Cable fault analysis method and system based on edge cloud collaboration

Abstract

The invention discloses a cable fault analysis method and system based on edge cloud collaboration, and relates to the technical field of intelligent power grid data processing, and the method comprises the steps: building and training a full-quantity fault analysis model at a cloud end based on historical cable fault data; deriving a first light fault analysis model and a second light fault analysis model through model compression, and deploying the first light fault analysis model and the second light fault analysis model at an edge side and a cloud; real-time state data of a target cable are collected and input into the edge side model to obtain a preliminary diagnosis result, then the result is fed back to the cloud model for collaborative diagnosis, and a final fault analysis result is output, so that the technical effects of bandwidth optimization utilization, diagnosis efficiency improvement and diagnosis accuracy enhancement are achieved.

Description

Technical Field

[0001] This invention relates to the field of smart grid data processing technology, and in particular to a cable fault analysis method and system based on edge cloud collaboration. Background Technology

[0002] As the core carrier of power transmission, the real-time monitoring of cable operating status and rapid fault diagnosis have become key technologies for ensuring the safe and stable operation of the power grid. Currently, cable fault analysis is mostly achieved through cloud computing or edge computing. Cloud computing utilizes the powerful computing capabilities and massive historical data in the cloud to build complex fault diagnosis models, achieving high diagnostic accuracy. Edge computing, on the other hand, pushes analytical capabilities down to the cable monitoring terminal, achieving low-latency response through local data processing. Meanwhile, with the advancement of smart grid transformation and construction, edge-cloud collaborative architectures are gradually being applied to the field of fault analysis, attempting to balance the dual requirements of real-time performance and accuracy through a combination of edge preprocessing and cloud-based deep analysis.

[0003] However, existing edge-cloud collaborative cable fault analysis technologies still have significant shortcomings: their collaborative mechanism is crude, with the transmission between the edge and the cloud mostly consisting of direct transmission of raw data or simple features, and cloud analysis and edge judgment are relatively isolated, resulting in technical problems such as wasted network bandwidth, low diagnostic efficiency, and impact on diagnostic accuracy. Summary of the Invention

[0004] This invention provides a cable fault analysis method and system based on edge cloud collaboration to solve the technical problems of network bandwidth waste, low diagnostic efficiency and reduced diagnostic accuracy in the prior art, and achieve the technical effects of optimized bandwidth utilization, improved diagnostic efficiency and enhanced diagnostic accuracy.

[0005] In a first aspect, the present invention provides a cable fault analysis method based on edge cloud collaboration, wherein the cable fault analysis method based on edge cloud collaboration includes: Based on historical cable fault data, a full fault analysis model is built and trained in the cloud.

[0006] Using model compression technology, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model. The first lightweight fault analysis model is deployed to the edge, and the second lightweight fault analysis model is deployed to the cloud.

[0007] Real-time status data of the target cable is collected and input into the first lightweight fault analysis model to obtain the first diagnostic result. The first diagnostic result is then fed back to the cloud for collaborative diagnosis to obtain the fault analysis result.

[0008] In one feasible implementation, a full-scale fault analysis model is built and trained in the cloud based on historical cable fault data, including: Obtain historical cable fault logs for the target cable specification, and extract those with true fault types as the initial fault sample set.

[0009] Define the type probability of each fault type based on the relative importance of each fault type.

[0010] Based on the time distance, the temporal correlation of each initial fault sample is calculated, and the temporal selection probability of each initial fault sample is defined accordingly.

[0011] Based on the type probability and the time sequence selection probability, iterative random selection without replacement is performed on the initial fault sample set until the selection result meets the preset sample size constraint, and the selection result is output as the historical cable fault data.

[0012] In one feasible implementation, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model using model compression technology, including: Based on a pre-defined list of first-order feature dimensions, the output dimensions of the full training logs of the full fault analysis model are pruned to obtain first training data, wherein the first training data includes hard labels of historical cable fault data and soft labels of the full training logs.

[0013] A lightweight fault analysis model is constructed based on the computing power constraints at the edge. The lightweight fault analysis model is defined as taking first-order features as input and fault existence probability and fault type distribution as output.

[0014] Based on the knowledge distillation method, the lightweight fault analysis model is trained and validated using the first training data to obtain the first lightweight fault analysis model.

[0015] The list of first-order feature dimensions includes a valley set and a peak set.

[0016] In one feasible implementation, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model using model compression technology, including: Obtain the verification dataset and extract key information from it.

[0017] The verification dataset is input into the first lightweight fault analysis model to obtain the first verification output, which is then combined with the key information extraction results to define the second training input data.

[0018] The verification dataset is input into the full-scale fault analysis model to obtain the full-scale verification output and define it as the second training output data, wherein the full-scale verification output includes soft labels.

[0019] Based on the knowledge distillation method, using the full fault analysis model as the teacher model, and combining the second training input data and the second training output data, a student model is constructed, trained, and validated.

[0020] When the verification results meet the preset model accuracy requirements, the student model is output as the second lightweight fault analysis model.

[0021] In one feasible implementation, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model using model compression technology, and the method further includes: Obtain the verification dataset and extract key information from it.

[0022] Using the first validation output of the first lightweight fault analysis model on the validation dataset as the input feature, and the hard labels of the historical cable fault data corresponding to the validation dataset as the output target, a second training dataset is constructed.

[0023] Based on the transfer learning method, the full fault analysis model is trained and fine-tuned using the second training dataset to obtain the second lightweight fault analysis model.

[0024] In one feasible implementation, real-time status data of the target cable is collected and input into the first lightweight fault analysis model to obtain a first diagnostic result, including: The real-time status data of the target cable is collected through the cable monitoring terminal on the edge side, wherein the real-time status data includes at least real-time waveform data.

[0025] The real-time status data is preprocessed, wherein the preprocessing includes at least one of ROI localization, image extraction, and image enhancement.

[0026] The preprocessed data is input into the first lightweight fault analysis model to obtain the first diagnostic result, wherein the first diagnostic result includes the probability of fault existence and the distribution of the first fault type.

[0027] In one feasible implementation, the first diagnostic result is fed back to the cloud for collaborative diagnostics to obtain fault analysis results, including: If the first diagnostic result does not meet the preset significance threshold, the diagnostic process is terminated, and the fault analysis result is fed back to the cloud as no cable fault.

