Wire and cable fault early warning system based on intelligent monitoring

Abstract

The invention relates to the technical field of power system monitoring, and discloses a wire and cable fault early warning system based on intelligent monitoring, which comprises a data sensing module, a multi-mode fusion module, a characteristic evolution module, an abnormal early warning module and a dynamic optimization module. The data sensing module collects multi-source heterogeneous data, the multi-modal fusion module processes the data to generate a spatial-temporal feature matrix, the feature evolution module extracts cable degradation features, the abnormity early warning module performs fault early warning based on the cable degradation features, and the dynamic optimization module optimizes system parameters by using federal learning. In addition, the system also comprises a digital twin mapping and topology analysis module for assisting decision making and enhancing positioning. According to the invention, real-time monitoring, accurate fault early warning and system performance optimization of the operation state of the wire and cable are realized, the stability and reliability of power transmission are improved, and the system has the advantages of comprehensive multi-source data acquisition, efficient data processing, accurate early warning, data privacy protection and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring, and in particular to a wire and cable fault early warning system based on intelligent monitoring. Background Art

[0002] In modern society, the stability and reliability of power supply are crucial to industrial production, residents' lives, and other aspects. As the key carriers of power transmission, the operating status of wires and cables directly affects the normal operation of the power system. However, traditional wire and cable fault monitoring and early warning methods have many limitations. Early manual inspection methods were inefficient and unable to achieve real-time monitoring. Inspectors needed to conduct regular inspections along the cable lines, which not only consumed a lot of manpower, material resources, and time, but also made it difficult to detect potential faults within the cables. When cables are long and distributed over a wide area, manual inspections are difficult to achieve comprehensive and detailed detection. Minor fault hazards are easily overlooked, resulting in the inability to detect and address faults in a timely manner, which in turn affects the continuity of power supply.

[0003] With technological advancements, some monitoring methods have begun to rely on a single type of sensor for data collection, such as monitoring only cable temperature or current. However, the operating status of wires and cables is affected by a combination of factors, and monitoring a single parameter cannot fully reflect the actual condition of the cable. For example, relying solely on temperature monitoring may not promptly detect potential faults caused by factors such as insulation aging and changes in mechanical stress. In some cases, cable temperature may not show obvious abnormalities at the beginning of a fault, but other parameters may have already changed.

[0004] Some existing monitoring systems also lack data processing and analysis capabilities. Faced with large amounts of monitoring data, they are unable to effectively integrate and deeply mine it, making it difficult to extract information that accurately reflects cable fault characteristics. Furthermore, traditional threshold judgment methods are too rigid and cannot be dynamically adjusted based on the actual operating conditions of the cable, making them prone to false positives or missed positives. The normal operating parameter range of a cable varies under different environmental conditions and operating conditions. Without adaptive threshold adjustment, it is impossible to accurately determine whether the cable is in a faulty state.

[0005] Furthermore, traditional approaches to system optimization and maintenance often require the centralized collection and processing of large amounts of data, which not only increases the burden of data transmission and storage but also poses the risk of data privacy leaks. Furthermore, when the number of monitoring points is large, centralized data processing leads to low computational efficiency and fails to meet real-time requirements. The rapid development of smart grids is placing higher demands on the intelligence, efficiency, and security of wire and cable fault warning systems. Existing monitoring and warning technologies are struggling to meet these demands, and a new technical solution is urgently needed to address these issues. Summary of the Invention

[0006] The purpose of the present invention is to provide a wire and cable fault early warning system based on intelligent monitoring to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a wire and cable fault early warning system based on intelligent monitoring, the system comprising:

[0008] It includes data perception module, multimodal fusion module, feature evolution module, anomaly warning module and dynamic optimization module;

[0009] The data perception module is used to collect multi-source heterogeneous data on the operating status of wires and cables in real time, specifically to obtain cable surface temperature data, internal current data, insulation layer voltage fluctuation data, ambient temperature and humidity data, and mechanical stress data by deploying a distributed sensor network, and synchronously transmit the multi-source heterogeneous data to the multimodal fusion module;

[0010] The multimodal fusion module is used to perform spatiotemporal alignment and fusion processing on multi-source heterogeneous data. Specifically, it uses a dynamic time warping algorithm to align time series data with different sampling frequencies, and generates a unified spatiotemporal feature matrix through a feature weighted fusion method based on an attention mechanism. The spatiotemporal feature matrix is ​​sent to the feature evolution module.

[0011] The feature evolution module is used to extract cable degradation trends and potential fault characteristics. Specifically, it constructs a multi-scale convolutional recursive network, which includes a spatiotemporal convolution submodule, a gated recurrent unit submodule, and a feature pyramid aggregation submodule. By performing multi-scale feature extraction and cross-layer information fusion on the spatiotemporal feature matrix, a cable degradation evolution feature map is generated.

