Real-time evaluation system for power cable operating state based on multi-modal deep learning

Abstract

The application relates to the technical field of power equipment state monitoring, and particularly discloses a power cable operation state real-time evaluation system based on multi-modal deep learning, which synchronously collects load, partial discharge and temperature strain data of a cable through a current-voltage sensor, an ultrahigh frequency sensor and a distributed optical fiber sensor; multi-source data are subjected to feature extraction and cross-dimension fusion to construct a multi-dimensional feature matrix; a pre-trained deep learning model is used to analyze the internal correlation between features, and a cable health degree score, a fault risk level and a defect type identification result are output; and a load adjustment, loop switching or accurate maintenance strategy is automatically generated and executed according to the evaluation result; the application solves the problems of inaccurate evaluation and early warning lag caused by isolated data analysis of a traditional monitoring method, and realizes real-time and accurate evaluation and intelligent closed-loop operation and maintenance of the cable operation state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring, and particularly relates to a power cable operation state real-time evaluation system based on multi-modal deep learning. BACKGROUND

[0002] As a key carrier of urban power grid power transmission, the reliability of the operation state of the power cable directly affects the safety and stability of the power supply system. At present, cable state monitoring mainly relies on periodic inspection and single parameter threshold alarm, such as partial discharge monitoring, optical fiber temperature measurement and other technologies have been applied. However, these traditional methods have obvious limitations: first, partial discharge, load current, temperature strain and other parameters are usually collected and analyzed independently, lacking multi-parameter collaborative analysis means, and it is difficult to reveal the real state under the coupling of "electric-thermal-mechanical" multi-physical field; second, the existing evaluation methods are mostly based on simple threshold judgment or shallow machine learning models, and the identification ability of the evolution law of the cable state under complex working conditions is limited, which cannot effectively warn potential failures; in addition, the monitoring and operation and maintenance decision-making are disconnected, and the operation and maintenance personnel are difficult to obtain accurate state evaluation results and disposal suggestions. With the increase of cable service life and the continuous improvement of urban power grid load density, the traditional method has been unable to meet the demand of accurate cable state evaluation and risk warning. SUMMARY

[0003] The purpose of the present application is to provide a power cable operation state real-time evaluation system based on multi-modal deep learning to solve the problems in the above background.

[0004] The purpose of the present application can be realized by the following technical solutions:

[0005] The power cable operation state real-time evaluation system based on multi-modal deep learning comprises:

[0006] A data acquisition module is used for synchronously acquiring multi-source monitoring data in the operation process of the power cable and directly combining the multi-source monitoring data into a multi-modal data set; the multi-source monitoring data comprises a load change sequence, a partial discharge waveform, and a distributed temperature and strain distribution along the length of the cable;

[0007] A feature extraction module is used for feature extraction of the multi-modal data set, comprising: a load fluctuation feature vector is extracted from the load change sequence by using a sliding time window;

[0008] The partial discharge waveform is decomposed by 5 layers based on db4 wavelet basis to separate noise; based on the denoised partial discharge waveform, the threshold value of the mean value plus 3 times the standard deviation of the waveform amplitude is used to identify the effective discharge pulse; in a 1-minute statistical period, the amplitude distribution is counted in logarithmic coordinates, the phase distribution is counted in 18 intervals according to the power frequency period, and the three-dimensional statistical feature map is formed by combining the total number of pulses per minute;

[0009] According to the spatial characteristics of the cable laying path, a multi-level grid division method is adopted to divide the cable into spatial grids with different precisions along the length direction;

[0010] The spatial continuity of the temperature and strain distribution along the cable length in each spatial grid is analyzed, the temperature gradient and strain change rate between adjacent grids are calculated, the temperature abnormal rise area and strain mutation area are identified as the candidate position area of the local thermal mechanical combined fault of the cable, and the feature parameters of the temperature gradient, strain mutation degree and spatial distribution range of the candidate position area are proposed;

[0011] The feature fusion module fuses the extracted multi-modal features across dimensions, establishes the correlation between the load fluctuation features, discharge statistical features and temperature strain features, and generates a unified multi-dimensional feature matrix;

[0012] The state evaluation module inputs the multi-dimensional feature matrix into the pre-trained cable operation state evaluation model, analyzes the internal relationship and evolution law between the multi-modal features, and simultaneously outputs the comprehensive health degree score, fault risk level and defect type recognition result of the cable;

[0013] The intelligent decision module automatically formulates the operation and maintenance strategy according to the cable operation state evaluation result, automatically executes the load adjustment, standby circuit switching or issues the maintenance instruction containing the positioning information when the fault risk level exceeds the set threshold.

