Relay state prediction and fault early warning method and system based on deep learning

Abstract

The invention discloses a relay state prediction and fault early warning method and system based on deep learning, and the method comprises the steps: S1, building a constraint condition of a generative adversarial network based on a relay physical model, and forming an enhanced fault waveform signal according with a physical rule through adversarial training; s2, receiving a real-time current and voltage signal and a mechanical vibration signal, and extracting an electric signal feature vector by using a time sequence convolutional network; s3, inputting the combined feature tensor into the lightweight assessment model, and outputting a health degree scoring signal; s4, responding to the meta-learning activation instruction, loading historical data of equipment to construct a parameter optimization set, performing online fine tuning on the early warning model based on a meta-learning framework, and generating a fault determination parameter; and S5, analyzing real-time signal characteristics according to the fine-tuned judgment parameters, and outputting graded early warning signals to a monitoring terminal. According to the method, the problems of early state prediction and accurate early warning of the relay under small sample fault data can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnosis and prediction of relay faults, and in particular to a method and system for relay state prediction and fault early warning based on deep learning. Background Art

[0002] Relays are critical power control components, and faults such as contact sticking and coil aging can easily cause electrical accidents. Currently, the industry primarily uses threshold methods or machine learning for condition monitoring, but these methods face three major bottlenecks: First, the scarcity of actual fault samples leads to insufficient generalization of supervised learning models. Public research shows that a certain relay model has fewer than 1,000 fault data points over its entire lifecycle. Second, manual feature extraction struggles to capture implicit correlations between electrical signals and mechanical vibrations, such as the cross-modal coupling between contact bounce vibration and arc current. Third, fixed threshold warning mechanisms cannot adapt to the degradation characteristics of different devices, often resulting in false positives or missed warnings.

[0003] While recent research has employed generative adversarial networks to augment samples, the generated data violates physical laws, leading to model failure. For example, a paper reported that the contact welding current waveform generated by one paper failed to satisfy the energy conservation equation. Furthermore, the simple fusion of time series models and computer vision models is susceptible to interference from asynchronous sensor sampling. One experiment showed that a 0.1 millisecond delay can reduce the correlation between vibration and current features by 37%. Furthermore, traditional early warning models require retraining for new equipment deployments, taking an average of over 72 hours on industrial sites. Therefore, there is an urgent need to build an intelligent early warning system that integrates physical laws, adapts to equipment characteristics, and possesses incremental learning capabilities. Summary of the Invention

[0004] In view of the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a deep learning-based relay state prediction and fault warning method and system to solve the problem of early relay state prediction and accurate warning under small sample fault data. The present invention creates high-fidelity fault samples through a physically constrained generative adversarial network, combines multimodal spatiotemporal feature fusion technology to extract deep correlations between relay electro-mechanical signals, and uses a lightweight evaluation model to dynamically output health scores. When the score is abnormal, a meta-learning framework is triggered to quickly adapt device characteristics, realizing incremental learning-based graded warning of contact anomalies and coil degradation, breaking through the traditional method's reliance on fault samples and fixed thresholds.

[0005] The present invention provides a relay state prediction and fault warning method based on deep learning, comprising: S1: Constructing constraints for a generative adversarial network based on the relay physical model, and generating enhanced fault waveform signals that conform to physical laws through adversarial training; S2: receive real-time current voltage signals and mechanical vibration signals, extract electric signal feature vectors using a time sequence convolution network, simultaneously generate mechanical feature vectors through a mechanical feature extraction network, adopt an adaptive attention mechanism to dynamically fuse the two types of feature vectors to form a joint feature tensor; S3: input the joint feature tensor into a lightweight evaluation model to output a health degree score signal, and generate a meta-learning activation instruction when the score signal first falls below a preset threshold; S4: in response to the meta-learning activation instruction, load device historical data to construct a parameter optimization set, and perform online fine-tuning on the early warning model based on a meta-learning framework to generate device customized fault judgment parameters; S5: analyze real-time signal features according to the fine-tuned judgment parameters, and output a graded early warning signal to a monitoring terminal when abnormal contact features or coil degradation trends are identified.

