Contact network anomaly detection method based on generative adversarial network and LSTM coding and decoding

Abstract

The invention relates to the technical field of safety monitoring of rail transit traction power supply systems, in particular to an overhead line system anomaly detection method based on a generative adversarial network and LSTM coding and decoding, which comprises the following steps: S1, synchronously acquiring multi-physics field data; s2, obtaining a mixed feature vector; s3, obtaining a single frame anomaly score; s4, obtaining a robust anomaly score; s5, obtaining a comprehensive risk score of a single node in all the nodes; step S6, obtaining a verified risk score Rv of the single node; step S7, obtaining the residual life RL of a single node; step S8, carrying out continuous evolution on the residual life of the single node; and step S9, ending the current frame flow, sliding the clock window, and continuing the next frame. According to the method, the data quality of the contact network is improved, fault samples are enriched, the zero sample detection capability of the model on unknown faults is improved, residual life prediction is carried out on progressive degradation, and the operation and maintenance cost of the contact network is reduced.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for rail transit traction power supply systems, and in particular to a method for detecting catenary anomalies based on generative adversarial networks and LSTM encoding and decoding. Background Technology

[0002] The overhead contact line is the artery of the high-speed railway traction power supply system. Its service environment is harsh, and failures are highly unpredictable; a failure directly jeopardizes train operation safety. Existing technologies primarily rely on threshold alarms for single physical quantities (such as vibration or infrared), which have the following drawbacks: 1) Inconsistent spatiotemporal references among sensor data lead to misalignment and missed alarms; 2) Data quality cannot be quantified online, resulting in low-quality data being directly fed into the model, leading to a "garbage in, garbage out" scenario; 3) Model training relies on a large number of fault samples, while high-speed railway fault samples are scarce, leaving the model without zero-sample detection capability for unknown faults; 4) It cannot predict the remaining lifetime (RUL) of progressive degradation, forcing maintenance departments to replace faults periodically or repair them after the fact, resulting in high operation and maintenance costs. Therefore, there is an urgent need for an overhead contact line anomaly detection and lifetime prediction technology that can integrate virtual and real data, self-evolve, and provide causal explanation.

[0003] Chinese patent application document with publication number CN116363106A discloses a method based on SCA The GANomaly model-based method for fault detection of high-speed railway catenary components includes: In the data preprocessing stage, images of normal high-speed railway catenary components are first acquired, and then the dataset of normal high-speed railway catenary components is expanded using DCGAN data augmentation techniques; In the training stage, the expanded dataset of normal high-speed railway catenary components is fed into SCA (Supervisory Classification). The GANomaly network is trained, and after training, the data distribution of normal catenary component images is obtained. During the testing phase: images of catenary components for the test samples are acquired and then input into the trained SCA network. In the GANomaly network, after obtaining the output image, it is compared with the test sample to obtain a difference score. It is then determined whether the difference score is greater than a threshold K. If it is greater, the image of the catenary component of the test sample is considered to have a fault; otherwise, the image of the catenary component of the test sample is considered to be normal. However, this scheme still has problems: 1) The spatiotemporal reference of the data from each sensor is inconsistent, leading to misalignment and missed reports; 2) Data quality cannot be quantified online, and low-quality data is directly entered into the model, resulting in garbage in and garbage out; 3) Model training depends on a large number of fault samples, but fault samples in high-speed railways are scarce, and the model lacks zero-sample detection capability for unknown faults; 4) It cannot predict the remaining useful life (RUL) of progressive degradation, and the maintenance department can only replace it periodically or repair it afterward, resulting in high operation and maintenance costs. Summary of the Invention

[0004] To address these issues, this invention provides a contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding. This method overcomes the following problems in existing technologies: inconsistent spatiotemporal references of data from various sensors, leading to misalignment and missed reports; inability to quantify data quality online, resulting in low-quality data directly entering the model and causing garbage data to be imported and exported; reliance on a large number of fault samples for model training, while high-speed railway fault samples are scarce, resulting in a lack of zero-sample detection capability for unknown faults; and inability to predict the remaining lifespan of progressive degradation, requiring maintenance departments to replace faults periodically or repair them after the fact, leading to high operation and maintenance costs.

[0005] To achieve the above objectives, this invention provides a contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding, comprising: Step S1: Synchronously acquire multiphysics data and obtain the data quality score for each frame based on three-dimensional indicators to obtain multiphysics data that passes the score, and optimize the synchronous acquisition process. Step S2: Obtain physical prior features through finite element simulation, construct a meta-learning network, and use the meta-learning network to preprocess the multi-physics data scored in step S1 to obtain data features. Then, concatenate the physical prior features with the data features to obtain a hybrid feature vector. Step S3: Use a layered decoupled generative adversarial network and LSTM encoding and decoding to measure the anomalies in the hybrid feature vector, obtain the single-frame anomaly score, and output the contact network anomaly detection results. Step S4: The anomaly measurement process is reinforced by course learning and adversarial regularization to obtain a robust anomaly score, and the single-frame anomaly score of all nodes in step S3 is output based on the robust anomaly score. Step S5: Obtain the single-frame anomaly score of all nodes output in step S4, and obtain the comprehensive risk score of a single node among all nodes. Step S6: Use a three-layer verification system combining virtual and physical elements to verify and reinforce the comprehensive risk score R of a single node, and obtain the verified risk score Rv of the single node, and transmit it to step S7. Step S7: Estimate the remaining lifetime of a single node by performing a post-verification risk score, obtain the remaining lifetime RL of the single node, and transmit it to step S8. Step S8: Utilize cloud-edge-device elastic inference and federated lifelong learning to continuously evolve the remaining lifetime of a single node, obtain the evolved weight ΔW of the single node, and transmit it to step S9. Step S9: Automated iteration of the evolved weights of a single node is performed using the full-link monitoring dashboard and A / B test shadow mode to obtain the iterated model version number Ver of the single node. The current frame process ends, the clock window slides, and the process continues in the next frame.

[0006] Further, in step S1, the data quality score Q for each frame is calculated, and Q is set as Q = ω1·SNR + ω2·Stability + ω3·Integrity; In step S1, when obtaining multiphysics data that has passed the scoring and optimizing the synchronous acquisition process, the quality score Q of each frame of data is compared with the preset acquisition optimization parameter Q0, wherein: When Q > Q0, the multiphysics data corresponding to the quality score of the frame is used as the multiphysics data that passes the scoring, and the synchronous acquisition process is not optimized. Otherwise, the multiphysics data corresponding to the quality score of that frame will be regarded as multiphysics data that failed the scoring, and the multiphysics data will be re-acquired synchronously in order to optimize the synchronous acquisition process. If the synchronous acquisition process is optimized three times in a row, mark it as "sensor abnormal" and report it to the edge server.

[0007] Further, in step S2, the ANSYS finite element table Pprior=[f1,…,f6,φ1,…,φ6]∈R corresponding to the same dropper model is... 54 As a priori physical characteristic; In step S2, when constructing the meta-learning network, the physical prior features of three actual scenarios—tunnel, bridge, and plain—historical multiphysics data with passing scores, environmental parameters corresponding to the historical multiphysics data with passing scores, and the optimal preprocessing strategy are used as the training set for the meta-learning network. The initial meta-learning network with network parameters Φ is pre-trained offline based on the training set to obtain the pre-trained weights Φ*. The network parameters of the initial meta-learning network are then updated to the pre-trained weights Φ* to obtain the constructed meta-learning network Meta-Pret. In step S2, when preprocessing the multiphysics data that has passed the scoring, the multiphysics data Fclean that has passed the scoring is used as the data input, and the environmental parameters θenv=[humidity, wind speed, temperature] corresponding to the multiphysics data that has passed the scoring are used as the driving input. The preprocessing strategy θpret=[waveletbase,decomplevel] is obtained through the meta-learning network Meta-Pret. First, discrete wavelet decomposition is performed on the multiphysics data Fclean that has passed the scoring according to the preprocessing strategy θpret to obtain the subband coefficients Cwavelet. Then, statistical features are extracted from Cwavelet to obtain the data features Dfeat∈R²¹. 4 The data feature Dfeat and the physical prior feature Pprior are concatenated column-wise to obtain the mixed feature vector Hvec∈R². 68 .