[0014] As a further scheme of the application, the generation process of the multi-dimensional feature matrix is:

[0015] The load fluctuation feature vector, three-dimensional statistical feature spectrum and temperature strain feature parameter are dynamically aligned according to the collection time stamp;

[0016] The time correlation strength of the load fluctuation feature and the discharge statistical feature, the spatial corresponding relationship of the discharge statistical feature and the temperature strain feature, and the dynamic response relationship of the temperature strain feature and the load fluctuation feature are calculated to establish a multi-dimensional feature correlation network;

[0017] The load fluctuation feature, the discharge statistical feature and the temperature strain feature are weighted and reorganized according to their correlation strength;

[0018] The reorganized features are arranged according to the time sequence and the spatial distribution of two dimensions to generate a multi-dimensional feature matrix containing the space-time characteristics.

[0019] As a further scheme of the application, the construction process of the multi-dimensional feature correlation network is:

[0020] By comparing load fluctuation characteristics with discharge statistical characteristics within the same time window, and analyzing the synchronicity between load change trends and changes in the frequency and amplitude of discharge pulses, the time correlation strength is calculated.

[0021] The discharge statistical characteristics and temperature strain characteristics are matched in a grid according to the actual location of the measurement points. By analyzing the spatial distribution relationship between the discharge activity location and the temperature candidate location region, the spatial correspondence is calculated.

[0022] By comprehensively analyzing the temporal correlation strength and spatial correspondence, and by tracking the response characteristics of temperature changes to load fluctuations and the influence pattern of load changes on temperature distribution, a dynamic response relationship is established.

[0023] Based on the analysis results of time correlation strength, spatial correspondence and dynamic response relationship, a multi-dimensional feature correlation network reflecting the interaction between load fluctuation characteristics, discharge statistical characteristics and temperature strain characteristics is constructed.

[0024] As a further aspect of the present invention: the comprehensive health score, fault risk level, and defect type identification results of the output cable specifically include:

[0025] The multidimensional feature matrix is ​​input into the cable operation status assessment model constructed based on a deep spatiotemporal feature extraction network. By analyzing the spatiotemporal correlation patterns between load fluctuation features, discharge statistics features and temperature strain features in the feature matrix, key features characterizing the cable operation status are extracted.

[0026] Based on the extracted key features, a multi-task output structure is used to simultaneously calculate the comprehensive health score, determine the fault risk level, and identify the defect type.

[0027] The calculated comprehensive health score, fault risk level, and defect type identification results are used to form the final cable operation status assessment result.

[0028] As a further aspect of the present invention: the automatic formulation of operation and maintenance strategies based on the cable operation status assessment results specifically includes:

[0029] The comprehensive health score, fault risk level, and defect type identification results are logically integrated to form a unified final status conclusion.

[0030] Based on the defect type and risk level in the final status conclusion, generate initial operation and maintenance instructions that include load adjustment target values, backup circuit switching sequence or precise location maintenance information;

[0031] Verify the grid security and logical consistency of the initial operation and maintenance instructions, including checking whether the load adjustment range exceeds the limit, whether the standby circuit status is ready, and whether the maintenance location is valid;

[0032] Execute verified operation and maintenance commands, and calculate the strategy execution effect based on the re-collected multimodal data after execution. If the effect is lower than the set standard, automatically adjust the command parameters and re-verify the execution.

[0033] The beneficial effects of this invention are:

[0034] (1) By fusing three types of heterogeneous data—load, partial discharge, and temperature strain—and constructing a multi-dimensional feature correlation network, this method solves the technical problem that traditional single-parameter monitoring cannot reflect the coupling effect of multiple physical fields such as "electricity-heat-mechanics." This method can identify weak precursor features of typical defects such as insulation aging, partial discharge, and mechanical damage from complex data, upgrading fault identification from "threshold alarm" to "trend prediction," and reducing the risk of sudden faults.

[0035] (2) Based on the state assessment results of deep neural networks, the system can automatically generate and dynamically optimize operation and maintenance strategies, and realize intelligent decision-making and execution of operations such as load adjustment, circuit switching, and precise maintenance. This solution transforms the traditional passive maintenance that relies on human experience into data-driven proactive operation and maintenance, and at the same time shortens the response time when a fault occurs, thereby improving the reliability and efficiency of power grid operation and maintenance. Attached Figure Description

[0036] The invention will now be further described with reference to the accompanying drawings.