[0012] The abnormal warning module is used to identify abnormal cable status and generate graded warning signals. Specifically, based on the cable degradation evolution feature map, it adopts an adaptive threshold decision model driven by deep reinforcement learning to dynamically adjust the fault judgment boundary and output warning level and location information;

[0013] The dynamic optimization module is used to optimize system parameters based on historical warning data. Specifically, it introduces a federated learning framework to update the parameters of the multi-scale convolutional recursive network and the adaptive threshold decision model through distributed model aggregation while protecting data privacy.

[0014] Preferably, in the data perception module, the distributed sensor network adopts a heterogeneous node deployment strategy, specifically including:

[0015] High-frequency sampling node, collecting current transient pulse data at a frequency of 1000 times per second;

[0016] The medium frequency sampling node collects cable surface temperature distribution data every 10 seconds;

[0017] Low-frequency sampling nodes collect ambient temperature, humidity, and mechanical stress data every 5 minutes;

[0018] The multi-source heterogeneous data is locally cached and timestamped by the edge computing node and then transmitted to the multimodal fusion module.

[0019] Preferably, the dynamic time warping algorithm in the multimodal fusion module specifically adopts improved constraints, including:

[0020] Define the sliding window size as 3 times the current data sampling period;

[0021] A local path cost function is introduced, which combines the weighted sum of Euclidean distance and Manhattan distance as a similarity measure;

[0022] The feature weighted fusion method of the attention mechanism specifically calculates the entropy weight of each modal data to generate an adaptive fusion weight matrix, which is expressed as follows:

[0023]

[0024] Among them, S i is the entropy value of the i-th type of data, τ is the temperature coefficient, and N is the total number of modes.

[0025] Preferably, the spatiotemporal convolution submodule in the multi-scale convolutional recurrent network adopts a combined structure of a hole convolution kernel and an expansion convolution kernel, with the hole rate set to [1, 2, 4], to capture the short-term local features and long-range dependencies of the cable data; the gated recurrent unit submodule introduces a time gating mechanism to dynamically adjust the contribution of historical states and current inputs through gating weights; the feature pyramid aggregation submodule adopts a bidirectional cross-scale connection to upsample and downsample the output feature maps of different convolutional layers and then splice them into a fused feature vector.

[0026] Preferably, the deep reinforcement learning-driven adaptive threshold decision model specifically includes:

[0027] State space: It is composed of the statistical quantities of the cable degradation evolution characteristic map, including kurtosis, skewness and Hurst index;

[0028] Action space: defined as the continuous adjustment interval of the warning threshold [α_min, α_max];

[0029] Reward function: It is designed as the weighted inverse of the number of false positives and the number of missed negatives. The formula is:

[0030]

[0031] Among them, N FP is the number of false alarms, N FN is the number of missed reports, w1 and w2 are dynamically adjusted penalty coefficients;

[0032] The model is trained online using a proximal policy optimization algorithm to update policy network parameters in real time.

[0033] Preferably, the federated learning framework adopts a hierarchical aggregation mechanism, specifically including:

[0034] Local client: trains the sub-model on each edge computing node and uses differential privacy technology to add Gaussian noise to the gradient;

[0035] Regional server: Performs weighted averaging of model parameters of clients within its jurisdiction, with the weight determined by the proportion of client data volume;

[0036] Central coordinator: Aggregates model parameters from all regional servers, generates a global model version, and records model update logs through blockchain technology.

[0037] Preferably, the system further includes a digital twin mapping module for constructing a virtual mirror image of the cable physical entity, and the specific steps include:

[0038] Establish a three-dimensional electromagnetic-thermal coupling simulation model of the cable based on the finite element analysis method;

[0039] Inputting the cable degradation evolution characteristic map into the simulation model to predict the insulation layer aging rate and local hot spot distribution;

[0040] By comparing the actual monitoring data with the simulation results, the cable health index is generated and fed back to the abnormal warning module.

[0041] Preferably, in the digital twin mapping module, a transfer learning strategy is used to optimize the simulation model, specifically:

[0042] Pre-training phase: training a basic finite element model on a standard cable dataset;

[0043] Fine-tuning stage: Using the degradation evolution feature map of the actual cable, the feature distribution of the simulation data and the measured data is aligned through the domain adversarial network.

[0044] Preferably, the system further includes a topology analysis module for processing the structural relevance of the cable network, specifically:

[0045] Constructing graph-structured data of cable connection relationships, where nodes represent cable segments and edges represent electrical connections and physical proximity relationships;

[0046] Graph attention network is used to extract topological features and calculate the probability of fault propagation between nodes;

[0047] The topological features and the cable degradation evolution feature map are tensor-spliced ​​and input into the anomaly warning module to enhance positioning accuracy.

