A reliability prediction method and system for power cables

Abstract

This invention relates to the field of power system equipment operation and maintenance technology, and provides a method and system for reliability prediction of power cables. The method includes: acquiring time-series electrical parameters and non-time-series environmental parameters of the power cable; extracting and filtering features to obtain time-series feature vectors and target environmental feature vectors, and then fusing them; acquiring multiple initial remaining service life labels, and fusing them based on multi-dimensional confidence weights to obtain corrected remaining service life labels; updating the prediction model parameters based on the fused feature vectors and corrected labels, and outputting the performance degradation prediction result. This invention effectively reduces label noise and significantly improves prediction accuracy through heterogeneous feature separation and multi-source label confidence weight correction.

Description

Technical Field

[0001] This application relates to the field of power system equipment operation and maintenance technology, and in particular to a reliability prediction method and system for power cables. Background Technology

[0002] In the field of power transmission, medium and high voltage power transmission cables, as the core carriers of power grid transmission and distribution, operate in high-voltage, high-current environments for extended periods. Their insulation performance undergoes irreversible degradation due to factors such as thermal aging, environmental corrosion, and overload, primarily manifested as increased dielectric loss, decreased insulation resistance, and abnormal internal current, posing safety hazards. Therefore, predicting the reliability of power transmission cables hinges on accurately assessing the degradation trend of their insulation electrical performance. To ensure the safe operation of the power grid, it is typically necessary to predict the reliability of power transmission cables.

[0003] In existing technologies, data-driven prediction methods are commonly used. However, these methods often directly mix time-series electrical parameters (such as dielectric loss and insulation resistance) and non-time-series environmental parameters (such as temperature and humidity) into the model, failing to fully consider the differences in the properties of the data. This results in incomplete feature extraction and makes it difficult to accurately predict the true performance degradation of cables.

[0004] Therefore, there is an urgent need to provide a method and system for predicting the reliability of power cables. Summary of the Invention

[0005] To address the problem of insufficient feature extraction in existing technologies for predicting the degradation of electrical insulation properties of power cables, this application provides a reliability prediction method and system for power cables.

[0006] To achieve the above objectives, this application adopts the following technical solution: On one hand, this application provides a reliability prediction method for power cables, comprising: acquiring time-series electrical parameters and non-time-series environmental parameters of the power cable; extracting features from the time-series electrical parameters to obtain a time-series feature vector, and filtering the non-time-series environmental parameters to obtain a target environmental feature vector; fusing the time-series feature vector with the target environmental feature vector to obtain a fused feature vector; acquiring multiple initial remaining service life labels of the power cable, and weighting and fusing the multiple initial remaining service life labels based on multi-dimensional confidence to obtain a corrected remaining service life label; updating the model parameters of the prediction model based on the fused feature vector and the corrected remaining service life label, and outputting the performance degradation prediction result of the power cable through the updated prediction model.

[0007] Its beneficial effects are that by separating and processing time-series electrical parameters and non-time-series environmental parameters, and performing feature extraction and screening separately, it avoids incomplete feature extraction caused by heterogeneous data mixing; at the same time, by weighting and fusing multi-source labels with multi-dimensional confidence, it effectively quantifies and reduces label noise, thereby significantly improving the prediction accuracy of cable insulation electrical performance degradation.

[0008] As one implementation, the step of extracting features from the time-series electrical parameters to obtain a time-series feature vector includes: extracting features from the time-series electrical parameters using a time-series convolutional network to obtain the time-series feature vector; the step of filtering the non-time-series environmental parameters to obtain a target environmental feature vector includes: calculating the correlation coefficient between the non-time-series environmental parameters and the performance indicators of the power cable, and using environmental parameters whose absolute value of the correlation coefficient is greater than a preset threshold as the target environmental feature vector.

[0009] Its beneficial effects lie in the fact that extracting temporal features through temporal convolutional networks can effectively capture the temporal dependencies in long sequences; and filtering non-temporal features through correlation coefficients eliminates redundant information and reduces model complexity. Specifically, the correlation coefficient is the correlation coefficient between non-temporal environmental parameters and the performance indicators of power cables (such as dielectric loss, insulation resistance, and other insulation electrical performance indicators), thereby ensuring that the selected environmental parameters are strongly correlated with the insulation state of the cables.

[0010] In one implementation, the temporal convolutional network includes multiple layers of residual units, each residual unit including dilated convolutional units and residual connections; the dilation coefficient of the dilated convolutional units increases exponentially with the number of network layers; the residual connections add the input of each residual unit to the convolutional output of that residual unit.

[0011] Its beneficial effects are that by expanding the receptive field through dilating the convolutional units, it can capture long-range dependencies; by preventing gradient vanishing through residual connections, it allows stacking more layers to extract deeper features.

[0012] As one implementation, the multidimensional confidence level includes a time decay factor, a trend stability factor, and a user credibility factor; the time decay factor is determined based on the time difference between the data source time and the failure point time of the initial remaining useful life tag; the trend stability factor is determined based on the data trend stability corresponding to the initial remaining useful life tag; and the user credibility factor is determined based on the user's permission level or historical operation accuracy rate that generated the initial remaining useful life tag.

[0013] Its beneficial effect lies in the fact that it comprehensively considers time, trends and human factors to construct a multi-dimensional confidence evaluation system, which can more comprehensively quantify label quality.

[0014] In one implementation, the time decay factor is a monotonically decreasing function with respect to the time difference; the user credibility factor is determined based on a preset user permission level mapping table, which includes the correspondence between user permission levels and user credibility factors.

[0015] Its beneficial effect is that, through specific function properties and mapping relationships, the calculation method of confidence degree is further clarified, thereby enhancing the feasibility of the scheme.

[0016] As one implementation, the method further includes: for samples that meet the near-failure requirement, using the remaining lifetime percentage as a soft label instead of the hard classification label.

[0017] Its beneficial effect is that, through soft labeling, the model can learn the gradual process of performance degradation, thus solving the problem of blurred classification boundaries for near-fault samples.

[0018] As one implementation method, the prediction model is a long short-term memory network.

[0019] Its beneficial effect lies in the fact that by utilizing the gating mechanism of long short-term memory networks, it is possible to effectively learn complex temporal dependencies in fusion features.

[0020] As one implementation, the method further includes: calculating a health index of the power cable by weighting the ratio of the predicted parameters to the initial parameters in the performance degradation prediction results; and triggering a maintenance warning when the health index is lower than a preset threshold.

[0021] Its beneficial effect is that it quantifies the current status of the cable through a health index and realizes automatic early warning, providing a scientific basis for operation and maintenance decisions.

[0022] On the other hand, this application also provides a reliability prediction system for power cables, comprising: a data acquisition module for acquiring time-series electrical parameters and non-time-series environmental parameters of the power cable; a feature processing module for extracting features from the time-series electrical parameters to obtain a time-series feature vector, and filtering the non-time-series environmental parameters to obtain a target environmental feature vector; a feature fusion module for fusing the time-series feature vector with the target environmental feature vector to obtain a fused feature vector; a label correction module for acquiring multiple initial remaining service life labels of the power cable, and performing weighted fusion of the multiple initial remaining service life labels based on multi-dimensional confidence to obtain a corrected remaining service life label; and a prediction module for updating the model parameters of the prediction model based on the fused feature vector and the corrected remaining service life label, and outputting the performance degradation prediction result of the power cable through the updated prediction model.

[0023] The above system achieves the same technical effect as the method embodiment through the coordinated work of its various modules.

[0024] In addition, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the reliability prediction method as described in any of the preceding claims.