[0037] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figure 1 As shown, this invention is a real-time evaluation system for the operating status of power cables based on multimodal deep learning, comprising:

[0040] The data acquisition module is used to synchronously acquire multi-source monitoring data during the operation of power cables and directly combine them into a multi-modal dataset; the multi-source monitoring data includes load change sequences, partial discharge waveforms, and distributed temperature and strain distributions along the cable length;

[0041] The feature extraction module is used to extract features from multimodal datasets, including: extracting load fluctuation features from load change sequences, separating discharge pulses from partial discharge waveforms and calculating discharge statistical features, and identifying temperature and strain characteristic parameters of abnormally rising temperature regions and abruptly changing strain regions from temperature and strain distributions.

[0042] The feature fusion module performs cross-dimensional fusion of the extracted multimodal features, establishes the correlation between load fluctuation features, discharge statistical features and temperature strain features, and generates a unified multidimensional feature matrix.

[0043] The condition assessment module inputs the multi-dimensional feature matrix into the pre-trained cable operation condition assessment model. By analyzing the intrinsic relationship and evolution law between multi-modal features, it outputs the cable's comprehensive health score, fault risk level, and defect type identification results.

[0044] The intelligent decision-making module automatically formulates operation and maintenance strategies based on the cable operation status assessment results. When the fault risk level is identified to exceed the set threshold, it automatically performs load adjustment, backup circuit switching, or issues maintenance instructions containing precise location information.

[0045] In the data acquisition module, the data acquisition process of this invention simultaneously acquires multi-source monitoring data during the operation of the power cable using three types of sensors. The load change sequence of the power cable is acquired in real time through a current and voltage sensor. This sensor is directly connected to the conductive core of the cable and measures the current flowing through the cable and the cable-to-ground voltage, forming a continuously changing load data sequence over time.

[0046] Partial discharge waveforms are captured using an ultra-high frequency sensor. This sensor, employing the coupling capacitor principle, is installed at cable joints or on the cable itself to detect electromagnetic wave signals generated by cable insulation defects. The sensor converts the captured electromagnetic wave signals into time-domain waveform data, fully recording the waveform characteristics of each discharge pulse, including pulse rise time, pulse width, and amplitude information.

[0047] Temperature and strain distribution along the cable length are measured using distributed optical fiber sensors. These sensors embed the sensing fiber directly into the cable sheath or lay it close to the cable surface. Using optical time-domain reflectometry (OTDR), the temperature measurements at different locations along the entire cable are obtained by analyzing the intensity changes of backscattered light in the fiber. Simultaneously, the mechanical strain measurements at various points on the cable are obtained by analyzing the frequency shift of the backscattered light, thus forming temperature and strain field data distributed along the cable length.

[0048] In the feature extraction module, when extracting features from the multimodal dataset, the load change sequence is processed first. A sliding time window method is used to divide the load sequence into continuous time periods. The window length is dynamically adjusted according to the load change characteristics, and automatically shortened when the load change rate exceeds a set threshold. Within each time window, the short-term fluctuation amplitude and long-term trend of the load data are calculated. The short-term fluctuation amplitude is obtained by calculating the maximum absolute value of the load difference between adjacent sampling points within the window; the long-term trend is obtained by calculating the ratio of the load difference between the start and end times of the window to the time span of that window.

[0049] When extracting features from partial discharge waveforms, multi-scale noise separation is first performed. Using wavelet transform, the discharge waveform is decomposed into 5 levels using the db4 wavelet basis. The standard deviation of the detail coefficients at each level is calculated, and coefficients with a standard deviation less than a set threshold are set to zero. For the denoised waveform, based on the statistical distribution of waveform amplitude, the mean and standard deviation of the waveform data are calculated, and the threshold is set to the mean plus three times the standard deviation. Valid discharge pulses are identified using the mean amplitude plus three times the standard deviation as the threshold, and the peak value, timestamp, and power frequency phase of the valid discharge pulses are recorded. Within each 1-minute statistical period, pulse features are integrated as follows: Amplitude distribution: The pulse peak value is divided into 10 equally spaced intervals on a logarithmic coordinate system, and the number of pulses in each interval is counted. Repetition frequency: The total number of pulses within that minute is counted as the discharge frequency index. Phase distribution: The power frequency period is divided into 18 phase intervals (each interval 20°), and the number of pulses distributed in each interval is counted. Finally, a 10 (amplitude) × 1 (frequency) × 18 (phase) three-dimensional statistical feature map is formed.

