the embodiment is based on an online sequence extreme learning machine, which is an incremental model and can be quickly adapted through local updating. Therefore, after triggering model optimization, the cable defect recognition model is further optimized through incremental updating.

[0115] Specifically, the cable defect recognition model is optimized through incremental updating, including:

[0116] The newly added multi-source feature samples are continuously received according to a preset data block size, and the newly added multi-source feature samples are input into the cable defect recognition model.

[0117] The corresponding newly added hidden layer output is obtained through forward propagation, and the output layer weight of the cable defect recognition model is adjusted in combination with the defect label corresponding to the newly added multi-source feature samples, so as to optimize the cable defect recognition model.

[0118] In order to avoid model fluctuation caused by single updating or irregular sample receiving, the newly added multi-source feature samples are continuously received according to a preset data block size. The newly added multi-source feature samples received are consistent with the feature format in the defect sample set. Then, the newly added multi-source feature samples are input into the current cable defect recognition model as the original input data for incremental learning. It should be noted that the newly added multi-source feature samples are continuously collected, and are not collected after triggering model optimization, so as to avoid update delay caused by temporary collection after triggering optimization.

[0119] After the newly added multi-source feature samples are input into the model, the corresponding hidden layer output is calculated through forward propagation. Specifically, the newly added multi-source feature samples are first mapped to the hidden layer by using the current input layer weight, and then the hidden layer output matrix is calculated through an activation function in combination with the corresponding hidden layer bias.

[0120] The role of the output layer weight is to establish a mapping relationship between the hidden layer output and the defect label. After the newly added hidden layer output is calculated, the deviation between the hidden layer output of the newly added multi-source feature samples and the corresponding defect label is obtained in combination with the defect label corresponding to the newly added multi-source feature samples in the data block. Then, the new output layer weight is solved by using a least square method or other optimization algorithm, so as to minimize the error between the newly added hidden layer output and the label, and complete the incremental updating of the cable defect recognition model.

[0121] In order to further ensure the safe operation of the cable, after obtaining the cable operation state, an adaptive cable operation and maintenance strategy is further established in combination with historical operation and maintenance data, so as to provide suggestions for cable operation and maintenance.

[0122] The cable operation and maintenance strategy is established according to the cable operation state, including:

[0123] The operation and maintenance target of the current cable is identified according to the cable operation state, and the corresponding operation and maintenance measures and operation and maintenance frequency are matched according to the operation and maintenance target.

[0124] Obtaining historical operation and maintenance data of the current cable, and determining the execution parameters of the operation and maintenance measures in combination with the cable operation state;

[0125] According to the change trend of the cable operation state score, the operation and maintenance frequency is corrected, and the cable operation and maintenance strategy is established in combination with the execution parameters of the operation and maintenance measures.

[0126] According to the cable operation state, the operation and maintenance target is determined, the higher the state risk is, the more the target is inclined to eliminate the emergency hidden danger, and the lower the risk is, the more the target is focused on prevention and stable maintenance.

[0127] According to the target, the corresponding operation and maintenance measures are matched, the risk is high, the processing measures corresponding to the defect can be selected in combination with the previous defect identification result, the risk is low, and the routine maintenance is mainly used. At the same time, the basic operation and maintenance frequency is set according to the risk urgency, the higher the risk is, the more intensive the frequency is.

[0128] The matched operation and maintenance measures are template contents, and the specific execution parameters can be set by referring to the historical operation and maintenance data and the current cable operation state.

[0129] And considering that the cable operation state is dynamically changing, it is difficult to adapt to the defect change situation with fixed frequency, therefore, the change trend of the cable operation state score is further judged to determine the defect evolution of the cable, so as to correct the operation and maintenance frequency, if the score continues to decline, the state appears to be deteriorating, the operation and maintenance frequency needs to be shortened, if the score is stable or rising, it proves that the operation state of the cable may be affected by the operation and maintenance operation, and the operation state is improved to a certain extent, and the frequency can be appropriately prolonged. On the basis of the corrected operation and maintenance frequency, the final cable operation and maintenance strategy is constructed in combination with the set execution parameters.