[0025] The technical solution provided in this application has the following beneficial effects: Firstly, it improves the prediction accuracy of electrical performance degradation of power cable insulation. At the feature level, Pearson correlation coefficient is used to screen non-time-series environmental parameters, eliminating redundant noise; a temporal convolutional network (TCN) is used to efficiently extract features of time-series electrical parameters reflecting insulation status, making up for the deficiency of insufficient feature extraction in existing methods. At the label level, the multi-source RUL label correction mechanism comprehensively considers time, trend, and human factors, effectively quantifying and reducing label noise, providing high-quality supervision signals for the model, and solving the core problem of limited prediction accuracy due to high label noise from the data source. Secondly, it enhances the robustness and generalization ability of the model. Through techniques such as residual connection and batch normalization, overfitting of the model is effectively prevented; the multi-source label fusion mechanism enables the model to integrate the advantages of different data sources, enhancing the model's robustness against the defects of a single data source. Thirdly, it enables scientific and systematic predictive maintenance. A complete technical solution has been developed, encompassing data acquisition, feature screening, feature extraction, label correction, model training, and early warning triggering. This solution is particularly suitable for insulation condition assessment and predictive maintenance of medium- and high-voltage power transmission cables, promoting a shift from reactive repair to predictive maintenance, effectively preventing sudden accidents, and ensuring equipment operational safety. A health index is used to quantitatively assess the degree of degradation in cable insulation electrical performance, providing an intuitive and actionable basis for operation and maintenance decisions. Attached Figure Description

[0026] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0027] Figure 1 A flowchart illustrating a reliability prediction method for power cables provided in this application is shown.

[0028] Figure 2 A block diagram of a reliability prediction system for power cables provided in this application is shown. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0031] To facilitate understanding, let's first take one application scenario involved in this application as an example for explanation: In many regions, such as the power grid of a city in southern China, medium and high voltage power transmission cables (such as 10kV and 35kV) serve as the core carriers of power transmission and distribution, operating in high-voltage and high-current environments for extended periods. As the service life increases, the insulation performance of these cables undergoes irreversible degradation due to factors such as thermal aging, environmental corrosion, and overload. This degradation is mainly manifested in increased dielectric loss, decreased insulation resistance, and abnormal internal current, posing safety hazards.

[0032] According to the preventive testing procedures of the power industry (such as DL / T 596 "Preventive Testing Procedures for Power Equipment"), the dielectric loss tangent and insulation resistance value are indicators for evaluating the insulation condition of power transmission cables. However, traditional periodic power outage testing methods suffer from long testing intervals and cannot reflect changes in insulation condition in real time, necessitating an online predictive maintenance method. Existing data-driven prediction methods, when applied to predicting the degradation of electrical performance of power cable insulation, mainly face the following technical challenges: (1) Insufficient feature extraction due to mixed processing of heterogeneous data: Existing methods often directly mix time-series electrical parameters (such as dielectric loss and insulation resistance) and non-time-series environmental parameters (such as temperature and humidity) into the model without fully considering the differences in the properties of the data. Time-series electrical parameters contain rich time-dependent information and need to capture long-term historical degradation trends; while non-time-series environmental parameters are more often used as background conditions or related factors, and may contain a lot of redundant information. If the two are mixed, on the one hand, it will lead to insufficient extraction of time-series features, making it difficult for the model to effectively learn long-term dependencies; on the other hand, redundant environmental noise will interfere with the training of the model and reduce the prediction accuracy.

[0033] (2) The quality of training labels varies greatly, with significant human-induced noise: The remaining useful life (RUL) labels used for model training often come from diverse sources and vary in quality. Although offline destructive test data is relatively accurate, it deviates from actual working conditions and has poor timeliness; online monitoring data trend extrapolation is easily affected by noise, resulting in uncertainty in the results; the generation, recording, and verification of labels are highly dependent on manual processes, and the operations of maintenance personnel with different permissions and experience levels vary significantly, leading to inconsistent raw data quality and human-induced noise. Existing methods do not effectively quantify and integrate the differences in the quality of these multi-source labels, resulting in the training model being misled by low-quality labels and having poor prediction accuracy.

[0034] Taking the aging diagnosis and monitoring method and system for power cables disclosed in CN117829002A as an example, it solves the problem of inaccurate power cable aging monitoring in existing technologies by constructing an initial degradation model and multi-level data processing, combining a spatiotemporal feature map convolutional long short-term memory model and dynamic simulation of transfer functions, identifying false negative samples, and employing a maximum likelihood unbiased estimation algorithm. This achieves more accurate power cable condition assessment and lifespan prediction, improving the reliability and safety of cables. However, it also fails to perform separate feature extraction and environmental redundancy screening for the two types of heterogeneous data, making it difficult for the model to fully capture the long-term degradation trend of electrical parameters and susceptible to environmental noise interference. Secondly, the accuracy of its maximum likelihood unbiased estimation algorithm is highly dependent on the accuracy of the input labels, and it lacks a systematic quantification and correction mechanism for the diversity of label sources (offline experiments, online extrapolation, manual annotation) and quality differences (time decay, trend fluctuations, user credibility), making it difficult to effectively suppress the misleading influence of human noise on model training.

[0035] Based on this, to achieve effective separation and targeted processing of heterogeneous features in the prediction of electrical performance degradation of power cable insulation, and to further reduce the interference of multi-source label noise on model training from the data source, this application provides a method and system for reliability prediction of power cables. By separating heterogeneous features and applying confidence-weighted correction to multi-source labels, the accuracy of predicting electrical performance degradation of cable insulation is improved from the data source. The method is described first, followed by the system.

[0036] Example 1

[0037] like Figure 1 As shown, this embodiment provides a reliability prediction method for power cables, including the following steps: Step S100: Obtain the time-series electrical parameters and non-time-series environmental parameters of the power cable.

[0038] Specifically, the time-series electrical parameters referred to in this embodiment are electrical operation data that change continuously over time or are collected at a fixed frequency, exhibiting significant time dependence and sequential characteristics. For example, time-series electrical parameters may include dielectric loss tangent, insulation resistance value, internal current, and capacitive current, which directly reflect the insulation state. These electrical performance parameters are indicators for evaluating the insulation state of medium and high voltage power transmission cables and can directly reflect the real-time aging degree of the cable insulation material.

[0039] The non-time-series environmental parameters referred to in this embodiment are data characterizing the cable's operating environment status or maintenance records. These parameters typically do not have strict time-series characteristics and are more often discrete background information. For example, non-time-series environmental parameters may include ambient temperature, ambient humidity, and the number of maintenance and cleaning cycles. It should be understood that although specific parameter types have been listed above, in other embodiments, other types of electrical or environmental parameters may be selected based on actual monitoring conditions, as long as they can reflect the cable's operating status or environmental impact.

[0040] Step S200: Extract features from the time-series electrical parameters to obtain a time-series feature vector, and filter the non-time-series environmental parameters to obtain a target environmental feature vector.

[0041] This step is used to separate heterogeneous features. Related techniques often directly mix time-series electrical parameters with non-time-series environmental parameters as input into the model. This approach ignores the differences in data properties: time-series electrical parameters contain rich time-dependent information and need to capture long-term historical trends; while non-time-series environmental parameters serve more as background conditions or related factors, and may contain a large amount of redundant information unrelated to cable performance degradation. Mixing the two leads to insufficient extraction of time-series features, making it difficult for the model to effectively learn long-term dependencies; furthermore, redundant environmental noise interferes with model training, reducing prediction accuracy.

[0042] Therefore, this embodiment adopts a separate processing strategy: for time-series electrical parameters, a dedicated feature extraction network (such as a time-series convolutional network) is used for deep feature mining to capture the degradation patterns in their time series; for non-time-series environmental parameters, a filtering process is performed to remove redundant information and retain strongly correlated features. This strategy enables targeted processing of data with different properties, significantly improving the effectiveness of feature representation.

[0043] Step S300: The temporal feature vector is fused with the target environment feature vector to obtain a fused feature vector.