[0008] Furthermore, in step S3, when performing anomaly measurement on the mixed feature vector, the mixed feature vector Hvec∈R² output in step S2 is used. 68 As the sole input, after being processed in parallel by the three parallel paths of the hierarchical decoupled generative adversarial network HD-GAN (GT, GD, Fuse-Att), the reconstructed frame x' is output, and the single-frame anomaly score As is calculated; In step S3, when obtaining the single-frame anomaly score, it is calculated as As = β·‖x x'‖2+(1 β)·(DT(x'T)+DD(x'D)) is calculated, where β=0.7 is the anomaly fusion weight, and ||x x'‖2 is the reconstruction error term, and (DT+DD) is the discrimination confidence term; If As>0.65, then mark the single-frame anomaly of the mixed feature vector and output the catenary anomaly detection result as the current frame multi-physics data anomaly corresponding to the mixed feature vector; Otherwise, mark the single frame of the mixed feature vector as normal, and output the catenary anomaly detection result as the current frame multiphysics data corresponding to the mixed feature vector as normal.

[0009] Furthermore, in step S4, when strengthening the anomaly measurement process, the single-frame anomaly score As output in step S3 is used as the monitoring object. The Curri-SCH course learning scheduler is used to increase the difficulty in three stages: wind speed <2m / s→2-10m / s→>10m / s. In each stage, an ε-perturbation is applied to the multiphysics data Fclean that has passed the scoring to obtain the robust anomaly score RW. RW is set as ||HD-GAN(Fclean+ε)|| HD-GAN(Fclean)‖2, where: When RW < δ, the disturbance test is passed, and the original single-frame anomaly score As is output. Otherwise, the disturbance test is deemed to have failed, the original single-frame anomaly score As is not output, and a new frame of multiphysics data Fclean that has passed the score is retrieved from step S1, and steps S2-S4 are repeated. If the perturbation test fails three times in a row, the course is marked as abnormal and the process is rolled back to S3 to retrain the GT and GD parameters.

[0010] Further, in step S5, a dynamic graph G=(V,E) is constructed using the single-frame anomaly scores of all nodes output in step S4 within the same clock window as the source, where V represents the 134 sets of four-in-one sensor nodes arranged in S1, and E represents the spatial distance ≤15m; the node embedding ei is updated through 2 layers of GraphSAGE+GAT, and the edge mutation Δeij=‖ei is calculated. ej‖2, for each edge eij, we have obtained Δeij=‖ei ej‖2, take its normalized value △'eij, △'eij=Δeij / (Δeij+τ), pre-generate edge weight wij=1 / (Δs+1) offline, Δs is the mileage difference along the contact line, obtain the graph structure anomaly score Ag of a single node among all nodes, set Ag=Σ{j∈N(i)}[△'eij·wij] / Σ{j∈N(i)}wij, that is, for all neighboring nodes j of a single node i among all nodes, the graph structure anomaly score Ag of the node is obtained by weighting the normalized mutation and edge weight, Ag∈[0,1); output the comprehensive risk score R=DSFuse(As,Ag) through the DS fusion unit, where: If R>0.65, then mark the node as having exceeded the risk limit, i.e., the area within a 15m spatial distance from the node constitutes the local catenary unit, and transmit the comprehensive risk score R of the node and the local catenary unit to step S6. Otherwise, the risk of the local contact network unit consisting of the area within 15m of the node is marked as not exceeding the limit, and the comprehensive risk score R of the node is transmitted to step S6; In step S5, the conflict factor K is calculated based on the mass distribution of the node's comprehensive risk score R; K is set as K = ∑(m1(A1)·m2(B1)), where A1∩B1 = m1 is the mass distribution of As evidence, and m2 is the mass distribution of Ag evidence; K∈[0,1], set the tolerance threshold = 0.1; The output of the comprehensive risk score R for this node is optimized based on the conflict factor K, where: When K < 0.1, the output of the comprehensive risk score R for that node is not optimized; When K≥0.1, the output of the comprehensive risk score R for this node is optimized. Instead of using the DS fusion machine to output the comprehensive risk score R, the output of the comprehensive risk score R is changed to a weighted average minimum guarantee, and R=0.5·As+0.5·Ag is set. If K≥0.1 for 3 consecutive times, mark the graph as abnormal and report it to the edge server. Then, retrieve the next frame of new multiphysics data Fclean that has passed the score from step S1 and repeat steps S2-S5.

[0011] Furthermore, in step S6, when verifying and strengthening the comprehensive risk score R of a single node, the comprehensive risk score R value within the same node and the same clock window is used as the source, and three layers of verification are performed sequentially: Simpack high-fidelity simulation layer verification: Using the stiffness of the dropper corresponding to the node as a variable, a 101-level stiffness-decreasing fault library is generated. The comprehensive risk score R value of the node is used to replay the simulation. If R>0.65 and the simulation truth value is a fault, then the simulation is recorded as a hit; otherwise, it is recorded as a false alarm. PGD ​​white-box adversarial attack layer verification: Using the mixed feature vector Hvec of the node as input, apply ε=0.03 perturbation. If the false negative rate of HD-GAN for the node is ≤5%, do not backtrack. Otherwise, backtrack to step S3 to retrain GT and GD for 3 epochs. On-site backtracking layer verification: Compare the comprehensive risk score R value of the node with the results of subsequent manual maintenance. If the hit rate is ≥91.5%, do not backtrack the steps; otherwise, backtrack to step S3 and retrain GT and GD for 3 Epochs. In step S6, when obtaining the post-verification risk score Rv for a single node, Rv is set to R·(1 UI), where UI is the correction coefficient, set UI=max(0,1). |nf| / 3), where nf is the number of simulation misses; if the corrected Rv>0.65, then mark the node verification as abnormal and transmit it to S7; otherwise, mark the node verification as normal and transmit it to S7.

[0012] Furthermore, in step S7, when estimating the remaining lifetime of a single node, two levels of prediction are performed sequentially using the verified risk score Rv within the same node and the same clock window as the source: Causal discovery layer: Taking the risk score Rv sequence {Rv(t)} verified over the past 30 days of the node as input, a causal graph C-graph is constructed using the PC-stable algorithm to identify the top-3 root cause features. The top-3 root cause features consist of nf, PGD missed detection flag, and maintenance hit rate output by S6, forming a causal feature vector Cvec, where Cvec∈R³. Transformer sequence prediction layer: Using {Rv(t)} and Cvec as joint inputs, the risk score Rv(t+Δt) for the next 60 days is predicted through 4 layers of Transformer. The number of days that the failure threshold of 0.9 is first exceeded is recorded as the remaining lifetime RL of the node. In step S7, when the remaining lifetime RL of a single node is obtained, the following is set: When RL < 7 days, mark the node lifetime alarm and transmit the remaining lifetime RL and node lifetime alarm to S8; Otherwise, mark the node's lifetime as normal and transmit the remaining lifetime RL to S8; If the lifespan is less than 7 days for 3 consecutive times, it will be marked as "abnormal lifespan" and reported to the cloud center, triggering a front-end speed reduction suggestion.

