CN-121980247-A - Multi-branch fault detection method and system for electric energy metering assembly line

CN121980247ACN 121980247 ACN121980247 ACN 121980247ACN-121980247-A

Abstract

The invention discloses a multi-branch fault detection method and a multi-branch fault detection system for an electric energy metering assembly line, which relate to the technical field of fault detection and comprise the following steps of obtaining multi-source heterogeneous operation sensing data of the electric energy metering assembly line; the method comprises the steps of processing perception data through a multi-branch feature extraction network to generate multi-dimensional characterization features, wherein the multi-dimensional characterization features comprise processing CNN-transducer branches of visual images to generate visual features, processing spectrum perception branches of control signal information to generate control features, processing multi-layer perception machine branches of running state information to generate process state features, and obtaining enhanced joint features through self-adaptive gating network fusion. The intelligent fault detection device and the intelligent fault detection method are used for solving the problems of inaccurate intelligent fault detection and weak adaptability of the electric energy metering pipeline equipment.

Inventors

WANG JINZHI
Tang Beini
BAO ZHENFENG
YANG KAIXUAN
WANG YAOYING
SUN RUILI
CHEN ZIHAO
QIAN LIDONG
ZHANG JIANGBO
YANG YANG
CHENG XIA
WAN SHUN
CHEN SHUO
LIU KANGJIAN
ZHAO XIN

Assignees

国网安徽省电力有限公司肥东县供电公司
西安交通大学
国网安徽省电力有限公司合肥供电公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (8)

1. The multi-branch fault detection method for the electric energy metering pipeline is characterized by comprising the following steps of: the method comprises the steps of obtaining multisource heterogeneous operation sensing data of an electric energy metering pipeline, wherein the sensing data comprise visual images, control signal information and operation state information; Processing the perception data through a multi-branch feature extraction network to generate a multi-dimensional characterization feature, wherein the multi-dimensional characterization feature comprises a CNN-transducer branch for processing a visual image to generate a visual feature, a spectrum perception branch for processing control signal information to generate a control feature, a multi-layer perception machine branch for processing running state information to generate a process state feature, and a self-adaptive gating network fusion to obtain an enhanced joint feature; and verifying the reinforced joint characteristics, and selecting a stable loop to identify pipeline equipment faults and update multi-branch characteristics to extract network parameters according to verification results, or selecting an emergency loop to update network parameters and then identifying faults.
2. The electrical energy metering pipeline oriented multi-branch fault detection method of claim 1, wherein the visual feature generation method comprises: performing convolution and pooling operation on the visual image to obtain a visual feature map; dividing the visual feature map into a plurality of regional feature units and converting the regional feature units into a visual feature sequence; Based on the installation region characteristics of the detection device in the pipeline, region prior is performed on the visual feature sequence, and visual features are generated through an encoder.
3. The power metering pipeline oriented multi-branch fault detection method according to claim 1, wherein the control feature generation method comprises: extracting intermediate control features based on process beats and boundary constraints of control signal information; performing short-time Fourier transform on the intermediate control features to construct a spectrum sensing feature sequence; and (3) extracting control dependency relations from the frequency spectrum sensing characteristic sequence through an encoder to generate control characteristics.
4. The electrical energy metering pipeline oriented multi-branch fault detection method of claim 1, wherein the fusing results in an enhanced joint feature comprising: splicing the multi-dimensional characterization features to obtain spliced features; Inputting the spliced features into a gating attention network to generate dynamic attention weights corresponding to the characterization features; Weighting and summing the corresponding characterization features by using the dynamic attention weight to obtain initial fusion features; and carrying out residual connection on the primary fusion characteristic and the splicing characteristic, and obtaining the reinforced joint characteristic after linear projection.
5. The power metering pipeline oriented multi-branch fault detection method of claim 1, wherein the validating the enhanced joint feature comprises: Calculating the minimum mahalanobis distance between the reinforced joint feature and a failure feature cluster center in a preset failure library; if the mahalanobis distance exceeds a preset distance threshold, judging that the fault characteristic is the initial fault characteristic; Searching known fault characteristics for the initial fault characteristics through nearest neighbors, and generating confidence according to the similarity between the initial fault characteristics and the known fault characteristics; If the confidence coefficient is higher than the preset confidence coefficient threshold value, verifying the fault characteristic as the inferable first-occurrence fault characteristic.
6. The power metering pipeline oriented multi-branch fault detection method as claimed in claim 5, wherein selecting a stable loop to identify pipeline equipment faults and update multi-branch feature extraction network parameters or selecting an emergency loop to update network parameters to identify faults according to verification results comprises: if the reinforced joint characteristics are verified to be known fault characteristics, a stable loop is selected, wherein the fault is identified through a classifier, the reinforced joint characteristics are temporarily stored in a cache queue, and when the cache queue is full, the network parameters are updated based on elastic weight consolidation constraint; And if the enhanced joint features are verified to be the inferable initial fault features, selecting an emergency loop, namely temporarily storing the enhanced joint features into a training queue, inferring a fault main body by combining a knowledge graph, offline training the multi-branch feature extraction network and the classifier based on the training queue, and updating online network parameters in a hot switching mode.
7. The electrical energy metering pipeline oriented multi-branch fault detection method according to claim 6, wherein the combining the knowledge graph reasoning fault body comprises: calculating the matching degree of the reinforced joint feature combination in the training queue in the knowledge graph known path, wherein the matching degree comprises coverage features, sequence features and confidence degrees; and inputting the matching degree into a preset sorting learning model, generating a path sorting score, and outputting a fault main body corresponding to the path with the highest score.
8. A system using the power metering pipeline oriented multi-branch fault detection method of any one of claims 1-7, comprising: the data perception module is used for acquiring multi-source heterogeneous operation perception data of the electric energy metering pipeline; The branch feature extraction module is used for processing the perceived data through a multi-branch feature extraction network to generate multi-dimensional characterization features, and comprises a CNN-converter branch for processing a visual image to generate visual features, a spectrum perception branch for processing control signal information to generate control features, a multi-layer perception machine branch for processing running state information to generate process state features, and a self-adaptive gating network fusion to obtain enhanced joint features; And the branch identification module is used for verifying the reinforced joint characteristics, selecting a stable loop to identify the equipment faults of the assembly line and updating the multi-branch characteristics to extract network parameters according to the verification result, or selecting an emergency loop to update the network parameters and then identifying the faults.