[0006] In an embodiment of the present application, in S1: based on the electromagnetic characteristic parameters and mechanical motion characteristic parameters of the relay, a physical constraint layer is constructed, which is embedded into the output end of the generator of the generative adversarial network, and the discriminator is used for adversarial training of the real fault waveform feature distribution and the generated signal, so that the generated enhanced fault waveform signal meets the requirements of physical law constraint and statistical distribution authenticity, and the signal will be one of the input signal sources of the multi-source feature fusion step.

[0007] In an embodiment of the present application, in S2: the time sequence convolution network adopts an expansion convolution structure to extract deep time sequence dependent features of the current voltage signal, the mechanical feature extraction network captures the frequency domain mutation mode of the vibration signal through a residual connection module, and the adaptive attention mechanism generates dynamic weight coefficients by calculating the spatio-temporal correlation degree of the electric signal feature vector and the mechanical feature vector, and finally forms a joint feature tensor containing a weighted fused multi-dimensional state representation vector.

[0008] In an embodiment of the present application, in S3: the lightweight evaluation model adopts a multi-head self-attention mechanism to construct a feature interaction module, performs time sequence slicing processing on the joint feature tensor and calculates the health degree score of each slice, and then generates the final health degree score signal through sliding weighted average, and triggers the meta-learning activation instruction when the fluctuation amplitude of the score signal in three consecutive sampling periods exceeds the preset tolerance range and the mean value is lower than the threshold.

[0009] In an embodiment of the present application, in S4: when constructing the parameter optimization set from the device historical data, the historical running segments similar to the current health degree score signal drop pattern are preferentially selected, the meta-learning framework generates fast adaptive parameters by simulating the fault evolution process based on the historical segments in the internal loop, and verifies the effectiveness of the parameters using real-time signals in the external loop to output device customized fault judgment parameters.

[0010] In one embodiment of the present invention, in S5: contact abnormality feature detection uses waveform morphology analysis technology to identify current zero-crossing delay and arc duration anomalies, and coil degradation trend analysis is achieved by comparing the Mahalanobis distance change rate of the real-time inductance feature vector and the standard parameter vector. When any detection result of the two exceeds the safety limit, a first-level warning signal is generated. When both exceed the limit at the same time, a second-level warning signal is generated, and a third-level warning signal is added when the contact temperature rises abnormally.

[0011] In one embodiment of the present invention, the physical constraint layer construction process includes: decomposing the relay electromagnetic attraction process into a static attraction characteristic stage and a dynamic motion characteristic stage, introducing magnetic circuit equivalent parameters to constrain the current-flux conversion relationship in the static stage, and introducing the mass-spring damping equation to constrain the displacement-acceleration conversion relationship in the dynamic stage, and generating a fault waveform signal that conforms to the electromechanical coupling law through alternating optimization of the two-stage constraint conditions.

[0012] In one embodiment of the present invention, the adaptive attention mechanism is implemented as follows: a cross-modal correlation matrix is ​​established between the electrical signal feature vector and the mechanical feature vector, the feature contribution of each time step is calculated through learnable parameters, and then the mechanical signal feature vector is time-axis aligned and compensated according to the contribution distribution, and finally a joint feature tensor is generated that eliminates the sensor sampling delay error.

[0013] In one embodiment of the present invention, the health score signal verification mechanism includes: when the meta-learning activation instruction is triggered, the health data two hours before the current score signal is generated is retained as a verification set, and after the incremental model adaptation step is completed, the verification set is input into the fine-tuning model. If the deviation rate between the newly generated score signal and the original signal exceeds the allowable threshold, the model rollback mechanism is activated and a calibration alarm signal is issued.