[0028] According to the preset model update constraints, the new first model parameters obtained by federated learning are synchronized from the cloud to the edge, and the first lightweight fault analysis model is updated accordingly.

[0029] The model update constraints include: Periodic update constraint, which is used to define the period length for periodically updating the first lightweight fault analysis model.

[0030] Sudden update constraint, which defines the amount of change in model parameters when the first lightweight fault analysis model is suddenly updated beyond the periodic update constraint.

[0031] In one feasible implementation, feeding back the first diagnostic result to the cloud for collaborative diagnostics and obtaining fault analysis results further includes: If the first diagnostic result meets the preset significance threshold, the edge analysis result is fed back to the cloud as a suspected cable fault.

[0032] The edge device extracts key information from the real-time status data and merges the first diagnostic result and the key information extraction result into the cloud.

[0033] The cloud inputs the first diagnostic result and the key information extraction result into the second lightweight fault analysis model to obtain and output the fault analysis result, wherein the fault analysis result includes at least the fault type.

[0034] In one feasible implementation, the temporal correlation of each initial fault sample is calculated based on the time distance, and the temporal selection probability of each initial fault sample is defined accordingly, including: Calculate the distance between the timestamp of each initial fault sample and the current time, and obtain the time distance.

[0035] By combining a preset nonlinear mapping relationship, the time distance is mapped to the timing selection probability of each initial fault sample.

[0036] The nonlinear mapping relationship is based on any one of the following: exponential decay relationship, piecewise constant, and decay relationship.

[0037] Secondly, the present invention also provides a cable fault analysis system based on edge cloud collaboration, wherein the cable fault analysis system based on edge cloud collaboration includes: The cloud-based model training module is used to build and train a full-scale fault analysis model in the cloud based on historical cable fault data.

[0038] The lightweight model derivation and deployment module is used to derive a first lightweight fault analysis model and a second lightweight fault analysis model from the full fault analysis model through model compression technology, and deploy the first lightweight fault analysis model to the edge side and deploy the second lightweight fault analysis model to the cloud.

[0039] The real-time cable fault diagnosis and collaboration module is used to collect real-time status data of the target cable, input it into the first lightweight fault analysis model, obtain the first diagnosis result, and feed back the first diagnosis result to the cloud to perform collaborative diagnosis and obtain fault analysis results.

[0040] This invention discloses a cable fault analysis method and system based on edge-cloud collaboration, comprising: constructing and training a full fault analysis model in the cloud based on historical cable fault data; deriving a first lightweight fault analysis model and a second lightweight fault analysis model from the full fault analysis model using a model compression method, deploying the first lightweight fault analysis model on the edge side and the second lightweight fault analysis model on the cloud; collecting real-time operating status data of the target cable, inputting the real-time operating status data into the first lightweight fault analysis model to obtain preliminary diagnostic results, and feeding the preliminary diagnostic results back to the cloud for collaborative analysis to obtain the final fault analysis results. The cable fault analysis method and system based on edge-cloud collaboration disclosed in this invention solves the technical problems of network bandwidth waste, low diagnostic efficiency, and reduced diagnostic accuracy, achieving the technical effects of optimized bandwidth utilization, improved diagnostic efficiency, and enhanced diagnostic accuracy. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the cable fault analysis method based on edge cloud collaboration of the present invention. Figure 2 This is a schematic diagram of the cable fault analysis system based on edge cloud collaboration of the present invention.

[0042] The components represented by each number in the attached diagram are explained below: Cloud-based model training module 11, lightweight model derivation and deployment module 12, and real-time cable fault diagnosis and collaboration module 13. Detailed Implementation

[0043] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0044] Example 1, as Figure 1This is a flowchart illustrating the cable fault analysis method based on edge cloud collaboration according to the present invention, wherein the cable fault analysis method based on edge cloud collaboration includes: S100: Based on historical cable fault data, a full fault analysis model is built and trained in the cloud.

[0045] Specifically, historical cable fault data is information accumulated through monitoring and maintenance, including cable operating status (especially waveform), fault type, fault location, environmental conditions, and maintenance records. Ideally, this includes time-series data and tagged data. The cloud, on the other hand, is a remote computing and storage environment provided by a cloud computing platform, possessing high computing power and well-suited for large-scale data processing and model training.

[0046] Specifically, the full-scale fault analysis model is an intelligent model trained using all historical data to identify, classify, and predict cable faults. Preferably, this full-scale fault analysis model is built based on machine learning or deep learning algorithms and has the ability to identify and generalize multiple fault modes.

[0047] Preferably, the full-scale fault analysis model uses both first-order and second-order features obtained from historical cable fault data analysis as input. Specifically, first-order features refer to basic statistical or structural features extracted directly from the original waveform signals of historical cable fault data, such as discrete height sequences, peaks, and troughs, which can reflect information such as local extrema and trends of the signal. Among them, discrete height sequences refer to the sequence data composed of discrete sampled values ​​of physical quantities such as current during cable operation. Troughs are local minimum points in the sequence, and peaks are local maximum points, which together reveal the periodicity and abnormal change patterns of the signal.

[0048] Specifically, second-order features are higher-order information such as the rate of change and abrupt change points, which are calculated further based on first-order features. They can be used to identify the location of sudden jumps in a signal, thereby helping to capture potential fault precursors or structural changes.

[0049] For example, firstly, a multi-type sensor network deployed at cable joints, terminals, and critical paths in the smart grid is used to collect historical status data covering at least one year of operation. The sampling frequency is adjustable from 10kHz to 1MHz. The collected raw waveform data undergoes noise reduction filtering, baseline correction, and normalization preprocessing to construct a structured dataset. Each data sample contains synchronous multi-channel waveform data with a time window length of 1 second and the corresponding fault label. Subsequently, a full-scale fault analysis model based on a Transformer encoder-decoder architecture is built on a cloud GPU cluster using a TensorFlow or PyTorch deep learning framework. The model input dimensions correspond to the number of first-order and second-order features, the input sequence length is set to 10,000 sampling points, the number of encoder layers is set to 12, and the number of attention heads is set to [missing information]. 16. The hidden layer dimension is set to 1024. The decoder output layer uses the Softmax activation function to map to the fault category space. The fault categories include at least the subcategories of short-circuit faults, grounding faults, partial discharge, insulation aging, mechanical damage, and composite faults. The training process uses the AdamW optimizer with an initial learning rate of 0.0001, a cosine annealing scheduling strategy, a batch size of 256, and a training cycle of 200 rounds. Dropout regularization is used to prevent overfitting with a coefficient of 0.3. An early stopping strategy is used on the validation set to monitor model performance. After training, if the model achieves a fault type recognition accuracy of over 95% and a fault location accuracy of over 90% on the test set, it is considered a teacher model with comprehensive fault pattern recognition capabilities, providing a knowledge source for subsequent model compression and distillation.

