Accurate remaining useful life (RUL) prediction of rolling bearings is essential for predictive maintenance and reliable operation of industrial rotating machinery. However, variations in operating conditions introduce substantial differences in vibration characteristics and degradation patterns, limiting the applicability of models trained under individual operating conditions. This study investigates a multi-source transfer learning approach for bearing RUL prediction across different operating conditions. A framework termed deep residual shrinkage network with multi-source weighted screening (DRSN-MW) was developed by integrating a deep residual shrinkage network with multi-source domain screening, subdomain alignment, and physical monotonicity constraints. An adaptive feature extraction module was constructed to capture degradation characteristics from vibration signals, while weighted extension correlation and dynamic time warping (DTW) were combined to identify source domains with similar degradation patterns. Degradation-stage partitioning and entropy-weighted local maximum mean discrepancy (LMMD) were employed to align source and target features within corresponding degradation stages. Physical monotonicity and decision consistency constraints were further incorporated to regulate the predicted RUL trajectories. The framework was evaluated using the 2012 Prognostics and Health Management (PHM2012) bearing dataset and Xi’an Jiaotong University–Sumyoung (XJTU-SY) bearing dataset under single-source and multi-source transfer scenarios. Average mean absolute errors (MAEs) of 0.126 and 0.129 and root mean square errors (RMSEs) of 0.150 and 0.156 were obtained on the two datasets, respectively. On the PHM2012 dataset, the proposed method reduced the average root mean square error by 33.6% compared with the strongest baseline in the reported comparison. The results indicate that combining source-domain screening with degradation-stage alignment and physical constraints can improve RUL prediction across operating conditions. The proposed framework provides a data-driven approach to bearing health prognostics and predictive maintenance in industrial rotating machinery.

Industrial intelligence relies on the integration of industrial software, interconnected production systems, digital twins, and artificial intelligence (AI) to support data-driven monitoring, predictive analysis, and operational decision-making. Despite advances in these technologies, differences in technological maturity, system interoperability, and engineering trustworthiness remain obstacles to their coordinated deployment. This study investigates the relationships among the principal technologies underlying industrial intelligence, with particular attention to their functional dependencies, implementation constraints, and capacity to support reliable industrial operations. A structured literature review was conducted using a corpus comprising 55 peer-reviewed publications and 29 supplementary technical and institutional sources. Industrial software, the industrial Internet, digital twins and digital threads, and industrial AI were examined through a common framework covering seven dimensions of technological and engineering readiness. The comparative analysis identified two recurring categories of deployment constraints. Industrial software and the industrial Internet were primarily constrained by the accumulation of domain-specific knowledge, legacy-system compatibility, and heterogeneous data interoperability. For digital twins and industrial AI, the review identified additional challenges concerning model verification, uncertainty assessment, and the integration of intelligent decisions into industrial processes. The analysis further identified interdependent requirements for data consistency, mechanism-based modeling, computational resources, and closed-loop decision execution. These findings indicate that the engineering readiness of industrial intelligence depends not only on the maturity of individual technologies but also on the reliability of their interactions within integrated production systems. The study establishes a system-level perspective for assessing industrial intelligence technologies and provides a structured basis for identifying integration bottlenecks, evaluating engineering trustworthiness, and directing research toward verifiable industrial decision-making.

Bearing fault diagnosis remains challenging when models trained on one set of machines are applied to equipment with different bearing geometries, operating speeds, and vibration characteristics. Conventional evaluation protocols often overestimate cross-machine performance, while prediction uncertainty becomes difficult to calibrate under machine-level distribution shift. This study investigates whether geometry-anchored fault features and machine-level conformal calibration can jointly improve diagnostic accuracy and prediction-set reliability on previously unseen machines. A structured review was conducted to identify methodological gaps, followed by controlled simulations involving eight machines with different bearing geometries, shaft speeds, sampling rates, structural resonances, and noise conditions. Three feature representations and four calibration schemes were evaluated using a leave-one-machine-out protocol. The geometry-anchored representation achieved a mean classification accuracy of 0.908 $\pm$ 0.014, compared with 0.585 $\pm$ 0.016 and 0.575 $\pm$ 0.023 for frequency-domain and order-domain features, respectively. Random window splitting overestimated the accuracy of these baselines by 0.25 and 0.29. At a nominal coverage level of 90%, machine-level conformal calibration achieved mean coverage between 0.903 and 0.931, compared with 0.642–0.879 for conventional split calibration. With geometry-anchored features, the calibrated prediction sets contained 0.99 classes on average, whereas the other representations required 2.77–2.96 classes. Conservative calibration increased worst-machine coverage from 0.751 to 0.866, although individual-machine coverage remained variable. The results indicate that physically indexed fault features and machine-level calibration address complementary aspects of cross-machine diagnosis: the former improves fault discrimination, while the latter supports more reliable marginal coverage under machine shift. The study establishes a controlled evaluation framework for examining diagnostic generalization and uncertainty calibration at the machine level, providing a methodological basis for subsequent validation using experimental and industrial bearing data.