Rolling bearing is the core load-bearing component of rotating machinery, and accurate assessment of its degradation trend and prediction of its remaining useful life (RUL) provide key supports for condition-based operation and maintenance of equipment. Existing deep learning-driven fault prediction frameworks generally suffer from difficulties in defining health indicators, insufficient mining of long-term degradation features, and inadequate RUL prediction accuracy, making them poorly suited to industrial monitoring conditions with complex noise and scarce samples. To address these technical weaknesses, a RUL prediction method for rolling bearings was proposed by integrating multi-scale feature screening and a bidirectional residual long short-term memory (Bi-ResLSTM) network. The method consists of three main procedures. First, multi-scale hybrid features are extracted from the vibration signals collected from the equipment, highly sensitive features are screened by dynamic correlation window analysis, and multi-source feature denoising and fusion are completed by using a deep denoising autoencoder (DDA), so as to construct a health indicator curve that intuitively reflects the degradation state of the bearing. Second, the reconstruction error of DDA is used to identify the occurrence time of early bearing faults and accurately determine the starting point of RUL prediction. Finally, the constructed health indicator time-series curve is fed into the Bi-ResLSTM network to deeply capture the degradation evolution pattern of the indicator over time, and the RUL of the bearing is quantitatively calculated according to the pre-specified failure threshold of the equipment. The experimental results show that, compared with the traditional particle filter life prediction algorithm and the basic single-layer long short-term memory (LSTM) prediction algorithm, the overall prediction percentage error of the proposed method decreases by 15.57 and 5.75 percentage points, respectively, indicating that it has higher prediction accuracy and anti-interference robustness. It is fully demonstrated that the proposed scheme can deliver stable and accurate prediction even when only limited training samples of faults are available on site, and has high engineering application potential for fault prediction and preventive maintenance of industrial rotating machinery.