Bearing health management approach with reliability health indicators and enhanced fitness prediction

Abstract The construction and prediction of health indicators (HIs) are crucial for maintaining mechanical equipment containing rolling bearings. Conventional HI construction methodologies often rely on either extensive feature engineering or indirect signal transformation, which may hinder adaptability in diverse operating conditions. To address this, we propose a linear encoding composite degradation-aware CNN-Transformer network (LECDA-Net). It constructs the HI directly from raw multi-axis vibration signals without the need for handcrafted features or complex physical models, and, through linear positional encoding and a composite degradation-aware loss function, ensures that the constructed HI is not only consistent with data-driven patterns but also aligns with physical degradation laws. First, vibration data from the X and Y axes of the bearings undergo low-pass filtering. These filtered signals are then individually subjected to linear positional encoding and fed into LECDA-Net for HI construction. The constructed HI from each axis are subsequently fused through weighted aggregation to produce the overall bearing HI. Comparative experiments show that, compared with convolutional neural network (CNN)-Transformer, CNN-Transformer-DA, deep convolutional autoencoder, convolutional autoencoder, PCA, and DDA methods, LECDA-Net achieves remarkable performance improvements. The constructed HI demonstrates clear advantages in monotonicity, correlation, and overall composite performance. Following the successful construction of a reliable bearing HI, health index prediction based on historical data plays a key role in proactive equipment maintenance. Although existing prediction methods can achieve high accuracy, the evaluation of prediction goodness of fit has received only limited attention. Therefore, we designed an AuTanh-bidirectional long short-term memory (BiLSTM) method, introducing channel-wise adaptive learnable parameters to address the issue that traditional activation functions use scalar parameters for all channels, resulting in poor linear fitting. This method incorporates channel-wise learnable parameters in the AuTanh layer, enhancing the model’s fitting ability through a flexible nonlinear activation mechanism. Compared with BiLSTM, BiLSTM-DA, LSTM, gated dual attention unit, GAU, GAHAU, and temporal convolutional network models, the proposed AuTanh-BiLSTM model demonstrates lower error and superior fitting performance.