[0050] When extracting features from temperature and strain distribution data, the first step is spatial mesh generation. Based on the cable laying path, the cable is divided into basic mesh units of 1 meter length. For key areas such as cable joints and bends, the mesh is further refined to finer meshes of 0.2 meters length. For each mesh unit, the temperature gradient between it and its adjacent meshes is calculated using the following formula: ;in, Represents the temperature gradient. This indicates the current grid temperature value. Indicates the temperature values ​​of adjacent grid cells. This represents the grid spacing. The degree of strain abrupt change is quantified by calculating the standard deviation between the strain value of a grid cell and the strain values ​​of its adjacent cells. When identifying candidate location regions, preset temperature gradient thresholds and strain abrupt change thresholds are used; grid cells exceeding these thresholds are marked as candidate location regions. The feature parameters extracted for candidate location regions include: the maximum temperature gradient value within the region, the average strain abrupt change, the continuous length of the candidate location region, and the distance from the center of the candidate location region to the cable starting point. When integrating the extracted features, time alignment is performed first. Using whole minutes as the time base, feature data from different timestamps are unified to the same time point using linear interpolation. Next, feature standardization is performed, calculating the mean and standard deviation for each feature dimension using the z-score standardization method. ;in These are the original eigenvalues. This is the mean of this feature dimension. Standard deviation, This represents the standardized feature values. The standardized feature values ​​form a unified multimodal feature set, which includes 3 dimensions of the load fluctuation feature vector, 10×1×18 dimensions of the three-dimensional statistical feature map, and 4 dimensions of the temperature strain feature, for a total of 187 feature dimensions.

[0051] In the generation of the load fluctuation feature vector, the length of the sliding time window is optimized based on the load change characteristics. For the load data within each time window, short-term fluctuation amplitude and long-term trend are calculated separately. The short-term fluctuation amplitude is obtained by calculating the 95th quantile of the load difference between adjacent sampling points within the window; this method effectively eliminates interference from abrupt changes. The long-term trend is calculated using linear regression to obtain the slope of load change over time. The calculation results for the current window are compared with historical data from the same period over the past 30 days. The Mahalanobis distance between the current feature value and the historical mean is calculated, ultimately generating a load fluctuation feature vector with time-series correlation characteristics.

[0052] The construction of the three-dimensional statistical feature map includes the following detailed steps: First, multi-scale noise separation is performed on the partial discharge waveform using a wavelet threshold denoising method, with the threshold ratio of the detail coefficients at each layer set to 0.2. For the denoised waveform, the discrimination threshold is automatically adjusted based on the statistical distribution characteristics of the waveform amplitude, with the threshold set to the mean of the waveform data plus four times the standard deviation. After identifying valid discharge pulses, the amplitude distribution, repetition frequency, and phase characteristics of the discharge pulses are statistically analyzed within a 1-minute statistical period. The amplitude distribution is divided into 10 intervals using logarithmic coordinates, the repetition frequency is calculated based on the total number of pulses per minute, and the phase characteristics are divided into 18 intervals based on the power frequency period. Finally, these statistical features are integrated according to a three-dimensional coordinate system of amplitude-frequency-phase to form a complete three-dimensional statistical feature map.

[0053] During feature standardization, appropriate standardization methods are employed for different types of features. For load fluctuation features, a minimum-maximum standardization method is used to scale the feature values ​​to the [0,1] interval. For three-dimensional statistical feature maps, logarithmic transformation followed by z-score standardization is used to eliminate the influence of dimensions. For temperature-strain features, physical meaning-based standardization methods are employed, such as dividing the temperature gradient by the maximum allowable value and dividing the degree of strain abrupt change by the material yield strength. Through unified standardization, the comparability of feature values ​​with different physical meanings is ensured, laying the foundation for subsequent feature fusion and analysis.

[0054] In the feature fusion module, dynamic alignment of multimodal features is performed first. The load fluctuation feature vector, 3D statistical feature map, and temperature strain feature parameters are precisely matched according to the acquisition timestamps. For data with inconsistent timestamps, cubic spline interpolation is used to unify the data to the same time base. During interpolation, a 1-minute time unit is used to ensure that all feature data have the same temporal resolution. The load fluctuation feature vector contains 3 feature dimensions; the 3D statistical feature map contains 180 feature dimensions; and the temperature strain feature parameters contain 4 feature dimensions, for a total of 187 feature dimensions participating in subsequent fusion processing.

