A cable state fault prediction method based on multi-source data fusion

Abstract

The application relates to the technical field of power equipment state monitoring, and discloses a cable state fault prediction method based on multi-source data fusion. The method constructs a digital twin model by acquiring multi-source data, and performs real-time simulation calculation on a theoretical health baseline changing with working conditions. The baseline is compared with measured data to generate a thermoelectric coupling matrix of quantitative deviation, and a deterioration state is diagnosed according to the thermoelectric coupling matrix. When the deviation exceeds a threshold value, a gradient boosting decision tree model is applied for attribution analysis, and a visual spectrum diagram associated with partial discharge and a root cause is generated. Finally, a defect type is determined and a high-risk state is predicted by intelligently identifying a high-risk visual mode on the spectrum diagram. The application filters out working condition interference through a dynamic benchmark, performs causal diagnosis, solves the problem of high false alarm rate in the prior art, and significantly improves prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, specifically to a cable condition fault prediction method based on multi-source data fusion. Background Technology

[0002] Power cables are the core arteries of power transmission and distribution networks, and their operational reliability is directly related to power grid security and socio-economic stability. To avoid huge losses caused by sudden failures, shifting cable maintenance from passive repair to proactive predictive health management has become an inevitable trend in the industry.

[0003] Existing cable fault prediction methods largely rely on monitoring single physical quantities, such as analyzing only partial discharge or temperature data. These methods typically employ static threshold alarms, ignoring the dynamic impact of load fluctuations and environmental changes on cable condition. Because they fail to effectively integrate multi-source information and establish dynamic health assessment benchmarks, they struggle to distinguish between weak precursor signals caused by insulation degradation and signal fluctuations caused by changes in normal operating conditions. Therefore, they suffer from low prediction accuracy and high false alarm and false negative rates.

[0004] Therefore, in complex operating conditions and changing environments, it is crucial to deeply integrate multi-source data to establish a dynamic health assessment model, thereby accurately and early identifying the true signs of insulation degradation in power cables and overcoming the shortcomings of inaccurate predictions in existing technologies.

[0005] To address this, a cable condition fault prediction method based on multi-source data fusion is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a cable condition fault prediction method based on multi-source data fusion. By constructing a digital twin model that is synchronized with the physical cable in real time, a dynamic health status benchmark is established, and a gradient boosting decision tree model is combined to solve the problems of low prediction accuracy and high false alarm rate caused by the lack of deep fusion and dynamic analysis capabilities of multi-source information in existing technologies.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A cable condition fault prediction method based on multi-source data fusion includes:

[0009] Real-time acquisition of load current, environmental data, electrical data, and temperature measurement data of power cables; construction of a digital twin model based on the multi-physics properties of the cable;

[0010] Based on the digital twin model, load current and environmental data are input, and the theoretical health baseline of the cable under the current operating conditions is calculated through multiphysics simulation. The theoretical health baseline includes theoretical temperature distribution and theoretical electrical signal level. The actual electrical data and temperature data are compared with the theoretical health baseline to generate a thermoelectric coupling matrix that quantifies the difference between the two.

[0011] The insulation degradation status of the cable is diagnosed based on the thermoelectric coupling matrix. When the value of the thermoelectric coupling matrix exceeds the threshold set according to the statistical analysis of the cable's historical data, a gradient boosting decision tree model is applied to quantify causal attribution, obtain the contribution degree, and generate a partial discharge spectrum that correlates the partial discharge signal with the root cause.

[0012] Based on the partial discharge spectrum, unique visual patterns generated by different defects on the partial discharge spectrum are identified. When a high-risk pattern of irreversible damage is identified, the type of insulation defect is determined; and it is predicted whether the cable is in a high-risk fault state.

[0013] Preferably, the steps for constructing the digital twin model specifically include:

[0014] A geometric model of the cable is established, defining its physical dimensions, structural layers, and laying path. A physical model of the cable is also established, integrating material property parameters of the conductor, insulation layer, sheath, and surrounding environmental medium, including thermal conductivity, thermal resistivity, dielectric constant, and electrical conductivity. A behavioral model of the cable is established, embedding multiphysics coupling equations describing the internal heat conduction and electromagnetic field distribution of the cable. A synchronization link is established between the three models and real-time data, linking the constructed geometric model, physical model, and behavioral model with the real-time acquired operational data to form a dynamic mirror reflecting the physical state of the cable.

[0015] Preferably, the step of calculating the theoretical health baseline specifically includes:

[0016] The real-time acquired load current is used as the internal heat source input for the heat conduction equation in the digital twin model, and the environmental data is used as the external boundary condition for the heat conduction equation. The heat conduction equation is solved by finite element analysis and / or equivalent thermal circuit method to obtain the theoretical temperature distribution of the cable under the current operating conditions.

