Abstract: To address the reduced monotonicity of the Health Indicator (HI) caused by the “self-healing” phenomenon during the degradation of coal mine main fan bearings, as well as the insufficient accuracy of Remaining Useful Life (RUL) prediction under complex operating conditions, an anti-self-healing HI construction method and a multi-branch parallel RUL prediction model are proposed. For anti-self-healing HI construction, the different capabilities of time-domain, frequency-domain, and time-frequency-domain features in characterizing bearing degradation are first analyzed, and multidimensional features reflecting vibration energy, impact intensity, spectral structure, and multiscale energy evolution are extracted. The energies around the characteristic frequencies and their harmonics associated with outer-race, inner-race, rolling-element, and cage faults are then used to divide the degradation process into different stages. Feature activity curves are constructed, and an anti-self-healing HI is generated by combining Principal Component Analysis (PCA), stage-adaptive weighting, and approximate isotonic smoothing, thereby reducing the influence of the “self-healing” phenomenon on degradation characterization. For RUL prediction, the time-domain, frequency-domain, and time-frequency-domain features are modeled separately according to their respective characteristics using a Temporal Convolutional Network with a Convolutional Block Attention Module (TCN−CBAM), a harmonic-attention-enhanced Long Short Term Memory network (LSTM), and a Vision Transformer (ViT) enhanced with energy-distribution and time-frequency state-space modules. An HI-guided dynamic fusion mechanism is subsequently employed to integrate predictive degradation information from multiple feature domains. Meanwhile, Physics-Informed Neural Network (PINN) constraints are designed for disturbances unrelated to bearing degradation, including power-supply harmonics, blade-passing frequency components, and lubrication-state fluctuations, to mitigate the influence of external operating-condition disturbances on RUL prediction. The proposed methods are evaluated using the University of Cincinnati Intelligent Maintenance Systems (IMS) bearing dataset, Xi’an Jiaotong University–Sumyoung Technology (XJTU−SY) rolling bearing accelerated life test dataset, and a field dataset collected from a coal mine main fan. In the HI comparison experiments, the proposed HI outperforms individual features and the HIs constructed using PCA, a Denoising Autoencoder (DAE), and a Variational Autoencoder (VAE) in terms of monotonicity and correlation across multiple datasets. Specifically, its monotonicity and correlation reach 0.988 and 0.975, respectively, on the IMS dataset, and 0.919 and 0.948, respectively, on the field dataset. In the RUL prediction experiments, the proposed model achieves a Mean Absolute Error (MAE) of 2.16, a Root Mean Square Error (RMSE) of 2.98, and a coefficient of determination (R²) of 0.964 on the field dataset, outperforming LSTM, TCN, ViT, Graph Attention Network (GAT), Mamba, and diffusion-based models. Finally, ablation experiments verify the contributions of the HI-guided mechanism and the PINN constraints to the prediction model. The results demonstrate that the proposed methods improve the capability of the HI to characterize the actual degradation trend of bearings and enhance the accuracy and stability of RUL prediction for coal mine main fan bearings under complex operating conditions.