Brake uneven wear of high-speed train intelligent monitoring using an ensemble model based on multi-sensor feature fusion and deep learning

• Published in: Engineering failure analysis 2022-07, Vol.137, p.106219, Article 106219
• Main Authors: Zhang, Min, Zhang, Xunjie, Mo, Jiliang, Xiang, Zaiyu, Zheng, Pengwei
• Format: Article
• Language: English
• Subjects: Braking system; CNN-BiGRU; Ensemble model; Intelligent monitoring; Multi-source feature fusion; Uneven wear
• ISSN: 1350-6307; 1873-1961
• DOI: 10.1016/j.engfailanal.2022.106219

•The friction blocks are produced for different uneven wear conditions, and multi-source signals of the friction blocks are collected.•A new model based on CNN and BiGRU, which exploits the spatial feature extraction capability of CNN and advantages of BiGRU in extracting temporal features, is proposed.•A novel ensemble model using a grid search algorithm is proposed.•Comprehensive experiments are designed and performed to prove the effectiveness of the proposed algorithm. Uneven wear is an inevitable wear phenomenon in the braking system of high-speed trains. To avoid the hidden dangers of the aggravation of uneven wear to the safe operation of trains, it is necessary to monitor the uneven wear state in the braking system in real time. The signals caused by the friction interface present high nonlinearity and potentially chaotic characteristics, which require higher performance for monitoring models. In addition, the complex operating conditions of the friction interface and multiple sources of vibration excitation during braking can reduce the monitoring accuracy with only a single sensor. To address these issues, we propose a deep learning-based ensemble model for intelligent monitoring of the brake’s uneven wear condition. First, an integrated model based on convolutional neural network and bidirectional gated recurrent unit was developed to learn spatial and temporal knowledge from the training set of vibration noise, vibration acceleration and friction coefficient. Second, the initial weights were assigned to each model, and the validation set of each signal was used to determine the best weights combined with a grid search algorithm. The models were integrated to achieve multi-source information fusion to improve the accuracy of the final decision. Finally, the ensemble model exploited multi-source feature information from different perspectives. It established the mapping relationship between feature space and sample label space and obtained the diagnosis results through the test set. The proposed model effectively unified automatic feature extraction and information fusion, thus enhancing its intelligence and self-adaptability. Various experimental results showed that the proposed method was highly accurate and stable as well as could effectively monitor the various uneven wear conditions of braking systems.