[0044] Specifically, after obtaining the time-series feature vector and the target environment feature vector respectively, the two are fused. The fusion can be achieved through vector concatenation, weighted summation, or other feature fusion techniques. For example, the time-series feature vector and the target environment feature vector can be concatenated dimensionally to form a comprehensive feature vector that includes both time-series degradation information and environmental background information. This fused feature vector retains the time-dependent characteristics of electrical parameters while incorporating the influence of environmental factors, thus providing a more comprehensive characterization of the current operating status of power cables.

[0045] Step S400: Obtain multiple initial remaining service life tags of the power cable, and perform weighted fusion of the multiple initial remaining service life tags based on multidimensional confidence to obtain the corrected remaining service life tags.

[0046] This step is used to correct the multidimensional confidence labels. In practical applications, the labels used to train the prediction model (i.e., Remaining Useful Life (RUL) labels) often come from diverse sources and vary in quality. For example, some labels come from offline destructive testing, which, while relatively accurate, deviates from actual operating conditions; some labels are derived from trend extrapolation of online monitoring data and are susceptible to noise interference; and some labels are labeled by maintenance personnel based on experience, resulting in subjective errors. Human noise or system noise can seriously mislead model training. Related technologies often ignore the differences in label quality, treating all labels equally, which limits prediction accuracy.

[0047] This embodiment introduces a multi-dimensional confidence evaluation mechanism to quantify the quality of each initial remaining useful life label. The multi-dimensional confidence score considers multiple dimensions, such as the label's temporal attributes, data trend stability, and source reliability. Based on the calculated multi-dimensional confidence scores, multiple initial labels are weighted and fused: labels with high confidence scores are assigned higher weights, and labels with low confidence scores are assigned lower weights. This weighted fusion reduces the interference of low-quality labels on model training, improves the quality of training samples from the data source, and solves the problem of low prediction accuracy caused by inconsistent quality of multi-source labels.

[0048] Step S500: Based on the fused feature vector and the corrected remaining service life label, update the model parameters of the prediction model, and output the performance degradation prediction result of the power cable through the updated prediction model.

[0049] Specifically, the fused feature vector is used as input, and the corrected remaining lifetime label is used as a supervision signal to train the prediction model and update the model parameters. The prediction model can employ a neural network model, such as a Long Short-Term Memory (LSTM) network, to learn the complex mapping relationship between the fused feature vector and performance degradation.

[0050] After training, the fused feature vectors acquired in real time are input into the updated prediction model, which then outputs the performance degradation prediction results for power cables, such as the predicted remaining service life and health status score. Through the above steps, this embodiment achieves end-to-end optimization from data acquisition, feature separation and processing, label correction to model training and prediction, improving the accuracy and robustness of power cable reliability prediction.

[0051] Furthermore, existing power cable reliability prediction technologies assume that monitoring data is complete and stable. Whether based on data-driven prediction models or physical mechanism-based degradation models, current methods assume that the input end can reliably acquire all time-series electrical parameters, such as insulation resistance, and directly perform feature extraction and model training based on this. However, in actual power grid operation and maintenance scenarios, this assumption deviates significantly from engineering reality. Cable lines of different voltage levels, years of operation, and regions exhibit vastly different levels of online monitoring equipment configuration: newly built smart lines may deploy a full range of sensors, while many older lines, remote substations, or early renovation projects only have basic insulation resistance monitoring devices installed. Existing technologies do not provide any adaptive solutions for this gradient in monitoring capabilities, leading to a dilemma where prediction systems work on high-configuration lines but malfunction on low-configuration lines during actual deployment.

[0052] Therefore, before obtaining the time-series electrical parameters and non-time-series environmental parameters of the power cable, the method may further include: Send a status query command to the sensor interface to detect the availability of the data acquisition device corresponding to the timing electrical parameters; The data integrity score is determined based on the ratio of the number of available time-series electrical parameter types to the preset number of complete parameter types. The prediction model path is selected based on the data integrity score. When the full monitoring condition is met, the full parameter prediction model is invoked to perform reliability prediction. When the limited monitoring condition is met and at least the insulation resistance value can be obtained, the simplified prediction model is invoked to perform a simplified prediction process based on the insulation resistance value and the non-time-series environmental parameters. When no monitoring condition is met, the prediction process is terminated and a message indicating insufficient monitoring capability is output to the user equipment. The system continuously monitors the status of the data acquisition device during the prediction process, and dynamically switches the prediction model path when a change in the device status is detected. When the system switches to the limited monitoring conditions, it outputs prediction accuracy risk warning information to the user device.

[0053] Therefore, this embodiment incorporates sensor availability into the decision variable of the prediction process. It actively sends query commands to the sensor interface and quantifies the score. In scenarios where monitoring resources are limited and manual verification is required, existing technologies often directly report errors or output meaningless results. This embodiment provides a path selection between a complete model and a simplified model, achieving flexible adaptation of the prediction service to scenarios with varying monitoring capabilities.

[0054] Example 2

[0055] This embodiment, based on embodiment 1, provides a detailed explanation of the specific implementation of feature extraction and filtering in step S200.

[0056] For the feature extraction of time-series electrical parameters, this embodiment specifically uses a time-series convolutional network to extract the features of time-series electrical parameters and obtain time-series feature vectors.

[0057] Specifically, the temporal convolutional network comprises multiple layers of residual units, each layer including dilated convolutional units and residual connections. The dilation coefficient of the dilated convolutional unit increases exponentially with the number of network layers. For example, in a preferred embodiment, the dilation coefficient of the first layer residual unit can be set to 1, the second layer to 2, the third layer to 4, and so on, exhibiting an exponential growth. Through this exponentially increasing dilation coefficient, the receptive field of the convolutional kernel can quickly cover the entire input sequence, effectively capturing long-term degradation trends in temporal electrical parameters without increasing the number of network parameters or computational complexity. For example, for long-period data such as insulation resistance values ​​that decrease slowly over time, dilated convolution can establish correlations across multiple time steps, avoiding the problem of limited receptive fields in traditional convolutional networks.

[0058] Residual connections add the input of each residual unit to the convolutional output of that residual unit. Specifically, residual connections introduce cross-layer identity mappings, allowing gradients to be losslessly propagated to previous layers during backpropagation. This mechanism effectively solves the gradient vanishing problem that easily occurs during the training of deep networks, making it possible to build deeper temporal convolutional networks. This enables the extraction of deeper and more abstract temporal degradation features, further improving the robustness of feature representation.

[0059] The process of filtering the non-time-series environmental parameters to obtain the target environmental feature vector includes: Calculate the correlation coefficient between the non-time-series environmental parameters and the performance indicators of power cables, and use the environmental parameters whose absolute value of the correlation coefficient is greater than a preset threshold as the target environmental feature vector.

[0060] Specifically, this embodiment preferably uses the Pearson correlation coefficient as an indicator to measure correlation. The Pearson correlation coefficient measures the degree of linear correlation between two variables, and its value ranges from -1 to 1. The closer the absolute value is to 1, the stronger the correlation. In the calculation process, the Pearson correlation coefficient between each non-time-series environmental parameter (such as ambient temperature, humidity, maintenance frequency, etc.) and the electrical performance indicators of cable insulation (such as dielectric loss, insulation resistance) is first calculated. Then, a preset threshold is set, for example, 0.7. When the absolute value of the correlation coefficient of an environmental parameter is greater than 0.7, it is determined that the parameter has a strong correlation with cable performance and is retained as an element in the target environmental feature vector; otherwise, the parameter is removed.

