Chengying Zhao

The degradation process of bearings is characterized by pronounced nonlinearity and uncertainty, making it difficult to establish explicit dynamic models that accurately describe their degradation behavior. Moreover, most existing methods provide only deterministic remaining useful life (RUL) estimates and lack effective quantification of predictive uncertainty, thereby limiting their reliability in maintenance planning and operational decision-making. To overcome these limitations, this paper proposes a trustworthy RUL prediction method for bearings based on a Bayesian physics-informed neural network (BPINN). This method effectively integrates the physics-informed architecture with a Bayesian framework, thereby fully leveraging the enhancement effect of physical information on model interpretability. Simultaneously, it utilizes the probabilistic modeling capability of the Bayesian framework to provide uncertainty estimation for RUL prediction. Experimental results indicate that it can provide RUL predictions with associated uncertainty information, which offers additional decision-support information for predictive maintenance.

During long-term operation, the accumulation of material damage in bearings leads to transitions in failure mechanisms, which are reflected by pronounced time-varying characteristics in vibration responses. Conventional single-model prediction approaches are inadequate to simultaneously capture the slow degradation behavior in the early stage and the rapid deterioration in the later stage. To address this issue, a multi-stage bearing performance degradation digital twin modeling framework based on ensemble support vector regression (SVR) is proposed. First, a five-degree-of-freedom bearing fault dynamic model is developed to generate simulated vibration signals, and the defect size over the entire bearing life cycle is inversely identified through bidirectional mapping between simulated and measured vibration signals. Subsequently, a dual-feature synergistically driven degradation stage partitioning strategy is proposed to divide the bearing degradation process into multiple stages. Based on the identified stages, an SVR model pool is constructed to establish stage-wise mapping relationships between bearing defect size and vibration features. Finally, the predicted defect sizes are fed back into the bearing fault dynamic model to generate updated simulated vibration signals, thereby forming a closed-loop iterative updating mechanism of the bearing digital twin. Both the XJTU-SY bearing dataset and a specially constructed dataset are employed to verify the proposed method, which demonstrates exceptional accuracy in lifecycle-wide degradation forecasting.

Automated detection of surface defects in construction machinery has become a critical factor in ensuring equipment safety and stable operation. However, in practical applications, construction machinery often operates under complex illumination conditions, where defect areas exhibit significant variations across different illumination scenarios. Additionally, the scarcity of high-quality annotated samples presents challenges to the robustness and transferability of traditional deep learning methods. In this paper, a cross-modal few-shot segmentation network (CFSNet) is proposed to enhance surface defect segmentation performance under complex illumination conditions. The method utilizes Retinex decomposition to enhance the reflection information of RGB images, which effectively suppresses illumination interference and improves feature characterization. At the same time, infrared (IR) modal information is incorporated to provide stable structural priors. A semantic matching module and a feature guidance enhancement module are proposed to achieve multi-modal fusion and robust representation across varying illumination conditions. Under conditions of sparse annotated samples, CFSNet improves detection accuracy and generalization capabilities across illumination scenarios through structural alignment and semantic completion strategies. The experimental part constructs a construction machinery surface defect dataset containing multiple illumination conditions. The results show that the proposed method achieves the best performance compared to the state-of-the-art methods for 1-shot and 5-shot tasks, which validates its adaptability and practical value in complex illumination environments.

Accurate prediction of the remaining useful life (RUL) of rolling bearings is essential for predictive maintenance and safe operation of rotating machinery. However, bearing degradation vibration signals exhibit complex spatial cross-domain coupling and time-varying characteristics, which make it difficult for existing methods to effectively align graph construction with fault evolution and fully integrate graph spatial information with temporal vibration signals. To address these issues, this paper proposes a spatial-temporal dual-branch fusion neural network for bearing RUL prediction. First, a dual-sensitive feature collaborative processing mechanism is adopted to accurately identify the first prediction time (FPT), thereby extracting valid degradation data for subsequent prediction. Then, graph-structured data are constructed according to both feature correlation and temporal correlation. Based on this, a graph attention network (GAT) is employed to mine hidden associations among degradation features, while a temporal branch composed of a temporal convolutional network (TCN) and a Transformer encoder is introduced to capture local degradation fluctuations and long-term temporal dependencies from vibration sequences. Furthermore, a bidirectional cross-attention fusion mechanism (Bi-CA) is designed to achieve deep interactive fusion of spatial and temporal features for RUL prediction. Comparative experiments on the self-designed bearing dataset and the XJTU-SY dataset demonstrate that the proposed method achieves more accurate and stable RUL prediction than representative methods, especially under complex operating conditions. These results verify the effectiveness of the proposed spatial-temporal collaborative modeling framework for bearing prognostics.

Fluctuations in rolling bearing vibration signals, induced by stage-dependent degradation dynamics and stochastic disturbances under varying operating conditions, pose significant challenges for remaining useful life (RUL) prediction. Consequently, this nonstationarity masks subtle early degradation cues, hinders stage-transition modeling, and often causes purely data-driven methods to produce physically implausible RUL decay trajectories. Moreover, interpretability remains limited when learned degradation representations are weakly connected to physically meaningful decay dynamics. To overcome these challenges, a physics-informed temporal-enhancing neural network (PI-TENN) is developed for estimating the RUL of rolling bearings across multi-stage degradation regimes. Specifically, a temporal-enhancing degradation feature fusion module is constructed by coupling a multi-layer long short-term memory (LSTM) temporal dependency encoder with an attention-enhancing block. In this way, the multi-layer LSTM encoder captures long-range temporal dependency, while the attention-enhancing block suppresses noise-dominated time steps and highlights degradation-relevant patterns, thereby producing latent degradation representations that are stable across degradation stages. Furthermore, a universal partial differential equation (PDE) is constructed to explicitly relate the latent degradation representation to the decay rate of RUL, which enables the learning of stage-wise RUL evolution patterns across distinct degradation stages. Meanwhile, model optimization is guided by a data–physics fusion loss integrating data fitting, PDE residual, and monotonicity regularization, so that predictive accuracy and physically consistent stage-wise decay are jointly ensured. Finally, experiments on full-life-cycle steel and ceramic bearing datasets demonstrate improved accuracy, robustness under complex operating conditions, and enhanced physical consistency compared with representative baselines.

