Multi-state sensing cable fault prediction method and system

Abstract

The invention specifically relates to a multi-state sensing cable fault prediction method and system, and relates to the technical field of power equipment state monitoring and fault early warning, and the method comprises the steps: constructing a data fitting and physical constraint combined PINN model of a dual-drive network architecture; training the PINN model by using preset data until the PINN model converges; and on the basis of the trained PINN model, receiving real-time preprocessed data and then outputting an RUL predicted value. According to the invention, through multi-dimensional data acquisition processing and physical and data dual-drive PINN model design, the core problems that a traditional system is poor in data quality and the model is easy to violate physical rules are solved; pINN is embedded into an Arrhenius thermal aging equation, double loss calculation of data loss and physical loss is adopted, dynamic weight is matched, a model is forced to follow the law of temperature rise and life shortening, and RUL prediction precision and credibility are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring and fault early warning, and in particular to a cable fault prediction method and system with multi-state perception. BACKGROUND

[0002] As the core power transmission carrier of the power system, the operation reliability of the cable directly determines the power supply safety of the power grid. Current cable fault management relies on post-maintenance or regular maintenance mode, and the existing monitoring system has significant technical defects:

[0003] On the one hand, the data collection dimension is single, and multiple focuses on dynamic electrical signals such as traveling wave and partial discharge, without integrating static attribute data such as cable model, insulation material type, and operation time, which cannot build differentiated aging models for different types of cables, resulting in insufficient prediction universality and accuracy;

[0004] On the other hand, the mainstream fault prediction model is mainly driven by pure data (such as deep belief network and convolutional neural network), without embedding the physical mechanism of cable aging, which is easy to output results that violate objective laws due to data noise or sample distribution bias, such as conductor temperature rise but predicted remaining useful life extension, which seriously affects the credibility of the model.

[0005] Therefore, a cable fault prediction method and system with multi-state perception are needed to address the above-mentioned problems. SUMMARY

[0006] The purpose of the present application is to solve the above problems and propose a cable fault prediction method and system with multi-state perception.

[0007] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0008] A cable fault prediction system with multi-state perception includes a multi-dimensional state perception collection and processing module, a fault risk assessment and result interpretation module, and an operation and maintenance decision assistance output module, and also includes a PINN prediction model training and reasoning module, which specifically includes:

[0009] A PINN model with a dual-driven network architecture combining data fitting and physical constraints is constructed;

[0010] The PINN model is trained with preset data until the PINN model converges;

[0011] Based on the trained PINN model, the real-time preprocessed data is received to output the RUL prediction value.

[0012] Preferably, the multi-dimensional state perception collection and processing module is configured to:

[0013] Collecting dynamic time series data, including cable conductor absolute temperature , operating voltage , and ambient relative humidity .

[0014] And static attribute data, including cable model, operating age, insulation material type, laying method.

[0015] The collected raw data is cleaned, converted and feature extracted to generate standardized feature vectors conforming to the input format of the PINN model.

[0016] Preferably, the PINN model construction process includes:

[0017] Based on the input and output definition and network structure design of the PINN model, the network topology construction from feature mapping to RUL prediction is completed, while embedding the cable aging physical equation;

[0018] The network adopts a fully connected time series adaptive architecture with an input layer, a hidden layer and an output layer;

[0019] The Arrhenius thermal aging equation calculation module is embedded in the network to provide real-time aging rate benchmark values for physical loss calculation, and the specific logic is as follows:

[0020] For cables of different insulation material types, the key parameters of the Arrhenius equation are pre-stored;

[0021] Based on the cable conductor absolute temperature , the Arrhenius equation is substituted to calculate the aging rate in real time.

[0022] Integrate the automatic differentiation interface of the deep learning framework, record the mapping relationship between the RUL prediction value and the timestamp in the input time series features through the context manager during the network forward propagation process;

[0023] When needs to be calculated, the first-order derivative of the RUL prediction value with respect to time is automatically solved, and the derivative result is directly used for subsequent physical loss calculation.