[0055] When establishing a multi-dimensional feature association network, the temporal correlation strength between load fluctuation characteristics and discharge statistical characteristics is first calculated. The Pearson correlation coefficient method is used, and the formula is: ;in Indicates the first Load fluctuation characteristic values ​​at each time point This represents the corresponding discharge statistical characteristic value. and They represent the characteristic means, This represents the length of the time series. The calculated time correlation strength value ranges from -1 to 1, with a larger absolute value indicating a stronger correlation.

[0056] Calculate the spatial correspondence between discharge statistical characteristics and temperature strain characteristics. Establish a mapping relationship between cable length coordinates and measurement points based on the actual sensor placement locations. For each discharge statistical characteristic measurement point, find the nearest temperature strain characteristic measurement point. The formula for calculating the spatial correspondence is: ;in Indicates the first The discharge measurement point and the first Spatial distance between temperature measurement points Indicates spatial correspondence. This is the distance attenuation coefficient. This represents an exponential function with base e, set to 10 meters for the cable length. The closer this value is to 1, the stronger the spatial correspondence.

[0057] Then, a dynamic response relationship between temperature statistical characteristics and load fluctuation characteristics is established. By analyzing the response characteristics of temperature statistical characteristics after load changes, the dynamic response coefficient is calculated. A time-delay cross-correlation analysis method is used to find the maximum correlation between load fluctuations and temperature changes and its corresponding time delay. The formula for calculating the dynamic response coefficient is: ;

[0058] in Indicates the first Temperature statistical characteristics in time The value, Indicates time arrive The average value of the temperature statistical characteristics within. Indicates the first Individual load fluctuation characteristics in time The value, Indicates the characteristics of load fluctuation over time arrive The average value within, This represents the time delay parameter, with a value ranging from 0 to 30 minutes. This represents the dynamic response coefficient.

[0059] Based on the analysis results of the three types of associations mentioned above, a complete multi-dimensional feature association network was constructed. Network nodes represent various feature dimensions, and edge weights are determined by the association strength. For temporal association strength, a threshold of 0.6 was set, retaining associations with temporal association strength values ​​greater than this threshold; for spatial correspondence, a threshold of 0.7 was set, retaining associations with temporal association strength values ​​greater than this threshold; and for dynamic response relationships, a threshold of 0.5 was set, retaining associations with temporal association strength values ​​greater than this threshold. Through association analysis in these three dimensions, a feature association network containing 187 nodes and possessing clear physical meaning was established.

[0060] When performing feature-weighted reorganization, the weight allocation is determined based on the degree centrality of each node in the feature association network, specifically including:

[0061] In the constructed feature association network, each feature node represents a feature dimension, and the weight of each edge represents the association strength between two features. For a network containing 187 feature nodes, the feature nodes... The weighted degree centrality is obtained by summing the weights of all edges connected to it. To prevent the weights from being too large or too small and affecting the recombination effect, and to ensure that the sum of the weights of all features is 1, the weighted degree centrality needs to be normalized to obtain the final weight of each feature dimension. : ;in Indicates the first The degree value of a feature node in the network, i.e., the number of edges connected to that node. Multiplying the original feature values ​​by their corresponding final weights generates a weighted new feature representation; this process is called weighted reorganization. For features with weaker associations, their weights are appropriately reduced; for features with stronger associations, their weights are increased accordingly, thus differentiating the importance of features.

[0062] Finally, when generating the multidimensional feature matrix, the weighted and recombined features are arranged according to two dimensions: time series and spatial distribution. The time dimension is arranged in the order of acquisition time to retain complete temporal evolution information; the spatial dimension is arranged in the order of the actual location of the cable measurement points to maintain spatial continuity. The final generated multidimensional feature matrix has a dimension of number of time points × 187 (for example, if T evaluation cycles are monitored, T rows of data will be generated, so the number of time points equals T, and the multidimensional feature matrix dimension is T rows × 187 columns), where 187 includes 3 load fluctuation features, 180 discharge statistical features, and 4 temperature strain features. This matrix contains both the dynamic information of each feature changing over time and retains the spatial correlation characteristics between features, providing a complete feature representation for subsequent condition assessment.

[0063] During feature reorganization, special attention is paid to the synergistic relationship between different feature dimensions. For feature dimensions with strong positive correlation, the consistency of their changing trends is maintained during reorganization; for feature dimensions with negative correlation, their complementary characteristics are highlighted during reorganization. Through this refined reorganization method, the generated multidimensional feature matrix can better reflect the overall characteristics of the cable's operating status.