[0130] Another aspect of the embodiment also provides a medium-voltage cable defect identification system based on multi-source feature parameters, as shown in Figure 2 , comprising:

[0131] A multi-source data processing module 1 is used for collecting multi-source data and corresponding defect data of the cable under different operation conditions, performing feature extraction on the multi-source data, obtaining multi-source feature parameters, constructing a defect sample set in combination with the corresponding working condition type and defect data;

[0132] A model construction module 2 is connected with the multi-source data processing module, used for initializing the model parameters of the online sequence extreme learning machine based on the parameter initialization constraint of the fault prior, and training the online sequence extreme learning machine according to the defect sample set to obtain a cable defect identification model;

[0133] The defect identification module 3 is connected with the model construction module, collects multi-source data of the current cable operation, obtains the defect identification result of the cable according to the cable defect identification model, and evaluates the cable operation state through fuzzy evaluation according to the defect identification result and the operation condition of the current cable.

[0134] The strategy construction module 4 is connected with the defect identification module, and is used for establishing the cable operation and maintenance strategy according to the cable operation state.

[0135] The identification system further comprises:

[0136] The model optimization module 5 is connected with the model construction module and the defect identification module respectively, is used for comparing the cable operation state evaluation result and the defect identification result to generate a feedback signal of state deviation, and when the feedback signal triggers model optimization, the cable defect identification model constructed is optimized through incremental update.

[0137] The multi-source data processing module, the model construction module, the defect identification module, the strategy construction module and the model optimization module are all data processing devices with data processing and analysis capabilities, such as microprocessors and calculators.

[0138] The multi-source data processing module is internally provided with related algorithm programs corresponding to feature extraction and defect sample construction, and has an interface for accessing an external data platform to obtain required multi-source data and defect data.

[0139] The model construction module is provided with related algorithm programs of parameter initialization constraint setting of corresponding fault priori and online sequence extreme learning machine training, and can realize preliminary establishment of the cable defect identification model.

[0140] The defect identification module is internally provided with related algorithm programs for fuzzy evaluation of the cable operation state, and also has an interface for accessing an external data platform to obtain required data.

[0141] The strategy construction module is internally provided with a cable operation and maintenance strategy database storing operation and maintenance measures under different operation states, and also has an interface for accessing an external data platform to obtain required historical operation and maintenance data.

[0142] The model optimization module is provided with related algorithm programs of feedback signal generation of state deviation and incremental update, and also has an interface for accessing an external data platform to receive multi-source data. In addition, the model optimization module can be further connected with the multi-source data processing module to obtain new multi-source feature samples by using the feature extraction capability of the multi-source data processing module, or can be internally provided with corresponding feature extraction algorithms.

[0143] The above-described embodiments are only the preferred ones of the present application, and do not limit the present application in any form, and other variants and modifications can be made without departing from the technical solutions recited in the claims.

Claims

1. A method for identifying defects of a medium-voltage cable based on multi-source characteristic parameters, characterized in that, The method comprises the following steps: Collecting multi-source data and corresponding defect data of the cable under different operating conditions, extracting features from the multi-source data, obtaining multi-source feature parameters, and constructing a defect sample set in combination with the corresponding operating condition types and defect data; Initializing the model parameters of the online sequence extreme learning machine based on the parameter initialization constraint of the fault prior, and training the online sequence extreme learning machine according to the defect sample set to obtain a cable defect recognition model, wherein Screening the multi-source feature parameters before the cable defect occurs in the defect sample set to obtain fault precursor features; According to the defect type, the corresponding fault precursor features are statistically analyzed to obtain the distribution rule and sensitive interval of each defect type; Based on the distribution rule and the sensitive interval, the parameter constraints of the input layer weight and the hidden layer bias of the online sequence extreme learning machine are set; Initializing the model parameters of the online sequence extreme learning machine according to the parameter constraints, and setting the corresponding parameter adjustment amplitude boundary, training the online sequence extreme learning machine according to the defect sample set, and iteratively optimizing the model parameters to obtain the cable defect recognition model; Collecting the multi-source data of the current cable operation, obtaining the cable defect recognition result according to the cable defect recognition model, and evaluating the cable operation state through fuzzy evaluation according to the defect recognition result and the operating condition of the current cable; Establishing a cable operation and maintenance strategy according to the cable operation state; The method comprises the following steps: Dividing the defect sample set into a training subset and a validation subset, inputting the multi-source feature parameters of the training subset into the initialized online sequence extreme learning machine, calculating the hidden layer output through forward propagation, and solving the output layer weight through the least square method to obtain a basic cable defect recognition model; According to the validation subset, the model performance of the basic cable defect recognition model is evaluated, if the model performance evaluation result does not reach the preset condition, the output layer weight is iteratively adjusted based on the prediction error of the validation subset, until the model performance reaches the preset condition, and the cable defect recognition model is obtained.