[0048] Preferably, the dynamic optimization module further integrates a genetic algorithm for multi-objective parameter optimization, specifically including:

[0049] The fitness function is defined as the Pareto frontier of warning accuracy and calculation delay;

[0050] Adopting the elite retention strategy and non-dominated sorting selection operator;

[0051] The hyperparameter combination of the federated learning framework, including learning rate, aggregation frequency, and privacy budget parameters, is optimized through cross-mutation operations.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The data perception module of the present invention, through a distributed sensor network and heterogeneous node deployment strategy, can collect multi-source heterogeneous data such as cable surface temperature, internal current, insulation layer voltage fluctuations, ambient temperature and humidity, and mechanical stress in real time. Nodes with different sampling frequencies work together. High-frequency sampling nodes accurately capture the rapid changes in current transient pulses, medium-frequency sampling nodes promptly obtain dynamic information on the cable surface temperature distribution, and low-frequency sampling nodes effectively monitor the long-term changing trends of environmental factors. This design comprehensively and meticulously reflects the operating status of the cable, providing a rich and accurate data foundation for subsequent fault analysis. Compared with traditional single-parameter monitoring, it greatly improves data integrity and reliability.

[0054] The multimodal fusion module uses a dynamic time warping algorithm and a feature weighted fusion method based on an attention mechanism to perform spatiotemporal alignment and fusion processing on multi-source heterogeneous data. Improved constraints in the dynamic time warping algorithm, such as defining an appropriate sliding window size and combining Euclidean distance and Manhattan distance similarity metrics, can more accurately align time series data with different sampling frequencies. An adaptive fusion weight matrix calculated based on entropy weights enables each modal data to be adaptively fused based on its information content. The resulting unified spatiotemporal feature matrix better reflects the actual state of cable operation. The multi-scale convolutional recurrent network constructed by the feature evolution module effectively extracts cable degradation trends and potential fault characteristics through the collaborative work of spatiotemporal convolution submodules, gated recurrent unit submodules, and feature pyramid aggregation submodules. The resulting cable degradation evolution feature map provides a key basis for fault warning, improving the accuracy and comprehensiveness of fault feature extraction.

[0055] The abnormal warning module is based on the cable degradation evolution feature map and adopts an adaptive threshold decision model driven by deep reinforcement learning. By rationally defining the state space, action space, and reward function, this model can dynamically adjust the fault determination boundary. The state space is composed of statistics from the cable degradation evolution feature map, which comprehensively reflects the cable's operating status characteristics. The action space is a continuous adjustment interval for the warning threshold, allowing the model to find the optimal warning threshold based on different situations. The reward function is based on the weighted inverse of the number of false alarms and missed alarms, effectively guiding the model to optimize the warning strategy. Through online training using a proximal policy optimization algorithm, the model can update the policy network parameters in real time, continuously improving the accuracy and timeliness of warnings. At the same time, it outputs warning levels and location information, providing clear guidance for fault handling, reducing troubleshooting time, and improving the emergency response capabilities of the power system.

[0056] The dynamic optimization module introduces a federated learning framework and employs a hierarchical aggregation mechanism. Local clients train sub-models on edge computing nodes and use differential privacy techniques to protect data privacy. Regional servers perform a weighted average of model parameters across clients within their jurisdiction. The central coordinator aggregates and generates a global model version, using blockchain technology to log updates. This enables distributed model training and optimization while ensuring data privacy. A genetic algorithm is further integrated for multi-objective parameter optimization, using the Pareto frontier of early warning accuracy and computational latency as the fitness function. Hyperparameter combinations within the federated learning framework, such as the learning rate, aggregation frequency, and privacy budget, are optimized through an elite retention strategy, a non-dominated sorting selection operator, and crossover mutation operations, improving the overall performance and adaptability of the system.

[0057] The digital twin mapping module establishes a three-dimensional electromagnetic-thermal coupled simulation model of the cable based on finite element analysis. By inputting the cable degradation evolution characteristic map, it predicts the insulation aging rate and local hotspot distribution. Comparing this with actual monitoring data, it generates a cable health index and feeds it back to the anomaly warning module, providing a more comprehensive reference for fault warning and assisting in decision-making. The topology analysis module constructs graph-structured data of cable connection relationships, extracts topological features using a graph attention network, and calculates the probability of fault propagation between nodes. The topological features are then concatenated with the cable degradation evolution characteristic map tensor and fed into the anomaly warning module, effectively enhancing fault location accuracy and improving the monitoring and early warning capabilities for complex cable network faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a working principle diagram of the wire and cable fault early warning system of the present invention;

[0059] Figure 2 Flowchart for heterogeneous node deployment and data transmission of data perception modules;

[0060] Figure 3This is the schematic diagram of the key algorithm of the multimodal fusion module. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] See also Figure 1-3 The present invention provides a technical solution: a wire and cable fault early warning system based on intelligent monitoring. The system is mainly composed of a data perception module, a multimodal fusion module, a feature evolution module, an abnormality early warning module and a dynamic optimization module. The modules work together to achieve effective early warning of wire and cable faults.