[0064] In the condition assessment module, the multidimensional feature matrix is ​​input into a pre-trained cable operation condition assessment model for condition analysis. This model employs a deep neural network structure, comprising an input layer, three hidden layers, and an output layer. The input layer has 187 nodes, consistent with the dimension of the multidimensional feature matrix. The first hidden layer contains 128 nodes, using the modified linear unit activation function; the second hidden layer contains 64 nodes, using the hyperbolic tangent activation function; and the third hidden layer contains 32 nodes, using the sigmoid activation function. The output layer contains 6 nodes, corresponding to the health score, fault risk level, and the identification results of four main defect types. The model parameters are trained using historical data, employing an adaptive moment estimation algorithm with a learning rate of 0.001 and 500 training cycles.

[0065] A weighted comprehensive evaluation method is used to calculate the overall cable health score. First, each feature dimension in the multidimensional feature matrix is ​​normalized so that its value ranges from 0 to 1. Then, weights are assigned based on the degree of influence of each feature on the cable's health status. These weights are determined through expert experience and historical data analysis. The total weight for load-related features is 0.3, for discharge-related features it is 0.4, and for temperature and strain-related features it is 0.3. The health score is calculated as the weighted sum of each feature value multiplied by its corresponding weight. The final score ranges from 0 to 100, with a higher score indicating a better cable health status.

[0066] When determining the fault risk level, a risk assessment system based on multi-dimensional characteristics is established. Risk levels are divided into three categories: low risk, medium risk, and high risk. Low risk corresponds to a health score greater than 80, with all key characteristic indicators within the normal range; medium risk corresponds to a health score between 60 and 80, or one key characteristic indicator showing an anomaly; high risk corresponds to a health score below 60, or two or more key characteristic indicators showing anomalies simultaneously. Key characteristic indicators include maximum discharge amplitude, maximum temperature gradient, and degree of strain mutation, etc. The abnormal thresholds for these indicators are dynamically adjusted according to cable specifications and operating environment.

[0067] When identifying defect types, a feature pattern library of four typical defects is established. The first type is insulation aging defects, characterized by discharge pulse phase distribution concentrated near the positive and negative peaks of the power frequency cycle, accompanied by a uniform temperature increase. The second type is partial discharge defects, characterized by high discharge pulse repetition rate, dispersed amplitude distribution, and weak correlation with load changes. The third type is mechanical damage defects, characterized by significant abrupt strain changes, localized hot spots in temperature distribution, and indistinct discharge characteristics. The fourth type is connection defects, characterized by obvious hot spots in temperature distribution and high synchronization between discharge pulses and load changes. By calculating the similarity between the current feature and each defect pattern, the nearest neighbor classification method is used to determine the defect type.

[0068] During the condition assessment process, a verification mechanism for the assessment results should be established. The reliability of the assessment results is verified by analyzing the consistency among different characteristic indicators. For example, when the health score indicates high risk, it is necessary to check whether discharge characteristics, temperature characteristics, and strain characteristics simultaneously show abnormalities; when the defect type is identified as insulation aging, it is necessary to verify whether the discharge phase distribution and temperature changes conform to the typical characteristics of insulation aging. If there are contradictions among different characteristic indicators, the feature extraction and fusion process needs to be re-examined to ensure the accuracy of the assessment results.

[0069] The cable operation status assessment model is updated online based on the latest monitoring data. Each time a new monitoring data sample is obtained, the model parameters are fine-tuned using that sample, with the update magnitude controlled by the learning rate. During the model update process, the statistical characteristics of historical data are retained while adapting to the latest changes in cable condition. This online update mechanism enables the cable operation status assessment model to track the evolution trend of cable condition, promptly reflect subtle changes in cable health, and ensure the timeliness and accuracy of the status assessment results.

[0070] A confidence evaluation method for state assessment results was established. The confidence evaluation is based on three aspects: the completeness of feature data, the consistency between features, and the stability of model output. The completeness of feature data has a weight of 0.4, obtained by calculating the proportion of valid feature data to the total number of features; the consistency between features has a weight of 0.3, obtained by analyzing the correlation between different feature indicators; and the stability of model output has a weight of 0.3, obtained by statistically analyzing the degree of change in recent assessment results. The final confidence score ranges from 0 to 1. When the confidence score is below 0.7, the assessment result requires manual review.