[0024] Preferably, the PINN model is trained with preset data until the PINN model converges, specifically including:

[0025] The preprocessed data is divided into labeled data set and unlabeled data set, and the ratio satisfies ≈500:1; and the data loss and the physical loss are calculated respectively;

[0026] Data loss The calculation process comprises: inputting 12-dimensional input features into the PINN network to obtain RUL prediction values for each sample in the labeled training set , and the true label ;

[0027] During each batch training, 32 labeled training set samples are synchronously sampled, and the average data loss of the batch is calculated as the data-driven error feedback;

[0028] Physical loss The calculation process comprises: for each sample in the unlabeled sample, first calculate the aging rate , and then calculate it through the automatic differentiation interface; and based on the physical relationship , the physical loss is the mean square error of the deviation between the derivative calculation value and the theoretical value.

[0029] Pre-set weight After that, during each batch training, the data loss and the physical loss are calculated respectively, and the data loss and the physical loss are fused to obtain the total loss.

[0030] Preferably, the determination process of the weight is as follows:

[0031] The physical consistency error is defined as the proportion of samples in the validation set whose RUL derivative with respect to time ≥0, and is normalized to obtain ;

[0032] The data fitting error is defined as the mean absolute error of the validation set, and is normalized to obtain ;

[0033] Since the weight needs to increase with the increase of and decrease with the increase of , while ∈[0.1,0.3];

[0034] A linear combination function about the weight is obtained:

[0035]

[0036] Wherein,

[0037] is the basic weight;

[0038] , are adjustment coefficients;

[0039] This is a truncation function;

[0040] In each verification phase, the following steps are performed:

[0041] A1: Calculate the current validation set and ;

[0042] A2: Yes , Normalization is performed to obtain and ;

[0043] A3: Substitute into the weighting function to calculate the weights. This is then updated in the training process for the next round of total loss.

[0044] Preferably, training the PINN model with preset data until the PINN model converges specifically includes:

[0045] Using the Adam optimizer, the batch training size is 256 records / batch. Rate of change over 100 consecutive rounds < Training is stopped at this point, and a stable model file is generated.

[0046] Preferably, the fault risk assessment and result interpretation module is configured to classify cable fault risk levels based on the RUL prediction value output by the PINN model, and provide an interpretable explanation of the prediction results in conjunction with the physical mechanism.

[0047] Preferably, the operation and maintenance decision support output module is configured to combine fault prediction results with power grid operation and maintenance needs, output feasible operation and maintenance suggestions, and realize data storage and system interaction, so as to solve the problem of disconnect between prediction results and operation and maintenance, and support power grid condition-based maintenance decisions.

[0048] A multi-state sensing cable fault prediction method includes the following steps:

[0049] Dynamic time-series data and static attribute data are collected, and after cleaning, feature extraction and Z-Score standardization, a 12-dimensional standardized feature vector is generated.

[0050] A fully connected time-series adaptation network driven by both data fitting and physical constraints is constructed, and the Arrhenius thermal aging equation module and automatic differentiation interface are embedded to calculate the aging rate and the derivative of RUL with respect to time.

[0051] The data is divided into labeled and unlabeled categories, and the data loss and physical loss are calculated separately, using dynamically adjusted weights. The total loss is fused, and the Adam optimizer is used for training.

[0052] Based on the PINN model trained, the preprocessed cable real-time feature vector is received, the model is loaded and the feature vector is input, and the current time cable remaining useful life is directly output;

[0053] According to the RUL prediction value, the low, medium and high fault risk levels are divided, and the prediction result explainability is given in combination with the Arrhenius law and the RUL derivative relationship.

[0054] To sum up, due to the adoption of the above technical scheme, the beneficial effects of the present application are:

[0055] 1. The present application solves the core problems of poor data quality and model easily violating physical laws of traditional systems through multi-dimensional data acquisition processing and physical and data dual-driven PINN model design; PINN embeds the Arrhenius thermal aging equation, adopts data loss and physical loss double loss calculation, and is matched with dynamic weight, so as to force the model to follow the law of life shortening with temperature rise, and improve the RUL prediction accuracy and reliability.

