TY  - EJOU
AU  - Kang, Jaeyong 
AU  - Kim, Chul-Su 
AU  - Kang, Jeong Won 
AU  - Gwak, Jeonghwan 

TI  - Recurrent Autoencoder Ensembles for Brake Operating Unit Anomaly Detection on Metro Vehicles
T2  - Computers, Materials \& Continua

PY  - 2022
VL  - 73
IS  - 1
SN  - 1546-2226

AB  - The anomaly detection of the brake operating unit (BOU) in the brake systems on metro vehicle is critical for the safety and reliability of the trains. On the other hand, current periodic inspection and maintenance are unable to detect anomalies in an early stage. Also, building an accurate and stable system for detecting anomalies is extremely difficult. Therefore, we present an efficient model that use an ensemble of recurrent autoencoders to accurately detect the BOU abnormalities of metro trains. This is the first proposal to employ an ensemble deep learning technique to detect BOU abnormalities in metro train braking systems. One of the anomalous cases on metro vehicles is the case when the air cylinder (AC) pressures are less than the brake cylinder (BC) pressures in certain parts where the brake pressures increase before coming to a halt. Hence, in this work, we first extract the data of BC and AC pressures. Then, the extracted data of BC and AC pressures are divided into multiple subsequences that are used as an input for both bi-directional long short-term memory (biLSTM) and bi-directional gated recurrent unit (biGRU) autoencoders. The biLSTM and biGRU autoencoders are trained using training dataset that only contains normal subsequences. For detecting abnormalities from test dataset which consists of abnormal subsequences, the mean absolute errors (MAEs) between original subsequences and reconstructed subsequences from both biLSTM and biGRU autoencoders are calculated. As an ensemble step, the total error is calculated by averaging two MAEs from biLSTM and biGRU autoencoders. The subsequence with total error greater than a pre-defined threshold value is considered an abnormality. We carried out the experiments using the BOU dataset on metro vehicles in South Korea. Experimental results demonstrate that the ensemble model shows better performance than other autoencoder-based models, which shows the effectiveness of our ensemble model for detecting BOU anomalies on metro trains.
KW  - Anomaly detection; brake operating unit; deep learning; machine learning; signal processing

DO  - 10.32604/cmc.2022.023641