[0061] It should be understood that the threshold of 0.7 is only a preferred example. In practical applications, the threshold can be adjusted by looking up a table according to the requirements of data quality and prediction accuracy. For example, for scenarios with high data quality and low noise, the threshold can be appropriately reduced to 0.5 or 0.6 to retain more potential marginally relevant features; for scenarios with high data noise, the threshold can be increased to 0.8 to ensure that the selected features have high reliability.

[0062] By using the Pearson correlation coefficient filtering method described above, redundant environmental parameters that are unrelated to or weakly correlated with the degradation of cable insulation electrical performance, such as the number of maintenance and cleaning operations in certain specific scenarios, can be effectively eliminated. This not only reduces the dimensionality of the input data and alleviates the computational burden on the subsequent prediction model, but also eliminates the interference of redundant noise on model training. This allows the prediction model to focus on the key environmental factors that truly affect the reliability of cable insulation, thereby significantly improving prediction accuracy.

[0063] This embodiment achieves targeted processing of heterogeneous data through the combined application of temporal convolutional networks and Pearson correlation coefficient screening: the former uses dilated convolution and residual connections to deeply mine the long-term dependency features of temporal data, while the latter uses statistical methods to accurately screen key influencing factors of non-temporal data for constructing a fusion feature vector.

[0064] Example 3

[0065] Based on Example 1, this embodiment provides a detailed explanation of the specific calculation method of multidimensional confidence in step S400 and the process of label weighted fusion.

[0066] Specifically, the multidimensional confidence level includes time decay factor, trend stability factor and user credibility factor, which quantify the credibility of the initial remaining useful life label from the dimensions of time, data quality and source reliability, respectively, and together constitute a three-dimensional label quality evaluation system.

[0067] The time decay factor is determined based on the time difference between the data source time of the initial remaining useful life tag and the time of the failure point.

[0068] It should be understood that the operating status of power cables changes dynamically over time. Data closer to the time of the fault point can better reflect the true state of the cable before the fault. Conversely, data from a longer time ago will have less reference value for predicting the current fault due to changes in the cable's operating environment and load conditions.

[0069] Therefore, in this embodiment, the time decay factor is set as a monotonically decreasing function with respect to the time difference. In a preferred embodiment, the time decay factor can be characterized by an exponential decay function, for example... ,in The time difference between the data source time and the time of the fault point. This represents the decay coefficient. By using exponential decay, the objective law of data value rapidly decaying over time can be simulated, allowing labels closer to the fault point to receive higher confidence weights, thereby improving the sensitivity of the prediction model to recent state changes.

[0070] The trend stability factor is determined based on the trend stability of the data corresponding to the initial remaining useful life label.

[0071] Specifically, the trend stability factor measures the degree of volatility of the data sequence on which the label is based. If the data sequence fluctuates wildly and is noisy, the generated label has low reliability; if the data sequence is stable and has a clear trend, the label has high reliability. In one specific implementation, trend stability can be quantified by calculating the variance or standard deviation of the data sequence. For example, for an initial label obtained by extrapolating the trend of online monitoring data, the variance of its source data sequence can be calculated. The smaller the variance, the more stable the data trend, and the higher the corresponding trend stability factor. This design effectively eliminates unreliable labels caused by sudden sensor interference or short-term abnormal fluctuations, further improving the quality of training data.

[0072] The user credibility factor is determined based on the user's permission level or historical operation accuracy when generating the initial remaining useful life tag.

[0073] It is important to note that the user credibility factor referred to in this embodiment is a technical quantitative parameter for data source quality. In actual power operation and maintenance systems, tag generation and confirmation are often performed by maintenance personnel at different levels. Tags confirmed by senior engineers or experienced experts are generally more accurate than those confirmed by junior maintenance personnel; or, some data sources have a high accuracy record after multiple historical verifications. This embodiment introduces a user credibility factor to transform this source reliability into a calculable numerical weight. Specifically, the user credibility factor can be determined based on a preset user permission level mapping table, which includes the correspondence between user permission levels and user credibility factors. For example, a mapping table can be pre-stored, mapping senior experts to a credibility factor of 0.95, intermediate engineers to 0.8, and junior maintenance personnel to 0.6. Alternatively, it can be dynamically calculated based on historical operation accuracy; for example, if a data source has achieved an accuracy rate of 98% in the past 100 tag generation iterations, its credibility factor is set to 0.98. Through this technical means, the data quality differences caused by human factors are quantified, effectively reducing the interference of human noise on model training.

[0074] After calculating the above three factors, this embodiment obtains the corrected remaining useful life label through weighted fusion. Specifically, the multidimensional confidence level can be determined as the product of the time decay factor, the trend stability factor, and the user confidence factor, i.e. ,in To assess the overall confidence level, The time decay factor, As a trend stability factor, The user confidence factor is used. Then, for multiple initial remaining service life tags of the same cable, a weighted average is performed using the overall confidence level to obtain the corrected remaining service life tag. Through the above process, this embodiment achieves the fusion of multi-source labels, enabling high-quality labels to play a dominant role in training while suppressing the influence of low-quality labels, thereby significantly improving the training accuracy of the prediction model from the data source.

[0075] Example 4

[0076] This embodiment, based on Embodiment 1, provides a detailed explanation of the training details and model structure of the prediction model.

[0077] In actual power cable operation, cable performance degradation is a continuous and gradual process, rather than an abrupt binary state of normal and fault. However, traditional prediction methods often use hard classification labels, such as labeling normal samples as 0 and faulty samples as 1. This approach has significant drawbacks: for samples that meet the near-fault requirement, their state is between normal and faulty. The hard classification label forces them into a certain category, ignoring their gradual degradation information, making it difficult for the model to learn the smooth transition process of performance degradation, thus affecting prediction accuracy. Therefore, this embodiment provides a preferred implementation method: for samples that meet the near-fault requirement, the remaining lifetime percentage is used as a soft label instead of the hard classification label.

[0078] Specifically, in this embodiment, a near-failure sample refers to a sample whose remaining lifetime percentage is lower than a preset threshold (e.g., 20%). For such samples, a simple 0 or 1 label is no longer used; instead, the remaining lifetime percentage is used as the label value. For example, if a sample's remaining lifetime percentage is 20%, its label is set to 0.2; if the remaining lifetime percentage is 50%, the label is set to 0.5.

[0079] It should be understood that the above values ​​are merely examples and can be flexibly set according to cable type and fault mode in practical applications. By introducing soft labels, the model can learn the gradual trajectory of the cable transitioning from a normal state to a fault state, thereby more accurately capturing the dynamic process of performance degradation, solving the problem of blurred classification boundaries caused by hard classification labels, and significantly improving the model's ability to identify critical states.

[0080] After obtaining the corrected remaining lifetime label (or soft label) and the fused feature vector, a suitable prediction model needs to be selected for training. This embodiment specifically uses a Long Short-Term Memory (LSTM) network as the prediction model. This is because, in the application scenario of this application, the LTM network introduces a gating mechanism, which can effectively solve the gradient vanishing or gradient explosion problems that traditional recurrent neural networks easily encounter when processing long sequences, making it suitable for processing time-dependent fused feature vectors.

[0081] Specifically, the basic units of a Long Short-Term Memory (LSTM) network include a forget gate, an input gate, and an output gate. The forget gate determines how much information from the previous time step's unit state is retained in the current time step. For example, regarding the historical ambient temperature information of a cable, if the temperature changes abruptly at the current time step, the forget gate can selectively forget some historical temperature information to adapt to the new environmental conditions. The input gate determines how much of the current input information (i.e., the fused feature vector) is stored in the unit state. For example, when an abnormal downward trend in the insulation resistance value is detected, the input gate can control the writing of this key feature into the unit state, updating the memory. The output gate determines how much information from the current unit state is output to the next layer. Through the synergistic effect of these three gating units, the LSM network can selectively remember key historical degradation information and forget irrelevant noise interference, thereby accurately learning the complex mapping relationship between the fused feature vector and cable performance degradation.