[0056] 2. The present application solves the pain points of traditional model black box output and difficult landing of prediction results through explainable output and deep connection with operation and maintenance decision; in the risk assessment link, the low / medium / high risk levels are divided according to the RUL threshold, the prediction causes are explained in combination with the Arrhenius law and the RUL derivative relationship, and the risk and physical factor correlation are visualized, so that the operation and maintenance personnel can understand the prediction logic; in the operation and maintenance decision link, differentiated suggestions are generated based on the risk level and cable importance, and are pushed to the operation and maintenance end through multiple channels; the power grid is supported to shift from regular maintenance to condition-based maintenance, the operation and maintenance resource mismatch is reduced, the fault outage rate is reduced, and the engineering practical value is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0057] In the following description of exemplary embodiments in conjunction with the drawings, more details, features and advantages of the present application are disclosed, in which:

[0058] Figure 1 The PINN prediction model training and inference module structure diagram of the present application;

[0059] Figure 2 The method flowchart of the present application. DETAILED DESCRIPTION

[0060] Several embodiments of the present application will be described in more detail below with reference to the accompanying drawings so as to enable those skilled in the art to implement the present application. The present application can be embodied in many different forms and purposes and should not be limited to the embodiments described herein. These embodiments are provided to make the present application comprehensive and complete, and to fully convey the scope of the present application to those skilled in the art. The embodiments do not limit the present application.

[0061] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and / or the present

[0062] Embodiment 1

[0063] BEST MODE FOR CARRYING OUT THE INVENTION Figure 1 and the accompanying drawings which are incorporated in and constitute a part of this specification, illustrate embodiments of the application and serve to explain the principles of the application. Figure 2 The detailed description explains embodiments of the application, with reference to the accompanying drawings.

[0064] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and serve to explain the principles of the application. Figure 1 A multi-state perception cable fault prediction system structure block diagram is provided for the embodiments of the application, which shows the connection relationship between the multi-dimensional state perception acquisition and processing module and the operation and maintenance decision auxiliary output module, and labels the main function interaction process of each module.

[0065] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and serve to explain the principles of the application. Figure 2 A multi-state perception cable fault prediction method flow chart is provided for the embodiments of the application, which shows the complete steps from collecting dynamic time series data and static attribute data to giving the explainability of the prediction result.

[0066] In this embodiment, it includes:

[0067] The multi-dimensional state perception acquisition and processing module is configured to:

[0068] Collect dynamic time series data, including cable conductor absolute temperature (through distributed optical fiber temperature sensor, sampling frequency 1 / 5 minutes), operating voltage (through voltage transformer, sampling frequency 1 / minute) and environmental relative humidity (through temperature and humidity sensor deployed in the cable trench, sampling frequency 1 / 10 minutes), using edge computing gateway, transmitting real-time data to the system background through 5G / optical fiber, controlling the delay within 100ms, avoiding the time series distortion of time series data;

[0069] and static attribute data, including cable model, operation life, insulation material type (such as cross-linked polyethylene, ethylene-propylene rubber), laying method (directly buried, pipe gallery) and other static characteristics; data entry supports manual batch import (connecting with power grid equipment account system) and code scanning input (reading through the unique identification code of the cable body), ensuring the accurate association of attribute data and dynamic data.

[0070] The collected raw data is cleaned, converted and feature extracted to generate standardized feature vectors that meet the input format of the PINN model, solving the problems of noise interference, dimension mismatch and time sequence feature inconspicuousness of raw data.

[0071] Data cleaning: For key time series data such as temperature and voltage, use adjacent time interpolation method (short-term missing, missing duration < 30 minutes) or similar cable data migration completion method (long-term missing); based on 3σ principle to identify abnormal data (such as temperature rising more than 10K), combined with physical mechanism verification (such as judging whether it is a sensor failure or real overload), the fault data is excluded or corrected.

[0072] Feature extraction and standardization: For time series data such as temperature and voltage, use sliding window method to extract statistical features (such as mean, maximum, fluctuation variance in 1 hour window), convert time series data to fixed dimension features, adapt to PINN model input layer structure; static feature encoding: for classification attributes such as cable model and insulation material type, use one-hot encoding to convert to numerical features; data standardization: for all features (such as temperature range 293K-353K, voltage range 10kV-220kV), use Z-Score standardization to avoid affecting PINN model training convergence speed due to dimension difference of features.