[0063] The data perception module deploys a distributed sensor network to collect real-time, multi-source, heterogeneous data on the operating status of wires and cables, including cable surface temperature, internal current, insulation voltage fluctuations, ambient temperature and humidity, and mechanical stress data. Once collected, this data is synchronously transmitted to the multimodal fusion module.

[0064] After receiving multi-source heterogeneous data, the multimodal fusion module first uses the dynamic time warping algorithm to align the time series data of different sampling frequencies, and then generates a unified spatiotemporal feature matrix through a feature weighted fusion method based on the attention mechanism, and then sends the matrix to the feature evolution module.

[0065] The feature evolution module constructs a multi-scale convolutional recurrent network, which includes a spatiotemporal convolution submodule, a gated recurrent unit submodule, and a feature pyramid aggregation submodule. By extracting multi-scale features from the spatiotemporal feature matrix and fusing cross-layer information, it generates a cable degradation evolution feature map.

[0066] The abnormal warning module is based on the cable degradation evolution feature map and adopts an adaptive threshold decision model driven by deep reinforcement learning to dynamically adjust the fault judgment boundary and then output the warning level and positioning information.

[0067] The dynamic optimization module introduces a federated learning framework to update the parameters of the multi-scale convolutional recursive network and the adaptive threshold decision model through distributed model aggregation while protecting data privacy.

[0068] The present invention will be further described below in conjunction with Examples 1 to 5:

[0069] Example 1:

[0070] This embodiment aims to elaborate on the heterogeneous node deployment strategy and data transmission method of the distributed sensor network in the data perception module in detail, so as to achieve efficient collection and preliminary processing of multi-source heterogeneous data.

[0071] In the data perception module, the distributed sensor network adopts a heterogeneous node deployment strategy. High-frequency sampling nodes collect current transient pulse data at a frequency of 1000 times per second. This is because current transient pulse data changes rapidly and contains a large amount of critical information reflecting the cable's operating status. The high sampling frequency accurately captures these transient changes, providing detailed data support for subsequent analysis. For example, when a fault such as a momentary short circuit or poor contact occurs within the cable, the current transient pulse will exhibit abnormal fluctuations. High-frequency sampling can promptly detect these anomalies.

[0072] The medium-frequency sampling node collects cable surface temperature distribution data every 10 seconds. Cable surface temperature is a key indicator of cable operation. While its changes are slower than those of transient current pulses, its dynamic changes must be monitored promptly. A 10-second sampling interval satisfies the need to monitor temperature trends without creating excessive data redundancy and storage and transmission pressures caused by overly frequent sampling.

[0073] Low-frequency sampling nodes collect ambient temperature, humidity, and mechanical stress data every five minutes. Ambient temperature, humidity, and mechanical stress typically change slowly, and a five-minute sampling frequency is sufficient to capture their changing patterns. For example, a slow rise in ambient temperature may affect the heat dissipation performance of a cable, while long-term mechanical stress may damage the cable structure. Low-frequency sampling ensures the acquisition of valid information while reducing the system's sampling costs and data processing burden.

[0074] Collected multi-source, heterogeneous data is locally cached and timestamped by edge computing nodes. Edge computing nodes possess sufficient computing and storage capabilities, and local data caching mitigates the risk of data loss due to network transmission failures. Timestamping also provides accurate time information for the data, facilitating subsequent spatiotemporal alignment within the multimodal fusion module. For example, in subsequent processing, data collected by different nodes can be accurately aligned to the same moment based on timestamps, providing a foundation for analyzing the comprehensive operating status of the cable at a specific moment.

[0075] Example 2:

[0076] In the multimodal fusion module, the dynamic time warping algorithm employs improved constraints. First, the sliding window size is defined as three times the current data sampling period. This setting is based on an analysis of data with different sampling frequencies. For example, high-frequency and low-frequency data have short sampling periods and dense data points, while low-frequency data have long sampling periods and sparse data points. If the sliding window is too small, important information may be missed when aligning high-frequency and low-frequency data. If the sliding window is too large, excessive unnecessary data will be introduced, increasing the computational effort and affecting the alignment effect. Setting the sliding window size to three times the current data sampling period effectively balances computational efficiency while ensuring alignment accuracy. For example, for high-frequency current transient pulse data, the short sampling period necessitates a smaller sliding window, enabling precise alignment of adjacent data points. For low-frequency ambient temperature and humidity data, a larger sliding window covers a sufficient data range, ensuring that a suitable alignment path can be found.