[0014] The present invention also includes a relay state prediction and fault warning system based on deep learning, including: The generation module constructs the constraints of the generative adversarial network based on the relay physical model, and forms an enhanced fault waveform signal that conforms to the laws of physics through adversarial training; The coordination module receives real-time current and voltage signals and mechanical vibration signals, uses a temporal convolutional network to extract electrical signal feature vectors, and simultaneously generates mechanical feature vectors through a mechanical feature extraction network. An adaptive attention mechanism is used to dynamically fuse the two feature vectors to form a joint feature tensor. The comparison module inputs the joint feature tensor into the lightweight evaluation model and outputs a health score signal. When the score signal falls below a preset threshold for the first time, a meta-learning activation instruction is generated. The pre-performance module loads the device historical data construction parameter optimization set in response to the meta-learning activation instruction, performs online fine-tuning on the early warning model based on the meta-learning framework, and generates device customized fault judgment parameters; The analysis module analyzes real-time signal characteristics according to the fine-tuned judgment parameters, and outputs a graded early warning signal to a monitoring terminal when abnormal contact characteristics or coil degradation trends are identified.

[0015] The relay state prediction and fault early warning method and system based on deep learning provided by the application create high-fidelity fault samples through a physically constrained generative adversarial network, extract deep correlations of relay electro-mechanical signals by combining multi-modal spatio-temporal feature fusion technology, dynamically output health scores using a lightweight evaluation model, trigger the meta-learning framework to quickly adapt to device characteristics when the score is abnormal, realize graded early warning of contact abnormalities and coil degradation based on incremental learning, and break through the dependence on fault samples and fixed thresholds of traditional methods. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0017] Figure 1 The method flowchart of the relay state prediction and fault early warning method based on deep learning; Figure 2 The system architecture diagram of the relay state prediction and fault early warning system based on deep learning. DETAILED DESCRIPTION

[0018] The embodiments of the application will be described below through specific concrete examples. Those skilled in the art can easily understand other advantages and effects of the application from the content disclosed in the specification. The application can also be implemented or applied through other different specific embodiments, and each detail in the specification can be modified or changed based on different views and applications without departing from the spirit of the application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0019] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the application in a schematic manner, and the diagrams only show the components related to the application, not the number, shape and size of the components when actually implemented. The actual implementation of each component may be a random change, and the component layout pattern may be more complex.

[0020] In the following description, numerous specific details are discussed in order to provide a thorough explanation of embodiments of the application. However, it will be apparent to one of ordinary skill in the art that embodiments of the application can be practiced without these specific details. In other instances, well-known structures and devices are not described in detail in order to avoid obscuring embodiments of the application.

[0021] Referring to Figures 1-2 , a deep learning-based relay state prediction and fault warning method and system of the application is shown. The deep learning-based relay state prediction and fault warning method of the application comprises: S1: constructing constraint conditions of a generative adversarial network based on a relay physical model, and forming an enhanced fault waveform signal conforming to physical laws through adversarial training; S2: receiving real-time current and voltage signals and mechanical vibration signals, extracting electrical signal feature vectors using a time sequence convolution network, simultaneously generating mechanical feature vectors through a mechanical feature extraction network, and dynamically fusing the two types of feature vectors using an adaptive attention mechanism to form a joint feature tensor; S3: inputting the joint feature tensor into a lightweight evaluation model to output a health score signal, and generating a meta-learning activation instruction when the score signal first falls below a preset threshold; S4: in response to the meta-learning activation instruction, loading device historical data to construct a parameter optimization set, and online fine-tuning the warning model based on a meta-learning framework to generate device-customized fault judgment parameters; S5: analyzing real-time signal features according to the fine-tuned judgment parameters, and outputting a graded warning signal to a monitoring terminal when abnormal contact features or coil degradation trends are identified.