Description

Multi-branch fault detection method and system for electric energy metering assembly line Technical Field The invention relates to the technical field of fault detection, in particular to a multi-branch fault detection method and system for an electric energy metering pipeline. Background In modern production lines, the metering equipment is a core node for quality control and compliance production, and faults of the metering equipment not only can lead to product metering errors and rising defective rate, but also can cause whole line shutdown of the production line to cause large-scale production loss, so that the metering equipment is required to have real-time fault sensing capability. The CNN-transducer complementarily fuses the local space feature extraction advantage of the CNN with the global dependency modeling capability of the transducer, and combines local detail and global profile. The multi-branch feature extraction is a feature extraction strategy in deep learning, and is characterized in that a feature extraction process is split into a plurality of parallel branch branches, each branch can adopt different convolution kernel sizes, step sizes, channel numbers or extraction modes to respectively capture features with different scales and different dimensions in input data, and the features extracted by each branch are integrated into a unified feature representation in the modes of splicing, fusing and the like, so that the comprehensiveness and diversity of feature extraction are considered, and the feature capturing capability of a model on complex data is improved. For example, the invention patent application with publication number CN118779737A discloses a method and a system for managing and controlling faults of an electric energy metering multi-verification pipeline, and the method constructs a complete system comprising fault prediction, positioning, diagnosis, self-healing and global scheduling. According to the technical scheme, time sequence and image features are extracted through LSTM and CNN respectively, then, XGBoost is used for fault prediction after simple splicing, fault positioning depends on a pre-constructed equipment network diagram and a probability propagation model, SVM model is adopted for fault diagnosis, and finally, cross-pipeline scheduling is carried out through a genetic algorithm. The prior art scheme has the following defects: 1. The feature fusion mode is simple, namely the feature fusion is carried out on the multi-mode data only by adopting a splicing mode, and the inherent correlation and dynamic importance difference among different mode features cannot be fully considered, so that the characterization capability of the fusion features is limited, and the accuracy of subsequent fault detection is affected. 2. The adaptability and the interpretability of the model to new faults are insufficient, the fault detection and diagnosis model (XGBoost, SVM) in the prior art essentially belongs to a traditional machine learning or integrated learning model, the new fault mode or new fault characteristic recognition capability which does not appear in a training set is weak, and an effective online verification and model updating mechanism is lacked, so that the system reliability is challenged. 3. The fault detection links are isolated, priori knowledge is not fully utilized, namely the fault detection links, the fault positioning links and the fault diagnosis links in the prior art are relatively fractured, and the prior knowledge of equipment installation positions, process logic and the like is not effectively integrated when the fault detection model is judged, so that false alarm can be caused. In the existing fault detection technology of the electric energy metering pipeline equipment, the problems of insufficient multi-mode feature fusion and inaccurate detection exist, and the adaptability to new fault features is poor. The present invention proposes a solution to the above-mentioned problems. Disclosure of Invention In order to overcome the defects in the prior art, the invention provides a multi-branch fault detection method and a multi-branch fault detection system for an electric energy metering pipeline, which are used for extracting multi-dimensional characteristics by constructing a multi-branch characteristic extraction network, the reinforced joint characteristics are obtained through fusion and verified, and the branch loop is selected to identify the equipment faults of the production line, so that the problems of inaccurate intelligent detection and weak adaptability of the equipment faults of the electric energy metering production line are solved. In order to achieve the above purpose, the present invention provides the following technical solutions: A multi-branch fault detection method for an electric energy metering assembly line comprises the following steps of obtaining multi-source heterogeneous operation sensing data of the electric