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

Abstract

The method comprises the following steps: S1, constraint conditions of a generative adversarial network are constructed based on a relay physical model, and an enhanced fault waveform signal conforming to physical laws is formed through adversarial training; S2, real-time current and voltage signals and mechanical vibration signals are received, and an electric signal feature vector is extracted using a time sequence convolution network; S3, a joint feature tensor is input into a lightweight evaluation model, and a health degree score signal is output; S4, in response to a meta-learning activation instruction, device historical data are loaded to construct a parameter optimization set, an early warning model is fine-tuned online based on a meta-learning framework, and fault determination parameters are generated; and S5, real-time signal features are analyzed according to the fine-tuned determination parameters, and a graded early warning signal is output to a monitoring terminal. The present application can solve the problem of early state prediction and accurate early warning of relays under small sample fault data.

Description

TECHNICAL FIELD

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

[0002] As a key power control component, the contact adhesion and coil aging of the relay can easily cause electrical accidents. The current industry mainly uses threshold method or machine learning for state monitoring, but there are three major bottlenecks: first, the lack of actual fault samples leads to insufficient generalization ability of the supervised learning model, and public research shows that the fault data of a certain type of relay in the whole life cycle is less than one thousand; second, artificial feature extraction is difficult to capture the implicit correlation between electrical signals and mechanical vibrations, such as the cross-modal coupling characteristics of contact bounce vibration and arc current; third, the fixed threshold early warning mechanism cannot adapt to the degradation characteristics of different devices, and false positives or false negatives often occur.

[0003] In recent years, although some research has used generative adversarial networks to expand samples, the generated data violates the physical law, leading to model failure, such as the contact welding current waveform generated by a certain document does not satisfy the energy conservation equation. At the same time, the simple fusion of time series models and computer vision models is easily disturbed by asynchronous sampling of sensors, and a certain experiment shows that a 0.1 millisecond time delay can reduce the vibration-current feature correlation by 37%. In addition, the traditional early warning model needs to be retrained when deploying new devices, and the average time consumption in industrial field is more than 72 hours. Therefore, it is urgent to build an intelligent early warning system that integrates physical laws, adapts to device characteristics, and has incremental learning ability. SUMMARY

[0004] In view of the shortcomings of the above prior art, the purpose of the present application is to provide a relay state prediction and fault early warning method and system based on deep learning, to solve the problems of early state prediction and accurate early warning of relays under small sample fault data. The present application creates high-fidelity fault samples through physically constrained generative adversarial networks, extracts deep correlations of relay electrical-mechanical signals by combining multi-modal spatio-temporal feature fusion technology, and uses a lightweight evaluation model to dynamically output a health score. When the score is abnormal, trigger the meta-learning framework to quickly adapt to the device characteristics, realize the grading early warning of contact abnormalities and coil degradation based on incremental learning, and break through the dependence of traditional methods on fault samples and fixed thresholds.

[0005] The present application provides a relay state prediction and fault early warning method based on deep learning, comprising:

[0006] S1: constructing constraint conditions of the generative adversarial network based on the physical model of the relay, and forming enhanced fault waveform signals conforming to the physical law through adversarial training;

[0007] S2: Receives real-time current and voltage signals and mechanical vibration signals, uses a temporal convolutional network to extract electrical signal feature vectors, and generates mechanical feature vectors through a mechanical feature extraction network. An adaptive attention mechanism is used to dynamically fuse the two types of feature vectors to form a joint feature tensor.

[0008] S3: Input the joint feature tensor into the lightweight evaluation model and output a health score signal. When the score signal first falls below a preset threshold, generate a meta-learning activation instruction.

[0009] S4: Respond to the meta-learning activation command, load historical data of the device to build a parameter optimization set, fine-tune the early warning model online based on the meta-learning framework, and generate customized fault judgment parameters for the device;

[0010] S5: Analyze the real-time signal characteristics based on the fine-tuned judgment parameters. When abnormal contact characteristics or coil degradation trend are detected, output a graded early warning signal to the monitoring terminal.

