CN-122020090-A - Circuit breaker residual life prediction method and system based on deep learning

CN122020090ACN 122020090 ACN122020090 ACN 122020090ACN-122020090-A

Abstract

The invention discloses a method and a system for predicting the residual life of a circuit breaker based on deep learning, and belongs to the technical field of intelligent operation and maintenance of circuit breakers. The invention adopts the breaker health index construction method combining multiple constraint condition self-supervision learning, realizes that the health index can stably, continuously and monotonically decrease to reflect the degradation of the breaker, maintains consistency under similar working conditions, provides reliable degradation reference for the residual life prediction, adopts the breaker residual life prediction method combining physical degradation mechanism constraint, realizes that the health index is tightly combined with each degradation mechanism, and ensures that the residual life prediction is more objective, reasonable and accords with engineering practice by carrying out physical constraint guidance on the long-term degradation trend of the breaker.

Inventors

WANG TONGKUI

Assignees

玉环玮斯电气股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (9)

1. The method for predicting the residual life of the circuit breaker based on deep learning is characterized by comprising the following steps of: Step S1, acquiring a state to obtain a breaker state sample sequence; S2, depth characteristic representation, namely obtaining a depth characteristic vector of the circuit breaker; Step S3, constructing a health index, namely, adopting a breaker health index construction method combining multiple constraint condition self-supervision learning according to a breaker depth feature vector to obtain a breaker health index sequence, wherein the method comprises the following steps of step S31 feature mapping modeling, step S32 degradation monotonic constraint, step S33 time smoothness constraint, step S34 state consistency constraint and step S35 health index combined self-supervision optimization; Step S4, predicting the residual life of the circuit breaker, namely adopting a method for predicting the residual life of the circuit breaker by combining physical degradation mechanism constraint according to a circuit breaker state sample sequence, a circuit breaker depth feature vector and a circuit breaker health index sequence, and obtaining the predicted information of the residual life of the circuit breaker, wherein the method comprises the following steps of step S41 degradation physical mechanism modeling, step S42 physical information embedding modeling, step S43 time sequence prediction modeling and step S44 predicted information generation; Step S41, modeling a degradation physical mechanism, namely constructing a degradation physical mechanism model aiming at three typical degradation forms of electrical wear, mechanical fatigue and thermal aging of a circuit breaker contact according to a circuit breaker state sample sequence, calculating to obtain degradation increment corresponding to each degradation mechanism based on the degradation physical mechanism model, and carrying out weighted fusion to obtain a physical degradation quantity used for representing the physical degradation degree of the circuit breaker; the degradation physical mechanism model comprises a contact electric wear model, a mechanical fatigue accumulation damage model and a thermal aging degradation model.
2. The method for predicting the residual life of the circuit breaker based on deep learning of claim 1 is characterized in that in step S31, the feature mapping modeling is specifically implemented by constructing a health index mapping model composed of a feature coding network and a health index regression head, performing feature mapping on depth feature vectors of the circuit breaker to generate health indexes, wherein the feature coding network is used for performing nonlinear feature extraction and semantic compression on the depth feature vectors of the circuit breaker, and the health index regression head is used for mapping coding features output by the feature coding network into the health indexes.
3. The method for predicting the residual life of the circuit breaker based on deep learning of claim 2 is characterized in that in the step S32, the degradation monotonic constraint is specifically that a monotonic non-increasing constraint is applied to a health index sequence of the same circuit breaker in a time dimension by constructing a monotonic loss function, when a health index corresponding to a subsequent time window is higher than a health index corresponding to a previous time window, a penalty is applied to a training process of a health index mapping model, and unreasonable rise of the health index is restrained, so that the evolution trend of the health index accords with the actual degradation rule of the circuit breaker; in step S33, the temporal smoothness constraint is used to limit the severe fluctuation of the health index in a short time scale, specifically, by constructing a smoothness loss function, the smoothness constraint is applied to the variation amplitude of the health index in the adjacent time window; in step S34, the state consistency constraint is used to ensure that the health indexes have a consistent response relationship to the running state change, and by constructing a state consistency loss function, the health index consistency constraint is applied to samples with similar running states, so that the samples with higher running state similarity are more strongly constrained to the corresponding health index differences.
4. The method for predicting the residual life of a circuit breaker based on deep learning of claim 3, wherein in step S35, the health index is combined with self-supervision optimization for comprehensively guiding the training process of the health index mapping model, specifically, by performing weighted fusion on the monotonicity loss function, the smoothness loss function and the state consistency loss function, constructing a self-supervision combined objective function, training the health index mapping model by minimizing the self-supervision combined objective function, and enabling the trained health index mapping model to output the circuit breaker health index sequence.
5. The method for predicting the residual life of the circuit breaker based on deep learning of claim 4 is characterized by comprising the following steps of S42, modeling physical information in an embedded mode, namely, obtaining degradation state characteristics by inputting a circuit breaker health index sequence and a depth characteristic vector into a variation encoder together, and simultaneously introducing regularization constraint terms based on a degradation physical mechanism model to construct a physical regularization constraint loss function so as to enable physical degradation dimensions to be consistent with the physical degradation quantity; Step S43, time sequence prediction modeling is carried out, namely, a multi-scale time sequence prediction block is constructed according to degradation state characteristics, short-term operation behaviors, periodic load changes and long-term degradation trends are subjected to joint modeling, and an event-triggered attention gating mechanism is introduced in the modeling process, so that a future health index track of the circuit breaker is obtained; the event-triggered attention gating mechanism is specifically used for adaptively improving attention weight when a contact state mutation, an abnormal vibration mode or a heavy current break event is detected; And S44, generating prediction information, namely performing statistical analysis on future health index tracks of the circuit breaker, taking a time step corresponding to the first time when the health index drops to a failure threshold value as a failure time step, calculating residual life probability distribution by counting failure samples of different tracks, and generating the residual life prediction information of the circuit breaker.
6. The method for predicting the residual life of a circuit breaker based on deep learning of claim 5, wherein in step S2, the depth feature representation is specifically that a circuit breaker state sample sequence is input into a depth neural network model, local time domain feature extraction is performed by adopting a convolutional neural subnet for waveform signal data, time sequence feature extraction is performed by adopting a cyclic neural subnet for non-waveform signal data, and output features of all the subnets are fused through a feature fusion layer, so that a circuit breaker depth feature vector is obtained.
7. The method for predicting the residual life of a circuit breaker based on deep learning of claim 6, wherein in step S1, the state acquisition is specifically that original state data is obtained by acquiring information of an electrical state, a mechanical state and an environmental state of the circuit breaker in the operation process, and then the original state data is subjected to time synchronization, abnormal data elimination and normalization processing, and is segmented according to the operation period of the circuit breaker to obtain a circuit breaker state sample sequence.
8. The breaker remaining life prediction system based on deep learning is used for realizing the breaker remaining life prediction method based on deep learning as claimed in any one of claims 1 to 7, and is characterized by comprising a breaker multi-source sensing module, a feature extraction module, a breaker health evaluation module and a prediction output module.
9. The deep learning-based breaker residual life prediction system of claim 8, wherein the breaker multisource perception module is used for acquiring states, acquiring a breaker state sample sequence through the state acquisition, and sending the breaker state sample sequence to the feature extraction module and the prediction output module; the feature extraction module is used for depth feature characterization, obtaining a depth feature vector of the circuit breaker through the depth feature characterization, and sending the depth feature vector of the circuit breaker to the circuit breaker health evaluation module and the prediction output module; The circuit breaker health evaluation module is used for constructing a health index, obtaining a circuit breaker health index sequence through the health index construction, and sending the circuit breaker health index sequence to the prediction output module; and the prediction output module is used for predicting the residual life and obtaining the predicted information of the residual life of the circuit breaker through the prediction of the residual life.