[0073] The PINN prediction model training and inference module is configured as follows: build a PINN model with a dual-driven network architecture combining data fitting and physical constraints; train the PINN model with pre-set data until the PINN model converges; based on the trained PINN model, receive real-time preprocessed data and output RUL prediction value, that is, trigger the inference process by receiving preprocessed real-time feature vectors, load the trained PINN model, input the feature vectors and directly output the current time RUL prediction value, which includes:

[0074] The PINN model construction process includes:

[0075] Based on the input and output definition and network structure design of the PINN model, the network topology from feature mapping to RUL prediction is constructed, and the cable aging physical equation is embedded;

[0076] The network adopts a fully connected time sequence adaptive architecture with input layer, hidden layer and output layer;

[0077] Input layer: designed as a 12-dimensional feature vector, corresponding to the input requirements of dynamic time sequence features and static attribute features, with the following specific dimension splitting:

[0078] Dynamic time series features (6 dimensions): composed of the sliding window statistical features output by the data preprocessing module, including cable conductor temperature 1-hour mean, temperature maximum, temperature fluctuation variance, operating voltage 1-hour mean, voltage fluctuation variance, ambient humidity 1-hour mean, all are standardized numerical features (Z-Score standardized range [-3, 3]);

[0079] Static attribute features (6 dimensions): obtained by static attribute encoding, including cable model (one-hot encoding 3 dimensions: XLPE, PVC, others), insulation material type (one-hot encoding 2 dimensions: cross-linked polyethylene, ethylene-propylene rubber), service life (1-dimensional numerical feature, standardized input), to ensure the dimensionality of static attributes and dynamic time series features.

[0080] Hidden layer: 4 layers of fully connected layers (to adapt to the smoothing fitting needs of time series features), the number of neurons in each layer decreases by 64, 32, 16, 8 (to avoid overfitting), and the activation function is uniformly selected as ReLU (to alleviate the gradient vanishing problem and perform more stably in the mapping of non-negative features such as temperature and voltage); Dropout layers (dropout rate = 0.1) are added between layers to further suppress overfitting and ensure the generalization ability of the model on small sample data.

[0081] Output layer: single neuron design, output dimension is 1, directly corresponding to the current time cable remaining useful life (RUL) prediction value, unit is month; the output layer does not use activation function (to avoid the RUL prediction value being truncated due to the value range limitation of the activation function, such as ReLU filtering negative values, but the physical meaning of negative derivative needs to be preserved in RUL calculation).

[0082] Embedding Arrhenius thermal aging equation calculation module inside the network to provide real-time aging rate benchmark value for physical loss calculation, the specific logic is as follows:

[0083] For cables of different insulation material types, pre-store Arrhenius equation key parameters (pointing to pre-exponential factor A, activation energy Ea, and frequency factor f0) , for example:

[0084] XLPE (cross-linked polyethylene) cable: , ;

[0085] PVC (polyvinyl chloride) cable: , ;

[0086] Automatically match the corresponding parameters through the insulation material type feature of the input layer without human intervention.

[0087] Based on the absolute temperature of the cable conductor Substituting into the Arrhenius equation: Real-time calculation of aging rate (Unit: 1 / month, the original 1 / year unit needs to be converted to 1 / month to match the RUL output unit); Boltzmann constant is fixed during the calculation process. This ensures the accuracy of physical calculations.

[0088] The automatic differentiation interface, integrated with a deep learning framework, records the RUL prediction value and its value during network forward propagation via a context manager. The mapping relationship of timestamps in the input time series features;

[0089] When calculation is required At that time, the first derivative of the RUL prediction value with respect to time is automatically calculated, and the derivative result is directly used for subsequent physical loss calculation;

[0090] To ensure the accuracy of derivative calculation, the timestamp uses continuous time encoding (e.g., 2025-01-01 00:00 is encoded as the value 1735689600, in seconds), avoiding derivative calculation errors caused by discrete time encoding.