[0071] In the intelligent decision-making module, after obtaining the comprehensive health score (denoted as H, ranging from 0-100), fault risk level (L, with values ​​of low, medium, and high), and defect type identification result (D, such as insulation aging, partial discharge, mechanical damage, and connection defects) output by the status assessment module, logical verification is first performed to form the final assessment result. The verification rules include: if L is "high risk," then H < 60 and at least two key characteristics (such as discharge amplitude and temperature gradient) must simultaneously exceed limits; if D is identified as "insulation aging," then it needs to be verified that the discharge phase is concentrated within ±30 degrees of the power frequency peak and the temperature shows a uniform upward trend. After all rules are passed, H, L, D, and their corresponding confidence levels are encapsulated into the final cable operating status assessment result R.

[0072] The process of automatically formulating operation and maintenance strategies based on the assessment result R is as follows. First, it is determined whether the fault risk level L exceeds the set threshold (the threshold is set as medium risk, i.e., L is triggered when it is "medium" or "high"). If it does not exceed the threshold, monitoring continues; if it does exceed the threshold, the strategy generation stage begins. Strategy generation is based on the defect type D and the current load conditions: For "insulation aging" defects, a strategy centered on load adjustment is generated, with the adjustment target value being 80% of the current rated load; for "partial discharge" defects, a strategy centered on switching to the backup circuit is generated, and the switching action is set to a no-current-carrying moment; for "mechanical damage" or "connection" defects, an immediate repair strategy is generated, along with location information calculated from the center location of the abnormal temperature strain characteristic area (e.g., 35.2 meters from the cable starting point).

[0073] Before executing the strategy, logical consistency verification is performed. Verification includes: the load adjustment strategy must ensure the target load is not lower than the lower limit for safe grid operation (e.g., the decrease does not exceed 30% of the current load); the circuit switching strategy must verify whether the standby circuit is currently idle; and the location information of the maintenance command must be within the total cable laying length. After successful verification, the strategy is issued for execution.

[0074] During implementation, establish a strategy effectiveness index. The calculation formula is as follows: ;in, and These represent the health scores before and after the strategy implementation. The rate of decrease of key anomalous features (such as maximum discharge amplitude), This is a preset proportional coefficient (with a value of 0 or 1) to determine whether load or voltage fluctuations tend to stabilize. The effect evaluation cycle is 5 minutes. If the effect index E is below 0.6 for two consecutive cycles, the strategy is deemed ineffective, and the parameter adjustment program is automatically initiated, such as further reducing the load adjustment target value by 5% or rescheduling the maintenance time window. This process forms a closed loop of "evaluation-decision-execution-feedback" to ensure continuous adaptation to dynamic changes in cable conditions.

[0075] The working principle of this invention is as follows: Load time-series data, partial discharge waveforms, and temperature-strain distribution data of the cable are simultaneously acquired using current and voltage sensors, ultra-high frequency sensors, and distributed optical fiber sensors. Feature extraction is performed on the multi-source monitoring data, including analyzing load fluctuation characteristics using a sliding time window, extracting statistical features of discharge pulses through multi-scale noise separation and dynamic threshold comparison, and identifying temperature and strain candidate location region features based on spatial grid partitioning. The extracted multi-modal features are then fused across dimensions, and a feature association network is constructed by calculating time correlation strength, spatial correspondence, and dynamic response relationships to generate a unified multi-dimensional feature matrix. This matrix is ​​input into a cable operation status assessment model, which outputs cable health scores, fault risk levels, and defect type identification results. Finally, based on the assessment results, operation and maintenance strategies are automatically generated and executed, including load adjustment, backup circuit switching, or precise location maintenance, forming a closed-loop intelligent operation and maintenance system from status perception to decision execution.

