A New Deep Tensor Autoencoder Network for Unsupervised Health Indicator Construction and Degradation

For metro vehicles, wheel performance degradation is inevitable due to the continuous wear of tread and rim. It is of significant value to apply machine learning techniques to evaluate the wheel’s degradation state with vibration signals, as it can lessen expert dependence and optimize repair strategy. In the real metro operation process, however, the wheels’ vibration signal is easily disturbed by irregular noise from many factors, such as load, road condition, temperature, etc., and lacks essential tendency characteristics for the intelligent state evaluation. To solve this problem, this article proposes a new unsupervised degradation state evaluation method for metro wheels against irregular noise disturbance. First, a new deep tensor autoencoder (TAE) network is proposed based on the merit of tensor representation in exploring the intrinsic information from raw signals. By extracting the core tensor from hidden features via Tucker decomposition, a new decoding process is rebuilt and a tensorized tendency regularizer is designed with the core tensor. Moreover, a new training algorithm with alternating minimization scheme is constructed to seek the optimal tensor representation and unsupervised feature extraction. Second, a health indicator (HI) of the degradation process is built with the extracted features. A HI-based state assessment algorithm is further proposed to adaptively determine the interval of state change to substitute the first/second warning in the current maintenance strategy. A set of comparative experiments are conducted using the wheel data from Beijing Subway. The results demonstrate that TAE can extract features with good monotonicity and tendency from the raw signals under irregular noise disturbance. More importantly, the obtained warning positions can precisely match the wheel diameter’s change in the actual repair record, which offers an easy-to-deploy and intelligent solution for the health management of metro wheels.