Abnormal identification and evolution modeling of rail transit signal equipment in big data environment

Aiming at the problems that traditional threshold alarm of rail transit signal equipment is difficult to capture early weak anomalies, insufficient fusion of multi-source heterogeneous data and unclear fault evolution mechanism, this paper proposes a big data-driven anomaly identification and evolution modeling method. Firstly, a multi-source data fusion and feature enhancement module is constructed. Hadoop architecture is used to store PB-level monitoring data, multi-head convolutional neural network (CNN) is used to extract spatio-temporal features, and the average volatility index is introduced to characterize the stability of equipment. Combined with semi-supervised collaborative training, the problem of label scarcity is alleviated. Secondly, a multi-head CNN-RNN hybrid anomaly detection model is designed, with G-means and average accuracy as indicators, and the recognition ability of unbalanced data is improved through multi-task joint learning objectives. Finally, a dynamic fault propagation graph (DFPG) model is established. Based on graph attention networks, the intensity of fault propagation between devices is quantified. By integrating a Hidden Markov Model-Particle Filter hybrid model (HMM-PF), probabilistic prediction of remaining useful life (RUL) is achieved. The experiment is based on 13.1TB real data of a subway in a city from 2019 to 2022. The results show that the F1-score of this method is 87.8%-89.0%, the G-mean is over 89%, the average absolute error of RUL prediction is only 3.6 days, and 82.6% of the faults are forewarned for an average of 42 days, and the false alarm rate is reduced by 27.5 percentage points compared with the traditional method.