[0082] This embodiment further improves prediction accuracy by combining soft labeling with a long short-term memory network, addressing both label optimization and model architecture. Soft labels enable the model to capture gradual degradation processes, while the long short-term memory network ensures the effective utilization of long-sequence information. Together, they achieve accurate prediction of the reliability status of power cables.

[0083] Example 5

[0084] This embodiment, based on Embodiment 1, provides a detailed explanation of the application and output of the prediction results. Specifically, after outputting the performance degradation prediction results of the power cable through the updated prediction model, the method of this embodiment further includes: calculating a health index of the power cable based on a weighted average of the ratio of the predicted parameters in the performance degradation prediction results to the initial parameters; and triggering a maintenance warning when the health index is lower than a preset threshold.

[0085] Specifically, performance degradation prediction results typically include several key parameters that characterize the electrical performance of cable insulation, such as the predicted dielectric loss tangent, insulation resistance value, and internal current. To intuitively quantify the current health status of the cable, this embodiment constructs a comprehensive health index calculation model. The core idea of ​​this model is to compare the predicted parameters with the initial parameters of the cable when it is put into operation (i.e., the baseline values ​​in a healthy state), and reflect the degree of insulation electrical performance degradation through the ratio.

[0086] In a preferred embodiment, the health index can be calculated as follows: Health Index = (Predicted dielectric loss / Initial dielectric loss) × W1 + (Initial insulation resistance / Predicted insulation resistance) × W2 + (Predicted internal current / Initial internal current) × W3.

[0087] In this formula, W1, W2, and W3 are preset weighting coefficients, with values ​​ranging from 0 to 1, and the sum of W1, W2, and W3 is 1. It should be understood that the ratio term in the above formula fully considers the physical characteristics of different parameters in the application scenario of this application: for dielectric loss and internal current, an increase in their values ​​usually means a decrease in insulation performance; therefore, the predicted value / initial value is used. The larger this ratio, the worse the health status, and the smaller the corresponding health index component should be (or it can be inverted). For insulation resistance, a decrease in its value means a decrease in insulation performance; therefore, the initial value / predicted value is used. The larger this ratio, the more severe the decrease in predicted resistance value relative to the initial value. Through this differentiated ratio calculation, it is ensured that the contribution direction of each component to the health index is consistent. The specific values ​​of the weighting coefficients W1, W2, and W3 can be adaptively adjusted according to the cable type, operating environment, and historical fault data. For example, for cables sensitive to insulation aging, the weight W2 corresponding to insulation resistance can be appropriately increased; for cables with heavy loads, the weight W3 corresponding to internal current can be appropriately increased. By using weighted calculations, multiple performance indicators can be combined to obtain a comprehensive and objective health index, avoiding the one-sidedness of evaluation based on a single indicator.

[0088] After obtaining the health index, this embodiment further compares it with a preset threshold to determine whether a maintenance warning should be triggered. The preset threshold is a critical value for measuring whether the cable's health status is in a dangerous range, and its specific value can be set according to operation and maintenance experience and safety standards. For example, in a specific embodiment, the preset threshold can be set to 0.6. When the calculated health index is lower than 0.6, the cable is determined to be in a sub-healthy or fault-risk state, and the maintenance warning mechanism is automatically triggered.

[0089] The maintenance early warning mechanism goes beyond simple alarm prompts; it can further generate decision reports containing detailed information. These reports may include current health index values, specific values ​​of various predicted parameters, deviation percentages from initial values, estimated remaining useful life (RUL), and recommended maintenance measures (such as "immediate shutdown for maintenance" or "insulation performance re-inspection within the next week"). This embodiment does not limit these measures. Through this visualized health index and decision report, maintenance personnel can intuitively and quickly grasp the operating status of cables, thereby developing scientific and reasonable maintenance plans. This achieves a shift from traditional reactive repairs to predictive maintenance, effectively preventing sudden power outages and ensuring the safe and stable operation of the power system.

[0090] Example 6

[0091] See Figure 2This embodiment provides a reliability prediction system for power cables, which implements the reliability prediction method for power cables described in any one of embodiments 1 to 5. The system integrates data acquisition, feature processing, label correction, and prediction and early warning functions through a modular architecture design, using software or hardware, making it easy to deploy and implement in power operation and maintenance platforms or edge computing devices.

[0092] Specifically, the reliability prediction system for power cables includes a data acquisition module, a feature processing module, a feature fusion module, a tag correction module, and a prediction module. It should be understood that these modules can be implemented as software functional units, such as program code stored in memory and executed by a processor; or as hardware circuits, such as logic circuits in a field-programmable gate array (FPGA) or application-specific integrated circuit (ASIC). The following provides a detailed description of each module in conjunction with its specific functions.

[0093] The data acquisition module is used to acquire time-series electrical parameters and non-time-series environmental parameters of power cables. This module serves as the system's input, responsible for communicating with external sensor networks or databases. At the hardware level, the data acquisition module can connect via a data bus to online monitoring devices (such as dielectric loss sensors and current transformers) and environmental monitoring sensors (such as thermometers and hygrometers) deployed on the power cables. At the software level, this module is equipped with a data interface driver, capable of reading sensor data at a preset sampling frequency (e.g., once per hour) and performing preliminary data cleaning, such as removing obvious outliers and filling in missing values, thereby providing high-quality raw data for subsequent processing.

[0094] The feature processing module is used to extract features from the time-series electrical parameters to obtain time-series feature vectors, and to filter the non-time-series environmental parameters to obtain target environmental feature vectors. This module is the core unit for data governance in the system. It integrates a temporal convolutional network (TCN) algorithm and a correlation analysis algorithm. For time-series electrical parameters, the feature processing module calls a pre-trained TCN model, uses dilated convolutional units to capture long-distance dependency features, and outputs a high-dimensional time-series feature vector. For non-time-series environmental parameters, the feature processing module executes a Pearson correlation coefficient calculation program, retaining environmental parameters with an absolute correlation coefficient greater than a preset threshold (e.g., 0.7) and removing redundant parameters. Through the differentiated processing of the feature processing module, effective separation and feature mining of heterogeneous data are achieved.

[0095] The feature fusion module is used to fuse the temporal feature vector with the target environment feature vector to obtain a fused feature vector. This module is responsible for integrating the two types of features output by the feature processing module.

[0096] In one specific implementation, the feature fusion module uses vector concatenation to link the temporal feature vector with the target environment feature vector in terms of dimension, forming a fused feature vector containing complete state information. This fused feature vector includes both the temporal evolution of electrical parameters and the influence weights of environmental factors, enabling it to comprehensively characterize the current operating state of the power cable.

[0097] The label correction module is used to acquire multiple initial remaining service life labels for the power cables, and then weighted and fused these initial remaining service life labels based on multi-dimensional confidence scores to obtain corrected remaining service life labels. This module is a key unit for improving the quality of training data in the system. The label correction module internally stores a multi-dimensional confidence score calculation model, including a time decay function, trend stability calculation logic, and a user permission level mapping table. When initial labels are received from different data sources, the module automatically extracts the timestamps, source data sequences, and operator information of the labels, and calculates the time decay factor, trend stability factor, and user confidence factor respectively to obtain a comprehensive confidence score. Subsequently, this confidence score is used to perform a weighted average of multiple initial labels to output the corrected labels. This process effectively suppresses the interference of low-quality labels on model training.