[0050] Through the above process, the ability to build high-precision fault analysis models based on full historical data can be achieved, fully utilizing the powerful computing and storage resources of the cloud to improve model training efficiency and generalization ability, and providing a reliable model foundation for the subsequent construction of lightweight models on the edge. At the same time, the unified modeling approach based on the cloud effectively avoids the problem of inconsistent model training quality across different regions, enhancing the overall stability and maintainability of the system.

[0051] In some embodiments, based on historical cable fault data, a full fault analysis model is built and trained in the cloud, prior to which the following steps are taken: Obtain historical cable fault logs for the target cable specification and extract those with true fault types as the initial fault sample set; define the type probability of each fault type based on the relative importance of each fault type; calculate the temporal correlation of each initial fault sample based on the time distance and define the temporal selection probability of each initial fault sample accordingly; based on the type probability and the temporal selection probability, perform iterative random selection without replacement on the initial fault sample set until the selection result meets the preset sample size constraint, and output the selection result as the historical cable fault data.

[0052] Specifically, historical cable fault logs refer to cable fault event archives extracted from the smart grid's operation and maintenance management center and recorded in chronological order. They include information such as fault occurrence time, cable specifications, fault type labeling, fault severity, and maintenance records. The initial fault sample set refers to the set of fault events that meet the authenticity requirements and are selected from the original logs. Its purpose is to serve as the basic data space for subsequent sampling.

[0053] Specifically, relative importance of faults refers to the prior weight values ​​of various fault types defined based on their impact on the safe operation of the power grid, their frequency of occurrence, and maintenance costs; while type probability refers to the probability distribution obtained after normalization based on relative importance of faults, which is used to guide the selection of samples to favor critical fault types.

[0054] Specifically, temporal correlation refers to the degree of temporal relevance between a fault sample and the current moment, reflecting the timeliness value of the sample; temporal selection probability is a probability weight calculated based on temporal correlation, used to prioritize the selection of samples that are more recent and have higher information value.

[0055] Specifically, firstly, historical fault logs matching the target cable specifications are extracted from the power operation and maintenance database, and records where the fault type is marked as true are selected as the initial fault sample set. Then, based on the impact of different fault types on system operation (e.g., short circuit > grounding > partial discharge), corresponding weights are assigned, and the results are normalized to form a type probability distribution. For example, the weights for short circuits can be configured as 0.5, grounding as 0.3, and partial discharge as 0.2. Next, the distance between each initial fault sample and the current time on the time axis is calculated, and the temporal correlation is estimated using a preset function or mapping relationship, then normalized to a temporal selection probability, making recently occurring fault samples more representative. Finally, a weighted random sampling algorithm is used. In each iteration, samples are drawn from the initial fault sample set without replacement based on the product of the type probability and the temporal selection probability as a comprehensive weight, until the cumulative sample size reaches a preset constraint. The corresponding selection results are then output as the historical cable fault dataset.

[0056] The type probability weighting mechanism based on the relative importance of faults described above balances the proportion of high-risk, low-frequency but impactful fault samples in the training set, thereby enhancing the model's sensitivity and accuracy in identifying key faults. The time-series correlation weighting approach ensures the timeliness and real-world relevance of the training data, making the model more adaptable to the current operating environment of the power grid and the aging state of equipment. Overall, this approach avoids interference from imbalanced samples, redundancy, or time shifts in model training, improves the representativeness and diversity of the training data, and is beneficial for model training.

[0057] In some implementations, the temporal correlation of each initial fault sample is calculated based on the time distance, and the temporal selection probability of each initial fault sample is defined accordingly, including: Calculate the distance between the timestamp of each initial fault sample and the current time to obtain the time distance; combine the preset nonlinear mapping relationship to map the time distance to the time sequence selection probability of each initial fault sample; wherein, the nonlinear mapping relationship is based on any one of the exponential decay relationship, the piecewise constant, and the decay relationship.

[0058] Specifically, temporal distance refers to the time interval between the occurrence time of the initial fault sample and the current time (or the model training time point). Temporal correlation is used to reflect the reference value of the initial fault sample for the current model training. Temporal selection probability is a sampling probability obtained by mapping the temporal distance of samples through a certain functional relationship, and is used to guide more timely samples to be selected first during the sample selection process.

[0059] Specifically, a nonlinear mapping is a mathematical function that converts time distance into probability values. This nonlinear characteristic allows samples closer to the present time to receive a higher selection probability. Examples of nonlinear mappings include: exponential decay relationships, which rapidly reduce the weight of samples from further back in time; and piecewise constant relationships, such as a constant weight within a certain time window followed by a sharp drop or linear decay beyond that window.

[0060] Specifically, firstly, each sample in the initial fault sample set is traversed, and its fault occurrence timestamp field is read. Then, the absolute difference between the timestamp and the current system time is calculated to obtain a time distance set in days. Next, this time distance is mapped to a time series selection probability according to a preset nonlinear mapping function. For example, when using an exponential decay function, if λ = 0.005, the selection probability of a sample from 365 days ago is e. −0.005×365 ≈0.16, while the probability of a sample from 30 days ago is e. −0.005×30≈0.86, significantly higher; if a piecewise function is used, the time distance can be divided into four intervals: 0-90 days, 91-180 days, 181-365 days, and more than 365 days, and assigned constant probability values ​​of 0.9, 0.6, 0.3, and 0.1 respectively; the nonlinear mapping function can flexibly adapt to the "time sensitivity" requirements of different scenarios.

[0061] The above steps, through a nonlinear probability mapping mechanism, can transform time distance into quantifiable sample selection weights, enabling the training data construction process to adaptively prioritize recently occurring fault samples. This ensures that the temporal distribution of the training set shifts towards the current operating state, avoiding interference from outdated samples on the model's learning of current fault feature patterns. Consequently, the timeliness and representativeness of the acquired training samples are improved, helping the model to better reflect the current cable operating environment.

[0062] S200: Using model compression technology, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model. The first lightweight fault analysis model is deployed to the edge, and the second lightweight fault analysis model is deployed to the cloud.

