Improved Random Forest Based Anomaly Detection for Urban Rail Transits.

As the construction of urban rail transit is rapidly expanding, the maintenance costs associated with it are also increasing. However, traditional anomaly detection methods have limitations as they require setting thresholds for various indicators to monitor rail transit status. To address this issue, there is a pressing need for utilizing machine learning methods to detect possible anomalies in trains. This paper proposes an anomaly detection method based on random forest, which can estimate the abnormality score of the train's traction motor based on low-dimensional sensor data and insufficient labels collected by the motor sensor. This approach can identify potential anomalies of the motor and help maintenance personnel develop specific plans to address them. In addition to the proposed anomaly detection method, we have designed a train traction motor anomaly alarm system that effectively detects potential anomalies in the train traction motor covering data pre-processing and anomaly warning. Compared to deep learning models that have high training costs, our proposed method requires less training data, making it suitable for situations with limited training data. The results demonstrate that our method can significantly improve the accuracy of anomaly detection compared to traditional machine learning methods.