[0098] The prediction module updates the model parameters of the prediction model based on the fused feature vector and the corrected remaining useful life (RUL) label, and outputs the performance degradation prediction result of the power cable through the updated prediction model. This module is the core of the system's output. The prediction module deploys a Long Short-Term Memory (LSTM) network or other prediction models. During the training phase, the prediction module uses the fused feature vector as input and the corrected label as a supervision signal, updating the model parameters through a backpropagation algorithm. During the inference phase, the prediction module receives the real-time generated fused feature vector and outputs the cable's performance degradation prediction result, such as the remaining useful life (RUL) or health status score. Furthermore, the prediction module can integrate early warning logic; when the prediction result indicates that the cable health index is below a preset threshold, it automatically generates a maintenance decision report and pushes it to maintenance personnel through a human-machine interface.

[0099] Through the collaborative work of the above modules, the system provided in this embodiment realizes fully automated processing from data acquisition, feature engineering, tag management to prediction output, which not only significantly improves the accuracy of power cable reliability prediction, but also provides reliable technical support for the intelligent operation and maintenance of power systems.

[0100] Furthermore, the step of extracting features from the timing electrical parameters to obtain a timing feature vector includes: The time-series electrical parameters are feature-extracted using a time-series convolutional network to obtain the time-series feature vector; The process of filtering the non-time-series environmental parameters to obtain the target environmental feature vector includes: Calculate the correlation coefficient between the non-time-series environmental parameters and the performance indicators of power cables, and use the environmental parameters whose absolute value of the correlation coefficient is greater than a preset threshold as the target environmental feature vector.

[0101] Furthermore, the reliability prediction system also includes: The early warning triggering module is used to calculate the health index of the power cable by weighting the ratio of the predicted parameters to the initial parameters in the performance degradation prediction results; when the health index is lower than a preset threshold, a maintenance early warning is triggered.

[0102] Furthermore, the reliability prediction system also includes: The monitoring and evaluation module sends a status query command to the sensor interface to detect the availability of the data acquisition device corresponding to the timing electrical parameters; The path selection module determines a data integrity score based on the ratio of the number of actually obtainable time-series electrical parameter types to the preset number of complete parameter types. It then selects a prediction model path based on this score. Specifically, when the conditions for comprehensive monitoring are met, the complete parameter prediction model is invoked to perform reliability prediction. When the conditions for limited monitoring are met and at least the insulation resistance value is obtainable, a simplified prediction model is invoked, and a simplified prediction process is performed based on the insulation resistance value and the non-time-series environmental parameters. When no monitoring condition is met, the prediction process is terminated, and an insufficient monitoring capability warning message is output to the user equipment. The dynamic switching module continuously monitors the status of the data acquisition device during the prediction process and dynamically switches the prediction model path when a change in the device status is detected. When the operation is switched to the limited monitoring conditions, the module outputs prediction accuracy risk warning information to the user device.

[0103] Example 7

[0104] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the reliability prediction method for power cables as described in any of the foregoing embodiments of this application.

[0105] Specifically, a computer-readable storage medium can be any tangible medium capable of containing or storing a program that can be used by, or in conjunction with, an instruction execution system, apparatus, or device. In this embodiment, a computer-readable storage medium may include, but is not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. It should be understood that a computer-readable storage medium is not limited to the specific forms listed above; any medium capable of carrying computer program code and being readable by a processor should be covered within the scope of protection of this embodiment.

[0106] In practical applications, when a computer program is executed on a computer or processor, the program is loaded into the computer's main memory via a computer-readable storage medium, and the processor reads and executes the instructions within the program. The processor then executes the steps described in Examples 1 to 5 according to the program's instructions: The system acquires time-series electrical parameters and non-time-series environmental parameters of power cables; extracts features from the time-series electrical parameters and filters the non-time-series environmental parameters; performs feature fusion; applies weighted fusion correction to the initial remaining service life label based on multi-dimensional confidence; and updates the prediction model and outputs the prediction results.

[0107] By combining hardware and software in this embodiment, the technical solutions in the aforementioned method embodiments are solidified into a distributable and storable product form, which facilitates the deployment, transmission and commercial application of the software, thereby covering software distribution channels other than methods and systems, and providing comprehensive protection for the technical solutions.

[0108] Example 8

[0109] To more intuitively demonstrate the practical application effect of the technical solution of this application in the field of power transmission cables, this embodiment takes the online monitoring and reliability prediction of a 10kV power transmission cable in a city power grid as an example to explain in detail the reliability prediction method described in the above embodiment.

[0110] In this application scenario, a section of power transmission cable is selected as the monitoring object. This cable operates in a high-voltage, high-current environment for extended periods, representing a typical medium- to high-voltage power transmission cable maintenance scenario. First, step S100 is executed to acquire the time-series electrical parameters and non-time-series environmental parameters of the power cable. Specifically, sensors deployed at the cable joints continuously collect time-series electrical parameters at a frequency of once per hour, primarily including the dielectric loss tangent and insulation resistance value. Simultaneously, non-time-series environmental parameters are collected through an environmental monitoring terminal, including ambient temperature, ambient humidity, and maintenance records such as cumulative equipment runtime and maintenance / cleaning frequency.

[0111] Subsequently, step S200 is executed to extract features from the time-series electrical parameters to obtain a time-series feature vector, and to filter the non-time-series environmental parameters to obtain a target environmental feature vector. For the time-series electrical parameters, this embodiment inputs them into a pre-trained temporal convolutional network (TCN) for deep feature extraction. The TCN network uses dilated convolutional units to capture long-range dependencies. For example, the network identifies a slow decreasing trend in insulation resistance over the past 30 days; this feature is extracted and mapped as a key component in the high-dimensional time-series feature vector. For the non-time-series environmental parameters, this embodiment uses the Pearson correlation coefficient for filtering. The calculation results show that the absolute value of the correlation coefficient between ambient temperature and cable insulation performance indicators is 0.8, and the ambient humidity is 0.75, both greater than the preset threshold of 0.7, indicating that these two parameters are strongly correlated with cable aging and are retained as the target environmental feature vector. However, the absolute value of the correlation coefficient between the number of maintenance cleaning operations is only 0.3, indicating a weak linear correlation with the current insulation performance of the cable, possibly due to noise caused by lag or incomplete maintenance records, and is therefore discarded. This filtering process effectively reduces the dimensionality and noise interference of the input data.

[0112] Next, step S300 is executed to fuse the temporal feature vector with the target environment feature vector to obtain a fused feature vector. This embodiment uses vector concatenation to integrate the temporal feature vector output by the TCN with the filtered environmental feature vector, forming a comprehensive feature vector that includes electrical degradation trends and environmental background information.

[0113] In the tag processing stage, step S400 is executed to obtain multiple initial remaining service life (RUL) tags for the power cable. These initial RUL tags are then weighted and fused based on multidimensional confidence levels to obtain corrected RUL tags. In this embodiment, two initial RUL tags are obtained for the cable: one derived from trend extrapolation of online monitoring data, and the other from experience-based annotations by maintenance personnel. The multidimensional confidence calculation process is as follows: First, for the online monitoring extrapolation tag, because its data source time is relatively recent, the time decay factor is high; however, the data sequence exhibits some fluctuation, resulting in a moderate trend stability factor; and since this tag is automatically generated by the system, the user confidence factor is set to a moderate level. Second, for the tag annotated by maintenance personnel, although the data source time is slightly earlier, the time decay factor is slightly lower; however, this tag is confirmed by a senior engineer, whose historical operation accuracy is high. According to the preset user permission level mapping table, the user's confidence factor is assigned a higher value (e.g., 0.95). Finally, the comprehensive confidence calculation results show that the tag confirmed by the senior engineer has a significantly higher weight than the system extrapolation tag. Through weighted fusion, the corrected RUL labels are more aligned with expert experience, effectively correcting potential biases in system extrapolation. In practical applications, a predetermined number of initial RUL labels can be selected from the database in the order of most recent updates.