[0063] Specifically, through methods such as knowledge distillation, network pruning, parameter quantization, and structure search, the knowledge of the full fault analysis model can be transferred to a derived model with a more streamlined structure and lower computational complexity. The first lightweight fault analysis model is a lightweight model derived under the resource constraints of edge devices in the smart grid, used for rapid judgment and coarse classification of fault existence. The second lightweight fault analysis model is a refined diagnostic model with a lower degree of compression and a structural complexity between the full model and the first model. It is deployed in the cloud management and control center of the smart grid and is specialized for verifying and deepening the analysis of the initial edge judgment results.

[0064] This process enables a precise match between computing resources and task requirements. The first lightweight fault analysis model runs smoothly on edge microcontrollers or lightweight processors, providing millisecond-level basic response capabilities to meet the real-time requirements of fault perception and help reduce edge deployment costs. The second lightweight fault analysis model, while compressed, retains sufficient complexity and, with cloud computing power, can fully utilize the contextual information provided by the initial edge assessment, avoiding redundant calculations and focusing on feature refinement and anomaly pattern discovery. The collaborative work of the two models eliminates the need for continuous uploading of raw monitoring data, thereby reducing network traffic and optimizing bandwidth utilization.

[0065] In some embodiments, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model using model compression technology, including: Based on a preset list of first-order feature dimensions, the output dimensions of the full training logs of the full fault analysis model are pruned to obtain first training data, wherein the first training data includes hard labels for historical cable fault data and soft labels for the full training logs; a lightweight fault analysis model is constructed according to the computing power constraints of the edge, wherein the lightweight fault analysis model is defined as taking first-order features as input and outputting the probability of fault existence and the distribution of fault types; based on the knowledge distillation method, the lightweight fault analysis model is trained and validated in combination with the first training data to obtain the first lightweight fault analysis model; wherein the list of first-order feature dimensions includes a valley set and a peak set.

[0066] Specifically, the first-order feature dimension list refers to retaining the most basic and critical feature set in feature engineering. In other words, the first-order feature dimension is the original physical quantity most sensitive to faults, such as the peak and valley values ​​of current and voltage. Among them, the valley set and peak set represent the local minimum and maximum values ​​of the signal within the time window, used to capture abnormal fluctuations. Output dimension pruning refers to retaining only the output information related to the first-order features in the model output layer or intermediate layer.

[0067] Specifically, hard labels refer to the actual fault types already labeled in historical data, while soft labels refer to the probability distribution results of the full model's prediction of the samples, which contain richer category inference information.

[0068] Specifically, firstly, based on a pre-defined list of first-order feature dimensions, such as peak and valley values ​​of cable current and temperature abrupt changes, the intermediate outputs of the full-scale fault analysis model during training are pruned, retaining only the output dimensions associated with the aforementioned first-order features to generate the first training data. This training data includes not only hard labels of historical cable fault data, such as "short circuit" and "grounding," but also the predicted probability distribution of the full-scale model for each sample, i.e., soft labels. For example, the true label of sample A is "short circuit," while the full-scale model predicts it as "short circuit" 0.7, "grounding" 0.2, and "partial discharge" 0.1. Subsequently, based on the computing power constraints of the edge devices, lightweight neural network structures such as shallow convolutional networks and low-dimensional MLPs are configured, and first-order features are defined as inputs, with the fault existence probability (whether there is a fault) and fault type distribution (probabilities of each type) as the desired outputs. Next, a knowledge distillation method is used to use the soft labels in the first training data as teacher signals, combined with the hard labels, to train the aforementioned lightweight neural network structure, preserving as much of the full-scale model's discriminative ability as possible, ultimately obtaining the first lightweight fault analysis model.

[0069] Through the above process, the model structure can be compressed while retaining key judgment capabilities, making it suitable for edge computing scenarios and meeting the deployment requirements of resource-constrained equipment such as substations, distribution boxes, and inspection robots. Specifically, by using a first-order feature dimension list, the most representative input features are focused on, reducing input dimensionality and computational complexity; knowledge distillation technology introduces soft-label information, enabling the lightweight model to learn the "inter-class similarity" and "fuzzy boundaries" of the entire model during training, enhancing its generalization ability.

[0070] In some embodiments, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model using model compression technology, including: A validation dataset is obtained, and key information is extracted from the validation dataset. The validation dataset is input into the first lightweight fault analysis model to obtain a first validation output, which is then merged with the key information extraction result to define the second training input data. The validation dataset is input into the full fault analysis model to obtain a full validation output, which is then defined as the second training output data, wherein the full validation output includes soft labels. Based on the knowledge distillation method, the full fault analysis model is used as the teacher model, and the second training input data and the second training output data are combined to construct, train, and validate a student model. When the validation result meets the preset model accuracy requirements, the student model is output as the second lightweight fault analysis model.

[0071] Specifically, key information extraction is the process of extracting the most representative feature information for fault judgment from the original verification data. The key information obtained may include the fluctuation range, duration, and frequency of current and voltage.

[0072] Specifically, the full validation output is the prediction result obtained after inputting validation data into the full model, including soft labels, which are the probability distributions of the model for various types of faults. Knowledge distillation allows the student model (i.e., the first lightweight fault analysis model) to learn the prediction behavior of the teacher model (i.e., the full fault analysis model), especially the soft labels, thus improving the performance of the lightweight model.

[0073] Specifically, firstly, a validation dataset is prepared and key information is extracted, such as the peak and trough current values ​​and frequency change rate of each cable before and after a fault. Then, this validation data is input into the previously constructed first lightweight fault analysis model to generate the first validation output, which is the preliminary fault type prediction probability corresponding to the validation dataset. This output is then merged with the extracted key information to construct the second training input data. Simultaneously, the same validation data is input into the full fault analysis model to obtain the full validation output, i.e., high-precision soft-label output (e.g., the probability of validation sample A being "short circuit" is 0.82, and the probability of "grounding" is 0.12, etc.). Next, using the full model as the teacher model, training pairs are constructed by combining the second training input data and the full validation output to further construct and train the student model. Specifically, this student model is a lightweight version of the full model structure; by learning the output behavior of the full model, it can better fit the predictive capabilities of complex models. Finally, the student model is validated and tested. If the prediction accuracy, recall and other indicators on the new validation set meet the preset thresholds, the model can be considered to have good predictive ability, and the model is designated as the second lightweight fault analysis model.