[0091] The PINN model is trained using pre-defined data until it converges. This process includes:

[0092] The preprocessed data was divided into labeled datasets (source: historical fault cable records, accelerated aging test data) and unlabeled datasets (source: real-time collected normal cable data), with the ratio meeting the following requirements. ≈500:1; and calculate the data loss and physical loss separately;

[0093] Data loss The calculation process includes: for each sample in the labeled training set, inputting the 12-dimensional input features into the PINN network to obtain the RUL prediction value. , with real labels The mean square error is calculated, and its function is expressed as: ;

[0094] During each training batch, 32 labeled training set samples are sampled simultaneously (samples within a batch are stratified by cable type to avoid excessive data proportion of a single type), and the average data loss of that batch is calculated as data-driven error feedback.

[0095] Physical loss The calculation process includes: for each sample in the unlabeled sample set, first calculating the aging rate. (Based on temperature characteristics in the sample) Then calculate via the automatic differentiation interface. (the derivative of the RUL prediction with respect to time); and based on physical relationship (mathematical definition of RUL , the result obtained after taking the time derivative according to the basic theorem of calculus (the negative sign indicates that RUL is monotonically decreasing over time, and the decreasing rate is equal to the inverse of the aging rate), the physical loss is the mean square error of the deviation between the calculated value and the theoretical value, and the function is expressed as: (224 unlabeled samples per batch, 7:1 ratio with 32 labeled samples per batch, matching M:N data volume ratio);

[0096] If the physical loss of a batch is too large (such as > 100), temporarily increase the physical loss weight to force the model to learn the physical law and avoid the error result of RUL extension caused by temperature rise.

[0097] preset weight After each batch training, calculate the data loss and physical loss respectively, and fuse the data loss and physical loss to obtain the total loss, using .

[0098] The data set division and double loss calculation mechanism effectively solves the core problem of sample scarcity and model reliability in cable fault prediction. It divides the preprocessed data into labeled data set (historical fault records, accelerated aging experiment data) and unlabeled data set (normal cable real-time data) and matches them in a ratio of about 500:1, which can fully utilize the valuable fault sample to train the model data fitting ability, and also can strengthen the model generalization by means of a large amount of normal data;

[0099] The data loss calculation uses mean square error and stratified sampling of labeled samples to avoid model bias caused by excessive proportion of single cable model data. The physical loss is calculated based on the physical relationship between Arrhenius equation and RUL derivative, which ensures that the model output conforms to the objective law that cable aging temperature rise will shorten the life, and breaks the limitation that pure data-driven model may violate physical common sense.

[0100] The total loss fusion strategy and physical loss regulation mechanism further improve the precision and stability of model training. By presetting the weight to fuse the data loss and the physical loss, the balance between data fitting accuracy and physical consistency is achieved, so that the model has both precise adaptation ability to historical fault data and strict compliance with the physical mechanism of cable aging. When the physical loss of a batch is too large, temporarily increase the physical loss weight to force the model to learn the physical law, which can effectively avoid the false prediction caused by data noise or local sample anomaly (such as temperature rise but RUL extension), providing key guarantee for the model to output reliable RUL prediction value after convergence, and laying a foundation for model accuracy and credibility for the engineering application of cable fault prediction.

[0101] weight The determination process is as follows:

[0102] The physical consistency error is defined as the proportion of samples in the validation set whose RUL derivative with respect to time ≥0 (derivative non-negative violates the physical law that RUL decreases over time, the higher the proportion, the worse the physical consistency), and after normalization, we get , and ∈[0,1];

[0103] The data fitting error is defined as the mean absolute error (MAE) of the validation set (the larger the MAE, the worse the data fitting effect); and after normalization, we get , and ∈[0,1];

[0104] Since the weight needs to increase as increases (enhance data loss weight when data fitting is poor), and decrease as increases (enhance physical loss weight when physical consistency is poor), while ∈[0.1,0.3];

[0105] The linear combination function about weight is obtained:

[0106]

[0107] where

[0108] is the base weight (e.g. 0.2);

[0109] , are adjustment coefficients (set by relevant personnel, can take = 0.2, = 0.3, so that and have the same impact on );

[0110] is a truncation function to ensure that x is within the range [a, b] (here a = 0.1, b = 0.3);

[0111] Every validation phase (e.g. after every 100 training rounds), the following steps are performed:

[0112] A1: Calculate the derivative non-negative sample proportion) and (mean absolute error, MAE);

[0113] A2: Normalization, get , ; and ;

[0114] A3: Calculate weight by substituting weight function , and update to the training process for the next round of total loss .