[0017] Based on the ideal insulation dielectric parameters and electromagnetic field model defined in the digital twin model, and combined with the correction of material parameters by the theoretical temperature distribution, the reference electrical signal response that a healthy cable should have under this operating condition is calculated, which is taken as the theoretical electrical signal level.

[0018] Preferably, the step of generating the thermoelectric coupling matrix specifically includes:

[0019] The cable is divided into multiple discrete regions along its length in a digital twin model. For each discrete region, the actual measured temperature and electrical signal values, as well as the corresponding theoretical temperature and electrical signal values ​​given by the theoretical health baseline, are extracted. Within each discrete region, the thermodynamic and electrical deviations between the actual measured values ​​and the theoretical values ​​are calculated independently. The thermodynamic and electrical deviations calculated for all discrete regions are arranged in order of their physical location on the cable to construct a two-dimensional matrix, where one dimension of the matrix corresponds to the discrete regions of the cable, and the other dimension corresponds to different physical quantity deviations, forming the thermoelectric coupling matrix.

[0020] Preferably, the step of diagnosing the insulation degradation state of the cable based on the thermoelectric coupling matrix specifically includes:

[0021] Spatiotemporal pattern analysis is performed on the deviation values ​​in the thermoelectric coupling matrix, including identifying the amplitude of the deviation values, the spatial distribution characteristics along the cable length, and the evolution trend over time; by identifying whether there are spatially concentrated deviation clusters in the thermoelectric coupling matrix that continuously exceed the normal fluctuation range, it is determined whether there are local insulation defects.

[0022] By analyzing the overall statistical mean of the deviation values ​​in the matrix and its rate of increase over time, the general aging degree of the cable is assessed. Combining the judgment results of the local insulation defects with the assessment results of the general aging degree, the specific insulation deterioration state of the cable is finally determined.

[0023] Preferably, the step of generating the partial discharge spectrum specifically includes: extracting high-frequency partial discharge pulse signals from the electrical data, and obtaining the amplitude of each pulse and its precise occurrence time; synchronizing the occurrence time of each pulse with the phase of the power frequency AC voltage, and generating a two-dimensional coordinate data point of phase-amplitude for each pulse signal;

[0024] Using the gradient boosting decision tree model, the root cause of the deviation obtained from the thermoelectric coupling matrix analysis is associated with the two-dimensional coordinate data points of phase and amplitude; all associated two-dimensional coordinate data points are projected onto a two-dimensional statistical spectrum with phase and amplitude as coordinate axes, and the data points are visualized and distinguished according to the root cause, thereby generating the partial discharge spectrum.

[0025] Preferably, the step of identifying unique visual patterns generated by different defects on the partial discharge spectrum specifically includes: using a pre-trained convolutional neural network model to learn from a dataset containing n known insulation defect types and their corresponding standard partial discharge spectrum samples, thereby gaining the ability to identify visual patterns of different defects; inputting the real-time generated partial discharge spectrum as an image into the convolutional neural network model; and finally outputting a classification result by performing feature extraction and pattern matching on the input partial discharge spectrum, wherein the classification result is the identification of the unique visual pattern on the current partial discharge spectrum.

[0026] Preferably, the high-risk modes include electrical tree discharge mode, surface discharge mode, and internal air gap discharge mode.

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

[0028] 1. This invention establishes a health benchmark that dynamically changes with operating conditions by constructing a digital twin model and generating a thermoelectric coupling matrix. This method can effectively filter out false anomalies caused by normal load and environmental fluctuations, highlighting only the true deviations caused by insulation degradation. This significantly improves the signal-to-noise ratio for identifying early fault precursors, greatly reduces the false alarm rate and false negative rate of traditional methods, and improves the accuracy of prediction.

[0029] 2. This invention utilizes a gradient boosting decision tree model to generate causal attribution partial discharge spectra, associating abstract electrical signals with specific physical root causes. This method overcomes the shortcomings of traditional "black box" models where prediction results are difficult to understand, making the diagnostic process transparent and traceable. This greatly enhances the trust of maintenance personnel in the prediction results and provides a clear and reliable basis for subsequent maintenance decisions.

[0030] 3. This invention, through intelligent recognition of visual patterns in partial discharge spectra, can accurately determine specific defect types, such as distinguishing high-risk patterns like "electrical trees." This method not only predicts whether a fault will occur, but also clarifies the urgency and potential causes of the fault, achieving a leap from abnormal state detection to specific fault mode diagnosis, making predictions more targeted and instructive. Attached Figure Description

[0031] Figure 1 This is a flowchart of a cable condition fault prediction method based on multi-source data fusion proposed in this invention.