[0074] Through the above process, the accuracy of the lightweight model is enhanced, allowing it to approach the predictive capabilities of the full model as closely as possible while maintaining low computational resource consumption. The second lightweight fault analysis model not only achieves "knowledge transfer" from the full model but also learns the contextual discrimination logic of the first lightweight fault analysis model through soft labels, which helps improve the ability to identify complex fault modes.

[0075] In some embodiments, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model using model compression technology, and the method further includes: Obtain a validation dataset and extract key information from it; use the first validation output of the first lightweight fault analysis model on the validation dataset as the input feature and the hard label of the historical cable fault data corresponding to the validation dataset as the output target to construct a second training dataset; based on the transfer learning method, train and fine-tune the full fault analysis model in combination with the second training dataset to obtain the second lightweight fault analysis model.

[0076] Specifically, the first validation output is the output of the first lightweight fault analysis model on the validation dataset. Hard labels refer to the known real fault types in the validation dataset, labeled manually or by rules. The second training dataset is a new training dataset constructed using the first validation output as input features and hard labels as supervision targets.

[0077] Specifically, firstly, a validation dataset is acquired and key information is extracted. Next, this validation data is input into the previously trained first lightweight fault analysis model to obtain the first validation output. Then, this first validation output is used as a new input feature, paired with the hard-labeled elements in the validation dataset to construct a second training dataset. The purpose of this second training dataset is to provide a training sample set with "low-dimensional input + high-confidence labels." Then, using transfer learning, the full fault analysis model is used as the base model, freezing most of its network parameters and fine-tuning only some key layers, such as the output layer or intermediate fusion layers. The model is then trained using the second training dataset, allowing the adjusted, more streamlined full fault analysis model to adapt to the input distribution composed of the output features of the first lightweight model while maintaining its original capabilities. Finally, a second lightweight fault analysis model with a simpler structure and more efficient inference is generated.

[0078] For example, a validation dataset representing 15% of the total historical data is extracted. Key information is extracted from each sample, retaining waveform segments and core statistical features for 50ms before and after the fault, while removing redundant acquisition noise. Subsequently, the validation dataset is input into the first lightweight fault analysis model to obtain the fault existence probability vector and coarse classification probability distribution corresponding to each sample, serving as the first validation output (data dimension: 128-dimensional feature vector). Simultaneously, hard labels for the actual fault types corresponding to this validation dataset are extracted from the operation and maintenance management system, constructing binary encoded label vectors. Then, the first validation output and the hard labels are combined... The sample indices are matched one by one to form the second training dataset. Finally, the parameters of the feature extraction layer at the bottom of the full fault analysis model are fixed, and only the top three fully connected classification layers are opened. The Adam optimizer with a learning rate of 0.0005 is used to perform 30 rounds of iterative fine-tuning on the second training dataset. The preferred training loss function is designed as the sum of cross-entropy loss and KL divergence of the consistency of the first validation output distribution, with a weight ratio of 7:3. This allows the fine-tuned model to fit the true labels and maintain its ability to understand the output pattern of the first lightweight model, thus obtaining the second lightweight fault analysis model, which is deployed on a cloud container cluster.

[0079] Through the above process, a lightweight model structure is achieved via an alternative approach: reducing reliance on large amounts of labeled data through transfer learning, thereby improving the prediction accuracy of the simplified model based on the full fault analysis model in a specific input domain (i.e., the output space of the first lightweight model). Compared to traditional methods of training lightweight models from scratch, this approach utilizes existing model resources more efficiently, shortens the model development cycle, and helps improve the model's practicality and stability in edge environments.

[0080] S300: Collects real-time status data of the target cable, inputs it into the first lightweight fault analysis model, obtains the first diagnostic result, and feeds back the first diagnostic result to the cloud to perform collaborative diagnosis and obtain fault analysis results.

[0081] Specifically, the first diagnostic result is a multi-dimensional information package consisting of the fault existence probability value, the coarse classification label of the fault type, and the key feature vectors output by the first lightweight model after performing forward inference on the input real-time data. The edge side can encapsulate the first diagnostic result in a lightweight data format according to the preset communication protocol and transmit it to the remote cloud platform for active reporting via 4G / 5G or the power Internet of Things private network.

[0082] Specifically, collaborative diagnosis is the process by which the cloud-based second lightweight fault analysis model receives the first diagnostic result uploaded from the edge and combines it with its own stored historical diagnostic context and associated cable topology information to perform a detailed diagnostic analysis; the fault analysis result refers to the final output of a complete diagnostic report that includes fault confirmation status, precise fault type, fault location range, fault severity rating, etc.

[0083] In some embodiments, real-time status data of the target cable is collected and input into the first lightweight fault analysis model to obtain a first diagnostic result, including: The real-time status data of the target cable is collected through the cable monitoring terminal on the edge side of the smart grid, wherein the real-time status data includes at least real-time waveform data; the real-time status data is preprocessed, wherein the preprocessing includes at least one of ROI positioning, image extraction and image enhancement; the preprocessed data is input into the first lightweight fault analysis model to obtain a first diagnostic result, wherein the first diagnostic result includes the probability of fault existence and the distribution of a first fault type.

[0084] Specifically, real-time status data refers to the raw signals such as voltage, current, temperature, and partial discharge collected in real time during cable operation, especially real-time waveform data, such as current waveforms. Edge-side cable monitoring terminals are edge computing devices deployed close to the cable site, possessing data acquisition, preliminary processing, and model inference capabilities.

[0085] Specifically, preprocessing is the process of standardizing and enhancing the features of the original data. Among these processes, ROI (Region of Interest) is used to locate and identify key regions with obvious features in the data, image extraction is the process of extracting image segments from key time windows in the waveform image, and image enhancement can include filtering, contrast enhancement, edge enhancement, etc., to improve the model's recognition performance.

[0086] Specifically, the system first collects real-time cable operation status data, especially current waveform signals, through edge monitoring terminals such as embedded acquisition modules in distribution boxes or intelligent sensor nodes. Then, these raw waveform data undergo preprocessing: ROI positioning is used to automatically identify abrupt changes and abnormal segments in the waveform; image extraction converts these key segments into clean image formats; and further image enhancement, such as edge sharpening and noise suppression, is performed to improve the model's feature recognition capabilities. After preprocessing, the data is input into the first simplified fault analysis model to obtain the corresponding first diagnostic results, including the probability of a current fault and the distribution of various fault types, such as short circuit 0.65, grounding 0.25, and open circuit 0.10.

