Extracting degradation trends for roller bearings by using a moving-average stacked auto-encoder and a novel exponential function

Building a smooth degradation curve for a bearing can provide a good basis for predicting its remaining useful life, but the traditional models need to fuse multiple models. The stacked auto-encoder (SAE) can extract the potential features of the data from the frequency domain directly, but the oscillation of the original data diminishes the smoothness and monotonicity of the extracted degradation curve. Moreover, the degradation curve extracted by SAE shows slight oscillation, which needs to be further eliminated to improve its monotonicity. However, some methods, such as an exponentially weighted moving average, require parameter setting as outlier detection can only detect abnormal points for local curves. To solve these problems, we proposed a moving window-based stacked auto-encoder (MASAE) with an exponential function, which incorporates a slope local minimum point (ESLMP) to extract the degradation trends and improve its monotonicity. The experimental data results demonstrated the superiority of the proposed model.