[0032] Figure 2 This is a schematic diagram illustrating the steps of a cable condition fault prediction method based on multi-source data fusion proposed in this invention.

[0033] Figure 3 This is a structural diagram of the digital twin model proposed in this invention. Detailed Implementation

[0034] 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.

[0035] Please see Figures 1 to 2 This invention provides a cable condition fault prediction method based on multi-source data fusion, the technical solution of which is as follows:

[0036] A cable condition fault prediction method based on multi-source data fusion, such as Figures 1-2 As shown, it includes:

[0037] Real-time acquisition of load current, environmental data, electrical data, and temperature measurement data of power cables; construction of a digital twin model based on the multi-physics properties of the cable;

[0038] Based on the digital twin model, load current and environmental data are input, and the theoretical health baseline of the cable under the current operating conditions is calculated through multiphysics simulation. The theoretical health baseline includes theoretical temperature distribution and theoretical electrical signal level. The actual electrical data and temperature data are compared with the theoretical health baseline to generate a thermoelectric coupling matrix that quantifies the difference between the two.

[0039] The insulation degradation status of the cable is diagnosed based on the thermoelectric coupling matrix. When the value of the thermoelectric coupling matrix exceeds the threshold set according to the statistical analysis of the cable's historical data, a gradient boosting decision tree model is applied to quantify causal attribution, obtain the contribution degree, and generate a partial discharge spectrum that correlates the partial discharge signal with the root cause.

[0040] Based on the partial discharge spectrum, unique visual patterns generated by different defects on the partial discharge spectrum are identified. When a high-risk pattern of irreversible damage is identified, the type of insulation defect is determined; and it is predicted whether the cable is in a high-risk fault state.

[0041] Furthermore, such as Figure 3 As shown, the specific steps for constructing the digital twin model include:

[0042] A geometric model of the cable is established, defining its physical dimensions, structural layers, and laying path. A physical model of the cable is also established, integrating material property parameters of the conductor, insulation layer, sheath, and surrounding environmental medium, including thermal conductivity, thermal resistivity, dielectric constant, and electrical conductivity. A behavioral model of the cable is established, embedding multiphysics coupling equations describing the internal heat conduction and electromagnetic field distribution of the cable. A synchronization link is established between the three models and real-time data, linking the constructed geometric model, physical model, and behavioral model with the real-time acquired operational data to form a dynamic mirror reflecting the physical state of the cable.

[0043] Specifically, the multiphysics coupling equation uses the Joule heat loss calculated by the electromagnetic field model as the internal heat source of the heat conduction model, and feeds back the temperature distribution calculated by the heat conduction model to the electromagnetic field model to update material parameters such as conductivity that change with temperature in real time.

[0044] The synchronization link is implemented through an Industrial Internet of Things (IIoT) platform or SCADA system interface. Real-time operating data is input as dynamic boundary conditions and initial conditions of the behavioral model at a preset frequency, and the simulation results of the model are calibrated using measurement data to ensure the consistency between the digital twin model and the physical entity state.

[0045] This embodiment creates a high-fidelity and dynamic digital twin by systematically constructing geometric, physical, and behavioral models and establishing a synchronous link with real-time data. In particular, by embedding thermo-electric multiphysics coupling equations and calibrating using real-time data, the model can accurately simulate the complex physical processes of cables under real-world operating conditions, rather than simply presenting a static representation. This high degree of physical realism and consistency with the real world provides a solid and reliable foundation for subsequent accurate calculation of theoretical health baselines and identification of early signs of minor faults, fundamentally ensuring the accuracy and reliability of the entire prediction method.

[0046] Furthermore, the steps for calculating the theoretical health baseline specifically include:

[0047] The real-time acquired load current is used as the internal heat source input for the heat conduction equation in the digital twin model, and the environmental data is used as the external boundary condition for the heat conduction equation. The heat conduction equation is solved by finite element analysis and / or equivalent thermal circuit method to obtain the theoretical temperature distribution of the cable under the current operating conditions.

[0048] Based on the ideal insulation dielectric parameters and electromagnetic field model defined in the digital twin model, and combined with the correction of material parameters by the theoretical temperature distribution, the reference electrical signal response that a healthy cable should have under this operating condition is calculated, which is taken as the theoretical electrical signal level.

[0049] Specifically, the environmental data is applied according to the cable laying method. For direct-buried cables, the environmental data is soil temperature and humidity, which is used to set the far-field temperature boundary of the heat conduction equation. For cables in tunnels or cable trays, the environmental data is air temperature, humidity and wind speed, which is used to set the convective heat transfer boundary conditions.

[0050] The ideal insulation dielectric parameters refer to the design parameters of the cable at the time of manufacture or in a non-deteriorated state. The reference electrical signal response specifically includes the theoretical dielectric loss factor, the theoretical leakage current, and the background electrical noise level when there is no partial discharge.