[0022] As Figure 1As shown, the core method of the application includes a five-step signal processing chain: first, in the physical enhanced signal generation step, the physical constraint conditions for generating the generative adversarial network are established based on the electromagnetic characteristics and mechanical motion characteristics of the relay, and the physical law-compliant fault waveform signal is formed through the adversarial training of the generator and the discriminator, which will be one of the inputs of multi-source feature fusion; then, in the multi-source feature fusion step, the real-time current and voltage signals are input into the time sequence convolution network to extract their time sequence dependent features, and the mechanical vibration signals are captured through the residual network to extract the frequency domain mutation features, and the two types of feature vectors are dynamically weighted and fused through the adaptive attention mechanism to form a joint feature tensor; then, the health assessment signal triggers the step to input the joint feature tensor into the lightweight evaluation model, analyzes the feature interaction relationship through the multi-head self-attention mechanism, and outputs the health score signal, and when the score fluctuates continuously and the mean value is lower than the threshold, the meta-learning activation instruction is generated; then, the incremental model adaptation step responds to the instruction, filters the similar degradation patterns in the device historical data to construct the parameter optimization set, and updates the model through the double gradient update circuit for fine-tuning, freezes the feature layer in the first stage to update the decision layer to generate the primary parameters, and verifies the real-time signal in the second stage to generate the final fault judgment parameters; finally, the hierarchical early warning decision step analyzes the real-time signal based on the fine-tuned parameters, identifies the abnormal contact features through waveform morphology recognition, analyzes the coil degradation trend through inductance vector distance, and outputs the hierarchical early warning signal to the monitoring terminal according to the abnormal combination type, to realize the closed-loop processing from physical modeling to hierarchical decision.

[0023] Further, the constraint layer construction of the physical enhanced signal generation step specifically includes: decomposing the electromagnetic attraction process of the relay into two stages of static attraction force characteristics and dynamic motion characteristics, introducing the magnetic circuit equivalent parameter to constrain the current-magnetic flux conversion relationship in the static stage, introducing the mass-spring damping equation to constrain the displacement-acceleration conversion relationship in the dynamic stage, embedding the constraint conditions of the two stages into the generator output layer of the generative adversarial network, comparing the statistical distribution difference between the real fault waveform and the generated signal through the discriminator, alternately optimizing the generator parameters and the discriminator parameters in the adversarial training process, so that the generated signal meets the physical equation boundary conditions and the fault data distribution characteristics at the same time, and finally outputs the enhanced fault waveform signal with physical consistency, which will be a supplementary input source of the multi-source feature fusion step, and will be input into the time sequence convolution network together with the measured current and voltage signals for feature extraction. The multi-source feature fusion is specifically implemented as follows: the time sequence convolution network adopts a three-layer dilated convolution structure, and the dilated coefficient increases exponentially to expand the time sequence receptive field and capture the periodic arc feature of the current signal; the mechanical feature extraction network embeds a short-time Fourier transform layer in the residual block to extract the harmonic components of the vibration signal; the adaptive attention mechanism first establishes a time step correlation matrix of the electrical signal feature vector and the mechanical feature vector, calculates the contribution weight of each time step to the target time through a fully connected layer, realizes time axis compensation by cubic spline interpolation on the mechanical feature vector according to the weight distribution, and finally generates a spatiotemporally synchronized joint feature tensor according to the attention weight.

[0024] As Figure 1 shown, the multi-source feature fusion step uses an expanded convolutional structure to expand the receptive field and capture long-term temporal dependence patterns of the current-voltage signal; the mechanical feature extraction network builds a deep frequency domain feature extraction path through residual connection to identify the impact harmonic components in the vibration signal; the adaptive attention mechanism first establishes a cross-modal association matrix between electrical and mechanical signal features, calculates the contribution of each time step feature through learnable parameters, and compensates for the time axis shift of the mechanical signal feature vector to eliminate the sampling time delay, and finally dynamically weights the compensated feature vector to generate a joint feature tensor that retains the coupling relationship between the temporal evolution of the electrical signal and the frequency domain mutation characteristics of the mechanical signal. The lightweight evaluation model of the health evaluation signal triggering step uses a multi-head self-attention mechanism to build a feature interaction module, which first slices the joint feature tensor into fixed time windows, each slice is independently input into the feature interaction module to calculate the local health score, and then the sliding window weighted average algorithm is used to aggregate the scores to generate a global health score signal; when the fluctuation amplitude of the score signal in three consecutive sampling periods exceeds the preset tolerance and the moving average is below the threshold, a meta-learning activation instruction is generated; At the same time, a verification mechanism is set up: cache the health data of the previous two hours as a validation set after the instruction is triggered, input the validation set into the fine-tuning model after the incremental model adaptation is completed, and if the deviation rate of the new score signal from the original signal exceeds the limit, the model rollback is started and a calibration alarm signal is output to ensure the reliability of the evaluation.