[0087] Through the above process, it is possible to achieve rapid, accurate, and low-latency diagnosis of the target cable status, while reducing the latency and bandwidth pressure caused by data transmission to the cloud.

[0088] In some embodiments, the first diagnostic result is fed back to the cloud for collaborative diagnostics to obtain fault analysis results, including: If the first diagnostic result does not meet the preset significance threshold, the diagnostic process is terminated, and the fault analysis result indicating that the cable is fault-free is fed back to the cloud. According to the preset model update constraints, the cloud synchronously transmits the new first model parameters obtained based on federated learning to the edge, and updates the first lightweight fault analysis model accordingly. The model update constraints include: periodic update constraints, which define the period length for periodically updating the first lightweight fault analysis model; and burst update constraints, which define the amount of change in model parameters for burst updates of the first lightweight fault analysis model that exceed the periodic update constraints.

[0089] Specifically, the significance threshold is a preset confidence criterion used to determine whether the current diagnostic result has sufficient credibility. For example, it is set to consider a failure probability of less than 20% as a failure-free condition.

[0090] Specifically, model update constraints are strategies for controlling the frequency of model updates during federated learning. Periodic update constraints are set to synchronize the model at fixed time intervals, while burst update constraints are used to deal with sudden situations, such as a sudden drop in model performance or drastic fluctuations in parameters.

[0091] Specifically, firstly, the inference of the first lightweight fault analysis model is executed at the edge to obtain the first diagnostic result. If the probability of the fault in this result is lower than the set significance threshold, the result is judged as "no fault" and is fed back to the cloud for recording, without further diagnosis. If the diagnostic result is uncertain or the probability of the fault is high, a collaborative diagnostic mechanism is triggered, and the relevant data or diagnostic results are uploaded to the cloud. The cloud's stronger computing resources or more complex models are used for further analysis, and the final fault analysis result is output.

[0092] Furthermore, to maintain the continued effectiveness of the edge model, model parameters can be periodically or burstily synchronized from the cloud based on a federated learning mechanism. Periodic updates can be set to synchronize every 24 hours, while burst updates can be set to trigger an update immediately when model parameters change by more than 5%, ensuring that the edge model continuously adapts to environmental changes and data distribution shifts.

[0093] Through the above process, the synergistic integration of edge computing and cloud intelligence is achieved. On the one hand, the edge can independently complete diagnostic tasks in most cases, reducing communication latency and bandwidth consumption; on the other hand, when the diagnostic results are unclear, cloud-based collaborative diagnosis can be introduced in a timely manner, improving the accuracy and robustness of the overall system. Simultaneously, based on federated learning and model update constraint mechanisms, the dynamic update capability of the model is maintained, avoiding model degradation due to changes in data distribution, and ensuring long-term stable operation in complex operating environments.

[0094] In some embodiments, feeding back the first diagnostic result to the cloud to perform collaborative diagnosis and obtain fault analysis results further includes: If the first diagnostic result meets the preset significance threshold, the edge analysis result is fed back to the cloud as a suspected cable fault; the edge device extracts key information from the real-time status data and merges the first diagnostic result and the key information extraction result and uploads them to the cloud; the cloud inputs the first diagnostic result and the key information extraction result into the second lightweight fault analysis model, obtains and outputs the fault analysis result, wherein the fault analysis result includes at least the fault type.

[0095] Specifically, the edge analysis result refers to the fault existence judgment and coarse classification conclusion output by the first lightweight fault analysis model after performing local inference on the real-time state data; the key information extraction is the feature compression and dimensionality reduction of the original real-time state data by the edge side, which is used to extract concise waveform segments, statistical features and spectral components that characterize the essence of the fault.

[0096] Specifically, the collaborative diagnostic process includes: the edge decision unit determines whether the probability of a fault in the first diagnostic result exceeds a preset significance threshold. Preferably, this threshold is dynamically set according to the cable importance level and can be in the range of 0.5 to 0.8. If it exceeds the threshold, a feedback mechanism is triggered. At this time, the edge signal processing module immediately performs compression of the original data and extraction of second-order features on the real-time status data within the current 100ms sampling window to obtain a 64-dimensional key feature vector. For example, it includes the energy proportion of the 2-10MHz frequency band, waveform kurtosis coefficient, temperature gradient extremum, etc. Then, the 5 coarse classification probability distributions in the first diagnostic result are concatenated with the 64-dimensional feature vector to form a 69-dimensional data packet, which is pushed to the cloud message queue at QoS level 2 via the MQTT protocol. After receiving the data, the cloud parses the edge preliminary judgment information and compressed features, uses them as auxiliary inputs to the second lightweight fault analysis model, outputs fault analysis results containing subdivided categories, and returns them to the monitoring platform via RESTful API.

[0097] Through the aforementioned collaborative diagnostic mechanism, both bandwidth resources and diagnostic performance are optimized. Specifically, the saliency threshold mechanism ensures that only highly suspicious fault events trigger uploading, optimizing bandwidth utilization, while key information extraction further compresses the raw data. The second lightweight fault analysis model, receiving comprehensive input including edge initial judgment confidence and key feature information, can specifically focus on suspicious fault patterns for in-depth verification, enhancing diagnostic accuracy. Overall, the tight coupling and task relay between the two models helps improve the utilization rate of cloud computing resources, achieving green energy saving and cost reduction while ensuring the continuity of smart grid security monitoring.

[0098] In summary, the cable fault analysis method based on edge cloud collaboration provided by this invention has the following technical effects: By building and training a full-scale fault analysis model in the cloud based on historical cable fault data, a model compression method is used to derive a first lightweight fault analysis model and a second lightweight fault analysis model from the full-scale fault analysis model. The first lightweight fault analysis model is deployed on the edge side, and the second lightweight fault analysis model is deployed in the cloud. Real-time operating status data of the target cable is collected, and the real-time operating status data is input into the first lightweight fault analysis model to obtain preliminary diagnostic results. The preliminary diagnostic results are then fed back to the cloud for collaborative analysis to obtain the final fault analysis results, thereby achieving the technical effects of optimized bandwidth utilization, improved diagnostic efficiency, and enhanced diagnostic accuracy.