[0051] This embodiment calculates a theoretical health baseline that dynamically changes with operating conditions by combining real-time load and refined environmental data. This dynamic baseline accurately reflects the "ideal" state of the cable under specific current conditions, effectively distinguishing between abnormal signals caused by normal operating fluctuations (such as high-load temperature rise) and those caused by insulation degradation. This not only significantly reduces the false alarm rate caused by environmental or load changes but also significantly improves the detection sensitivity and diagnostic accuracy of early, subtle fault precursors, providing a reliable basis for accurate prediction.

[0052] Furthermore, the step of generating the thermoelectric coupling matrix specifically includes:

[0053] The cable is divided into multiple discrete regions along its length in a digital twin model. For each discrete region, the actual measured temperature and electrical signal values, as well as the corresponding theoretical temperature and electrical signal values ​​given by the theoretical health baseline, are extracted. Within each discrete region, the thermodynamic and electrical deviations between the actual measured values ​​and the theoretical values ​​are calculated independently. The thermodynamic and electrical deviations calculated for all discrete regions are arranged in order of their physical location on the cable to construct a two-dimensional matrix, where one dimension of the matrix corresponds to the discrete regions of the cable, and the other dimension corresponds to different physical quantity deviations, forming the thermoelectric coupling matrix.

[0054] Specifically, the division of the discrete regions is determined based on the spatial resolution of the distributed temperature sensing (DTS) system. For each region, the thermodynamic and electrical deviations are obtained by calculating the absolute difference and relative percentage difference between the actual measured values ​​and the theoretical values. Electrical signal values ​​measured by non-distributed sensors are assigned to the discrete region closest to the sensor.

[0055] This application achieves precise location of fault precursors by discretizing the cable state along its length and constructing a structured two-dimensional matrix. This thermoelectric coupling matrix not only quantifies the degree of anomaly at each location point, but also unifies the two physical deviations of heat and electricity in space, providing an intuitive and quantitative data foundation for subsequent spatiotemporal pattern analysis and accurate identification of local defects, significantly improving the precision and spatial resolution of the diagnosis.

[0056] Furthermore, the step of diagnosing the insulation degradation state of the cable based on the thermoelectric coupling matrix specifically includes:

[0057] Spatiotemporal pattern analysis is performed on the deviation values ​​in the thermoelectric coupling matrix, including identifying the amplitude of the deviation values, the spatial distribution characteristics along the cable length, and the evolution trend over time; by identifying whether there are spatially concentrated deviation clusters in the thermoelectric coupling matrix that continuously exceed the normal fluctuation range, it is determined whether there are local insulation defects.

[0058] By analyzing the overall statistical mean of the deviation values ​​in the matrix and its rate of increase over time, the general aging degree of the cable is assessed. Combining the judgment results of the local insulation defects with the assessment results of the general aging degree, the specific insulation deterioration state of the cable is finally determined.

[0059] Specifically, the spatiotemporal pattern analysis employs clustering algorithms such as DBSCAN to automatically identify deviation clusters. First, the thermodynamic and electrical deviation data in the thermoelectric coupling matrix are standardized using Z-scores. During clustering, Euclidean distance is used as the metric. Core parameters are determined based on the cable's physical characteristics and sensor resolution: the neighborhood radius eps is set to 3 times the discrete region length, and the minimum number of sample points min_samples is set to 5. Linear regression or time series prediction models are used to quantify the evolution trend. Further, the specific insulation degradation state is divided into four levels: healthy, attentive, warning, and alarm. The determination of these levels is achieved through a preset rule engine and / or a trained classification model. The rule engine or model uses the number, size, growth rate, and overall aging rate of deviation clusters as input features.

[0060] By employing automated clustering and parameterized settings, diagnosis is transformed from subjective judgment into objective and repeatable intelligent analysis. This method enables defect localization, trend assessment, and status classification, significantly improving the automation level and decision-making efficiency of diagnosis.

[0061] Furthermore, in the initial stage of cable healthy operation, multi-source data for at least N complete operating cycles are continuously collected, and a corresponding thermoelectric coupling matrix is ​​generated for each cycle. Statistical analysis is performed on each deviation element in all collected thermoelectric coupling matrices to calculate its mean and standard deviation. Using the 3σ principle, the mean of each type of deviation value plus three times the standard deviation is used as the upper limit boundary of its normal fluctuation range, and this is set as the alarm threshold for triggering the gradient boosting decision tree model to perform root cause analysis. At the same time, the simulation results of the digital twin model are calibrated using the measured data in this stage to minimize the theoretical and actual errors in the initial stage.