[0025] Further, the parameter optimization set of the incremental model adaptation step needs to filter the similar operation segments in the device historical data as the current health degradation pattern, and the meta-learning framework realizes rapid adaptation through a double gradient update loop: the internal loop freezes the feature extraction network parameters and only updates the early warning decision layer parameters, and generates primary adaptation parameters based on the device degradation path simulated by the historical segments; the external loop inputs the real-time signal into the model containing the primary parameters, constructs a loss function based on the difference between the prediction result and the actual detection result, and optimizes the feature extraction network and decision layer parameters to generate the final fault judgment parameters; The parameter will dynamically update the judgment boundary of the hierarchical early warning decision step. The contact abnormal feature detection of the hierarchical early warning decision step uses waveform morphology analysis technology to identify abnormalities by comparing the current zero-crossing delay time and arc duration with the standard waveform; the coil degradation trend analysis calculates the Mahalanobis distance change rate of the real-time inductance feature vector and the standard parameter vector; When only the contact abnormality or coil degradation single feature exceeds the limit, output a first-level warning signal; When the contact abnormality and coil degradation exceed the limit at the same time, output a second-level warning signal; If the contact temperature sensor detects abnormal temperature rise, output a third-level warning signal; All warning signals are transmitted to the monitoring terminal through an encrypted channel and trigger the audible and visual alarm.

[0026] AsFigure 1 As shown in the figure, the physical constraint layer is further refined: in the static suction characteristic stage, the equivalent parameters of the magnetic circuit are converted into current-flux conversion constraints, and the integral area of ​​the generated current waveform is limited by the flux conservation equation. In the dynamic motion characteristic stage, the mass-spring damping equation is discretized into displacement-acceleration differential constraints, limiting the frequency band energy distribution of the generated vibration signal. The two-stage constraints are implemented through an alternating optimization mechanism: the first round of training prioritizes satisfying the static constraints to generate the current waveform, and the second round of training calculates the dynamic constraints based on this current waveform to generate the vibration signal. The optimization is iteratively optimized until the generated signal satisfies both the electromagnetic-mechanical coupling equations. The specific process of cross-modal signal alignment achieved by the adaptive attention mechanism is as follows: first, the electrical signal feature vector and the mechanical signal feature vector are expanded by time step to construct a correlation matrix, where the matrix elements represent the similarity of different modal features at the same time. A learnable weight layer is used to calculate the contribution of each time step feature to the target time. Based on the contribution distribution, the mechanical signal feature vector is compensated for by time axis nonlinear interpolation to eliminate millisecond-level delays caused by sensor sampling frequency differences. The aligned feature vectors are then weighted using soft attention to generate a spatiotemporal fusion joint feature tensor.

[0027] Specifically, the operating rules of the health score verification mechanism are as follows: the timestamp T is recorded at the moment the meta-learning activation instruction is triggered, and all health scores in the period from T-120 minutes to T are extracted as the verification set; after the incremental model adaptation is completed, the verification set is immediately input into the fine-tuning model to calculate the point-by-point absolute deviation rate between the new score sequence and the original sequence; if the deviation rate of more than 20% of the data points is greater than 15%, the model adaptation is determined to have failed and the rollback mechanism is initiated: the model parameters before fine-tuning are restored, and a calibration alarm signal containing the device number and timestamp is sent to the monitoring terminal; this signal will trigger the manual intervention process. The physical constraint layer is constructed and refined as follows: the electromagnetic attraction process of the relay is decomposed into two stages: static suction characteristics and dynamic motion characteristics. In the static stage, the current-flux conversion relationship is constrained by the magnetic flux conservation equation to limit the integral area boundary of the generated current waveform; in the dynamic stage, the mass-spring damping equation is discretized into displacement-acceleration differential constraints to limit the frequency band energy distribution of the generated vibration signal; the two-stage constraint conditions are embedded in the generator output layer through an alternating optimization mechanism - the first round of training prioritizes satisfying the static constraints to generate the current waveform, and the second round of training calculates the dynamic constraints based on the current waveform to generate the vibration signal, and iterative optimization is carried out until the output signal satisfies the electromagnetic-mechanical coupling equation at the same time.