[0099] Example 2, as Figure 2 This is a schematic diagram of the cable fault analysis system based on edge cloud collaboration according to the present invention. For example, Figure 1The flowchart of the cable fault analysis method based on edge cloud collaboration of the present invention can be seen as follows: Figure 2 The structure shown is implemented.

[0100] Based on the same concept as the edge-cloud collaborative cable fault analysis method in the embodiments described above, the present invention also provides an edge-cloud collaborative cable fault analysis system comprising: The cloud-based model training module 11 is used to build and train a full-scale fault analysis model in the cloud based on historical cable fault data.

[0101] The lightweight model derivation and deployment module 12 is used to derive a first lightweight fault analysis model and a second lightweight fault analysis model from the full fault analysis model through model compression technology, and deploy the first lightweight fault analysis model to the edge side and deploy the second lightweight fault analysis model to the cloud.

[0102] The real-time cable fault diagnosis and collaboration module 13 is used to collect real-time status data of the target cable, input it into the first lightweight fault analysis model, obtain the first diagnosis result, and feed back the first diagnosis result to the cloud to perform collaborative diagnosis and obtain fault analysis results.

[0103] In some embodiments, the execution steps of the cloud-based model training module 11 include: Obtain historical cable fault logs for the target cable specification, and extract those with true fault types as the initial fault sample set.

[0104] Define the type probability of each fault type based on the relative importance of each fault type.

[0105] Based on the time distance, the temporal correlation of each initial fault sample is calculated, and the temporal selection probability of each initial fault sample is defined accordingly.

[0106] Based on the type probability and the time sequence selection probability, iterative random selection without replacement is performed on the initial fault sample set until the selection result meets the preset sample size constraint, and the selection result is output as the historical cable fault data.

[0107] In some implementations, the execution steps of the cloud-based model training module 11 also include: Calculate the distance between the timestamp of each initial fault sample and the current time, and obtain the time distance.

[0108] By combining a preset nonlinear mapping relationship, the time distance is mapped to the timing selection probability of each initial fault sample.

[0109] The nonlinear mapping relationship is based on any one of the following: exponential decay relationship, piecewise constant, and decay relationship.

[0110] In some embodiments, the execution steps of the lightweight model derivation and deployment module 12 include: Based on a pre-defined list of first-order feature dimensions, the output dimensions of the full training logs of the full fault analysis model are pruned to obtain first training data, wherein the first training data includes hard labels of historical cable fault data and soft labels of the full training logs.

[0111] A lightweight fault analysis model is constructed based on the computing power constraints at the edge. The lightweight fault analysis model is defined as taking first-order features as input and fault existence probability and fault type distribution as output.

[0112] Based on the knowledge distillation method, the lightweight fault analysis model is trained and validated using the first training data to obtain the first lightweight fault analysis model.

[0113] The list of first-order feature dimensions includes a valley set and a peak set.

[0114] In some embodiments, the execution steps of the lightweight model derivation and deployment module 12 further include: Obtain the verification dataset and extract key information from it.

[0115] The verification dataset is input into the first lightweight fault analysis model to obtain the first verification output, which is then combined with the key information extraction results to define the second training input data.

[0116] The verification dataset is input into the full-scale fault analysis model to obtain the full-scale verification output and define it as the second training output data, wherein the full-scale verification output includes soft labels.

[0117] Based on the knowledge distillation method, using the full fault analysis model as the teacher model, and combining the second training input data and the second training output data, a student model is constructed, trained, and validated.

[0118] When the verification results meet the preset model accuracy requirements, the student model is output as the second lightweight fault analysis model.

[0119] In some embodiments, the execution steps of the lightweight model derivation and deployment module 12 further include: Obtain the verification dataset and extract key information from it.

[0120] Using the first validation output of the first lightweight fault analysis model on the validation dataset as the input feature, and the hard labels of the historical cable fault data corresponding to the validation dataset as the output target, a second training dataset is constructed.

[0121] Based on the transfer learning method, the full fault analysis model is trained and fine-tuned using the second training dataset to obtain the second lightweight fault analysis model.

[0122] In some embodiments, the execution steps of the real-time cable fault diagnosis and coordination module 13 include: The real-time status data of the target cable is collected through the cable monitoring terminal on the edge side, wherein the real-time status data includes at least real-time waveform data.

[0123] The real-time status data is preprocessed, wherein the preprocessing includes at least one of ROI localization, image extraction, and image enhancement.

[0124] The preprocessed data is input into the first lightweight fault analysis model to obtain the first diagnostic result, wherein the first diagnostic result includes the probability of fault existence and the distribution of the first fault type.

[0125] In some embodiments, the execution steps of the real-time cable fault diagnosis and coordination module 13 further include: If the first diagnostic result does not meet the preset significance threshold, the diagnostic process is terminated, and the fault analysis result is fed back to the cloud as no cable fault.

[0126] According to the preset model update constraints, the new first model parameters obtained by federated learning are synchronized from the cloud to the edge, and the first lightweight fault analysis model is updated accordingly.

[0127] The model update constraints include: Periodic update constraint, which is used to define the period length for periodically updating the first lightweight fault analysis model.

[0128] Sudden update constraint, which defines the amount of change in model parameters when the first lightweight fault analysis model is suddenly updated beyond the periodic update constraint.

[0129] In some embodiments, the execution steps of the real-time cable fault diagnosis and coordination module 13 further include: If the first diagnostic result meets the preset significance threshold, the edge analysis result is fed back to the cloud as a suspected cable fault.

[0130] The edge device extracts key information from the real-time status data and merges the first diagnostic result and the key information extraction result into the cloud.

[0131] The cloud inputs the first diagnostic result and the key information extraction result into the second lightweight fault analysis model to obtain and output the fault analysis result, wherein the fault analysis result includes at least the fault type.

[0132] In some implementations, the process includes: calculating the distance between the timestamp of each initial fault sample and the current time, and obtaining the time distance.

[0133] By combining a preset nonlinear mapping relationship, the time distance is mapped to the timing selection probability of each initial fault sample.

[0134] The nonlinear mapping relationship is based on any one of the following: exponential decay relationship, piecewise constant, and decay relationship.

[0135] It should be understood that the focus of the embodiments mentioned in this specification is their difference from other embodiments. The specific embodiments in the aforementioned Embodiment 1 are also applicable to the cable fault analysis system based on edge cloud collaboration described in Embodiment 2. For the sake of brevity, they will not be elaborated further here.

