Remaining Useful Life Prediction of Rolling Bearings Based on Feature Screening and an Improved Loss Function

Rolling bearings are critical components in numerous industrial applications, and their Remaining Useful Life (RUL) prediction can enable proactive maintenance strategies that prevent unexpected failures, reduce operational costs, and enhance safety. However, most existing RUL prediction methods use a traditional loss function that does not account for the asymmetric risks of early predictions or late predictions, leading to suboptimal predictions that may incur severe consequences. To address this issue, this paper proposes an improved loss function, which can differentiate between different risk situations according to the error between the predicted RUL and the real RUL, and give different penalty weights, so that the model will tend to make a prediction lower than the real RUL of the bearing during the learning process, in order to avoid the serious consequences of the premature failure of the bearing. Specifically, this paper proposes a rolling bearing RUL prediction method using feature screening with an improved loss function in the following steps: (1) extracting time domain and frequency domain features from the bearing vibration signals and computing Integrated feature screening indicators to select the most relevant features that capture the bearing degradation trend; (2) training of encoder-decoder based on attention mechanism model using an improved loss function and filtered features; (3) fitting a linear function to the prediction outputs to obtain the final RUL estimates. The experimental results on the PHM2012 dataset demonstrate that the proposed method outperforms other state-of-the-art methods and baseline methods in terms of prediction accuracy and robustness.