[0028] like Figure 2As shown in the figure, the generation module constructs the constraints of the generative adversarial network based on the physical model of the relay, and forms an enhanced fault waveform signal that conforms to the physical laws through adversarial training; the coordination module receives the real-time current and voltage signals and the mechanical vibration signals, uses the temporal convolutional network to extract the electrical signal feature vector, and generates the mechanical feature vector through the mechanical feature extraction network, and uses the adaptive attention mechanism to dynamically fuse the two types of feature vectors to form a joint feature tensor; the comparison module inputs the joint feature tensor into the lightweight evaluation model and outputs the health score signal. When the score signal is lower than the preset threshold for the first time, a meta-learning activation instruction is generated; the rehearsal module responds to the meta-learning activation instruction, loads the equipment historical data to construct the parameter optimization set, and performs online fine-tuning on the early warning model based on the meta-learning framework to generate customized fault judgment parameters for the equipment; the analysis module analyzes the real-time signal characteristics according to the fine-tuned judgment parameters, and outputs a graded early warning signal to the monitoring terminal when abnormal contact characteristics or coil degradation trends are identified.

[0029] like Figure 2As shown, the relay as a key executive element in the power control system, its contact sticking, coil circuit failure may lead to major accidents, the traditional threshold early warning method relies on artificial experience to set fixed threshold, it is difficult to capture early implicit fault characteristics, and the existing machine learning method faces three major bottlenecks: the actual fault sample is scarce, which leads to the insufficient generalization ability of the model, the public data shows that the fault record of a certain type of relay in the whole life cycle is less than one thousand; Artificial feature engineering cannot effectively extract the deep coupling relationship between electrical signals and mechanical vibration, for example, the nonlinear relationship between the vibration spectrum change caused by contact bounce and arc current waveform; Fixed model is difficult to adapt to the degradation characteristics of different equipment, and the deployment of new equipment needs to collect training data again, which takes more than 72 hours in the industrial field on average. Although in recent years some researches have used generative adversarial networks to expand samples, but the generated data often violates the physical law, for example, the contact welding current waveform generated by a certain literature does not satisfy the energy conservation equation, which leads to the high model misjudgment rate of 42%. In addition, the sensor sampling time delay problem in multi-modal fusion has been ignored for a long time, experiments show that 0.1 millisecond time delay will make the vibration-current feature correlation degree decrease by 37%. In view of the above problems, the invention proposes a physical constraint fault data enhancement and incremental meta-learning early warning mechanism, generates high-fidelity fault samples by constructing a mechanical and electrical coupling constraint layer, eliminates the time delay of cross-modal signals by combining the space-time attention fusion technology, and designs a double gradient meta-learning framework to realize the rapid adaptation of new equipment within 7 minutes, finally forms a closed-loop early warning system.The implementation is achieved through a five-stage signal processing chain: firstly, in the physical enhanced signal generation step, a physical constraint layer of the generative adversarial network is constructed based on the electromagnetic attraction characteristics of the relay and the mechanical motion equation, the equivalent parameters of the magnetic circuit and the mass-spring damping equation are converted into the output constraint conditions of the generator, the statistical characteristics of the real fault data distribution and the generated signal are compared by the discriminator, and the enhanced fault waveform signal conforming to the physical law is output through the adversarial training; the signal is input as a supplementary input source into the multi-source feature fusion step, the real-time current and voltage signals are input into the time convolution network with an inflation convolution structure to extract long-term dependence features, the mechanical vibration signals pass through the residual network to capture the frequency domain impact components, the adaptive attention mechanism calculates the cross-modal correlation matrix of the electrical signal feature vector and the mechanical feature vector, and the joint feature tensor eliminating the time delay error is generated through the time axis nonlinear interpolation alignment according to the learnable weight; in the health evaluation signal triggering step, the joint feature tensor is processed by the multi-head self-attention module of the lightweight evaluation model, the local health score is calculated independently in each time window, and the 0-100 range score signal is generated by the sliding weighted average, and the meta-learning activation instruction is triggered when the score fluctuation of the continuous three sampling periods exceeds the tolerance and the mean value is lower than the threshold; the incremental model adaptation step responds to the instruction, filters the similar degradation mode segments in the device historical data to construct the parameter optimization set, and fine-tunes the dual gradient update loop - the internal loop freezes the feature extraction layer parameters and only updates the decision layer, generates the primary adaptation parameters based on the historical segment simulation degradation, and the external loop verifies the effectiveness of the parameters and optimizes the network weight to output the final fault judgment parameters; the hierarchical early warning decision step analyzes the real-time signal by using the fine-tuned parameters, detects the contact fault by detecting the current zero-crossing delay and arc duration anomaly through the waveform morphology recognition technology, analyzes the coil degradation trend combined with the inductance feature vector Mahalanobis distance change rate, outputs the first-level warning when a single feature exceeds the limit, outputs the second-level warning when the contact and coil features exceed the limit at the same time, and adds the third-level warning signal to the monitoring terminal if the temperature sensor detects abnormal temperature rise.