Description

Circuit breaker residual life prediction method and system based on deep learning Technical Field The invention belongs to the technical field of intelligent operation and maintenance of circuit breakers, and particularly relates to a method and a system for predicting the residual life of a circuit breaker based on deep learning. Background The method comprises the steps of predicting the residual life of the circuit breaker based on deep learning, automatically characterizing and timing modeling high-dimensional operation characteristics through a deep neural network by utilizing multi-source state data acquired by the circuit breaker in the operation process, constructing and restraining a physical degradation mechanism by combining a health index, and realizing accurate prediction of the residual life of the circuit breaker, thereby providing a reliable basis for optimizing the overhaul period of the circuit breaker, preventive maintenance arrangement and formulating an operation strategy of the circuit breaker, and further improving the operation safety and economy of a power system. In the existing breaker residual life prediction process, health state assessment is easily affected by noise and abnormal operation, so that the fluctuation of health index is severe in a short time, the continuity of the breaker degradation is difficult to truly reflect, meanwhile, the traditional method does not fully reflect the irreversibility of the degradation process, and the technical problem that inconsistent assessment results can be generated in a similar operation state is solved. Disclosure of Invention Aiming at the situation, in order to overcome the defects of the prior art, the invention provides a method and a system for predicting the residual life of the circuit breaker based on deep learning, creatively adopts a method for constructing the health index of the circuit breaker by combining multiple constraint conditions and self-supervision learning, realizes that the health index can stably, continuously and monotonically reflect the degradation of the circuit breaker, keeps consistency under similar working conditions, provides reliable degradation reference for predicting the residual life, creatively adopts a method for predicting the residual life of the circuit breaker by combining physical degradation mechanism constraint, realizes that the health index is tightly combined with each degradation mechanism, and leads the residual life prediction to be more objective, reasonable and accord with engineering practice by carrying out physical constraint guidance on the long-term degradation trend of the circuit breaker. The technical scheme adopted by the invention is that the method for predicting the residual life of the circuit breaker based on deep learning comprises the following steps: s1, collecting states; S2, depth characteristic characterization; S3, constructing a health index; and S4, predicting the residual life. Further, in step S1, the state acquisition is specifically that the original state data is obtained by acquiring the information of the electrical state, the mechanical state and the environmental state of the circuit breaker in the operation process, and then the original state data is subjected to time synchronization, abnormal data rejection and normalization processing, and is segmented according to the operation cycle of the circuit breaker, so as to obtain a circuit breaker state sample sequence. Further, in step S2, the depth feature characterization specifically includes inputting a breaker state sample sequence into a depth neural network model, performing local time domain feature extraction by adopting a convolutional neural sub-network for waveform signal data, performing time sequence feature extraction by adopting a cyclic neural sub-network for non-waveform signal data, and fusing output features of each sub-network through a feature fusion layer to obtain a breaker depth feature vector. Further, in step S3, the health index construction, specifically, according to the depth feature vector of the circuit breaker, adopts a circuit breaker health index construction method combining self-supervised learning of multiple constraint conditions to obtain a circuit breaker health index sequence, including the following steps: Step S31, feature mapping modeling, namely, feature mapping is carried out on depth feature vectors of the circuit breaker by constructing a health index mapping model consisting of a feature coding network and a health index regression head to generate a health index; the feature coding network is used for carrying out nonlinear feature extraction and semantic compression on depth feature vectors of the circuit breaker The health index regression head is used for mapping the coding features output by the feature coding network into health indexes; step S32, degradation monotone constraint is specifically that constraint that the health index sequence