[0115] A dynamic weight adjustment mechanism based on physical consistency error and data fitting error is constructed, which effectively solves the problem that traditional fixed weight is difficult to balance the model data fitting accuracy and the degree of compliance with physical laws. It defines the physical consistency error as the proportion of samples with non-negative time derivative of RUL (contrary to the law of life degradation), and the data fitting error as the average absolute error of the validation set. After normalization, the weight is dynamically calculated by a linear combination function , and is constrained in the range of 0.1-0.3;

[0116] When the data fitting effect is poor, it automatically increases to enhance the data loss weight, and when the physical consistency is poor, it automatically reduces to strengthen the physical loss weight, ensuring that the model can fit the real failure data in training, and not violate the objective physical law that the cable aging life decreases with time and accelerates with temperature rise, avoiding logical contradiction in the prediction results.

[0117] The weight update process provides real-time feedback and dynamic optimization capability for model training, significantly improving the stability and generalization of the model. The P and D of the validation set are recalculated and updated every time the validation is performed (such as every 100 training rounds), so that the weight can be self-adaptively adjusted with the model training process: for example, if the model deviates in physical consistency due to data noise at the beginning of training, it can be quickly corrected by reducing ; if the data fitting accuracy is insufficient at the later stage of training, it can be further optimized by increasing .

[0118] This dynamic regulation mechanism allows the model to always train in the balance between data-driven accuracy and physical mechanism constraints, ultimately ensuring that the PINN model converges to meet the data accuracy requirement of MAE < 5% on the validation set, while controlling the physical consistency error (proportion of samples with non-positive derivative) to < 2%, providing a RUL prediction basis with both accuracy and credibility for subsequent failure risk assessment and operation decision-making.

[0119] Train the PINN model with the preset data until the PINN model converges, which specifically includes:

[0120] Adam optimizer (initial learning rate , decayed with training epochs), batch size of 256, training stopped when the change rate of the total loss on the validation set was less than 0.001 for 100 consecutive epochs, and a stable model file was generated.

[0121] The batch training process is as follows:

[0122] Randomly sample 224 samples from the unlabeled dataset for each iteration, input the network to calculate the physical loss ;

[0123] Synchronously sample 32 samples from the labeled training set, input the network to calculate the data loss ;

[0124] Calculate the total loss , update all trainable parameters of the network through the backpropagation algorithm;

[0125] Calculate the data fitting MAE and the physical consistency error on the labeled validation set every 50 iterations, and record the performance changes of the validation set.

[0126] The convergence criteria are as follows:

[0127] Main criteria: when the total loss of the validation set changes by less than 0.001 for 100 consecutive iterations, and the data fitting MAE is less than 5% and the physical consistency error (the proportion of samples with positive RUL derivatives) is less than 2%; Early stopping mechanism: if the MAE of the validation set does not decrease (even increases) for 50 consecutive iterations, early stopping is triggered to avoid overfitting of the model;

[0128] Final training results: the model was trained for a total of 820 iterations, the MAE of the validation set was 4.2%, and the physical consistency error was 1.8%, meeting the preset accuracy requirements.

[0129] The fault risk assessment and result interpretation module is configured based on the RUL prediction value output by the PINN model, to divide the cable fault risk level and give an explainable explanation of the prediction result based on the physical mechanism, to solve the problem of black box output of traditional models and difficulty in tracing, to enhance the trust of operation and maintenance personnel, including:

[0130]

[0131] ​​The cable fault risk level classification step is: presetting three groups of threshold value ranges, each of which corresponds to a cable fault risk level; matching the output RUL prediction value with the three groups of threshold value ranges to obtain the cable fault risk level corresponding to the RUL prediction value, wherein the cable fault risk level includes low risk, medium risk and high risk; no special operation and maintenance is required for low risk, monthly review of state data is required for medium risk, and maintenance plan needs to be made for high risk, and power-off maintenance is preferentially arranged.