[0114] Finally, step S500 is executed: based on the fused feature vector and the corrected remaining service life label, the model parameters of the prediction model are updated, and the performance degradation prediction result of the power cable is output through the updated prediction model. The fused feature vector is input into a Long Short-Term Memory (LSTM) network, and the model outputs predicted performance parameters such as dielectric loss tangent and insulation resistance value.

[0115] Based on the prediction results, this embodiment further calculates the health index and triggers an early warning. Specifically, the health index of the power cable is obtained by weighting the ratio of the predicted parameters to the initial parameters in the performance degradation prediction results. Assuming the predicted dielectric loss tangent angle increases by 150% relative to the initial value, and the insulation resistance value decreases by 60% relative to the initial value, a weighted calculation is performed according to preset weighting coefficients, resulting in a final health index of 0.45. Since this health index is lower than the preset threshold of 0.6, the cable's insulation performance is determined to be in a sub-healthy state, posing a high risk of failure, and a maintenance early warning is triggered. A decision report is automatically generated and sent to the user's equipment (e.g., a mobile phone), prompting maintenance personnel: "The cable's health index is 0.45, lower than the safety threshold. It is recommended to arrange a re-inspection of the insulation performance within the next week, paying particular attention to abnormal temperatures at the joints." Based on this, the method provided in this application can eliminate redundant environmental noise, integrate expert experience to correct label bias, accurately predict the degradation trend of electrical performance of power transmission cable insulation, and quantify the output of health index, thus realizing the transformation from passive emergency repair to proactive predictive maintenance.

[0116] Example 9

[0117] This embodiment takes the insulation condition assessment and predictive maintenance of power transmission cables in a certain region as the application scenario, and provides the complete application process and technical effects of the technical solution of this application in the operation and maintenance of power transmission cables.

[0118] Application Scenario Overview: This power transmission cable serves as the main power supply line for an industrial area, enduring high voltage and high current loads over extended periods, and operating in an environment with significant temperature fluctuations (soil temperatures can reach over 40℃ in summer). Preventative testing for power cables includes insulation resistance measurement and dielectric loss tangent measurement. In recent years, this cable has repeatedly experienced overload operation, prompting maintenance personnel to monitor its insulation aging trend. However, traditional periodic power outage testing methods suffer from long testing intervals and cannot reflect real-time changes in insulation condition, necessitating an online predictive maintenance method.

[0119] Step S100: Obtain the time-series electrical parameters and non-time-series environmental parameters of the power cable. In this scenario, online monitoring devices deployed at cable terminals and joints continuously collect time-series electrical parameters every 30 minutes, including dielectric loss tangent (tanδ), insulation resistance value, internal current, and capacitive current. Simultaneously, non-time-series environmental parameters are collected, including ambient temperature, ambient humidity, historical overload counts, and maintenance records.

[0120] Step S200: Feature extraction is performed on the time-series electrical parameters to obtain a time-series feature vector, and the non-time-series environmental parameters are filtered to obtain a target environmental feature vector. For the time-series electrical parameters, dielectric loss tangent, insulation resistance, internal current, and capacitance current data for one consecutive month are input into a temporal convolutional network (TCN) for deep feature extraction. The TCN network captures the following key degradation trends through dilated convolutional units (the dilation coefficient increases exponentially with the number of layers): the dielectric loss tangent has shown a continuous upward trend over the past month, increasing from an initial 0.2% to 0.35%; the insulation resistance has shown a slow downward trend, decreasing from an initial 500MΩ to 320MΩ. These long-term degradation trends are effectively extracted by the TCN and mapped into key components in the high-dimensional time-series feature vector. For the non-time-series environmental parameters, Pearson correlation coefficients are used for filtering. The calculation results show that the absolute value of the correlation coefficient between ambient temperature and the tangent of dielectric loss is 0.82, the ambient humidity is 0.76, and the number of historical overloads is 0.71, all of which are greater than the preset threshold of 0.7 and are retained as target environmental feature vectors; while the absolute value of the correlation coefficient between the number of maintenance and repairs is only 0.25 and is discarded.

[0121] Step S300: The temporal feature vector is fused with the target environment feature vector to obtain a fused feature vector. The temporal feature vector output by TCN is concatenated and fused with the filtered environmental feature vector to form a comprehensive feature vector containing electrical degradation trends and environmental background information.

[0122] Step S400: Obtain multiple initial remaining service life (RUL) tags for the power cable. Weighted fusion of these tags based on multidimensional confidence levels yields a corrected RUL tag. In this scenario, three initial RUL tags are obtained: the first type originates from offline preventative test data (insulation resistance and dielectric loss values ​​measured during the previous power outage test, estimating RUL to be 8 months); the second type comes from trend extrapolation of online monitoring data (RUL estimated to be 5 months based on the data trend of the past 6 months); and the third type is based on the experience of senior engineers (RUL marked as 6 months based on a comprehensive judgment of on-site inspections and operation records). The calculation process for the multidimensional confidence level is as follows: For offline trial labels: the data source time is 12 months ago, and the time decay factor is low ( However, the experimental data itself is relatively accurate, and the calculated trend stability factor is of moderate quality. This tag was confirmed by an intermediate engineer, with a user credibility factor of 0.8. Overall confidence level. .

[0123] For online monitoring extrapolation labels: the data source time is only 1 month from the present, and the time decay factor is high ( However, the data series exhibits some fluctuations, and the calculated trend stability factor is moderately low. This tag was automatically generated by the system, with the user confidence factor set to 0.6. Overall confidence level. .

[0124] For senior engineers, the label should be: data source time is 3 months from the present, and the time decay factor is moderate. This label is based on a comprehensive judgment, and the calculated trend stability factor is relatively high. The annotator is a senior engineer with 20 years of experience; based on a pre-defined user permission level mapping table, the user confidence factor is 0.95. Overall confidence level. .

[0125] Based on the above comprehensive confidence level, the three initial labels are weighted and fused to obtain the corrected RUL label: Months. The revised RUL label is more aligned with the experience and judgment of senior engineers, effectively correcting for biases caused by the long time elapsed in offline test data and uncertainties caused by noise interference in online extrapolated data.

[0126] Step S500: Based on the fused feature vector and the corrected remaining service life label, update the model parameters of the prediction model, and output the performance degradation prediction result of the power cable through the updated prediction model. The fused feature vector is input into a Long Short-Term Memory (LSTM) network, and the model outputs predicted performance parameters such as dielectric loss tangent, insulation resistance, internal current, and capacitance current. This embodiment does not elaborate on the model training and updating process. For example, the fused feature vector is used as input to the LSTM network, and the corrected remaining service life label is used as a supervision signal. The network weights are iteratively updated through a backpropagation algorithm to minimize the mean square error loss between the predicted output and the corrected remaining service life label until the model converges. After training, the real-time collected fused feature vector is input into the updated prediction model to output the performance degradation prediction result of the power cable.

[0127] Based on the prediction results, a health index is calculated and an early warning is triggered. Specifically, the health index of the power cable is obtained by weighting the ratio of the predicted parameters to the initial parameters in the performance degradation prediction results. The prediction results show that the predicted value of the dielectric loss tangent is 0.38% (initial value is 0.2%, ratio 1.9); the predicted value of the insulation resistance is 280MΩ (initial value is 500MΩ, ratio 1.79); and the predicted value of the internal current is 1.3 times higher than the initial value (ratio 1.3). According to the preset weighting coefficients W1=0.4, W2=0.4, W3=0.2, the health index is calculated as follows: Health Index = (1 / 1.9) × 0.4 + (1 / 1.79) × 0.4 + (1 / 1.3) × 0.2 = 0.21 + 0.22 + 0.15 = 0.58. Because the health index is below the preset threshold of 0.6, the insulation performance of the 35kV power transmission cable is determined to be in a sub-healthy state, posing a high risk of failure, and an orange alert is triggered. The system automatically generates a decision report, which includes: current health index 0.58, predicted remaining service life of approximately 6 months, dielectric loss tangent angle increased by 90% compared to the initial value, insulation resistance value decreased by 44% compared to the initial value, and it is recommended to arrange an insulation performance review within the next 2 months, paying particular attention to abnormal temperature and overload records at cable joints, and prioritizing preventative testing of the cable during the next power outage maintenance.