[0011] In one embodiment of the present invention, in S1: a physical constraint layer is constructed based on the electromagnetic characteristic parameters and mechanical motion characteristic parameters of the relay. The constraint layer is embedded into the generator output of the generative adversarial network. The discriminator performs adversarial training on the distribution of real fault waveform features and the generated signal, so that the generated enhanced fault waveform signal simultaneously meets the requirements of physical law constraints and statistical distribution authenticity. This signal will be used as one of the input signal sources for the multi-source feature fusion step.

[0012] In one embodiment of the present invention, in S2: the temporal convolutional network uses a dilated convolutional structure to extract deep temporal dependency features of current and voltage signals, the mechanical feature extraction network captures the frequency domain abrupt change mode of vibration signals through a residual connection module, and the adaptive attention mechanism generates dynamic weight coefficients by calculating the spatiotemporal correlation between the electrical signal feature vector and the mechanical feature vector. The final joint feature tensor contains a weighted fusion multidimensional state representation vector.

[0013] In one embodiment of the present invention, in S3: the lightweight evaluation model uses a multi-head self-attention mechanism to construct a feature interaction module, performs temporal slicing on the joint feature tensor and calculates the health score of each slice, and then generates the final health score signal by sliding weighted average. When the fluctuation amplitude of the score signal in three consecutive sampling periods exceeds the preset tolerance range and the mean is lower than the threshold, the meta-learning activation instruction is triggered.

[0014] In one embodiment of the present invention, in S4: when constructing a parameter optimization set from historical equipment data, historical operation segments similar to the current health score signal decline pattern are selected first. The meta-learning framework constructs a dual-gradient update loop. In the inner loop, it simulates the fault evolution process based on historical segments to generate fast adaptation parameters. In the outer loop, it uses real-time signals to verify the validity of the parameters and outputs customized fault judgment parameters for the equipment.

[0015] In one embodiment of the present invention, in S5: the abnormal contact feature detection uses waveform morphology analysis technology to identify abnormal current zero-crossing delay and arc duration; the 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 either detection result exceeds the safety boundary, a first-level warning signal is generated; when both exceed the limit simultaneously, a second-level warning signal is generated; and a third-level warning signal is added when the contact temperature rises abnormally.

[0016] In one embodiment of the present invention, the physical constraint layer construction process includes: decomposing the relay electromagnetic engagement process into a static attraction characteristic stage and a dynamic motion characteristic stage; introducing equivalent parameters of the magnetic circuit to constrain the current-magnetic flux conversion relationship in the static stage; introducing mass-spring damping equations 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.

[0017] In one embodiment of the present invention, the adaptive attention mechanism is implemented as follows: a cross-modal correlation matrix between the electrical signal feature vector and the mechanical feature vector is established; the feature contribution of each time step is calculated through learnable parameters; and the mechanical signal feature vector is compensated for time axis alignment according to the contribution distribution. Finally, a joint feature tensor that eliminates sensor sampling delay error is generated.

[0018] In one embodiment of the present invention, the health score signal verification mechanism includes: when the meta-learning activation instruction is triggered, retaining the health data of the two hours before the current score signal generation time as a verification set; after the incremental model adaptation step is completed, inputting the verification set into the fine-tuning model; if the deviation rate between the newly generated score signal and the original signal exceeds the allowable threshold, then the model rollback mechanism is activated and a calibration alarm signal is issued.

[0019] This invention also includes a deep learning-based relay state prediction and fault early warning system, comprising:

[0020] The generation module constructs constraints for the generative adversarial network based on the relay physical model, and generates enhanced fault waveform signals that conform to physical laws through adversarial training.

[0021] The coordination module receives real-time current and voltage signals and mechanical vibration signals. It uses a temporal convolutional network to extract electrical signal feature vectors and a mechanical feature extraction network to generate mechanical feature vectors. An adaptive attention mechanism is used to dynamically fuse the two types of feature vectors to form a joint feature tensor.

[0022] The comparison module inputs the joint feature tensor into the lightweight evaluation model and outputs a health score signal. When the score signal first falls below a preset threshold, a meta-learning activation instruction is generated.

[0023] The pre-training module responds to the meta-learning activation command, loads historical equipment data to construct a parameter optimization set, fine-tunes the early warning model online based on the meta-learning framework, and generates customized fault judgment parameters for the equipment.

[0024] The analysis module analyzes real-time signal characteristics based on the fine-tuned judgment parameters. When abnormal contact characteristics or coil degradation trends are detected, it outputs graded early warning signals to the monitoring terminal.