[0062] This embodiment, based on health period data, scientifically sets alarm thresholds and calibrates the model, improving the accuracy and predictive reliability of the dynamic benchmark, effectively reducing the false alarm rate, and ensuring the sensitivity of early weak fault identification.

[0063] This application transforms the diagnosis of insulation degradation from subjective judgment relying on human experience to objective and repeatable intelligent analysis by introducing automated clustering algorithms such as DBSCAN and specific parameter settings. This method can not only accurately identify and locate spatially continuous local defect areas, but also assess the overall aging trend of the cable, achieving a comprehensive diagnosis of both local and global health status. Finally, a clear health level is output through a rule engine or classification model, providing maintenance personnel with intuitive, reliable, and quantifiable status assessment results, significantly improving the automation level of diagnosis and decision-making efficiency.

[0064] Furthermore, the step of generating the partial discharge spectrum specifically includes: extracting high-frequency partial discharge pulse signals from the electrical data, and obtaining the amplitude of each pulse and its precise occurrence time; synchronizing the occurrence time of each pulse with the phase of the power frequency AC voltage, and generating a two-dimensional coordinate data point of phase-amplitude for each pulse signal;

[0065] Using the gradient boosting decision tree model, the root cause of the deviation obtained from the thermoelectric coupling matrix analysis is associated with the two-dimensional coordinate data points of phase and amplitude; all associated two-dimensional coordinate data points are projected onto a two-dimensional statistical spectrum with phase and amplitude as coordinate axes, and the data points are visualized and distinguished according to the root cause, thereby generating the partial discharge spectrum.

[0066] The steps for associating the phase-amplitude two-dimensional coordinate data points are as follows: A time synchronization window is set, for example, 15 minutes. Within this window, the average SHAP contribution vector of each root cause to the thermoelectric coupling matrix deviation is calculated. All partial discharge pulse data points occurring within this window are assigned the same contribution vector as their attribute. When generating a spectrum for visualization, the color of each data point is determined by its attribute vector. Using the RGB color model, the top three root causes with the highest contribution are mapped to the intensities of the R, G, and B channels, respectively, thus visually demonstrating the combined effect of multiple causes through mixed colors.

[0067] Specifically, the extraction of high-frequency partial discharge pulse signals employs time-frequency analysis methods such as wavelet packet decomposition to separate effective signals from background noise. Further, the gradient boosting decision tree model is implemented by constructing a gradient boosting decision tree model. The input features of this model are operating parameters representing the root causes, and the output is the deviation value of the thermoelectric coupling matrix. During real-time diagnosis, the SHAP attribution algorithm is applied to interpret the single prediction result of the XGBoost model, calculating the contribution vector of each root cause to the current deviation. Within a preset time synchronization window, this contribution vector is assigned to all partial discharge pulses occurring within the window. When generating the spectrum, an RGB color model is used, mapping the contribution values ​​of the top three root causes with the highest contribution to the intensities of the R, G, and B color channels, respectively, thereby visualizing each discharge point on the spectrum through color mixing. This method, by constructing an XGBoost model and applying the SHAP algorithm, transforms fuzzy expert experience into an objective and quantitative attribution analysis process. It not only reveals the root causes of partial discharge but also, through an innovative RGB color mapping mechanism, intuitively presents the complex multi-factor influences on a single spectrum. This allows maintenance personnel to readily identify the dominant driving factors of the fault, elevating the partial discharge spectrum from a simple symptom monitoring tool into an intelligent analysis platform capable of deep causal diagnosis, providing unprecedented decision support for accurate and efficient predictive maintenance.

[0068] Specifically, a gradient boosting decision tree (XGBoost) model is first constructed to predict thermoelectric coupling bias. The input features of this model include: ... (e.g., quantified parameters representing the root causes such as soil moisture, ambient temperature, load rate, and years of operation). The model outputs statistical feature values ​​(such as mean and peak values) of key bias clusters in the thermoelectric coupling matrix.

[0069] A training sample set was constructed by collecting historical fault data and corresponding operating condition data. This sample set was then used to train the XGBoost model using supervised learning.

[0070] After model training is complete, when a deviation exceeding a threshold is detected in real time, the current input feature vector is fed into the trained XGBoost model. Subsequently, the SHAP algorithm is applied to interpret the model's prediction results for this time, thereby calculating the contribution of each input feature (root cause) to the current deviation output.

[0071] This application transforms fuzzy expert experience into an objective, quantitative attribution analysis process by constructing an XGBoost model and applying the SHAP algorithm. It not only reveals the root causes of partial discharge but also, through an innovative RGB color mapping mechanism, intuitively presents the complex multi-factor influences on a single spectrum. This allows maintenance personnel to readily identify the dominant driving factors of the fault, such as whether the discharge in a certain area is primarily caused by environmental factors or its own aging. This method elevates the partial discharge spectrum from a simple phenomenon monitoring tool to an intelligent analysis platform capable of deep causal diagnosis, providing unprecedented decision support for accurate and efficient predictive maintenance.