[0013] Furthermore, in step S8, when continuously evolving the remaining lifetime of a single node, the remaining lifetime RL within the same node and the same clock window is used as the priority weight, and three-level elastic inference and federated aggregation are executed sequentially: Initial screening of end-sided TinyGAN: Using the remaining lifetime RL of the node as the scheduling key, only nodes with RL < 7 days are processed; the frames of other nodes are skipped directly, and the initial screening gradient is output. Wtiny; Edge Std-GAN confirmation: Using the remaining lifetime RL of the node as the scheduling key, only nodes with RL < 7 days are confirmed, and the confirmation gradient is output. Wstd; In-depth analysis of Heavy-GAN in cloud centers: Using the remaining lifetime RL of the node as the scheduling key, in-depth analysis is performed only on nodes with RL < 7 days to generate node-specific weight increments ΔW.

[0014] Furthermore, in step S9, when automatically iterating the evolved weights of a single node, the evolved weights ΔW of the same node within the same clock window are used as the update source, and the monitoring-shadow-grayscale three-stage process is executed sequentially: Mon-Board monitoring layer: The node's 127 KPIs are tracked in real time, generating a node-level dashboard. If a KPI changes by more than ±2σ, the shadow mode is triggered in advance. Shadow Mode Layer: The new model B runs in parallel with the shadow for 7 days, only recording predictions and not linking with real alarms. If the false alarm rate of the node is significantly lower than 0.01 and the false alarm rate does not increase during the shadow period, it will be put into grayscale. A / B grayscale layers: Replace the old model A with a three-stage gray-scale release at the node level, from 20% to 50% to 100%. If the KPI still meets the target after the gray-scale release, a new version number Ver' is generated. The new version number Ver' is output as the iteration version number Ver of a single node, and the current frame process ends. The clock window slides, and the process continues in the next frame. Otherwise, the iteration version number Ver of a single node is output, the current frame process ends, the clock window slides, and the process continues in the next frame.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: The method, through step S1, ensures a network-wide synchronization error ≤50ns, early resampling of low-quality frames, and a training set purity >99%, thus solving the problem of garbage in / out. The method, through step S2, improves the AUC in small-sample scenarios by ≥12%, addressing the problem of scarce fault samples. The method, through step S3, performs HD-GAN layered decoupling + LSTM encoding / decoding to reduce the false alarm rate to ≤0.3%, addressing the problem of false alarms based on single physical quantity thresholds. The method, through step S4, is robust to strong disturbances such as wind swaying / icing, addressing the problem of false alarms due to environmental interference. The method, through step S5, improves the cascading fault detection rate by 42%, addressing the problem of single-point faults. The method addresses the issue of insufficient model credibility by employing a three-layer verification process (Simpack-PGD-backtracking) in step S6 to meet SIL-2 verification sufficiency. Step S7, using PC-stable and Transformer, ensures the remaining lifetime error is less than 15%, thus resolving the issue of only being able to correct errors after the fact. Step S8 saves uplink bandwidth by 70% and prevents data from leaving the domain, thus addressing data privacy and bandwidth concerns. Step S9, using Mon-Board and A / B shadowing, transforms the iteration process from "manual experimentation" to a "data-driven engineering workflow," thereby resolving the long-term model drift issue. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding in this embodiment. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0020] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Please see Figure 1 The diagram shown is a flowchart illustrating the contact network anomaly detection method based on generative adversarial networks and LSTM encoding / decoding in this embodiment. The method includes: Step S1: Synchronously acquire multiphysics data and obtain the data quality score of each frame based on the three-dimensional index to obtain multiphysics data that passes the score, and optimize the synchronous acquisition process. Step S2: Obtain physical prior features through finite element simulation, construct a meta-learning network, and use the meta-learning network to preprocess the multi-physics data scored in step S1 to obtain data features. Then, concatenate the physical prior features with the data features to obtain a hybrid feature vector. Step S3: Use a layered decoupled generative adversarial network and LSTM encoding and decoding to measure the anomalies in the hybrid feature vector, obtain the single-frame anomaly score, and output the contact network anomaly detection results. Step S4: The anomaly measurement process is reinforced by course learning and adversarial regularization to obtain a robust anomaly score, and the single-frame anomaly score of all nodes in step S3 is output based on the robust anomaly score. Step S5: Obtain the single-frame anomaly score of all nodes output in step S4, and obtain the comprehensive risk score of a single node among all nodes. Step S6: Use a three-layer verification system combining virtual and physical elements to verify and reinforce the comprehensive risk score R of a single node, and obtain the verified risk score Rv of the single node, and transmit it to step S7. Step S7: Estimate the remaining lifetime of a single node by performing a post-verification risk score, obtain the remaining lifetime RL of the single node, and transmit it to step S8. Step S8: Utilize cloud-edge-device elastic inference and federated lifelong learning to continuously evolve the remaining lifetime of a single node, obtain the evolved weight ΔW of the single node, and transmit it to step S9. Step S9: Automated iteration of the evolved weights of a single node is performed using the full-link monitoring dashboard and A / B test shadow mode to obtain the iterated model version number Ver of the single node. The current frame process ends, the clock window slides, and the process continues in the next frame.

[0022] Specifically, the method is applied to the operation and dispatch center terminal and maintenance team terminal of the catenary at the anchor section level of high-speed railways. This method improves the data quality of the catenary, enriches fault samples, enhances the model's ability to detect unknown faults without zero samples, predicts the remaining lifespan of progressive degradation, and reduces the operation and maintenance costs of the catenary. Specifically, step S1 ensures a network-wide synchronization error ≤50ns, early resampling of low-quality frames, and a training set purity >99% to address the problem of garbage in / out. Step S2 improves the AUC of small sample scenarios by ≥12% to address the problem of scarce fault samples. Step S3 performs HD-GAN layered decoupling + LSTM encoding and decoding to reduce the false alarm rate to ≤0.3% to address the problem of false alarms based on single physical quantity thresholds. Step S4 performs wind sway / icing analysis. The method is robust to strong perturbations to address the problem of false alarms caused by environmental interference. Step S5 improves the cascading fault detection rate by 42% to address the issue of single-point failure in detecting spatial propagation. Step S6 uses a three-layer verification process of Simpack-PGD-backtracking to meet SIL-2 verification sufficiency requirements, thus addressing the problem of insufficient model credibility. Step S7 uses PC-stable and Transformer to ensure that the remaining lifetime error is less than 15%, thus addressing the issue of only being able to repair after the fact. Step S8 saves uplink bandwidth by 70% and prevents data from leaving the domain, thus addressing the issues of data privacy and bandwidth. Step S9 uses Mon-Board and A / B shadowing to transform iteration from "manual experimentation" to "data-driven engineering process," thus addressing the problem of long-term model drift.

[0023] Specifically, in step S1, when synchronously acquiring multi-physics data, 134 sets of vibration-acoustic-infrared-strain four-in-one sensor nodes are arranged at 7.5m intervals within a 1km contact wire anchor section. The initial multi-physics data is acquired through the four-in-one sensor nodes, and the initial multi-physics data is synchronized using the IEEE1588v2 precision clock protocol to obtain multi-physics data, so that the synchronization error of the entire network is ≤50ns.

[0024] Specifically, the contact wire anchor section refers to a continuous section of contact wire set to 1 km in length with anchoring devices at both ends, with internal suspension wires arranged at equal intervals of 7.5 m. The vibration-acoustic-infrared-strain four-in-one sensor node refers to a VAIN node (Vibration-Acoustic-Infrared-StrainNode) with four built-in sensors. The IEEE1588v2 precision clock protocol refers to IEEE1588v2PTP, which is a time synchronization protocol.