[0030] The relay state prediction and fault warning method and system based on deep learning can create high-fidelity fault samples through the physical constraint generative adversarial network, extract deep correlations of relay electrical-mechanical signals combined with multi-modal spatio-temporal feature fusion technology, dynamically output health score using a lightweight evaluation model, trigger the meta-learning framework to quickly adapt to device characteristics when the score is abnormal, realize hierarchical early warning of contact anomalies and coil degradation based on incremental learning, and break through the dependence of traditional methods on fault samples and fixed thresholds.

[0031] Therefore, the relay state prediction and fault warning method and system based on deep learning can solve the problems of early state prediction and accurate warning of relays under small sample fault data.

[0032] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. A relay state prediction and fault warning method based on deep learning, characterized in that: include: S1: Constructing constraints for a generative adversarial network based on the relay physical model, and generating enhanced fault waveform signals that conform to physical laws through adversarial training; S2: Receives real-time current and voltage signals and mechanical vibration signals, extracts electrical signal feature vectors using a temporal convolutional network, and simultaneously generates mechanical feature vectors through a mechanical feature extraction network. Adaptive attention mechanism is used to dynamically fuse the two feature vectors to form a joint feature tensor. S3: Inputting the joint feature tensor into a lightweight evaluation model, outputting a health score signal, and generating a meta-learning activation instruction when the score signal falls below a preset threshold for the first time; S4: Responding to the meta-learning activation instruction, loading the device's historical data to construct a parameter optimization set, and fine-tuning the early warning model online based on the meta-learning framework to generate customized fault judgment parameters for the device; S5: Analyze the real-time signal characteristics according to the fine-tuned judgment parameters. When abnormal contact characteristics or coil degradation trends are identified, output graded warning signals to the monitoring terminal.

2. The relay state prediction and fault early warning method based on deep learning according to claim 1 is characterized in that, In S1: a physical constraint layer is constructed based on the electromagnetic characteristic parameters and mechanical motion characteristic parameters of the relay, and the constraint layer is embedded in the generator output end of the generative adversarial network. The real fault waveform feature distribution and the generated signal are adversarially trained through the discriminator, so that the generated enhanced fault waveform signal satisfies both the physical law constraints and the statistical distribution authenticity requirements. This signal will serve as one of the input signal sources of the multi-source feature fusion step.

3. The relay state prediction and fault early warning method based on deep learning according to claim 1 is characterized in that, In S2: the temporal convolutional network uses an expanded convolution structure to extract the deep temporal dependency features of the current and voltage signals, the mechanical feature extraction network captures the frequency domain mutation pattern of the vibration signal through the residual connection module, and the adaptive attention mechanism generates a dynamic weight coefficient by calculating the spatiotemporal correlation between the electrical signal feature vector and the mechanical feature vector. The final joint feature tensor contains the weighted fused multi-dimensional state representation vector.