[0072] Furthermore, the step of identifying unique visual patterns generated by different defects on the partial discharge spectrum specifically includes: using a pre-trained convolutional neural network model to learn a dataset containing n known insulation defect types and their corresponding standard partial discharge spectrum samples, thereby gaining the ability to identify visual patterns of different defects; inputting the real-time generated partial discharge spectrum as an image into the convolutional neural network model; and finally outputting a classification result by performing feature extraction and pattern matching on the input partial discharge spectrum, wherein the classification result is the identification of the unique visual pattern on the current partial discharge spectrum.

[0073] Specifically, the convolutional neural network model is preferably a mature network architecture such as ResNet or VGG trained using transfer learning methods. The dataset is constructed by combining samples generated from accelerated aging tests in the laboratory, samples extracted from historical failure cases, and samples annotated by on-site experts to ensure the model's generalization ability.

[0074] The classification result is specifically represented as a probability distribution vector, which quantifies the probability that the input partial discharge spectrum belongs to each known defect type.

[0075] This application, by applying attribution algorithms such as SHAP, is the first to quantitatively correlate partial discharge phenomena with their underlying physical causes. Through innovative visualization methods, it intuitively reveals the driving factors leading to discharge, such as environmental influences or inherent aging, on the spectrum. This elevates diagnosis beyond mere phenomenological description to a deeper causal analysis, transforming the partial discharge spectrum from a monitoring tool into an intelligent diagnostic platform. This provides unprecedented decision support for maintenance personnel to formulate precise and efficient intervention measures.

[0076] Furthermore, the high-risk modes include electrical tree discharge mode, surface discharge mode, and internal air gap discharge mode.

[0077] Specifically, the electrical tree discharge mode is considered high-risk because it forms an irreversible, continuously growing carbonized channel inside the insulation, which will eventually lead to complete insulation breakdown; the surface discharge mode is considered high-risk because it indicates severe degradation of the insulation surface or dielectric interface, which is prone to rapid development into a through flashover; the internal air gap discharge mode is considered high-risk when its discharge amplitude or repetition rate increases significantly over time, because this indicates that the internal cavity defects are expanding or its surface has been carbonized, thereby accelerating the overall insulation degradation process.

[0078] This application provides a solid physical basis for determining "high risk" by associating abstract discharge patterns with specific, irreversible physical degradation processes (such as the formation of carbonized channels and rapid development into flashover). This transforms fault prediction from a simple pattern classification into a deep diagnosis of fault urgency and development trends. Maintenance personnel can formulate maintenance strategies based on the actual physical hazard level of the defect, making the prediction results more instructive and credible.

[0079] This application first constructs a digital twin model and calculates a dynamic health baseline. This method effectively filters out interference from changes in operating conditions, significantly improving the accuracy and reliability of early fault identification and reducing the false alarm rate. By introducing interpretable artificial intelligence, it can not only discover problems but also reveal their root causes, making the diagnostic process transparent and reliable. Finally, through advanced pattern recognition of partial discharge spectra, this method achieves precise determination from "abnormal state" to "specific high-risk defect type," making the prediction results no longer vague alarms but highly instructive diagnostic conclusions containing clear fault modes, providing strong technical support for achieving accurate predictive maintenance.

[0080] Example 2

[0081] This embodiment aims to illustrate how the invention can be applied to a critical underground cable that powers a city’s central business district. The cable is approximately 2 kilometers long and is laid in the soil beneath a sidewalk.

[0082] The maintenance team first established a digital twin model of the cable. The geometric model accurately recorded its 2-kilometer laying path, burial depth (1.2 meters underground), and structural layers (copper conductor, XLPE insulation layer, PVC sheath, etc.). The physical model integrated the factory parameters of the materials, such as the dielectric constant and thermal conductivity of XLPE.

[0083] The model is connected to the field sensor network through an Industrial Internet of Things (IIoT) platform. The sensors include: distributed temperature sensing (DTS) fiber optic cables (1-meter spatial resolution) laid along the cable; high-frequency current transformers installed at both ends of the cable and at manholes in the middle (for collecting electrical data, especially high-frequency partial discharge signals); and soil temperature and humidity sensors deployed along the cable. The SCADA system provides real-time load current data.

[0084] In the initial stage of cable monitoring using the method of this invention, the system operated continuously for several months under different seasonal and load conditions. For example, on a hot summer workday, the load reached a peak of 450A in the afternoon, and the soil temperature was 28°C. After receiving this data, the digital twin model calculated, through multiphysics simulation, that the theoretical temperature distribution of a healthy cable should be 60-65°C along the cable, and the theoretical electrical signal level (such as leakage current) should be below 5μA. This calculation result became the "theoretical health baseline" at this moment.