[0025] Specifically, in step S1, the data quality score Q for each frame is calculated, and Q is set as Q = ω1·SNR + ω2·Stability + ω3·Integrity.

[0026] Specifically, ω1 is the first data quality weight parameter, ω2 is the second data quality weight parameter, and ω3 is the third data quality weight parameter, ω1+ω2+ω3=1. In this embodiment, ω1=0.5, ω2=0.3, and ω3=0.2 are set. SNR refers to wavelet entropy-weighted signal-to-noise ratio. In this embodiment, the SNR is calculated using the wavelet entropy weighting method based on each frame of multiphysics data. Stability refers to the 5ms sliding variance. In this embodiment, Stability is calculated using the 5ms sliding variance based on each frame of multiphysics data. Integrity refers to 1 - packet loss rate. In this embodiment, Integrity is calculated using packet loss rate quantization based on each frame of multiphysics data.

[0027] Specifically, in step S1, when obtaining multiphysics data that has passed the scoring and optimizing the synchronous acquisition process, the quality score Q of each frame of data is compared with the preset acquisition optimization parameter Q0, wherein: When Q > Q0, the multiphysics data corresponding to the quality score of the frame is used as the multiphysics data that passes the scoring, and the synchronous acquisition process is not optimized. Otherwise, the multiphysics data corresponding to the quality score of that frame will be regarded as multiphysics data that failed the scoring, and the multiphysics data will be re-acquired synchronously in order to optimize the synchronous acquisition process. If the synchronous acquisition process is optimized three times in a row, mark it as "sensor abnormal" and report it to the edge server.

[0028] Specifically, the preset acquisition optimization parameter Q0 refers to the preset value of the quality score for judging whether the multiphysics data passes the scoring. In this embodiment, the preset acquisition optimization parameter Q0 is set to 0.75. The method of marking "sensor abnormality" and reporting to the edge server is to report through the edge gateway MQTT message. The edge server refers to the industrial-grade computing node deployed in the nearest cabinet of the catenary anchor section.

[0029] Specifically, in step S2, the ANSYS finite element table Pprior=[f1,…,f6,φ1,…,φ6]∈R corresponding to the same dropper model is... 54 As a priori physical characteristic.

[0030] Specifically, the ANSYS finite element table Pprior refers to a 54-dimensional constant vector formed by straightening the first 6 natural frequencies and 6 mode shapes. f1,...,f6 refers to the first 6 natural frequencies (Hz) of the dropper. φ1,...,φ6 refers to the straightened components of the first 6 mode shape vectors. All samples of the same dropper model share the same mixed feature vector.

[0031] Specifically, in step S2, when constructing the meta-learning network, the physical prior features of three actual scenarios—tunnel, bridge, and plain—are used as the training set for the meta-learning network. This includes historical multi-physics data with passing scores, environmental parameters corresponding to the historical multi-physics data with passing scores, and the optimal preprocessing strategy. The initial meta-learning network with network parameters Φ is then pre-trained offline based on the training set to obtain the pre-trained weights Φ*. Finally, the network parameters of the initial meta-learning network are updated to the pre-trained weights Φ* to obtain the constructed meta-learning network Meta-Pret.

[0032] Specifically, in this embodiment, the meta-learning network Meta-Pret is set as a 5-layer MAML (Model-Agnostic Meta-Learning) network with an inner loop step size α=0.01 and an outer loop step size β=0.001.

[0033] Specifically, in step S2, when preprocessing the multiphysics data that has passed the scoring, the multiphysics data Fclean that has passed the scoring is used as the data input, and the environmental parameters θenv=[humidity, wind speed, temperature] corresponding to the multiphysics data that has passed the scoring are used as the driving input. The preprocessing strategy θpret=[waveletbase,decomplevel] is obtained through the meta-learning network Meta-Pret. First, discrete wavelet decomposition is performed on the multiphysics data Fclean that has passed the scoring according to the preprocessing strategy θpret to obtain the subband coefficients Cwavelet. Then, statistical features are extracted from Cwavelet to obtain the data features Dfeat∈R²¹. 4 The data feature Dfeat and the physical prior feature Pprior are concatenated column-wise to obtain the mixed feature vector Hvec∈R². 68 .

[0034] Specifically, the environmental parameter θenv corresponding to the multiphysics data that passed the scoring refers to the triplet environmental data of humidity, wind speed, and temperature given by the VAIN node environmental control sensor at the same clock stamp as the multiphysics data that passed the scoring. The waveletbase refers to the candidate wavelet basis. In this embodiment, the candidate wavelet basis is set to {db4, Coif3, Sym5}. The decomplevel refers to the number of candidate layers. In this embodiment, the number of candidate layers is set to {3, 4, 5}, with a total of 9 combinations. The subband coefficient Cwavelet refers to the subband coefficient obtained by performing DWT according to θpret. In this embodiment, the method for extracting statistical features from Cwavelet is set to 5-dimensional statistics of energy, entropy, kurtosis, mean, and variance, which are then combined into a 214-dimensional Dfeat.

[0035] Specifically, in step S3, when performing anomaly measurement on the mixed feature vector, the mixed feature vector Hvec∈R² output in step S2 is used. 68 As the sole input, after being processed in parallel by the three parallel paths of GT, GD, and Fuse-Att of the hierarchical decoupled generative adversarial network HD-GAN, the reconstructed frame x' is output, and the single-frame anomaly score As is calculated.

[0036] Specifically, the hierarchical decoupling refers to decomposing the original signal into two independent paths: macroscopic trend (low frequency) and local detail (high frequency), which are learned by GT and GD respectively, and then fused through Fuse-Att to achieve trend-detail decoupling modeling. HD-GAN refers to a hierarchical decoupling generative adversarial network composed of GT, GD, Fuse-Att, and corresponding discriminators DT and DD. GT is responsible for macroscopic trend encoding, outputting x'T. In this embodiment, GT is set as a 2-layer bidirectional LSTM with 128 hidden units, responsible for macroscopic trend encoding, i.e., LSTM encoding. GD is responsible for local detail encoding, outputting x'D. In this embodiment, GD is set as a 5-layer TCN with an expansion rate d=[1,2,4,8,16]. Fuse-Att refers to a single-head self-attention decoder, i.e., LSTM decoding, whose output x'=α⊙x'T+(1 α)⊙x'D, α is learnable, with an initial value of 0.7. x'T refers to the macro trend signal generated by GT, and x'D refers to the local detail signal generated by GD. DT and DD are both 3-layer CNN-leakyReLU-Sigmoid structures, which perform binary classification on the authenticity of trends / details respectively.

[0037] Specifically, in step S3, when obtaining the single-frame anomaly score, it is calculated as As = β·‖x x'‖2+(1 β)·(DT(x'T)+DD(x'D)) is calculated, where β=0.7 is the anomaly fusion weight, and ||x x'‖2 is the reconstruction error term, and (DT+DD) is the discrimination confidence term; If As>0.65, then mark the single-frame anomaly of the mixed feature vector and output the catenary anomaly detection result as the current frame multi-physics data anomaly corresponding to the mixed feature vector; Otherwise, mark the single frame of the mixed feature vector as normal, and output the catenary anomaly detection result as the current frame multiphysics data corresponding to the mixed feature vector as normal.

[0038] Specifically, x refers to the original frame Fclean within the same clock window, which is co-stamped with Hvec. In this embodiment, β is set to 0.7. The single-frame anomaly of the mixed feature vector refers to the node-level anomaly label written into the node metadata.