4. The relay state prediction and fault early warning method based on deep learning according to claim 1 is characterized in that, In S3: The lightweight evaluation model uses a multi-head self-attention mechanism to construct a feature interaction module, performs time-series slicing on the joint feature tensor and calculates the health score of each slice, and then generates the final health score signal through sliding weighted average. When the fluctuation amplitude of the score signal for three consecutive sampling periods exceeds the preset tolerance range and the mean is lower than the threshold, the meta-learning activation instruction is triggered.

5. The relay state prediction and fault early warning method based on deep learning according to claim 1 is characterized in that, In S4, when constructing a parameter optimization set based on historical data of the device, historical operation segments that are similar to the decline pattern of the current health score signal are preferentially screened. The meta-learning framework constructs a dual-gradient update loop, simulates the fault evolution process based on historical segments in the inner loop to generate fast adaptation parameters, and uses real-time signals in the outer loop to verify the validity of the parameters and output customized fault judgment parameters for the device.

6. The relay state prediction and fault early warning method based on deep learning according to claim 1 is characterized in that: In S5: contact abnormality feature detection uses waveform morphology analysis technology to identify current zero-crossing delay and arc duration anomalies, and coil degradation trend analysis is achieved by comparing the Mahalanobis distance change rate between the real-time inductance feature vector and the standard parameter vector. When any of the two detection results exceeds the safety limit, a first-level warning signal is generated. When both exceed the limit at the same time, a second-level warning signal is generated, and a third-level warning signal is added when the contact temperature rises abnormally.

7. The relay state prediction and fault early warning method based on deep learning according to claim 2 is characterized in that: The physical constraint layer construction process includes: decomposing the relay electromagnetic attraction process into a static attraction characteristic stage and a dynamic motion characteristic stage, introducing magnetic circuit equivalent parameters to constrain the current-flux conversion relationship in the static stage, and introducing a mass-spring damping equation to constrain the displacement-acceleration conversion relationship in the dynamic stage. Through the alternating optimization of the two-stage constraint conditions, a fault waveform signal that conforms to the electromechanical coupling law is generated.

8. The relay state prediction and fault early warning method based on deep learning according to claim 3 is characterized in that: The adaptive attention mechanism is implemented by establishing a cross-modal association matrix between the electrical signal feature vector and the mechanical feature vector, calculating the feature contribution of each time step through learnable parameters, and then performing time axis alignment compensation on the mechanical signal feature vector based on the contribution distribution, ultimately generating a joint feature tensor that eliminates the sensor sampling delay error.

9. The relay state prediction and fault early warning method based on deep learning according to claim 3 is characterized in that: The health score signal verification mechanism includes: when the meta-learning activation instruction is triggered, the health data two hours before the current score signal is generated is retained as a verification set. After the incremental model adaptation step is completed, the verification set is input into the fine-tuning model. If the deviation rate between the newly generated score signal and the original signal exceeds the allowable threshold, the model rollback mechanism is activated and a calibration alarm signal is issued.

10. A prediction and fault warning system using the relay state prediction and fault warning method based on deep learning according to any one of claims 1 to 9, characterized in that: include: A generation module, which constructs constraints for a generative adversarial network based on a physical model of a relay and forms an enhanced fault waveform signal that conforms to physical laws through adversarial training; A coordination module receives real-time current and voltage signals and mechanical vibration signals, extracts electrical signal feature vectors using a temporal convolutional network, generates mechanical feature vectors using a mechanical feature extraction network, and dynamically fuses the two feature vectors using an adaptive attention mechanism to form a joint feature tensor; a comparison module, wherein the comparison module inputs the joint feature tensor into a lightweight evaluation model, outputs a health score signal, and generates a meta-learning activation instruction when the score signal falls below a preset threshold for the first time; a rehearsal module, which responds to the meta-learning activation instruction, loads historical data of the device to construct a parameter optimization set, performs online fine-tuning on the early warning model based on the meta-learning framework, and generates customized fault judgment parameters for the device; The analysis module analyzes the real-time signal characteristics according to the fine-tuned judgment parameters. When abnormal contact characteristics or coil degradation trends are identified, it outputs a graded warning signal to the monitoring terminal.

Images