[0025] The present invention provides a method and system for relay state prediction and fault early warning based on deep learning. It creates high-fidelity fault samples through physical constraint generative adversarial networks, extracts deep correlations between relay electromechanical signals by combining multimodal spatiotemporal feature fusion technology, and dynamically outputs health scores using a lightweight evaluation model. When the score is abnormal, a meta-learning framework is triggered to quickly adapt to the characteristics of the equipment, realizing graded early warning of contact anomalies and coil degradation based on incremental learning, breaking through the dependence of traditional methods on fault samples and fixed thresholds. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a deep learning-based relay state prediction and fault early warning method.

[0028] Figure 2 This is a system architecture diagram of a deep learning-based relay state prediction and fault early warning system. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0030] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0031] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0032] Please see Figures 1-2 The figure shows the relay state prediction and fault early warning method and system based on deep learning of the present invention. The relay state prediction and fault early warning method based on deep learning of the present invention includes: S1: Constructing the constraint conditions of the generative adversarial network based on the relay physical model, and forming an enhanced fault waveform signal that conforms to the physical law through adversarial training; S2: Receiving real-time current and voltage signals and mechanical vibration signals, using a temporal convolutional network to extract electrical signal feature vectors, and simultaneously generating mechanical feature vectors through a mechanical feature extraction network, and using an adaptive attention mechanism to dynamically fuse the two types of 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 first falls below a preset threshold; S4: Responding to the meta-learning activation instruction, loading historical data of the equipment 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 equipment; S5: Analyzing real-time signal characteristics based on the fine-tuned judgment parameters, and when abnormal contact characteristics or coil degradation trends are identified, outputting a graded early warning signal to the monitoring terminal.

[0033] like Figure 1As shown, the core method of this invention comprises a five-step signal processing chain: First, in the physical enhancement signal generation step, physical constraints for a generative adversarial network are established based on the electromagnetic and mechanical motion characteristics of a relay. Through adversarial training between the generator and discriminator, a fault waveform signal conforming to physical laws is formed, which serves as one of the inputs for multi-source feature fusion. Next, in the multi-source feature fusion step, real-time current and voltage signals are input into a temporal convolutional network to extract their temporal dependent features, while the mechanical vibration signal is processed by a residual network to capture frequency domain abrupt changes. The two types of feature vectors are dynamically weighted and fused through an adaptive attention mechanism to form a joint feature tensor. Finally, in the health assessment signal triggering step, the joint feature tensor is input into a lightweight assessment model, and multi-head self-attention is used to... The force mechanism analyzes the interaction relationship of features and outputs a health score signal. When the score fluctuates continuously beyond the tolerance and the mean is lower than the threshold, a meta-learning activation instruction is generated. Then, the incremental model adaptation step responds to the instruction, selects similar degradation patterns in the equipment's historical data to construct a parameter optimization set, and uses a dual-gradient update loop to fine-tune the model. In the first stage, the feature layer is frozen and the decision layer is updated to generate primary parameters. In the second stage, real-time signals are used to verify and generate the final fault judgment parameters. Finally, the graded early warning decision step analyzes real-time signals based on the fine-tuned parameters, identifies abnormal contact features through waveform morphology, analyzes coil degradation trends by combining inductance vector distance, and outputs graded early warning signals to the monitoring terminal according to the abnormal combination type, realizing a closed-loop processing from physical modeling to graded decision-making.

[0034] Furthermore, the constraint layer construction of the physical enhancement signal generation step specifically includes: decomposing the relay electromagnetic attraction process into two stages: static attraction characteristics and dynamic motion characteristics. In the static stage, equivalent parameters of the magnetic circuit are introduced to constrain the current-magnetic flux conversion relationship, and in the dynamic stage, mass-spring damping equations are introduced to constrain the displacement-acceleration conversion relationship. These two-stage constraints are embedded into the generator output layer of the generative adversarial network. The discriminator compares the statistical distribution differences between the real fault waveform and the generated signal. During adversarial training, the generator parameters and discriminator parameters are alternately optimized so that the generated signal simultaneously satisfies the boundary conditions of the physical equations and the fault data distribution characteristics. Finally, an enhanced fault waveform signal with physical consistency is output. This signal will serve as a supplementary input source for the multi-source feature fusion step and will be input into the temporal convolutional network along with the measured current and voltage signals for feature extraction. The multi-source feature fusion is specifically implemented as follows: the temporal convolutional network adopts a three-layer dilated convolutional structure, with the dilation coefficient growing exponentially to expand the temporal receptive field and capture the periodic arc features 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 the time step correlation matrix between 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, performs cubic spline interpolation on the mechanical feature vector according to the weight distribution to achieve time axis compensation, and finally fuses the features according to the attention weights to generate a spatiotemporally synchronized joint feature tensor.