[0128] In practical applications, the prediction model includes a full model and a simplified model, which are adaptively switched based on the aforementioned monitoring capability assessment results, and adopt a modular shared architecture. The training steps described above are understood as the training steps for the full model, while the simplified prediction model operates under limited monitoring conditions, and its input only receives insulation resistance values ​​and non-time-series environmental parameters. The simplified prediction model shares the environmental feature screening logic and multi-dimensional confidence label correction mechanism of the full parameter prediction model to ensure the consistency of data governance standards; its time-series feature extraction sub-network is correspondingly lightweight, retaining one or two layers of dilated convolutional units, or using a sliding window statistical feature extraction method to generate low-dimensional time-series feature vectors to reduce model complexity and adapt to a single insulation resistance input. The simplified prediction model also uses a long short-term memory network as the prediction sub-network, mainly learning the mapping relationship between the rate of decrease in insulation resistance value and environmental factors, and outputting a simplified estimate of the remaining service life. The health index calculation of the simplified prediction model is based on the ratio of the predicted insulation resistance value to the initial insulation resistance value. When a change in sensor status is detected, the module can be dynamically switched to perform an upgrade or downgrade switch of the model path. During upgrade switching, historically cached insulation resistance data is time-aligned with newly connected dielectric loss, internal current, and other data before being input into the full parameter prediction model. During downgrade switching, the most recently available insulation resistance data is used as input to load a simplified prediction model. Under the transfer learning strategy, the simplified prediction model uses the pre-trained weights of the full parameter prediction model as initial values, freezes the parameters of shared modules, and fine-tunes independent modules to maintain the continuity of prediction results at the time of switching and avoid jumps in prediction values.

[0129] As can be seen from this embodiment, the technical solution of this application achieves the following technical effects in the operation and maintenance of power transmission cables: First, it accurately captures insulation degradation trends. The TCN network effectively identifies the long-term degradation trends of continuously increasing dielectric loss tangent and continuously decreasing insulation resistance through dilated convolution. These trends are highly consistent with the physical mechanisms of typical fault modes such as thermal aging and moisture absorption in power transmission cables, providing high-quality time-series feature input for the prediction model. Second, it scientifically integrates multi-source labels. The multi-dimensional confidence weighted correction mechanism effectively solves the problem of data quality differences among multiple sources in the operation and maintenance of power transmission cables: offline test data is accurate but has poor timeliness, online monitoring data has good timeliness but high noise, and expert experience annotation is comprehensive but subjective. Through quantitative evaluation of three dimensions—time decay, trend stability, and user credibility—the scientific integration of multi-source labels is achieved, and the corrected RUL label more reliably reflects the true aging state of the cable. Third, it provides quantitative evaluation and graded early warning. The calculation model of the health index directly corresponds to the core indicators of the insulation performance of power transmission cables (dielectric loss, insulation resistance, and internal current), making the prediction results clearly interpretable within the industry. The tiered early warning mechanism provides an intuitive and operable basis for operation and maintenance decisions, realizing the transformation from traditional periodic power outage testing to online predictive maintenance.

[0130] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application, such as using other types of temporal feature extraction networks instead of temporal convolutional networks, or using other correlation analysis methods instead of Pearson correlation coefficients, should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0131] This application describes the invention from the perspectives of purpose, performance, progress, and novelty, and it meets the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

Claims

1. A reliability prediction method for power cables, characterized by, include: Obtain the time-series electrical parameters and non-time-series environmental parameters of power cables; The timing electrical parameters are subjected to feature extraction to obtain a timing feature vector, and the non-timing environmental parameters are filtered to obtain a target environmental feature vector; The temporal feature vector is fused with the target environment feature vector to obtain a fused feature vector; Multiple initial remaining service life tags of the power cable are obtained, and the multiple initial remaining service life tags are weighted and fused based on multidimensional confidence to obtain the corrected remaining service life tags. Based on the fused feature vector and the corrected remaining service life label, the model parameters of the prediction model are updated, and the performance degradation prediction result of the power cable is output through the updated prediction model.

2. The reliability prediction method according to claim 1, characterized by, The step of extracting features from the timing electrical parameters to obtain a timing feature vector includes: The time-series electrical parameters are feature-extracted using a time-series convolutional network to obtain the time-series feature vector; The process of filtering the non-time-series environmental parameters to obtain the target environmental feature vector includes: Calculate the correlation coefficient between the non-time-series environmental parameters and the performance indicators of power cables, and use the environmental parameters whose absolute value of the correlation coefficient is greater than a preset threshold as the target environmental feature vector.

3. The reliability prediction method according to claim 2, characterized by, The temporal convolutional network includes multiple layers of residual units, each residual unit including dilated convolutional units and residual connections; the dilation coefficient of the dilated convolutional units increases exponentially with the number of network layers; the residual connections add the input of each residual unit to the convolutional output of that residual unit.

4. The reliability prediction method of claim 1, wherein The multidimensional confidence level includes a time decay factor, a trend stability factor, and a user credibility factor. The time decay factor is determined based on the time difference between the data source time of the initial remaining useful life tag and the time of the failure point. The trend stability factor is determined based on the data trend stability corresponding to the initial remaining useful life label. The user credibility factor is determined based on the user's permission level or historical operation accuracy rate that generated the initial remaining lifespan label.

5. The reliability prediction method according to claim 4, characterized by, The time decay factor is a monotonically decreasing function with respect to the time difference; the user credibility factor is determined based on a preset user permission level mapping table, which includes the correspondence between user permission levels and user credibility factors.

6. The reliability prediction method of claim 1, wherein The method further includes: The health index of the power cable is obtained by weighting the ratio of the predicted parameters to the initial parameters in the performance degradation prediction results; when the health index is lower than a preset threshold, a maintenance warning is triggered.

7. A reliability prediction system for power cables, characterized by include: The data acquisition module is used to acquire the time-series electrical parameters and non-time-series environmental parameters of power cables; The feature processing module is used to extract features from the time-series electrical parameters to obtain a time-series feature vector, and to filter the non-time-series environmental parameters to obtain a target environmental feature vector. The feature fusion module is used to fuse the temporal feature vector with the target environment feature vector to obtain a fused feature vector; The tag correction module is used to obtain multiple initial remaining service life tags of the power cable, and to perform weighted fusion of the multiple initial remaining service life tags based on multidimensional confidence to obtain the corrected remaining service life tags. The prediction module is used to update the model parameters of the prediction model based on the fused feature vector and the corrected remaining service life label, and output the performance degradation prediction result of the power cable through the updated prediction model.

8. The reliability prediction system according to claim 7, characterized in that, The step of extracting features from the timing electrical parameters to obtain a timing feature vector includes: The time-series electrical parameters are feature-extracted using a time-series convolutional network to obtain the time-series feature vector; The process of filtering the non-time-series environmental parameters to obtain the target environmental feature vector includes: Calculate the correlation coefficient between the non-time-series environmental parameters and the performance indicators of power cables, and use the environmental parameters whose absolute value of the correlation coefficient is greater than a preset threshold as the target environmental feature vector.

9. The reliability prediction system according to claim 7, characterized in that, Also includes: The early warning triggering module is used to calculate the health index of the power cable by weighting the ratio of the predicted parameters to the initial parameters in the performance degradation prediction results. When the health index falls below a preset threshold, a maintenance warning is triggered.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the program is executed by the processor, it implements the reliability prediction method as described in any one of claims 1 to 6.

Images