[0085] In the spring of its 16th year of operation, a tiny air gap defect formed inside the insulation layer of the cable located at 1.2 kilometers due to long-term mechanical stress and moisture intrusion.

[0086] The DTS system detected that the temperature in the vicinity of 1.2 km was consistently about 2-3°C higher than the theoretical predictions of the model. At the same time, the high-frequency current transformer also captured weak but persistent partial discharge pulse signals.

[0087] The system fills these deviations into the thermoelectric coupling matrix. In the matrix, the row corresponding to the 1.2 km position shows positive values ​​for both the "thermodynamic deviation" and "electrical deviation" columns, forming a clear "bright spot" or deviation cluster.

[0088] After three weeks of continuous monitoring, the values ​​in the deviation cluster consistently exceeded the 3σ alarm threshold set based on historical data statistics. The system automatically triggered a deep diagnostic process.

[0089] A pre-trained XGBoost model is invoked. Its input features include: current load rate (medium), soil moisture (85%), ambient temperature (22°C), and the service life of the cable segment (16 years).

[0090] After the XGBoost model predicts the severity level of the current bias, the SHAP algorithm immediately interprets this prediction. The analysis shows that the root cause contributing the most to the current bias is "increased dielectric loss (internal insulation degradation)," contributing 89%; followed by "external environmental disturbances," contributing only 11%.

[0091] The system collects all high-frequency partial discharge pulses from the past 15 minutes and generates phase-amplitude coordinates for each pulse. Then, based on the SHAP analysis results, these data points are "colored": since 89% of the causes point to internal degradation, the system uses an RGB color model, setting the intensity of the R channel, representing internal degradation, to a high value (e.g., 220), and setting other channels to low values, making these data points appear bright red on the spectrum. The generated spectrum shows a dense cluster of red points, concentrated at a specific phase angle after the voltage zero-crossing point.

[0092] This red causal attribution partial discharge spectrum was input as an image into a pre-trained ResNet convolutional neural network model.

[0093] The ResNet model extracts features by comparing the input spectral patterns with n standard defect patterns it has learned in the dataset.

[0094] The model outputs a probability distribution vector: "internal air gap discharge mode" has a probability of 94%, "electric tree discharge mode" has a probability of 4%, and "surface discharge mode" has a probability of 2%.

[0095] Based on the diagnostic result of "internal air gap discharge mode" and the trend of its discharge amplitude increasing over time, the system determined it to be a high-risk mode. The system immediately pushed a Level 2 alarm (early warning level) to the operations and maintenance center.

[0096] "Warning: A high-confidence (94%) internal air gap discharge has been identified at kilometer 1.2 of Line 2. The defect is in the development stage and there is a risk of evolving into an electrical tree, constituting a high-risk fault condition. It is recommended to arrange precise location and offline withstand voltage testing within one week."

[0097] Through this embodiment, the method successfully completed early warning, cause diagnosis, type determination, and risk assessment weeks or even months before a catastrophic cable breakdown failure, buying valuable processing time for maintenance personnel and avoiding a major accident that could have caused a large-scale power outage in the central business district.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cable state fault prediction method based on multi-source data fusion, characterized in that, The method comprises the following steps: Real-time acquisition of load current, environmental data, electrical data and temperature measurement data of the power cable; Constructing a digital twin model based on the multi-physical field properties of the cable; Based on the digital twin model, input the load current and environmental data, and calculate the theoretical health baseline of the cable under the current working condition through multi-physical field simulation. The theoretical health baseline includes theoretical temperature distribution and theoretical electrical signal level; Compare the actual electrical data and temperature measurement data with the theoretical health baseline to generate a thermoelectric coupling matrix quantifying the difference between the two; Diagnose the insulation degradation state of the cable based on the thermoelectric coupling matrix; Perform a spatiotemporal pattern analysis on the deviation values in the thermoelectric coupling matrix, including identifying the amplitude, spatial distribution characteristics along the cable length, and evolution trend over time of the deviation values. Determine whether there is a local insulation defect by identifying whether there is a deviation aggregation area in the thermoelectric coupling matrix that continuously exceeds the normal fluctuation range and is spatially concentrated; Evaluate the general aging degree of the cable by analyzing the overall statistical mean of the deviation values in the matrix and their rate of growth over time. Finally, determine the specific insulation degradation state of the cable by combining the results of the local insulation defect judgment and the evaluation of the general aging degree; When the numerical value of the thermoelectric coupling matrix exceeds the threshold value set based on historical cable data statistics, apply a gradient boosting decision tree model for quantitative causal attribution to obtain a contribution degree and generate a partial discharge spectrum that relates partial discharge signals to root causes; In the initial stage of cable health operation, continuously collect multi-source data for N complete operation cycles, and generate a corresponding thermoelectric coupling matrix for each cycle. Perform statistical analysis on each deviation element in all collected thermoelectric coupling matrices to calculate the mean and standard deviation. Use the 3σ principle to set the upper limit boundary of the normal fluctuation range as three times the standard deviation of each type of deviation value, and set it as the threshold value for triggering the gradient boosting decision tree model for root cause analysis; Based on the partial discharge spectrum, identify unique visual patterns on the partial discharge spectrum generated by different defects. When a high-risk pattern of irreversible damage is identified, determine the type of insulation defect. Predict whether the cable is in a high-risk failure state.