[0035] like Figure 1 As shown, in the multi-source feature fusion step, the temporal convolutional network uses a dilated convolutional structure to expand the receptive field and capture the long-term temporal dependence patterns of current and voltage signals; the mechanical feature extraction network constructs a deep frequency domain feature extraction path through residual connections to identify the shock harmonic components in the vibration signal; the adaptive attention mechanism first establishes a cross-modal correlation matrix between the electrical signal and mechanical signal features, calculates the feature contribution of each time step through learnable parameters, and performs time axis translation compensation on the mechanical signal feature vector according to the contribution distribution to eliminate sampling delay. Finally, the compensated feature vector is dynamically weighted and fused to generate a joint feature tensor, which retains the coupling relationship between the temporal evolution law of the electrical signal and the frequency domain abrupt change characteristics of the mechanical signal. The lightweight assessment model for the health assessment signal triggering step employs a multi-head self-attention mechanism to construct a feature interaction module. First, the joint feature tensor is sliced ​​according to a fixed time window. Each slice is independently input into the feature interaction module to calculate a local health score. Then, a sliding window weighted average algorithm is used to aggregate the scores and generate a global health score signal. When the fluctuation amplitude of the score signal over three consecutive sampling periods exceeds a preset tolerance and the moving average is below a threshold, a meta-learning activation instruction is generated. Simultaneously, a verification mechanism is established: after the instruction is triggered, the health data from the previous two hours is cached as a verification set. After the incremental model adaptation is completed, the verification set is input into the fine-tuning model. If the deviation rate between the new score signal and the original signal exceeds the limit, model rollback is initiated and a calibration alarm signal is output to ensure assessment reliability.

[0036] Furthermore, the parameter optimization set construction for the incremental model adaptation step requires screening historical data of the equipment for operational segments similar to the current health decline pattern. The meta-learning framework achieves rapid adaptation through a dual-gradient update loop: the inner loop freezes the parameters of the feature extraction network and updates only the parameters of the early warning decision layer, generating primary adaptation parameters based on the simulation of the equipment degradation path in historical segments; the outer loop inputs real-time signals into the model containing the primary parameters, constructs a loss function based on the difference between the predicted results and the actual detection results, and jointly optimizes the parameters of the feature extraction network and the decision layer to generate the final fault judgment parameters; these parameters will dynamically update the judgment boundary of the hierarchical early warning decision step. The contact anomaly detection step in the graded early warning decision-making process employs waveform morphology analysis technology. Anomalies are identified by extracting the current zero-crossing delay time and arc duration and comparing them with standard waveforms. Coil degradation trend analysis calculates the rate of change of the Mahalanobis distance between the real-time inductance feature vector and the standard parameter vector. A first-level early warning signal is output when only a contact anomaly or a single coil degradation feature exceeds the limit. A second-level early warning signal is output when both contact anomaly and coil degradation exceed the limit. If the contact temperature sensor detects an abnormal temperature rise, a third-level early warning signal is added. All early warning signals are transmitted to the monitoring terminal through an encrypted channel and trigger audible and visual alarms.