[0039] Specifically, in step S4, when strengthening the anomaly measurement process, the single-frame anomaly score As output in step S3 is used as the monitoring object. The Curri-SCH course learning scheduler is used to increase the difficulty in three stages: wind speed <2m / s→2-10m / s→>10m / s. In each stage, an ε-perturbation is applied to the multiphysics data Fclean that has passed the scoring to obtain the robust anomaly score RW. RW is set as ||HD-GAN(Fclean+ε)||. HD-GAN(Fclean)‖2, where: When RW < δ, the disturbance test is passed, and the original single-frame anomaly score As is output. Otherwise, the disturbance test is deemed to have failed, the original single-frame anomaly score As is not output, and a new frame of multiphysics data Fclean that has passed the score is retrieved from step S1, and steps S2-S4 are repeated. If the perturbation test fails three times in a row, the course is marked as abnormal and the process is rolled back to S3 to retrain the GT and GD parameters.

[0040] Specifically, in this embodiment, Curri-SCH is set as a three-stage data gate driven by the wind speed threshold. Stages 1 / 2 / 3 correspond to stable / wind swing / strong disturbance conditions, respectively. ε refers to the PGD step size, which is set to 0.02 in this embodiment. δ refers to the robustness upper limit, which is set to 0.005 in this embodiment. HD-GAN(Fclean+ε) refers to the reconstructed output after applying ε perturbation to Fclean. HD-GAN(Fclean) refers to the reconstructed output of the original Fclean. The rollback refers to freezing Fuse-Att and retraining only GT and GD for 3 epochs.

[0041] Specifically, in step S5, a dynamic graph G=(V,E) is constructed using the single-frame anomaly scores of all nodes output in step S4 within the same clock window as the source, where V represents the 134 sets of four-in-one sensor nodes arranged in S1, and E represents the spatial distance ≤15m; the node embedding ei is updated through 2 layers of GraphSAGE+GAT, and the edge mutation Δeij=‖ei is calculated. ej‖2, for each edge eij, we have obtained Δeij=‖ei ej‖2, take its normalized value △'eij, △'eij=Δeij / (Δeij+τ), pre-generate edge weight wij=1 / (Δs+1) offline, Δs is the mileage difference along the contact line, obtain the graph structure anomaly score Ag of a single node among all nodes, set Ag=Σ{j∈N(i)}[△'eij·wij] / Σ{j∈N(i)}wij, that is, for all neighboring nodes j of a single node i among all nodes, the graph structure anomaly score Ag of the node is obtained by weighting the normalized mutation and edge weight, Ag∈[0,1); output the comprehensive risk score R=DSFuse(As,Ag) through the DS fusion unit, where: If R>0.65, then mark the node as having exceeded the risk limit, i.e., the local catenary unit consisting of the area 15m away from the node, and transmit the comprehensive risk score R of the node and the local catenary unit to step S6. Otherwise, the risk of the local contact network unit consisting of the area 15m away from the node is marked as not exceeding the limit, and the comprehensive risk score R of the node is transmitted to step S6.

[0042] Specifically, all nodes refer to the four-in-one sensor nodes corresponding to the multiphysics data scored in step S1 within the same catenary anchor section; the same clock window refers to a 5ms sliding window, completely consistent with the IEEE1588v2 synchronization beat in step S1; GraphSAGE+GAT refers to a model containing 2 layers of GraphSAGE and 1 GAT head, with a sampling rate of 25 and an output of 64 dimensions; spatial distance refers to the mileage distance along the contact line between the complete sets of four-in-one nodes on the two suspension cables; ei refers to the 64-dimensional embedding of node i after GraphSAGE+GAT update; ej refers to the 64-dimensional embedding of node j after GraphSAGE+GAT update; and Δeij refers to ||ei||. The L2 distance of ej‖2, where node i refers to the node currently being analyzed among all nodes, and i refers to the label of the currently analyzed node among all nodes. This embodiment does not limit the labeling method, and those skilled in the art can freely set it according to the actual situation, as long as it meets the requirement of distinguishing the labels of all nodes. Node j refers to any node adjacent to node i, and j refers to the label of any node adjacent to node i among all nodes. In this embodiment, τ=0.3 is set. DSFuse refers to the DS fusion device, which is the Dempster-Shafer evidence theory. The pre-generated offline edge weight wij refers to the reciprocal normalized table of the mileage difference Δs along the contact line, and the mileage difference along the contact line refers to the GPS mileage difference between the two suspension wires.

[0043] Specifically, in step S5, the conflict factor K is calculated based on the mass distribution of the node's comprehensive risk score R; K is set as K = ∑(m1(A1)·m2(B1)), where A1∩B1 = m1 is the mass distribution of As evidence, and m2 is the mass distribution of Ag evidence; K∈[0,1], set the tolerance threshold = 0.1; The output of the comprehensive risk score R for this node is optimized based on the conflict factor K, where: When K < 0.1, the output of the comprehensive risk score R for that node is not optimized; When K≥0.1, the output of the comprehensive risk score R for this node is optimized. Instead of using the DS fusion machine to output the comprehensive risk score R, the output of the comprehensive risk score R is changed to a weighted average minimum guarantee, and R=0.5·As+0.5·Ag is set. If K≥0.1 for 3 consecutive times, mark the graph as abnormal and report it to the edge server. Then, retrieve the next frame of new multiphysics data Fclean that has passed the score from step S1 and repeat steps S2-S5.

[0044] Specifically, the mass distribution refers to the mass functions m1 and m2 of As and Ag, A1 refers to the proposition set of As evidence, B1 refers to the proposition set of Ag evidence, and the tolerance threshold is set to 0.1.

[0045] Specifically, in step S6, when verifying and strengthening the comprehensive risk score R of a single node, the comprehensive risk score R value within the same node and the same clock window is used as the source, and three layers of verification are performed sequentially: Simpack high-fidelity simulation layer verification: Using the stiffness of the dropper corresponding to the node as a variable, a 101-level stiffness-decreasing fault library is generated. The comprehensive risk score R value of the node is used to replay the simulation. If R>0.65 and the simulation truth value is a fault, then the simulation is recorded as a hit; otherwise, it is recorded as a false alarm. PGD ​​white-box adversarial attack layer verification: Using the mixed feature vector Hvec of the node as input, apply ε=0.03 perturbation. If the false negative rate of HD-GAN for this node is ≤5%, do not backtrack. Otherwise, backtrack to step S3 to retrain GT and GD for 3 epochs. On-site backtracking layer verification: Compare the comprehensive risk score R value of the node with the results of subsequent manual maintenance. If the hit rate is ≥91.5%, do not backtrack the steps; otherwise, backtrack to step S3 and retrain GT and GD for 3 Epochs.

[0046] Specifically, in this embodiment, the Simpack high-fidelity simulation layer is set as the Simpack dropper-pantograph coupling model of the multibody dynamics software Simpack; the 101-level stiffness reduction refers to the stiffness decreasing from 100% to 0% in 1% increments; the playback simulation refers to using the historical R value of the node to look up the simulation tag; in this embodiment, the PGD step size of the PGD white-box adversarial attack layer is set to 0.03, with 10 iterations; the field backtracking layer refers to the closed-loop comparison of manual inspection results and R values. This embodiment does not limit the method of obtaining the manual inspection results, and those skilled in the art can set it freely, as long as the upload requirement after obtaining the inspection results through manual inspection is met. For example, the inspection results can be set to be transmitted back through a handheld terminal APP to obtain the manual inspection results. The hit rate ≥91.5% refers to the single-node verification requirement that the manual inspection results are consistent with the closed-loop comparison of the R values.

