An unsupervised feature learning based health indicator construction method for performance assessment of machines

In order to assess the degradation process of machines, it is necessary to construct a suitable health indicator. Existing health indicators are mainly constructed with manually extracted features. Those manually extracted features are based on the rich domain knowledge of experts. However, domain knowledge may be difficult to obtain. In order to automatically construct a health indicator, an unsupervised feature learning based health indicator construction method is proposed in this paper. The proposed method mainly consists of three steps: Firstly, a multiscale convolutional autoencoder network is built where the network hyperparameters are optimized through a genetic algorithm. Then, acquired sensor signals are directly input into the constructed network to adaptively learn features. For enhancing the effective features and suppressing the useless ones, different weights are assigned to all features. At last, the relative similarity of learned features between the baseline sample data and the currently acquired sample data is calculated as a health indicator to represent the health condition of machines. The effectiveness of the proposed method is validated through two cases. In those case studies, two metrics, including trendability and scale similarity, are used to quantitatively compare the performance of the proposed method with some other state-of-the-art ones. Results demonstrate that the health indicator constructed with the proposed method is able to effectively identify the degradation process of machines and obtain better performance than those comparative ones.