2. The method for multi-source feature parameter based medium voltage cable defect identification according to claim 1, characterized in that, The method comprises the following steps: Setting a defect index and a working condition index, and configuring the corresponding index weight according to the working condition influence priority; According to the defect recognition result and the operating condition of the current cable, the index values are determined, and the indexes are normalized through a degradation degree model to construct a fuzzy membership matrix of each state corresponding to each index; According to the preset index coordination coefficient, the fuzzy measure value of each index and index combination is calculated, and the comprehensive membership degree of each state is calculated in combination with the fuzzy membership matrix of each state corresponding to each index; The comprehensive membership degrees of each state are weighted and summed to obtain the cable operation state score, and the cable operation state is identified according to the cable operation state score.

3. The method for multi-source feature parameter based medium voltage cable defect identification according to claim 2, characterized in that, The method comprises the following steps: According to the cable operation state, the operation and maintenance target of the current cable is identified, and the corresponding operation and maintenance measures and operation and maintenance frequency are matched according to the operation and maintenance target; Obtaining the historical operation and maintenance data of the current cable, and determining the execution parameters of the operation and maintenance measures in combination with the cable operation state; The operation and maintenance frequency is corrected according to the change trend of the cable operation state score, and a cable operation and maintenance strategy is established in combination with the execution parameters of the operation and maintenance measures.

4. The method for multi-source feature parameter based medium voltage cable defect identification according to claim 1, characterized in that, After the cable operation state is evaluated through fuzzy evaluation, the following is also performed: The cable operation state evaluation result and the defect identification result are compared, a feedback signal of state deviation is generated according to the comparison result, and a trigger adjustment judgment is performed according to the feedback signal of state deviation; When it is judged that the frequency proportion of state deviation exceeds a preset threshold, model optimization is triggered, and the cable defect identification model is optimized through incremental updating.

5. The method for multi-source feature parameter based medium voltage cable defect identification according to claim 4, characterized in that, The cable defect identification model is optimized through incremental updating, including: Continuously receiving newly added multi-source feature samples according to a preset data block size, and inputting the newly added multi-source feature samples into the cable defect identification model; The corresponding newly added hidden layer output is obtained through forward propagation, the output layer weight of the cable defect identification model is adjusted in combination with the defect label corresponding to the newly added multi-source feature sample, and the cable defect identification model is optimized.

6. The method for multi-source feature parameter based medium voltage cable defect identification according to claim 1, characterized in that, The multi-source data is feature extracted to obtain multi-source feature parameters, including: Identifying the data types of the multi-source data, and selecting corresponding feature dimensions according to the relevance of each data type to cable faults; According to the corresponding feature dimension matching extraction algorithm, the multi-source data of each data type is feature extracted to obtain multi-source feature parameters.

7. A system for medium voltage cable defect recognition based on multi-source feature parameters, for performing the recognition method according to any one of claims 1 to 6, characterized in that, Including: A multi-source data processing module is configured to collect multi-source data and corresponding defect data of the cable under different operating conditions, feature extract the multi-source data to obtain multi-source feature parameters, and construct a defect sample set in combination with the corresponding operating condition type and defect data; A model construction module is connected with the multi-source data processing module and is configured to initialize the model parameters of an online sequence extreme learning machine based on the parameter initialization constraint of the fault prior, train the online sequence extreme learning machine according to the defect sample set, and obtain a cable defect identification model; A defect identification module is connected with the model construction module, collects multi-source data of the current cable operation, obtains a defect identification result of the cable according to the cable defect identification model, and evaluates the cable operation state through fuzzy evaluation according to the defect identification result and the operating condition of the current cable; A strategy construction module is connected with the defect identification module and is configured to establish a cable operation and maintenance strategy according to the cable operation state.

8. The multi-source feature parameter based medium voltage cable defect identification system as claimed in claim 7, wherein, Further including: A model optimization module is connected with the model construction module and the defect identification module, configured to compare the cable operation state evaluation result and the defect identification result to generate a feedback signal of state deviation, and when the feedback signal triggers model optimization, the cable defect identification model is optimized through incremental updating.

Images