Performance degradation assessment of bearings is crucial for achieving high reliability in predictive maintenance of rotating machinery. In practical applications, the abrupt state transitions and highly fluctuating health indicators (HIs) of full-ceramic ball bearings render conventional threshold-based criteria unreliable, making early fault detection difficult. To address this, by calculating the degree of shock accumulation within the system, this article comprehensively utilizes fault propagation information across multiple time scales. The high fluctuations which hamper diagnosis are transformed into a physically meaningful HI, improving monotonicity and interpretability. The relationship between the proposed HI and other HIs is represented by a joint probability distribution, aiding in early fault detection associated with brittle-failure onset. The proposed method is validated by estimating appropriate failure threshold (FT) values for various HIs in real full-ceramic ball bearing degradation cases. The results indicate that the proposed method can establish a physically meaningful HI and develop a probability distribution criterion for FT under highly fluctuating HIs.

Addressing the limited adaptability of mechanical models and the deficiencies of data-driven methods in terms of noise robustness and interpretability, this paper proposes a physics and data hybrid driven online remaining useful life (RUL) prediction framework. By establishing a synergistic mapping among defect quantification, vibration features, and performance degradation, it achieves precise RUL prediction for rolling bearings. Specifically, this framework achieves robust quantification of defect size by establishing a bearing failure dynamics model and adopting a dual-objective feature matching method. By combining the quantified defect size with the high-trend time-frequency domain features as the input of the network, a deep synergy between physical degradation laws and vibration responses is achieved. On this basis, a multi-scale deep synergistic network is constructed to extract local transient features and global degradation trends of bearings. Furthermore, a physics-informed loss function with monotonicity constraints is designed to guide model optimization, ensuring that the prediction results conform to the physical mechanism of irreversible degradation of bearings. Experiments conducted on the XJTU-SY and self-built dataset demonstrate that the proposed method significantly outperforms mainstream models in terms of prediction accuracy. The results demonstrate that the deep integration of physical mechanisms and data features effectively enhances the characterization capability of degradation features, providing a novel perspective for the intelligent maintenance of industrial bearings.

To address the challenges of insufficient prediction accuracy and lack of interpretability caused by the dynamic competition of compound faults in rolling bearings, this article proposes a fault-type adaptive method for remaining useful life (RUL) prediction. It innovatively incorporates defect size acquisition, defect competition modeling, and physics-data collaboration mechanisms to build a comprehensive prediction system for compound degradation scenarios. Specifically, based on the dynamic model and measured signals, a multifeature quantification method is proposed to evaluate defect sizes, and the fault coupling factors are introduced to dynamically assess the impact of various defects on system degradation. Combining physical and data characteristics, an adaptive multiscale deep collaborative network (AF-MBGA) is designed to achieve online RUL prediction through dynamic weighting. The results based on the XJTU-SY public dataset and the experiments conducted independently show that this method exhibits excellent prediction accuracy and robustness under compound fault scenarios. By establishing a defect competition mechanism, this method depicts the dynamic switching of dominant relationships between compound defects, providing mechanism-level support for bearing health management.

Bearings are critical components that ensure the normal operation of rotating machinery predicting its remaining useful life (RUL) is essential for proactive maintenance and safeguarding production reliability. However, current bearing RUL prediction methods often face challenges such as incomplete acquisition of degradation information and physical inconsistency in prediction results. To resolve these problems, this paper presents a physics-guided multi-scale attention fusion network (PMSAFN) for RUL prediction of bearings. Specifically, a multi-scale attention fusion (MSAF) network is designed to simultaneously capture both global degradation trends and local fault evolution patterns from historical monitoring data of the bearing. By incorporating rotary position encoding, the proposed method can encode bearing features, which significantly enhances its ability to process long-sequence features. At the same time, a gated fusion mechanism is designed to dynamically integrate global and local features, ensuring a comprehensive representation of bearing health states. Furthermore, a physics-informed adaptive penalty loss function is constructed to guide the results to comply with the physical laws related to bearing degradation. Comparative experiments and ablation experiments based on the XJTU-SY dataset and PHM 2012 challenge dataset show that the PMSAFN network achieves better prognostic accuracy.

Surface defect detection in construction machinery is critical to ensure the safety and reliability of equipment operation. However, in full-scene detection, drastic changes in illumination conditions can seriously affect the image quality and reduce the accuracy of defect detection. To address this problem, a hybrid enhancement network (HENet) is proposed in this paper, which fuses visible, infrared and depth images to improve the detection performance of the model. The method enhances the complementary information expression ability of multi-modal images by constructing a cross-modal fusion strategy, and introduces a similarity metric mechanism to strengthen the stable perception of defect features. To verify the effectiveness of the method, a variable illumination multi-modal surface defect dataset is constructed and systematic comparative experiments are conducted. The average results show that HENet improves 3.39% and 0.29% in intersection over union (IoU) and foreground–background IoU metrics compared to the suboptimal method, respectively. In addition, cross-validation improved HENet by 6.92% and 6.14% in terms of mIoU in 1-shot and 5-shot settings, respectively. The experimental results verify the stable detection capability of the proposed model under multiple illumination conditions and its practical engineering application value.