[0037] likeFigure 1 As shown, the physical constraint layer is further refined: in the static attraction characteristic stage, the equivalent parameters of the magnetic circuit are transformed into current-magnetic flux conversion constraints, and the integral area of ​​the generated current waveform is limited by the magnetic flux conservation equation; in the dynamic motion characteristic stage, the mass-spring damping equation is discretized into displacement-acceleration differential constraints, which restrict the frequency band energy distribution of the generated vibration signal; the two-stage constraint conditions 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 the current waveform to generate the vibration signal, iterating until the generated signal simultaneously satisfies the electromagnetic-mechanical coupling equation. The specific process of the adaptive attention mechanism to achieve cross-modal signal alignment is as follows: first, the electrical signal feature vector and the mechanical signal feature vector are expanded according to the time step to construct an association matrix, and the matrix elements represent the similarity of different modal features at the same time; the contribution of each time step feature to the target time is calculated through a learnable weight layer; the mechanical signal feature vector is compensated by nonlinear interpolation of the time axis according to the contribution distribution to eliminate the millisecond-level delay caused by the difference in sensor sampling frequency; the aligned feature vector is used to generate a spatiotemporal fusion joint feature tensor through soft attention weight allocation.

[0038] Specifically, the operational details of the health score verification mechanism are as follows: A timestamp T is recorded the instant the meta-learning activation command is triggered, and all health scores within the period from T-120 minutes to T are extracted as the verification set; immediately after incremental model adaptation is completed, the verification set is input into the fine-tuned model, and the point-by-point absolute deviation rate between the new score sequence and the original sequence is calculated; if the deviation rate of more than 20% of the data points exceeds 15%, the model adaptation is deemed to have failed and a 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 a manual intervention process. The physical constraint layer is constructed in detail as follows: The electromagnetic engagement process of the relay is decomposed into two stages: static attraction characteristics and dynamic motion characteristics. In the static stage, the current-magnetic flux conversion relationship is constrained by the flux linkage conservation equation, which limits 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, which restrict the frequency band energy distribution of the generated vibration signal. The two-stage constraint conditions are embedded into the generator output layer through an alternating optimization mechanism. The first round of training prioritizes satisfying the static constraints to generate the current waveform. The second round of training calculates the dynamic constraints based on the current waveform to generate the vibration signal. The optimization is iterated until the output signal simultaneously satisfies the electromagnetic-mechanical coupling equation.

[0039] like Figure 2As shown, the module consists of a generation module, which constructs constraints for a generative adversarial network based on a relay physical model, and generates enhanced fault waveform signals that conform to physical laws through adversarial training; a coordination module, which receives real-time current and voltage signals and mechanical vibration signals, extracts electrical signal feature vectors using a temporal convolutional network, and generates mechanical feature vectors through a mechanical feature extraction network, and dynamically fuses the two types of feature vectors using an adaptive attention mechanism to form a joint feature tensor; a comparison module, which inputs the joint feature tensor into a lightweight evaluation model and outputs a health score signal, generating a meta-learning activation command when the score signal first falls below a preset threshold; a pre-exercise module, which responds to the meta-learning activation command, loads historical equipment data to construct a parameter optimization set, performs online fine-tuning of the early warning model based on the meta-learning framework, and generates customized fault judgment parameters for the equipment; and an analysis module, which analyzes real-time signal characteristics based on the fine-tuned judgment parameters, and outputs a graded early warning signal to the monitoring terminal when abnormal contact features or coil degradation trends are identified.