[0077] Secondly, we introduce the local path cost function and combine the weighted sum of Euclidean distance and Manhattan distance as a similarity measure. Euclidean distance can measure the absolute distance of data in space, while Manhattan distance focuses more on the sum of the differences in each dimension of the data. By combining the two, we can more comprehensively reflect the similarity between different modal data. Suppose two data points are X = (x1, x2, ..., x n ) and Y=(y1,y2,…,y n ), then the Euclidean distance Manhattan distance Local path cost function C(X,Y)=w1d E +w2d M , where w1 and w2 are weighting coefficients that can be adjusted based on the characteristics of different modal data. For example, for data with relatively stable fluctuations, the weight of Euclidean distance can be appropriately increased; for data with more drastic fluctuations, the weight of Manhattan distance can be increased.

[0078] The feature weighted fusion method based on the attention mechanism generates an adaptive fusion weight matrix by calculating the entropy weight of each modal data. The formula is Among them, S iis the entropy value of the i-th category of data, reflecting the amount of information contained in that data. A larger entropy value indicates greater uncertainty in the data and should be assigned a greater weight during the fusion process. τ is the temperature coefficient, which adjusts the smoothness of the weights and can generally be set based on empirical values, such as 1. N is the total number of modalities. The weight matrix calculated using this formula can adaptively adjust the fusion ratio based on the importance of each modal data, allowing the fused features to more accurately reflect the actual operating status of the cable. For example, when a localized overheating fault occurs in the cable, the entropy of the cable surface temperature data increases, and its corresponding weight also increases, thereby highlighting the influence of the temperature data in the fused features.

[0079] Example 3:

[0080] In the multi-scale convolutional recurrent network, the spatiotemporal convolution submodule uses a combination of dilated and atrous convolution kernels, with dilation ratios set to [1, 2, 4]. The dilated convolution kernel expands the receptive field of the convolution kernel without increasing the number of parameters or computational complexity, thereby capturing long-range dependencies in cable data. For example, dilated convolution can effectively capture long-term trends in cable operation, such as parameter changes caused by gradual aging of the insulation layer. The dilated convolution kernel further enhances the ability to extract features at different scales. Combining convolution kernels with different dilation ratios can simultaneously capture short-term local features and long-range dependencies in cable data. When the dilation ratio is 1, the convolution kernel behaves like a normal convolution kernel, primarily capturing local details. Attrition ratios of 2 and 4 capture a wider range of information at different scales, respectively.

[0081] The gated recurrent unit submodule introduces a time gating mechanism. This mechanism dynamically adjusts the contribution of historical states and current inputs through gating weights. Suppose the input of the gated recurrent unit is x t , the hidden state at the previous moment is h t-1 , update gate z t and reset gate r t The calculation formulas are:

[0082] z t =σ(W z x t +U z h t-1 )

[0083] r t =σ(W r x t +U r h t-1 )

[0084] Among them, σ is the sigmoid activation function, W z 、U z、W r 、U r is the weight matrix. Through the interaction of the update gate and reset gate, the gated recurrent unit selectively retains historical information and accepts new input information. When the cable's operating status is relatively stable, the update gate has a larger value, retaining more historical state information. When an abnormality occurs, the reset gate adjusts the information transfer so that the current input information can more effectively influence the hidden state, thereby better adapting to the dynamic changes in the cable's operating status.

[0085] The feature pyramid aggregation submodule utilizes bidirectional cross-scale connections. It upsamples and downsamples the output feature maps of different convolutional layers and then concatenates them into a fused feature vector. Upsampling amplifies low-resolution feature maps through methods like interpolation, while downsampling reduces the resolution of feature maps through methods like pooling. For example, a low-resolution feature map with rich detail information from a shallow convolutional layer is upsampled and concatenated with a high-resolution feature map with abstract semantic information from a deep convolutional layer. This fused feature vector incorporates both local, detailed features and global, abstract features of the cable data, more comprehensively reflecting the degradation and evolution of the cable and providing a more accurate basis for subsequent anomaly warnings.

[0086] Example 4:

[0087] In the adaptive threshold decision model driven by deep reinforcement learning, the state space is composed of statistical measures of the cable degradation evolution characteristic graph, including kurtosis, skewness, and the Hurst exponent. Kurtosis measures the peakedness of the data distribution. The kurtosis of the degradation evolution characteristic graph of a cable under normal operation and fault conditions can differ significantly. For example, when a cable experiences a fault such as partial discharge, the kurtosis of the characteristic graph may increase, reflecting an abnormal concentration in the data distribution. Skewness describes the asymmetry of the data distribution and can further supplement the data distribution information. The Hurst exponent characterizes the long-term memory and trend characteristics of a time series and is important for analyzing long-term trends in cable operating conditions. Using these statistics as a representation of the state space can more comprehensively reflect the cable's operating characteristics.

[0088] The action space is defined as the continuous adjustment interval [α_min, α_max] of the warning threshold. In practice, properly setting the warning threshold is crucial. If the threshold is set too high, it may result in missed alerts and fail to detect potential failures in a timely manner. If the threshold is set too low, it may generate excessive false positives, affecting system reliability. By setting the action space as a continuous adjustment interval, the model can dynamically find the optimal warning threshold based on different states during training.