[0047] Specifically, in step S6, when obtaining the post-verification risk score Rv for a single node, Rv is set to R·(1 UI), where UI is the correction coefficient, set UI=max(0,1). |nf| / 3), where nf is the number of simulation misses; if the corrected Rv>0.65, then mark the node verification as abnormal and transmit it to S7; otherwise, mark the node verification as normal and transmit it to S7.

[0048] Specifically, in step S7, when estimating the remaining lifetime of a single node, two levels of prediction are performed sequentially using the verified risk score Rv within the same node and the same clock window as the source: Causal discovery layer: Taking the risk score Rv sequence {Rv(t)} verified over the past 30 days of the node as input, a causal graph C-graph is constructed using the PC-stable algorithm to identify the top-3 root cause features. The top-3 root cause features consist of nf, PGD missed detection flag, and maintenance hit rate output by S6, forming a causal feature vector Cvec, where Cvec∈R³. Transformer sequence prediction layer: Using {Rv(t)} and Cvec as joint inputs, the risk score Rv(t+Δt) for the next 60 days is predicted by 4 layers of Transformers. The number of days that the failure threshold of 0.9 is first exceeded is recorded as the remaining lifetime RL of the node.

[0049] Specifically, PC-stable refers to the stable PC algorithm, and in this embodiment, its significance level is set to 0.01 and the maximum condition set is 3. C-graph refers to a directed acyclic graph with the risk score Rv after node-level verification and the three features output by S6 as vertices and causal relationships as edges. The Top-3 root cause features refer to the three causal features output by S6: the number of simulated missed cases nf, the PGD missed case flag, and the field backtracking hit rate. The 4-layer Transformer refers to a 4-layer Transformer sequence prediction network with 128 hidden layers, 4 heads, and 0.1 dropout. The risk score Rv(t+Δt) for the next 60 days refers to the risk score sequence after node-level verification for the next 60 days output by the Transformer. The failure threshold of 0.9 refers to the failure critical point, and in this embodiment, the failure threshold is set to 0.9. The remaining lifetime RL refers to the number of predicted days before the failure threshold of 0.9 is first exceeded.

[0050] Specifically, in step S7, when the remaining lifetime RL of a single node is obtained, the following is set: When RL < 7 days, mark the node lifetime alarm and transmit the remaining lifetime RL and node lifetime alarm to S8; Otherwise, mark the node's lifetime as normal and transmit the remaining lifetime RL to S8; If the RL is less than 7 days for 3 consecutive times, it will be marked as "abnormal lifespan" and reported to the cloud center, triggering a front-end speed reduction suggestion.

[0051] Specifically, the node lifespan alarm refers to a software tag indicating that the remaining lifespan of the corresponding local contact network unit is less than 7 days. The front-end speed reduction suggestion refers to a temporary speed limit instruction pushed from the cloud center to the dispatch center. The speed limit value is determined by the operation and maintenance rules. The front-end speed reduction suggestion only applies to the anchor section where the lifespan alarm node is located, and it is linked with the actual alarm.

[0052] Specifically, in step S8, when continuously evolving the remaining lifetime of a single node, the remaining lifetime RL within the same node and the same clock window is used as the priority weight, and three-level elastic inference and federated aggregation are executed sequentially: Initial screening of end-sided TinyGAN: Using the remaining lifetime RL of the node as the scheduling key, only nodes with RL < 7 days are processed; the frames of other nodes are skipped directly, and the initial screening gradient is output. Wtiny; Edge Std-GAN confirmation: Using the remaining lifetime RL of the node as the scheduling key, only nodes with RL < 7 days are confirmed, and the confirmation gradient is output. Wstd; In-depth analysis of Heavy-GAN in cloud centers: Using the remaining lifetime RL of the node as the scheduling key, in-depth analysis is performed only on nodes with RL < 7 days to generate node-specific weight increments ΔW.

[0053] Specifically, the priority weight refers to using the remaining lifetime (RL) value of the same node as the frame-level scheduling key value. The smaller the RL, the higher the priority, ensuring that nodes with lifetime alarms have priority in occupying edge-cloud computing power. The TinyGAN refers to a lightweight generative adversarial network with 0.8M parameters obtained by knowledge distillation. It is responsible for millisecond-level initial screening on the edge, only processing nodes with RL less than 7 days, and outputting the initial screening gradient. Wtiny, the Std-GAN mentioned refers to a 45M parameter standard generative adversarial network, responsible for edge confirmation, only processing nodes with RL less than 7 days, and outputting confirmation gradients. Wstd, where Heavy-GAN refers to a 210M parameter complete generative adversarial network, is responsible for deep analysis in the cloud center, only processes nodes with an RL of less than 7 days, and generates a node-specific weight increment ΔW. The RL<7-day priority refers to a frame-level scheduling queue. The TinyGAN / Std-GAN / Heavy-GAN three-level inference is only open to node frames with RL<7; other frames are skipped to save computing power. Wtiny refers to the gradient increment generated after the initial screening at the end. Wstd refers to the gradient increment generated after edge confirmation, and ΔW refers to the node-specific weight increment returned after aggregation by the cloud center FedAvg, which carries the version number and RL value and is used for the next cycle model update.

[0054] Specifically, in step S9, when automatically iterating the evolved weight of a single node, the evolved weight ΔW of a single node within the same node and the same clock window is used as the update source, and the monitoring-shadow-grayscale three stages are executed sequentially: Mon-Board monitoring layer: The node's 127 KPIs are tracked in real time, generating a node-level dashboard. If a KPI changes by more than ±2σ, the shadow mode is triggered in advance. Shadow Mode Layer: The new model B runs in parallel with the shadow for 7 days, only recording predictions and not linking with real alarms. If the false alarm rate of the node is significantly lower than 0.01 and the false alarm rate does not increase during the shadow period, it will be put into grayscale. A / B grayscale layers: Replace the old model A with a three-stage gray-scale release at the node level, from 20% to 50% to 100%. If the KPI still meets the target after the gray-scale release, a new version number Ver' is generated. The new version number Ver' is output as the iteration version number Ver of a single node, and the current frame process ends. The clock window slides, and the process continues in the next frame. Otherwise, the iteration version number Ver of a single node is output, the current frame process ends, the clock window slides, and the process continues in the next frame.