[0040] like Figure 2As shown, relays, as key actuators in power control systems, can cause major accidents due to faults such as contact adhesion and coil open circuits. Traditional threshold-based early warning methods rely on manual experience to set fixed thresholds, making it difficult to capture early, latent fault characteristics. Furthermore, existing machine learning methods face three major bottlenecks: the scarcity of actual fault samples leads to insufficient model generalization ability; publicly available data shows that there are fewer than a thousand fault records for a certain type of relay throughout its entire lifecycle; manual feature engineering cannot effectively extract the deep coupling relationship between electrical signals and mechanical vibrations, such as the nonlinear correlation between vibration spectrum changes caused by contact bounce and arc current waveforms; and fixed models are difficult to adapt to the degradation characteristics of different equipment, requiring the re-collection of training data for new equipment deployments, with an average time exceeding 72 hours in industrial settings. Although recent studies have used generative adversarial networks to expand samples, the generated data often violates physical laws; for example, the contact welding current waveform generated in one paper did not satisfy the energy conservation equation, resulting in a model misjudgment rate as high as 42%. In addition, the sensor sampling delay problem in multimodal fusion has long been neglected; experiments show that a 0.1 millisecond delay can reduce the correlation between vibration and current features by 37%. To address the aforementioned issues, this invention proposes a physical constraint-based fault data enhancement and incremental meta-learning early warning mechanism. It generates high-fidelity fault samples by constructing an electromechanical coupling constraint layer, eliminates cross-modal signal delay by combining spatiotemporal attention fusion technology, and designs a dual-gradient meta-learning framework to enable rapid adaptation of new devices within 7 minutes, ultimately forming a closed-loop early warning system.This is achieved through a five-level signal processing chain: First, in the physical enhancement signal generation step, a physical constraint layer of a generative adversarial network is constructed based on the electromagnetic attraction characteristics of relays and the mechanical motion equations. The equivalent parameters of the magnetic circuit and the mass-spring damping equation are transformed into generator output constraints. A discriminator compares the distribution of real fault data with the statistical characteristics of the generated signal, and after adversarial training, an enhanced fault waveform signal conforming to physical laws is output. This signal serves as a supplementary input source in the multi-source feature fusion step. Real-time current and voltage signals are input into a time-series convolutional network with an dilated convolution structure to extract long-term dependent features. The mechanical vibration signal captures the frequency domain impact component through a residual network. An adaptive attention mechanism calculates the cross-modal correlation matrix between the electrical signal feature vector and the mechanical feature vector. Based on learnable weights, nonlinear interpolation alignment of the time axis is performed to generate a joint feature tensor that eliminates time delay errors. In the health assessment signal triggering step, the joint feature tensor is sliced ​​by a multi-head self-attention module of a lightweight assessment model, and each time window is calculated independently. The system calculates a local health score and generates a score signal in the range of 0-100 using a sliding weighted average. When the score fluctuation exceeds the tolerance for three consecutive sampling periods and the mean is lower than the threshold, a meta-learning activation command is triggered. The incremental model adaptation step responds to this command by screening similar degradation pattern segments in the equipment's historical data to construct a parameter optimization set. A dual-gradient update loop is used for fine-tuning—the inner loop freezes the feature extraction layer parameters and only updates the decision layer, generating primary adaptation parameters based on historical segments to simulate degradation. The outer loop uses real-time signals to verify the validity of the parameters and jointly optimizes the network weights to output the final fault judgment parameters. The graded early warning decision step applies the fine-tuned parameters to analyze real-time signals. It uses waveform morphology recognition technology to detect abnormalities in current zero-crossing delay and arc duration to determine contact faults. It combines the inductance feature vector Mahalanobis distance change rate to analyze the coil degradation trend. When a single feature exceeds the limit, a first-level early warning is output. When both contact and coil features exceed the limit, a second-level early warning is output. If the temperature sensor detects an abnormal temperature rise, a third-level early warning signal is added to the monitoring terminal.

[0041] The present invention provides a deep learning-based relay state prediction and fault early warning method and system. It creates high-fidelity fault samples through physical constraint generative adversarial networks, extracts deep correlations between relay electromechanical signals by combining multimodal spatiotemporal feature fusion technology, and dynamically outputs health scores using a lightweight evaluation model. When the score is abnormal, a meta-learning framework is triggered to quickly adapt to the characteristics of the equipment, realizing incremental learning-based graded early warning of contact anomalies and coil degradation, breaking through the dependence of traditional methods on fault samples and fixed thresholds.

[0042] Therefore, the relay state prediction and fault early warning method and system based on deep learning of the present invention can solve the problem of early state prediction and accurate early warning of relays under small sample fault data.