[0089] The reward function is designed as the weighted inverse of the number of false positives and the number of missed negatives, and the formula is

[0090]

[0091] Among them, N FP is the number of false alarms, N FN is the number of missed alerts, and w1 and w2 are dynamically adjusted penalty coefficients. In the early stages of training, the penalty coefficients can be appropriately lowered to quickly explore different threshold settings. As training progresses, the penalty coefficients are gradually increased to improve model accuracy. This reward function design allows the model to continuously optimize its warning strategy during training, reducing the number of false alerts and missed alerts.

[0092] The model is trained online using a proximal policy optimization algorithm, updating the policy network parameters in real time. The proximal policy optimization algorithm is a highly efficient reinforcement learning algorithm that avoids the performance degradation often associated with excessive updates in traditional policy gradient algorithms by limiting the step size of policy updates. During training, the model selects actions based on its current state (i.e., adjusts the warning threshold), receives feedback based on a reward function, and then updates the policy network parameters using the proximal policy optimization algorithm. This allows the model to gradually learn the optimal warning strategy, improving its ability to identify and warn of cable anomalies.

[0093] Example 5:

[0094] This embodiment covers the hierarchical aggregation mechanism of the federated learning framework in the dynamic optimization module, the multi-objective parameter optimization of the integrated genetic algorithm, and the specific implementation of the digital twin mapping module and the topology analysis module to improve the overall performance and fault warning capabilities of the system.

[0095] In the dynamic optimization module, the federated learning framework adopts a hierarchical aggregation mechanism. The local client trains the sub-model on each edge computing node. To protect data privacy, differential privacy technology is used to add Gaussian noise to the gradient. Differential privacy technology adds a certain amount of noise to the gradient, making it difficult for attackers to infer the original data from the gradient information. Let the gradient be g, and the added Gaussian noise be ∈, which has a mean of 0 and a variance of σ 2 Gaussian distribution N(0,σ 2 ), then the gradient after adding noise is g ′ =g+∈. By adjusting the variance of the noise σ 2 , you can control the balance between the strength of privacy protection and the accuracy of model training.

[0096] The regional server performs a weighted average of the model parameters of the clients within its jurisdiction, with the weight determined by the client data volume. This allows clients with larger data volumes to contribute more to model updates, making better use of data from different clients. For example, if an edge computing node collects a large amount of data, it indicates that the cable operation data in the area where the node is located is more representative, and the corresponding model parameters will be given a higher weight during the regional server's aggregation process.

[0097] The central coordinator aggregates model parameters from all regional servers and generates a global model version, logging model updates via blockchain technology. The decentralized and tamper-proof nature of blockchain technology ensures the security and traceability of the model update process. A record of each model update is added to the blockchain, including information such as the update time, participating regional servers, and updated parameters, facilitating subsequent auditing and analysis of the model's evolution.

[0098] The dynamic optimization module further integrates a genetic algorithm for multi-objective parameter optimization. The fitness function is defined as the Pareto front of early warning accuracy and computational latency. Early warning accuracy reflects the accuracy of the system's cable fault warnings, while computational latency affects the system's real-time performance. The definition of the Pareto front allows for trade-offs between multiple objectives to find the optimal solution that meets different requirements. An elitist retention strategy and a non-dominated sorting selection operator are employed. The elitist retention strategy ensures that the best individuals in each generation are directly inherited to the next generation, preventing the loss of excellent solutions. The non-dominated sorting selection operator stratifies individuals in the population based on non-domination relationships, prioritizing individuals in the best strata for genetic operations, thereby improving the algorithm's search efficiency. Crossover and mutation operations are used to optimize the hyperparameter combination of the federated learning framework, including the learning rate, aggregation frequency, and privacy budget parameter. For example, the learning rate influences the convergence speed of model training, the aggregation frequency determines the frequency of model updates, and the privacy budget parameter controls the degree of privacy protection. Continuously adjusting these hyperparameters through the genetic algorithm can improve the overall performance of the system.

[0099] The system also includes a digital twin mapping module. This module is used to construct a virtual mirror image of the cable's physical entity and establish a three-dimensional electromagnetic-thermal coupling simulation model of the cable based on the finite element analysis method. The finite element analysis method divides the cable model into multiple small units and simulates the physical characteristics of the cable under different operating conditions by solving the electromagnetic and thermal equations of these units. The cable degradation evolution characteristic map is input into the simulation model to predict the aging rate of the insulation layer and the distribution of local hot spots. By comparing the actual monitoring data with the simulation results, the cable health index is generated and fed back to the abnormal warning module. For example, if the simulation results show that the insulation layer in a certain area is aging faster, and the temperature in the area in the actual monitoring data is also high, it means that there is a potential fault risk in the area, the health index will be reduced accordingly, and the abnormal warning module can issue an early warning based on this.