[0132] The explainability of the prediction result specifically includes:

[0133] Combining the Arrhenius law and the RUL derivative relationship to explain the reason for the change of RUL;

[0134] For example: the average conductor temperature in the past 72 hours is 333K (8K higher than last week), according to the Arrhenius law, the aging rate is increased by 2.3 times, which leads to a shortening of RUL by 1.8 months compared with the last prediction.

[0135] And through the broken line chart, the trend of RUL prediction value changing with time and the change curve of key influencing factors (temperature, voltage) are displayed, and the correlation between risk change and physical factors is intuitively presented.

[0136] The operation and maintenance decision assistance output module is configured to combine the fault prediction result with the power grid operation and maintenance demand, output the operation and maintenance suggestions that can be implemented, and realize data storage and system interaction, solve the problem that the prediction result is disconnected with the operation and maintenance, and support the power grid condition-based maintenance decision, specifically including:

[0137] The operation and maintenance suggestions are generated based on the risk level and the importance of the cable (such as whether it is a backbone line) to generate differentiated suggestions: high risk, backbone line: output suggestion to complete insulation state detection (such as partial discharge test) within 1 month, and arrange power-off maintenance within 3 months; medium risk, general line: output suggestion to collect partial discharge data once a month, and dynamically track the change of RUL;

[0138] Through multiple channels such as system interface, short message and power grid operation and maintenance APP, it is ensured that the operation and maintenance personnel can receive in time.

[0139] The data storage and interaction include: storing the original collected data, the preprocessed features, the PINN model training log and the RUL prediction result to the time series database (InfluxDB) and the relational database (MySQL), retaining 5 years of historical data for model iteration optimization; interfacing with the power grid EMS (energy management system) and PMS (equipment management system), supporting two-way data synchronization of prediction results and risk levels, realizing fault prediction, operation and maintenance execution and effect feedback.

[0140] Embodiment 2

[0141] Please refer to Figure 2 A multi-state-aware cable fault prediction method, comprising the following parts:

[0142] Collecting dynamic time series data such as cable conductor temperature and operating voltage, and static attribute data such as cable model and insulation material type, after cleaning (interpolation completion, abnormality elimination), feature extraction (sliding window statistics, one-hot encoding) and Z-Score standardization, a 12-dimensional standardized feature vector is generated;

[0143] Build a fully connected time series fitting network driven by data fitting and physical constraints, the input layer is a 12-dimensional feature vector, the hidden layer is set to 4 layers of gradient decreasing fully connected layers with Dropout layers, the output layer is a single neuron output RUL, and the Arrhenius thermal aging equation module and automatic differentiation interface are embedded to calculate the aging rate and the derivative of RUL with respect to time;

[0144] Divide the data into labeled (historical failure, accelerated aging data) and unlabeled (normal cable data) with a ratio of about 500:1, calculate the data loss (mean square error of labeled samples) and physical loss (mean square error of the deviation of the derivative of RUL with respect to time from the theoretical value of unlabeled samples) respectively, and fuse the total loss according to the dynamically adjusted weight Fusion total loss, trained with Adam optimizer;

[0145] Based on the trained PINN model, receive the preprocessed cable real-time feature vector, load the model and input the feature vector, and directly output the current time cable remaining life;

[0146] According to the RUL prediction value, divide the low, medium and high fault risk levels, give the prediction result explainability according to the Arrhenius law and the RUL derivative relationship, and generate differentiated operation and maintenance suggestions combined with the cable importance, push through multiple channels, store related data to the corresponding database, and interface with the power grid EMS and PMS systems.

[0147] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0148] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present application.