[0076] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A real-time evaluation system for the operating status of power cables based on multimodal deep learning, characterized in that, include: The data acquisition module is used to synchronously acquire multi-source monitoring data during the operation of power cables and directly combine them into a multi-modal dataset; the multi-source monitoring data includes load change sequences, partial discharge waveforms, and distributed temperature and strain distributions along the cable length; The feature extraction module is used to extract features from multimodal datasets, including: extracting load fluctuation feature vectors from load change sequences using a sliding time window; The partial discharge waveform is decomposed into five levels using the db4 wavelet basis to separate noise. Based on the denoised partial discharge waveform, the effective discharge pulse is identified by using the mean waveform amplitude plus three times the standard deviation as the threshold. Within a 1-minute statistical period, the amplitude distribution is statistically analyzed using logarithmic coordinates, and the phase distribution is statistically analyzed by dividing the power frequency period into 18 equal intervals. Combined with the total number of pulses per minute, a three-dimensional statistical feature map is formed. Based on the spatial characteristics of the cable laying path, a multi-level grid division method is adopted to divide the cable into spatial grids of different precision along the length direction; Spatial continuity analysis is performed on the distributed temperature and strain distribution along the cable length within each spatial grid. By calculating the temperature gradient and strain abrupt change between adjacent grids, regions with abnormal temperature rise and strain abrupt change are identified as candidate locations for local thermomechanical combined faults in the cable. Temperature and strain characteristic parameters of the candidate locations are extracted, including: the maximum temperature gradient value within the region, the average strain abrupt change, the continuous length of the candidate location region, and the distance from the center of the candidate location region to the cable starting point. The feature fusion module is used to establish the temporal correlation strength between load fluctuation characteristics and discharge statistical characteristics, the spatial correspondence between discharge statistical characteristics and temperature strain characteristics, and the dynamic response relationship between temperature strain characteristics and load fluctuation characteristics. It constructs a multi-dimensional feature association network and generates a unified multi-dimensional feature matrix, specifically including: By comparing load fluctuation characteristics with discharge statistical characteristics within the same time window, and analyzing the synchronicity between load change trends and changes in the frequency and amplitude of discharge pulses, the time correlation strength is calculated. The discharge statistical characteristics and temperature strain characteristics are matched in a grid according to the actual location of the measurement points. By analyzing the spatial distribution relationship between the discharge activity location and the temperature candidate location region, the spatial correspondence is calculated. By comprehensively analyzing the temporal correlation strength and spatial correspondence, and by tracking the response characteristics of temperature changes to load fluctuations and the influence pattern of load changes on temperature distribution, a dynamic response relationship is established. The condition assessment module inputs the multi-dimensional feature matrix into the pre-trained cable operation condition assessment model. By analyzing the intrinsic relationship and evolution law between multi-modal features, it outputs the cable's comprehensive health score, fault risk level, and defect type identification results. The intelligent decision-making module automatically formulates operation and maintenance strategies based on the cable operation status assessment results. When the fault risk level is detected to exceed the set threshold, it automatically performs load adjustment, backup circuit switching, or issues maintenance instructions containing location information.

2. The real-time evaluation system for the operating status of power cables based on multimodal deep learning according to claim 1, characterized in that, The process of generating the multidimensional feature matrix is ​​as follows: The load fluctuation feature vector, three-dimensional statistical feature map, and temperature strain feature parameters are dynamically aligned according to the acquisition timestamp; Based on the analysis results of time correlation strength, spatial correspondence and dynamic response relationship, a multi-dimensional feature correlation network reflecting the interaction between load fluctuation characteristics, discharge statistical characteristics and temperature strain characteristics is constructed. The load fluctuation characteristics, discharge statistical characteristics, and temperature strain characteristics are weighted and recombined according to their correlation strength; The recombined features are arranged according to two dimensions: time series and spatial distribution, to generate a multidimensional feature matrix containing spatiotemporal characteristics.

3. The real-time evaluation system for the operating status of power cables based on multimodal deep learning according to claim 1, characterized in that, The comprehensive health score, fault risk level, and defect type identification results of the output cable specifically include: The multidimensional feature matrix is ​​input into the cable operation status assessment model constructed based on a deep spatiotemporal feature extraction network. By analyzing the spatiotemporal correlation patterns between load fluctuation features, discharge statistics features and temperature strain features in the feature matrix, key features characterizing the cable operation status are extracted. Based on the extracted key features, a multi-task output structure is used to simultaneously calculate the comprehensive health score, determine the fault risk level, and identify the defect type. The calculated comprehensive health score, fault risk level, and defect type identification results are used to form the final cable operation status assessment result.

4. The real-time evaluation system for the operating status of power cables based on multimodal deep learning according to claim 1, characterized in that, The automatic formulation of operation and maintenance strategies based on cable operating status assessment results specifically includes: The comprehensive health score, fault risk level, and defect type identification results are logically integrated to form a unified final status conclusion. Based on the defect type and risk level in the final status conclusion, generate initial operation and maintenance instructions that include load adjustment target values, backup circuit switching sequence or precise location maintenance information; Verify the grid security and logical consistency of the initial operation and maintenance instructions, including checking whether the load adjustment range exceeds the limit, whether the standby circuit status is ready, and whether the maintenance location is valid; Execute verified operation and maintenance commands, and calculate the strategy execution effect based on the re-collected multimodal data after execution. If the effect is lower than the set standard, automatically adjust the command parameters and re-verify the execution.

Images