[0100] To optimize the simulation model, the digital twin mapping module employs a transfer learning strategy. During the pre-training phase, a basic finite element model is trained on a standard cable dataset to learn the general physical properties and patterns of cables. During the fine-tuning phase, the degradation evolution characteristic maps of actual cables are utilized to align the feature distributions of the simulated and measured data using a domain adversarial network. The domain adversarial network consists of a generator and a discriminator. The generator attempts to generate simulated data with a feature distribution similar to that of the measured data, while the discriminator distinguishes between the generated data and the measured data. Through adversarial training between the two, the simulation model can better adapt to the actual operation of cables, improving prediction accuracy.

[0101] The system also includes a topology analysis module. This module is used to process the structural relevance of the cable network and construct graph-structured data of cable connection relationships. Nodes represent cable segments, and edges represent electrical connections and physical proximity. This graph structure can intuitively reflect the topology of the cable network. A graph attention network is used to extract topological features and calculate the probability of fault propagation between nodes. The graph attention network calculates the attention weights between nodes, highlighting node information that is closely associated with the current node, thereby better extracting topological features. The topological features are tensor-concatenated with the cable degradation evolution feature map and input into the anomaly warning module to enhance positioning accuracy. For example, when a cable segment fails, the topology analysis module can quickly locate other cable segments that may be affected based on the fault propagation probability. Combined with the cable degradation evolution feature map, it can more accurately determine the scope and severity of the fault, thereby improving the accuracy and reliability of the anomaly warning.

[0102] In practical application scenarios, such as a power cable network within a large industrial park, the various modules of the system work closely together. Various sensors in the data perception module continuously collect cable operation data. Nodes with different sampling frequencies perform their respective functions, comprehensively acquiring information about the cable and its surrounding environment. This information is pre-processed by edge computing nodes before transmission. The multimodal fusion module processes this multi-source, heterogeneous data, providing a unified spatiotemporal feature matrix for subsequent analysis. The feature evolution module identifies cable degradation trends and potential fault characteristics, generating a graph that provides a key basis for anomaly warnings. The anomaly warning module utilizes an adaptive threshold decision model to determine cable anomalies and promptly issues graded warning signals. The dynamic optimization module leverages federated learning and genetic algorithms to continuously optimize system parameters and improve overall performance.

[0103] The digital twin mapping module creates a virtual image that simulates cable operation in real time, comparing simulated data with actual data to provide cable health assessments and assist in anomaly warnings. The topology analysis module processes the cable network structure, analyzes the likelihood of fault propagation, and enhances fault location accuracy. By working together, these modules enable the fault warning system to effectively monitor cable operation, proactively detect potential faults, ensure power supply reliability, and reduce production interruptions and economic losses caused by cable faults, playing a vital role in ensuring power supply security.

[0104] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0105] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The wire and cable fault early warning system based on intelligent monitoring is characterized by: It includes data perception module, multimodal fusion module, feature evolution module, anomaly warning module and dynamic optimization module; The data perception module is used to collect multi-source heterogeneous data on the operating status of wires and cables in real time, specifically to obtain cable surface temperature data, internal current data, insulation layer voltage fluctuation data, ambient temperature and humidity data, and mechanical stress data by deploying a distributed sensor network, and synchronously transmit the multi-source heterogeneous data to the multimodal fusion module; The multimodal fusion module is used to perform spatiotemporal alignment and fusion processing on multi-source heterogeneous data. Specifically, it uses a dynamic time warping algorithm to align time series data with different sampling frequencies, and generates a unified spatiotemporal feature matrix through a feature weighted fusion method based on an attention mechanism. The spatiotemporal feature matrix is ​​sent to the feature evolution module. The feature evolution module is used to extract cable degradation trends and potential fault characteristics. Specifically, it constructs a multi-scale convolutional recursive network, which includes a spatiotemporal convolution submodule, a gated recurrent unit submodule, and a feature pyramid aggregation submodule. By performing multi-scale feature extraction and cross-layer information fusion on the spatiotemporal feature matrix, a cable degradation evolution feature map is generated. The abnormal warning module is used to identify abnormal cable status and generate graded warning signals. Specifically, based on the cable degradation evolution feature map, it adopts an adaptive threshold decision model driven by deep reinforcement learning to dynamically adjust the fault judgment boundary and output warning level and location information; The dynamic optimization module is used to optimize system parameters based on historical warning data. Specifically, it introduces a federated learning framework to update the parameters of the multi-scale convolutional recursive network and the adaptive threshold decision model through distributed model aggregation while protecting data privacy.