[0149] It should be noted that, in the present document, relational terms are used to convey a relationship of one entity or action to another entity or action. For example, without necessarily implying any actual relationship or order between entities or actions, the terms "first," "second," "top" and "bottom" are used to name different entities and actions, and are used to distinguish one entity or action from another entity or action, without necessarily conveying any actual relationship or order between such entities or actions. Furthermore, the terms "comprise," "include," and "have," and variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises, includes, or has a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without limitation, an element preceded by "comprises... a," "includes... a," or "has... a" does not, without more constraints, foreclose the existence of additional identical elements other than the one listed or other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0150] It should be understood that the sequence numbers of the processes described above do not mean the execution sequence of the processes, and the execution sequence of the processes should be determined according to the functions and inherent logic of the processes, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0151] Those skilled in the art can clearly understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present document can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0152] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0153] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0154] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0155] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0156] The foregoing merely describes certain exemplary embodiments of this application by way of illustration. Obviously, modifications and alterations can be made by those skilled in the art without departing from the spirit and scope of the application. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of the claims of the present application.

Claims

1. A multi-state aware cable fault prediction system, characterized by, The method comprises the following steps: The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges.

2. A multi-state aware cable fault prediction system as claimed in claim 1, wherein, The PINN model is trained with preset data until the PINN model converges. Collecting dynamic time series data, including cable conductor absolute temperature , operating voltage , and ambient relative humidity ; The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges.

3. A multi-state aware cable fault prediction system as claimed in claim 1, wherein, The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. Based on the cable conductor absolute temperature , the Arrhenius equation is substituted to calculate the aging rate in real time ; an automatic differentiation interface of an integrated deep learning framework, recording the RUL prediction value and a mapping relationship of the timestamps in the input timing features; When the calculation of the first derivative of the RUL prediction with respect to time is required, the derivative result is directly used for subsequent physical loss calculations.

4. A multi-state aware cable fault prediction system as claimed in claim 3, wherein, The PINN model is trained with preset data until the PINN model converges. The preprocessed data is divided into a labeled data set and an unlabeled data set, and the ratio meets ≈500:1; and data loss and physical loss are calculated respectively; Data loss The computing process includes: for each sample in the labeled training set, input the 12-dimensional input feature into the PINN network to obtain the RUL prediction value , and the true label , the mean square error is calculated; The PINN model is trained with preset data until the PINN model converges. physical loss The computing process includes: for each sample in the unlabeled samples, first calculating the aging rate by the automatic differentiation interface; and based on the physical relationship, obtaining the physical loss as the mean square error of the deviation between the derivative calculation value and the theoretical value; Pre-set weight After that, the data loss and the physical loss are calculated respectively during each batch training, and the data loss and the physical loss are fused to obtain the total loss.

5. A multi-state aware cable fault prediction system as claimed in claim 4, wherein, weight The determination process is as follows: The physical consistency error is defined as The RUL derivative with respect to time in the validation set is defined as The proportion of samples with RUL derivative with respect to time ≥0 is normalized to obtain ; fitting error of data defined as the mean absolute error of the validation set, and after normalization ; As the weight increases, the power increases, the power decreases, while constraining ∈ [0.1, 0.3]; A linear combination function is obtained for the weights w = w1, w2,..., wn The PINN model is trained with preset data until the PINN model converges. is the base weight; , is an adjustment factor; is a step function; The PINN model is trained with preset data until the PINN model converges. A1 : compute the and ; A2: to , Normalization is performed to obtain and ; A3: Calculate weight by substituting weight function and update to the training process for the next round of total loss.

6. A multi-state aware cable fault prediction system as claimed in claim 4, wherein, The PINN model is trained with preset data until the PINN model converges. Using the Adam optimizer, the batch training size is 256 records / batch. Rate of change over 100 consecutive rounds < Training is stopped at this point, and a stable model file is generated.

7. A multi-state aware cable fault prediction system as claimed in claim 1, wherein, The PINN model is trained with preset data until the PINN model converges.

8. A multi-state aware cable fault prediction system as claimed in claim 1, wherein, The PINN model is trained with preset data until the PINN model converges.

9. A multi-state aware cable fault prediction method according to any one of claims 1-8, wherein, The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The data is divided into labeled and unlabeled, and the data loss and physical loss are calculated respectively, and the dynamic adjustment weight is used The total loss is fused, and the Adam optimizer is used for training; The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The PINN model is trained with preset data until the PINN model converges. The

Images