[0055] Specifically, the Mon-Board refers to a node-level end-to-end monitoring dashboard that tracks 127 KPIs for the node in real time. These 127 KPIs include 127 node-level frame-level metrics such as As, Ag, R, Rv, RL, ΔW, Ver, disturbance failure count, federated anomaly count, and lifetime alarm count. The node-level dashboard refers to an independent sub-panel generated by the Mon-Board for each VAIN node, displaying the real-time curves and threshold lines of the aforementioned 127 KPIs. A KPI mutation is defined as any KPI value differing from its 7-day average by more than ±2σ, where σ is the historical standard deviation. The new model B refers to an upgraded version of HD-GAN with ΔW, where only the GT / GD weights are updated, and the version number Ver' is greater than the old Ver. The shadow union... "Running" refers to the simultaneous operation of the new model B and the old model A. B only records predictions and does not participate in real alarm linkage. The real alarm refers to the real operation and maintenance alarm triggered by R>0.65 output in step S5. The "shadow period" refers to the 7-day shadow window, comparing only the differences in node-level As, R, and Rv between B and A. The significance of the node miss rate means that the node-level miss rate during the shadow period equals the proportion of B's ​​outputs that are not abnormal in the real abnormal frames, using a two-tailed t-test. The two-tailed t-test is a statistical hypothesis testing method used to determine whether there is a significant difference between the means of two sets of data, regardless of the direction of the difference. In this embodiment, it is used to compare the miss rates of the new model B and the old model A at the same node to determine whether the upgraded model is significantly better than the original model. "p<0.01" means that the significance level of the two-tailed t-test is 0.01. The statement that the false alarm rate does not increase means that the node-level false alarm rate during the shadow period is ≤ 105% of period A. The statement that the false alarm rate does not increase means that after the shadow verification is passed, the node-level 20%-50%-100% grayscale release queue is released. The statement that the grayscale ratio is controlled independently by a single VAIN node and is not mixed with the global anchor segment means that the 20%-50%-100% ratio refers to the three-stage ratio at the node frame level: 20% of frames in stage 1 use B, 50% in stage 2, and 100% in stage 3. The statement that the grayscale release means that the weights of the old model A are gradually replaced with ΔW of B at the node frame level, and the KPI is compared in real time. It can be rolled back at any time. The statement that the old model A refers to the HD-GAN (GT / GD) weights of the previous Ver version is released. The phrase "KPIs still meet targets after grayscale" means that after grayscale completion, the difference between the 127 KPIs of this node and period A is < ±1σ, and the false alarm rate and missed alarm rate do not deteriorate. "Ver" refers to the current model version number of the node, and "new version number Ver'" refers to the node-specific version number that automatically increments after successful grayscale completion, where Ver' = Ver + 1. "Node" refers to a single node among the 134 VAIN nodes in step S1. "Clock window sliding" means that the frame-level time window slides forward by one sampling step of 5ms, aligning with the IEEE1588v2 synchronization clock in step S1. "Continuing to the next frame" means that the frame metadata data packet carrying the new Ver' is archived, entering the next 5ms frame loop, and the process returns to step S1.

[0056] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in overhead contact lines based on generative adversarial networks and LSTM encoding and decoding, characterized in that, The method includes: Step S1: Synchronously acquire multiphysics data and obtain the data quality score of each frame based on the three-dimensional index to obtain multiphysics data that passes the score, and optimize the synchronous acquisition process. Step S2: Obtain physical prior features through finite element simulation, construct a meta-learning network, and use the meta-learning network to preprocess the multi-physics data scored in step S1 to obtain data features. Then, concatenate the physical prior features with the data features to obtain a hybrid feature vector. Step S3: Use a layered decoupled generative adversarial network and LSTM encoding and decoding to measure the anomalies in the hybrid feature vector, obtain the single-frame anomaly score, and output the contact network anomaly detection results. Step S4: The anomaly measurement process is reinforced by course learning and adversarial regularization to obtain a robust anomaly score, and the single-frame anomaly score of all nodes in step S3 is output based on the robust anomaly score. Step S5: Obtain the single-frame anomaly score of all nodes output in step S4, and obtain the comprehensive risk score of a single node among all nodes. Step S6: Use a three-layer verification system combining virtual and physical elements to verify and reinforce the comprehensive risk score R of a single node, and obtain the verified risk score Rv of the single node, and transmit it to step S7. Step S7: Estimate the remaining lifetime of a single node by performing a post-verification risk score, obtain the remaining lifetime RL of the single node, and transmit it to step S8. Step S8: Utilize cloud-edge-device elastic inference and federated lifelong learning to continuously evolve the remaining lifetime of a single node, obtain the evolved weight ΔW of the single node, and transmit it to step S9. Step S9: Automated iteration of the evolved weights of a single node is performed using the full-link monitoring dashboard and A / B test shadow mode to obtain the iterated model version number Ver of the single node. The current frame process ends, the clock window slides, and the process continues in the next frame.

2. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 1, characterized in that, In step S1, the data quality score Q for each frame is calculated, and Q is set as Q = ω1·SNR + ω2·Stability + ω3·Integrity; In step S1, when obtaining multiphysics data that has passed the scoring and optimizing the synchronous acquisition process, the quality score Q of each frame of data is compared with the preset acquisition optimization parameter Q0, wherein: When Q > Q0, the multiphysics data corresponding to the quality score of the frame is used as the multiphysics data that passes the scoring, and the synchronous acquisition process is not optimized. Otherwise, the multiphysics data corresponding to the quality score of that frame will be regarded as multiphysics data that failed the scoring, and the multiphysics data will be re-acquired synchronously in order to optimize the synchronous acquisition process. If the synchronous acquisition process is optimized three times in a row, mark it as "sensor abnormal" and report it to the edge server.

3. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 2, characterized in that, In step S2, the ANSYS finite element table Pprior=[f1,…,f6,φ1,…,φ6]∈R corresponding to the same dropper model is... 54 As a priori physical characteristic; In step S2, when constructing the meta-learning network, the physical prior features of three actual scenarios—tunnel, bridge, and plain—historical multiphysics data with passing scores, environmental parameters corresponding to the historical multiphysics data with passing scores, and the optimal preprocessing strategy are used as the training set for the meta-learning network. The initial meta-learning network with network parameters Φ is pre-trained offline based on the training set to obtain the pre-trained weights Φ*. The network parameters of the initial meta-learning network are then updated to the pre-trained weights Φ* to obtain the constructed meta-learning network Meta-Pret. In step S2, when preprocessing the multiphysics data that has passed the scoring, the multiphysics data Fclean that has passed the scoring is used as the data input, and the environmental parameters θenv=[humidity, wind speed, temperature] corresponding to the multiphysics data that has passed the scoring are used as the driving input. The preprocessing strategy θpret=[waveletbase,decomplevel] is obtained through the meta-learning network Meta-Pret. First, discrete wavelet decomposition is performed on the multiphysics data Fclean that has passed the scoring according to the preprocessing strategy θpret to obtain the subband coefficients Cwavelet. Then, statistical features are extracted from Cwavelet to obtain the data features Dfeat∈R²¹. 4 The data feature Dfeat and the physical prior feature Pprior are concatenated column-wise to obtain the mixed feature vector Hvec∈R². 68 .

4. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 3, characterized in that, In step S3, when performing anomaly measurement on the mixed feature vector, the mixed feature vector Hvec∈R² output in step S2 is used. 68 As the sole input, after being processed in parallel by the three parallel paths of the hierarchical decoupled generative adversarial network HD-GAN (GT, GD, Fuse-Att), the reconstructed frame x' is output, and the single-frame anomaly score As is calculated; In step S3, when obtaining the single-frame anomaly score, it is calculated as As = β·‖x x'‖2+(1 β)·(DT(x'T)+DD(x'D)) is calculated, where β=0.7 is the anomaly fusion weight, and ||x x'‖2 is the reconstruction error term, and (DT+DD) is the discrimination confidence term; If As>0.65, then mark the single-frame anomaly of the mixed feature vector and output the catenary anomaly detection result as the current frame multi-physics data anomaly corresponding to the mixed feature vector; Otherwise, mark the single frame of the mixed feature vector as normal, and output the catenary anomaly detection result as the current frame multiphysics data corresponding to the mixed feature vector as normal.

5. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 4, characterized in that, In step S4, when strengthening the anomaly measurement process, the single-frame anomaly score As output in step S3 is used as the monitoring object. The Curri-SCH course learning scheduler is used to increase the difficulty in three stages: wind speed <2m / s→2-10m / s→>10m / s. In each stage, an ε-perturbation is applied to the multiphysics data Fclean that has passed the scoring to obtain the robust anomaly score RW. RW is set as ||HD-GAN(Fclean+ε)|| HD-GAN(Fclean)‖2, where: When RW < δ, the disturbance test is passed, and the original single-frame anomaly score As is output. Otherwise, the disturbance test is deemed to have failed, the original single-frame anomaly score As is not output, and a new frame of multiphysics data Fclean that has passed the score is retrieved from step S1, and steps S2-S4 are repeated. If the perturbation test fails three times in a row, the course is marked as abnormal and the process is rolled back to S3 to retrain the GT and GD parameters.

6. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 5, characterized in that, In step S5, a dynamic graph G=(V,E) is constructed using the single-frame anomaly scores of all nodes output in step S4 within the same clock window as the source, where V represents the 134 sets of four-in-one sensor nodes arranged in S1, and E represents the spatial distance ≤15m. The node embedding ei is updated through 2 layers of GraphSAGE+GAT, and the edge mutation Δeij=‖ei is calculated. ej‖2, for each edge eij, we have obtained Δeij=‖ei ej‖2, take its normalized value △'eij, △'eij=Δeij / (Δeij+τ), generate edge weight wij=1 / (Δs+1) offline in advance, Δs is the mileage difference along the contact line, and obtain the graph structure anomaly score Ag of a single node among all nodes. Set Ag=Σ{j∈N(i)}[△'eij·wij] / Σ{j∈N(i)}wij, that is, for all neighboring nodes j of a single node i among all nodes, the graph structure anomaly score Ag of the node is obtained by weighted average of normalized mutation and edge weight, Ag∈[0,1); The DS fusion unit outputs a comprehensive risk score R = DSFuse(As,Ag), where: If R>0.65, then mark the node as having exceeded the risk limit, i.e., the area within a 15m spatial distance from the node constitutes the local catenary unit, and transmit the comprehensive risk score R of the node and the local catenary unit to step S6. Otherwise, the risk of the local contact network unit consisting of the area within 15m of the node is marked as not exceeding the limit, and the comprehensive risk score R of the node is transmitted to step S6; In step S5, the conflict factor K is calculated based on the mass distribution of the node's comprehensive risk score R; K is set as K = ∑(m1(A1)·m2(B1)), where A1∩B1 = m1 is the mass distribution of As evidence, and m2 is the mass distribution of Ag evidence; K∈[0,1], set the tolerance threshold = 0.1; The output of the comprehensive risk score R for this node is optimized based on the conflict factor K, where: When K < 0.1, the output of the comprehensive risk score R for that node is not optimized; When K≥0.1, the output of the comprehensive risk score R for this node is optimized. Instead of using the DS fusion machine to output the comprehensive risk score R, the output of the comprehensive risk score R is changed to a weighted average minimum guarantee, and R=0.5·As+0.5·Ag is set. If K≥0.1 for 3 consecutive times, mark the graph as abnormal and report it to the edge server. Then, retrieve the next frame of new multiphysics data Fclean that has passed the score from step S1 and repeat steps S2-S5.

7. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 6, characterized in that, In step S6, when verifying and strengthening the comprehensive risk score R of a single node, the comprehensive risk score R value within the same node and the same clock window is used as the source, and three layers of verification are performed sequentially: Simpack high-fidelity simulation layer verification: Using the stiffness of the dropper corresponding to the node as a variable, a 101-level stiffness-decreasing fault library is generated. The comprehensive risk score R value of the node is used to replay the simulation. If R>0.65 and the simulation truth value is a fault, then the simulation is recorded as a hit; otherwise, it is recorded as a false alarm. PGD ​​white-box adversarial attack layer verification: Using the mixed feature vector Hvec of the node as input, apply ε=0.03 perturbation. If the false negative rate of HD-GAN for this node is ≤5%, do not backtrack. Otherwise, backtrack to step S3 to retrain GT and GD for 3 epochs. On-site backtracking layer verification: Compare the comprehensive risk score R value of the node with the results of subsequent manual maintenance. If the hit rate is ≥91.5%, do not backtrack the steps; otherwise, backtrack to step S3 and retrain GT and GD for 3 Epochs. In step S6, when obtaining the post-verification risk score Rv for a single node, Rv is set to R·(1 UI), where UI is the correction coefficient, set UI=max(0,1). |nf| / 3), where nf is the number of simulation misses; if the corrected Rv>0.65, then mark the node verification as abnormal and transmit it to S7; otherwise, mark the node verification as normal and transmit it to S7.

8. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 7, characterized in that, In step S7, when estimating the remaining lifetime of a single node, two levels of prediction are performed sequentially using the verified risk score Rv within the same node and the same clock window as the source: Causal discovery layer: Taking the risk score Rv sequence {Rv(t)} verified over the past 30 days of the node as input, a causal graph C-graph is constructed using the PC-stable algorithm to identify the top-3 root cause features. The top-3 root cause features consist of nf output by S6, PGD missed detection flag, and maintenance hit rate, forming a causal feature vector Cvec, where Cvec∈R³. Transformer sequence prediction layer: Using {Rv(t)} and Cvec as joint inputs, the risk score Rv(t+Δt) for the next 60 days is predicted through 4 layers of Transformer. The number of days that the failure threshold of 0.9 is first exceeded is recorded as the remaining lifetime RL of the node. In step S7, when the remaining lifetime RL of a single node is obtained, the following is set: When RL < 7 days, mark the node lifetime alarm and transmit the remaining lifetime RL and node lifetime alarm to S8; Otherwise, mark the node's lifetime as normal and transmit the remaining lifetime RL to S8; If the lifespan is less than 7 days for 3 consecutive times, it will be marked as "abnormal lifespan" and reported to the cloud center, triggering a front-end speed reduction suggestion.

9. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 8, characterized in that, In step S8, when continuously evolving the remaining lifetime of a single node, the remaining lifetime RL within the same node and the same clock window is used as the priority weight, and three-level elastic inference and federated aggregation are executed sequentially: Initial screening of end-sided TinyGAN: Using the remaining lifetime RL of the node as the scheduling key, only nodes with RL < 7 days are processed; the frames of other nodes are skipped directly, and the initial screening gradient is output. Wtiny; Edge Std-GAN confirmation: Using the remaining lifetime RL of the node as the scheduling key, only nodes with RL < 7 days are confirmed, and the confirmation gradient is output. Wstd; In-depth analysis of Heavy-GAN in cloud centers: Using the remaining lifetime RL of the node as the scheduling key, in-depth analysis is performed only on nodes with RL < 7 days to generate node-specific weight increments ΔW.

10. The contact network anomaly detection method based on generative adversarial networks and LSTM encoding and decoding according to claim 9, characterized in that, In step S9, when automatically iterating the evolved weights of a single node, the evolved weights ΔW of the same node within the same clock window are used as the update source, and the monitoring-shadow-grayscale three-stage process is executed sequentially. Mon-Board monitoring layer: Track 127 KPIs of this node in real time and generate a node-level dashboard. If a KPI changes by more than ±2σ, trigger the shadow mode in advance. Shadow Mode Layer: The new model B runs in parallel with the shadow for 7 days, only recording predictions and not linking with real alarms. If the false alarm rate of the node is significantly lower than 0.01 and the false alarm rate does not increase during the shadow period, it will be put into grayscale. A / B grayscale layers: Replace the old model A with a three-stage gray-scale release at the node level, from 20% to 50% to 100%. If the KPI still meets the target after the gray-scale release, a new version number Ver' is generated. The new version number Ver' is output as the iteration version number Ver of a single node, and the current frame process ends. The clock window slides, and the process continues in the next frame. Otherwise, the iteration version number Ver of a single node is output, the current frame process ends, the clock window slides, and the process continues in the next frame.

Images