[0136] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A cable fault analysis method based on edge cloud collaboration, characterized in that, include: Based on historical cable fault data, a full fault analysis model is built and trained in the cloud. Using model compression technology, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model. The first lightweight fault analysis model is deployed to the edge, and the second lightweight fault analysis model is deployed in the cloud. Real-time status data of the target cable is collected and input into the first lightweight fault analysis model to obtain the first diagnostic result. The first diagnostic result is then fed back to the cloud for collaborative diagnosis to obtain the fault analysis result.

2. The cable fault analysis method based on edge cloud collaboration as described in claim 1, characterized in that, Based on historical cable fault data, a full-scale fault analysis model is built and trained in the cloud, including: Obtain historical cable fault logs for the target cable specification, and filter out those with true fault types to extract the initial fault sample set; Define the type probability of each fault type based on the relative importance of each fault type; Based on the time distance, the temporal correlation of each initial fault sample is calculated and obtained, and the temporal selection probability of each initial fault sample is defined accordingly. Based on the type probability and the time sequence selection probability, iterative random selection without replacement is performed on the initial fault sample set until the selection result meets the preset sample size constraint, and the selection result is output as the historical cable fault data.

3. The cable fault analysis method based on edge cloud collaboration as described in claim 1, characterized in that, Using model compression technology, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model, including: Based on a preset list of first-order feature dimensions, the output dimensions of the full training log of the full fault analysis model are pruned to obtain the first training data, wherein the first training data includes hard labels of historical cable fault data and soft labels of the full training log. A lightweight fault analysis model is constructed based on the computing power constraints at the edge. The lightweight fault analysis model is defined as taking first-order features as input and fault existence probability and fault type distribution as output. Based on the knowledge distillation method, the lightweight fault analysis model is trained and validated using the first training data to obtain the first lightweight fault analysis model. The list of first-order feature dimensions includes a valley set and a peak set.

4. The cable fault analysis method based on edge cloud collaboration as described in claim 3, characterized in that, Using model compression technology, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model, including: Obtain the verification dataset and extract key information from the verification dataset; The verification dataset is input into the first lightweight fault analysis model to obtain the first verification output, which is then combined with the key information extraction result and defined as the second training input data. The verification dataset is input into the full-scale fault analysis model to obtain the full-scale verification output and define it as the second training output data, wherein the full-scale verification output includes soft labels; Based on the knowledge distillation method, using the full fault analysis model as the teacher model, and combining the second training input data and the second training output data, a student model is constructed, trained, and verified. When the verification results meet the preset model accuracy requirements, the student model is output as the second lightweight fault analysis model.

5. The cable fault analysis method based on edge cloud collaboration as described in claim 3, characterized in that, Using model compression technology, a first lightweight fault analysis model and a second lightweight fault analysis model are derived from the full fault analysis model, and the model further includes: Obtain the verification dataset and extract key information from the verification dataset; Using the first validation output of the first lightweight fault analysis model on the validation dataset as the input feature, and the hard label of the historical cable fault data corresponding to the validation dataset as the output target, a second training dataset is constructed. Based on the transfer learning method, the full fault analysis model is trained and fine-tuned using the second training dataset to obtain the second lightweight fault analysis model.

6. The cable fault analysis method based on edge cloud collaboration as described in claim 1, characterized in that, Real-time status data of the target cable is collected and input into the first lightweight fault analysis model to obtain the first diagnostic result, including: The real-time status data of the target cable is collected through the cable monitoring terminal on the edge side, wherein the real-time status data includes at least real-time waveform data; The real-time status data is preprocessed, wherein the preprocessing includes at least one of ROI localization, image extraction, and image enhancement; The preprocessed data is input into the first lightweight fault analysis model to obtain the first diagnostic result, wherein the first diagnostic result includes the probability of fault existence and the distribution of the first fault type.

7. The cable fault analysis method based on edge cloud collaboration as described in claim 1, characterized in that, The first diagnostic result is fed back to the cloud for collaborative diagnostics to obtain fault analysis results, including: If the first diagnostic result does not meet the preset significance threshold, the diagnostic process is terminated, and the fault analysis result is fed back to the cloud as no cable fault. According to the preset model update constraints, the new first model parameters obtained by federated learning are synchronized from the cloud to the edge, and the first lightweight fault analysis model is updated accordingly. The model update constraints include: Periodic update constraint, which is used to define the period length for periodically updating the first lightweight fault analysis model; Sudden update constraint, which defines the amount of change in model parameters when the first lightweight fault analysis model is suddenly updated beyond the periodic update constraint.

8. The cable fault analysis method based on edge cloud collaboration as described in claim 1, characterized in that, The process of sending the first diagnostic result to the cloud for collaborative diagnosis and obtaining fault analysis results also includes: If the first diagnostic result meets the preset significance threshold, the edge analysis result is fed back to the cloud as a suspected cable fault. The edge device extracts key information from the real-time status data and merges the first diagnostic result and the key information extraction result and uploads them to the cloud. The cloud inputs the first diagnostic result and the key information extraction result into the second lightweight fault analysis model to obtain and output the fault analysis result, wherein the fault analysis result includes at least the fault type.

9. The cable fault analysis method based on edge cloud collaboration as described in claim 2, characterized in that, Based on the time distance, the temporal correlation of each initial fault sample is calculated, and the temporal selection probability of each initial fault sample is defined accordingly, including: Calculate the distance between the timestamp of each initial fault sample and the current time, and obtain the time distance; By combining a preset nonlinear mapping relationship, the time distance is mapped to the time sequence selection probability of each initial fault sample; The nonlinear mapping relationship is based on any one of the following: exponential decay relationship, piecewise constant, and decay relationship.

10. A cable fault analysis system based on edge cloud collaboration, characterized in that, The cable fault analysis method based on edge cloud collaboration as described in any one of claims 1-9 includes: The cloud-based model training module is used to build and train a full-scale fault analysis model in the cloud based on historical cable fault data. The lightweight model derivation and deployment module is used to derive a first lightweight fault analysis model and a second lightweight fault analysis model from the full fault analysis model through model compression technology, and deploy the first lightweight fault analysis model to the edge side and the second lightweight fault analysis model to the cloud. The real-time cable fault diagnosis and collaboration module is used to collect real-time status data of the target cable, input it into the first lightweight fault analysis model, obtain the first diagnosis result, and feed back the first diagnosis result to the cloud to perform collaborative diagnosis and obtain fault analysis results.

Images