2. The wire and cable fault early warning system based on intelligent monitoring according to claim 1 is characterized in that: In the data perception module, the distributed sensor network adopts a heterogeneous node deployment strategy, specifically including: High-frequency sampling node, collecting current transient pulse data at a frequency of 1000 times per second; The medium frequency sampling node collects cable surface temperature distribution data every 10 seconds; Low-frequency sampling nodes collect ambient temperature, humidity, and mechanical stress data every 5 minutes; The multi-source heterogeneous data is locally cached and timestamped by the edge computing node and then transmitted to the multimodal fusion module.

3. The wire and cable fault early warning system based on intelligent monitoring according to claim 2 is characterized in that: The dynamic time warping algorithm in the multimodal fusion module specifically adopts improved constraints, including: Define the sliding window size as 3 times the current data sampling period; A local path cost function is introduced, which combines the weighted sum of Euclidean distance and Manhattan distance as a similarity measure; The feature weighted fusion method of the attention mechanism specifically calculates the entropy weight of each modal data to generate an adaptive fusion weight matrix, which is expressed as follows: Among them, S i is the entropy value of the i-th type of data, τ is the temperature coefficient, and N is the total number of modes.

4. The wire and cable fault early warning system based on intelligent monitoring according to claim 1 is characterized in that: The spatiotemporal convolution submodule in the multi-scale convolutional recurrent network adopts a combination of a dilated convolution kernel and a hole convolution kernel, with the hole ratio set to [1, 2, 4], to capture the short-term local features and long-range dependencies of cable data. The gated recurrent unit submodule introduces a time gating mechanism to dynamically adjust the contribution of historical states and current inputs through gating weights. The feature pyramid aggregation submodule adopts a bidirectional cross-scale connection to upsample and downsample the output feature maps of different convolutional layers and then splice them into a fused feature vector.

5. The wire and cable fault early warning system based on intelligent monitoring according to claim 4 is characterized in that: The deep reinforcement learning-driven adaptive threshold decision model specifically includes: State space: It is composed of the statistical quantities of the cable degradation evolution characteristic map, including kurtosis, skewness and Hurst index; Action space: defined as the continuous adjustment interval of the warning threshold [α_min, α_max]; Reward function: It is designed as the weighted inverse of the number of false positives and the number of missed negatives. The formula is: Among them, N FP is the number of false alarms, N FN is the number of missed reports, w1 and w2 are dynamically adjusted penalty coefficients; The model is trained online using a proximal policy optimization algorithm to update policy network parameters in real time.

6. The wire and cable fault early warning system based on intelligent monitoring according to claim 1 is characterized in that: The federated learning framework adopts a hierarchical aggregation mechanism, which specifically includes: Local client: trains the sub-model on each edge computing node and uses differential privacy technology to add Gaussian noise to the gradient; Regional server: Performs weighted averaging of model parameters of clients within its jurisdiction, with the weight determined by the proportion of client data volume; Central coordinator: Aggregates model parameters from all regional servers, generates a global model version, and records model update logs through blockchain technology.

7. The wire and cable fault early warning system based on intelligent monitoring according to claim 1 is characterized in that: The system also includes a digital twin mapping module for building a virtual mirror image of the cable physical entity, which includes the following steps: Establish a three-dimensional electromagnetic-thermal coupling simulation model of the cable based on the finite element analysis method; Inputting the cable degradation evolution characteristic map into the simulation model to predict the insulation layer aging rate and local hot spot distribution; By comparing the actual monitoring data with the simulation results, the cable health index is generated and fed back to the abnormal warning module.

8. The wire and cable fault early warning system based on intelligent monitoring according to claim 7 is characterized in that: In the digital twin mapping module, a transfer learning strategy is used to optimize the simulation model, specifically: Pre-training phase: training a basic finite element model on a standard cable dataset; Fine-tuning stage: Using the degradation evolution feature map of the actual cable, the feature distribution of the simulation data and the measured data is aligned through the domain adversarial network.

9. The wire and cable fault early warning system based on intelligent monitoring according to claim 1 is characterized in that: The system also includes a topology analysis module for processing the structural relevance of the cable network, specifically: Constructing graph-structured data of cable connection relationships, where nodes represent cable segments and edges represent electrical connections and physical proximity relationships; Graph attention network is used to extract topological features and calculate the probability of fault propagation between nodes; The topological features and the cable degradation evolution feature map are tensor-spliced ​​and input into the anomaly warning module to enhance positioning accuracy.

10. The wire and cable fault early warning system based on intelligent monitoring according to claim 1, characterized in that: The dynamic optimization module further integrates a genetic algorithm for multi-objective parameter optimization, specifically including: The fitness function is defined as the Pareto frontier of warning accuracy and calculation delay; Adopting the elite retention strategy and non-dominated sorting selection operator; The hyperparameter combination of the federated learning framework, including learning rate, aggregation frequency, and privacy budget parameters, is optimized through cross-mutation operations.

Images