2. The cable state fault prediction method based on multi-source data fusion according to claim 1, characterized in that: The steps of constructing the digital twin model specifically include: Establishing a geometric model of the cable, which defines the physical dimensions, structural hierarchy, and laying path of the cable. Establishing a physical model of the cable, which integrates the material property parameters of the conductor, insulation layer, sheath, and surrounding environmental medium, including thermal conductivity, thermal resistivity, dielectric constant, and electrical conductivity. Establishing a behavior model of the cable, which embeds multi-physical field coupling equations describing the internal heat conduction and electromagnetic field distribution of the cable. Establishing a synchronization link between the three models and real-time data, correlating the constructed geometric model, physical model, and behavior model with real-time operation data to form a dynamic mirror reflecting the physical cable state.

3. The cable condition fault prediction method based on multi-source data fusion according to claim 1, characterized in that, The steps of calculating the theoretical health baseline specifically include: input the real-time acquired load current as an internal heat source of a heat conduction equation in the digital twin model, and input the environmental data as an external boundary condition of the heat conduction equation; solve the heat conduction equation by finite element analysis and / or equivalent thermal circuit method to obtain a theoretical temperature distribution of the cable under the current working condition; based on ideal insulation dielectric parameters and an electromagnetic field model defined in the digital twin model, and in combination with correction of the material parameters according to the theoretical temperature distribution, calculate a reference electrical signal response of a healthy cable under the working condition as the theoretical electrical signal level.

4. The cable state fault prediction method based on multi-source data fusion according to claim 1, characterized in that: The step of generating the thermoelectric coupling matrix specifically comprises: divide the cable into a plurality of discrete regions along the length direction in the digital twin model; for each discrete region, extract the actually measured temperature value and electrical signal value in the discrete region, and the corresponding theoretical temperature value and theoretical electrical signal value given by the theoretical health baseline; in each discrete region, independently calculate the thermodynamic deviation and electrical deviation between the actually measured value and the theoretical value; arrange all the thermodynamic deviations and electrical deviations calculated by the discrete regions in sequence according to the physical positions on the cable to form a two-dimensional matrix, wherein one dimension of the matrix corresponds to the discrete regions of the cable, and the other dimension corresponds to different physical quantity deviations, forming the thermoelectric coupling matrix.

5. The cable condition fault prediction method based on multi-source data fusion according to claim 1, characterized in that: The step of generating the partial discharge spectrum specifically comprises: extracting high-frequency partial discharge pulse signals from the electrical data, and obtaining the amplitude of each pulse and the accurate time of occurrence; synchronizing the occurrence time of each pulse with the phase of the power frequency alternating voltage to generate a phase-amplitude two-dimensional coordinate data point for each pulse signal; using the gradient boosting decision tree model, associate the deviation root cause analyzed from the thermoelectric coupling matrix with the phase-amplitude two-dimensional coordinate data points; project all the associated two-dimensional coordinate data points onto a two-dimensional statistical spectrum with phase and amplitude as coordinate axes, and visually distinguish the data points according to the root cause, thereby generating the partial discharge spectrum.

6. The cable condition fault prediction method based on multi-source data fusion according to claim 1, characterized in that, The step of identifying the unique visual patterns generated by different defects on the partial discharge spectrum specifically comprises: using a pre-trained convolutional neural network model to obtain the ability to identify different defect visual patterns by learning a data set containing n known insulation defect types and their corresponding standard partial discharge spectrum samples; input the real-time generated partial discharge spectrum into the convolutional neural network model as an image; finally output a classification result by feature extraction and pattern matching on the input partial discharge spectrum, and the classification result is the identification of the unique visual pattern on the current partial discharge spectrum.

7. The cable condition fault prediction method based on multi-source data fusion according to claim 1, characterized in that: The high-risk mode includes tree discharge mode, surface discharge mode and internal air gap discharge mode. The high-risk mode includes tree discharge mode, surface discharge mode and internal air gap discharge mode.

Images