[0043] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can 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 those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for relay state prediction and fault warning based on deep learning, characterized in that, The method comprises the following steps: S1: constructing constraint conditions of the 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 an electrical signal feature vector using a time convolution network, simultaneously generating a mechanical feature vector through a mechanical feature extraction network, and dynamically fusing the two types of feature vectors through an adaptive attention mechanism to form a joint feature tensor; in S2, the time convolution network uses an expansion convolution structure to extract deep time sequence dependent features of the current and voltage signals, the mechanical feature extraction network captures frequency domain mutation patterns of the vibration signals through a residual connection module, the adaptive attention mechanism generates dynamic weight coefficients by calculating the spatiotemporal correlation of the electrical signal feature vector and the mechanical feature vector, and the finally formed joint feature tensor contains a weighted fused multi-dimensional state representation vector; S3: inputting the joint feature tensor into a lightweight evaluation model to output a health degree score signal, and generating a meta-learning activation instruction when the score signal first falls below a preset threshold; in S3, the lightweight evaluation model uses a multi-head self-attention mechanism to construct a feature interaction module, calculates the health degree score of each slice through time slicing processing of the joint feature tensor, and 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 a preset tolerance range and the mean value is lower than the threshold; S4: in response to the meta-learning activation instruction, constructing a parameter optimization set using device historical data, and performing online fine-tuning of the early warning model based on a meta-learning framework to generate device customized fault judgment parameters; in S4, when the device historical data constructs the parameter optimization set, 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; S5: analyzing real-time signal features according to the fine-tuned judgment parameters, and outputting a hierarchical early warning signal to a monitoring terminal when abnormal contact features or coil degradation trends are identified. 2.The deep learning-based relay state prediction and fault warning method according to claim 1, characterized in that, In S1, the physical constraint layer is constructed based on electromagnetic characteristic parameters and mechanical motion characteristic parameters of the relay, the constraint layer is embedded into the output end of the generator of the generative adversarial network, the discriminator is used to perform adversarial training on the real fault waveform feature distribution and the generated signal, so that the generated enhanced fault waveform signal meets the requirements of physical law constraints and statistical distribution authenticity, and the signal will be used as one of the input signal sources in the multi-source feature fusion step. 3.The deep learning-based relay state prediction and fault warning method according to claim 1, characterized in that, In S5, the contact abnormality feature detection uses waveform morphology analysis technology to identify current zero-crossing delay and arc duration abnormality, and the coil degradation trend analysis is realized by comparing the Mahalanobis distance change rate of the real-time inductance feature vector and the standard parameter vector. When either of the two detection results exceeds the safety boundary, a first-level early warning signal is generated, when both exceed the limit, a second-level early warning signal is generated, and when the contact temperature abnormally rises, a third-level early warning signal is added. 4.The deep learning-based relay state prediction and fault warning method of claim 2, wherein, 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 parameter constraints in the current-magnetic flux conversion relationship in the static stage, introducing mass-spring damping equation constraints in the displacement-acceleration conversion relationship in the dynamic stage, and generating fault waveform signals conforming to the electromechanical coupling law through alternating optimization of the two-stage constraint conditions. 5.The deep learning-based relay state prediction and fault warning method according to claim 1, characterized in that, The adaptive attention mechanism implementation is: establishing a cross-modal association matrix of electrical signal feature vectors and mechanical feature vectors, calculating the feature contribution degree of each time step through learnable parameters, and then compensating the mechanical signal feature vector on the time axis according to the contribution degree distribution, and finally generating a joint feature tensor that eliminates the sampling time delay error of the sensor. 6.The deep learning-based relay state prediction and fault warning method according to claim 1, wherein, The health score signal verification mechanism includes: when the meta-learning activation instruction is triggered, the health data of the previous two hours before the current score signal generation time is retained as a verification set, the verification set is input into the fine-tuning model after the incremental model adaptation step is completed, and if the deviation rate of the newly generated score signal and the original signal exceeds the allowed threshold, the model rollback mechanism is started and a calibration alarm signal is issued.

7. A prediction and failure warning system using the deep learning-based relay state prediction and failure warning method according to any one of claims 1-6, characterized by, It includes: A generation module generates constraint conditions for the generative adversarial network based on a relay physical model, and forms an enhanced fault waveform signal conforming 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 time convolution network, generates mechanical feature vectors through a mechanical feature extraction network, and dynamically fuses the two types of feature vectors using an adaptive attention mechanism to form a joint feature tensor; A comparison module inputs the joint feature tensor into a lightweight evaluation model to output a health score signal, and generates a meta-learning activation instruction when the score signal first falls below a preset threshold; A rehearsal module loads device historical data to construct a parameter optimization set in response to the meta-learning activation instruction, and fine-tunes the early warning model online based on a meta-learning framework to generate device-specific fault judgment parameters; An analysis module analyzes real-time signal features based on the fine-tuned judgment parameters, and outputs a hierarchical early warning signal to a monitoring terminal when it identifies contact abnormality features or coil degradation trends.

Images