Ling Xiang

Wind turbine bearings operate under variable and complex working conditions, where weak degradation related features are usually submerged in strong background interference, making the early stage degradation identification challenging. To accurately characterize the degradation of wind turbine bearings, a novel coherent projection order interference field method is proposed. In this method, an order domain projection framework is first established by projecting the angular domain vibration signal onto a function subspace composed of dominant order fringes, where the degradation related components can be adaptively extracted. Within this projection framework, a hybrid weighting operator integrating cyclic spectral coherence and amplitude intensity is developed to optimize the projection process, which enhances periodic degradation signatures while suppressing irrelevant background noise. Furthermore, a historical energy sensitivity factor is introduced to construct a dynamic order interference field matrix, and the degradation evolution process is characterized by analyzing the texture variation of the interference field. With the progression of bearing degradation, the interference pattern gradually evolves from sparse low frequency fringes to dense high frequency fringes. Based on the resulting order interference representation, a high frequency weighted two-dimensional spectral distribution entropy is developed as the health indicator. Subsequently, a Chebyshev based threshold strategy is employed to automatically determine the degradation initiation point. Experimental investigations on full life wind turbine bearing datasets demonstrate that the proposed method can effectively reveal degradation evolution characteristics and accurately identify the incipient degradation stage, which indicates its potential for intelligent condition monitoring and predictive maintenance of wind turbine bearings.

Structural damage identification of wind turbine blades (WTBs) is crucial for reducing maintenance costs and ensuring operational reliability. Data-driven methods leveraging supervisory control and data acquisition (SCADA) data have advanced WTBs monitoring, but existing approaches remain limited by models that fail to capture subtle changes in blade damage characteristics. To address these gaps, a novel multi-scale deformable linear Transformer network (MDLTN) is proposed for structural damage identification of WTBs. In contrast to representative temporal Transformer models such as the Temporal Fusion Transformer and Informer, the proposed MDLTN introduces a multi-scale deformable attention mechanism with linear computational complexity, enabling adaptive modeling of scale-varying and damage-sensitive temporal features. Specifically, the proposed network can extract temporal features at various scales by applying a multi-scale time convolution module, and a deformable linear self-attention mechanism is proposed to adaptively capture both local and global temporal dependencies. Due to the limited availability of data across diverse blade regions and varying damage levels, a wind turbine model is established to obtain the blade failure state data, considering the complex wind condition environment based on GH Bladed. The response dataset collects data from different blade sections subjected to various damage degrees under a range of wind speed conditions. The SCADA data are provided by a wind farm, covering 16 distinct measured blade conditions. The analyzing results of two datasets based on the response data from simulated model and the real SCADA data from wind farm demonstrate that the proposed network achieves superior performance in terms of accuracy, robustness, and generalization capability.

Affected by complicated and harsh operating conditions, the faint fault characteristics of wind turbine yaw bearings are inevitably submerged by severe background noise. Consequently, the accurate fault diagnosis for wind turbine yaw bearings is always considered a challenging task. For the purpose of tackling this problem, a novel adaptive frequency band division method named EGOgram is proposed. Specifically, two key parts are designed for extracting fault components in this method. First, with the proposed EG indicator, an adaptive frequency band division strategy is developed by fusing spectral coherence analysis and multi-window demodulation. The traditional fixed band division often leads to signal over-decomposition or under-decomposition, and this strategy solves the issue. Furthermore, to derive the dual band product spectrum, the method element-wise multiplies the improved envelope spectra of two strongly correlated frequency bands, and the loss of fault characteristics of traditional single band demodulation is avoided. Finally, the EGOgram exhibits the superior fault characteristic extraction capability by processing both simulated and experimental signals, and achieves better diagnostic performance for wind turbine yaw bearings compared with other competing methods.

The intelligent fault diagnosis of rolling bearing is crucial for ensuring stable operation of rotating machinery. However, in industrial settings, the normal data far exceeds fault data, resulting in a long-tailed distribution. This issue causes existing intelligent diagnosis models to be biased toward the normal class, making it difficult to identify tail-class faults. To address this issue, a novel network is proposed named Feature-aware Dynamic Expert (FADE) for rolling bearing fault diagnosis under long-tailed data. Within this network, a feature-aware balanced sampler is constructed, which uses sample feature distinctness to quantify learning value and assigns high sampling weights to hard-to-classify tail-class samples. In proposed FADE network, a dynamic routing-based multi-expert fusion module is proposed, which utilizes a routing mechanism to assign expert weights at the sample level, enabling different experts to focus on specific fault patterns. A multi-objective optimization constrained loss is designed to dynamically adjust the learning pace for head and tail classes. Through the coordinated constraints of multiple losses, it achieves balanced learning for both head and tail classes and enables precise capture of fault features. Experimental results on the public dataset and laboratory dataset show that the proposed method achieves high fault diagnosis accuracies. Compared to other mainstream methods, the proposed FADE model demonstrates better overall performance in the task of intelligent fault diagnosis for bearings.

In intelligent fault diagnosis of rotating machinery, the domain adaptation under limited data is important for the model to capture transient impact features of variable operating conditions precisely. A confidence-gradient adaptive meta-learning (CGAML) method is proposed for few-shot fault diagnosis of cross-condition and cross-testbed scenarios in rotating machinery. Within this method, a dynamic confidence and gradient synergy adaptive weighting mechanism is proposed to adaptively reweight meta-tasks based on loss confidence and gradient consistency, thereby suppressing outlier interference during optimization. Meanwhile, a supervised contrastive regularization strategy is employed to optimize the feature manifold, enhancing intra-class compactness and inter-class separability. Furthermore, a base-learner incorporating a fused attention gated module is proposed, utilizing an adaptive gating unit to integrate channel-spatial and multi-head self-attention for capturing transient impact features from non-stationary signals. The proposed method is validated on three distinct rotor testbeds. The results show that, compared to other diagnostic methods, the proposed CGAML approach achieves highly competitive and overall superior accuracy in cross-condition and cross-testbed tasks, providing a new solution for transfer fault diagnosis of rotating machinery under significant domain shifts and equipment heterogeneity.

The high variability and uncertainty of meteorological data from wind farm measurement towers and weather stations limit wind power forecasting accuracy, thereby affecting operation and maintenance efficiency. Although multi-modal information fusion has demonstrated substantial potential for improving forecasting performance, existing approaches remain constrained by their limited capability to mine complex spatiotemporal dependencies under non-stationary (NS) conditions and by the difficulty of effectively integrating auxiliary textual information with time-series modeling. To address these limitations, a cross-modality fusion stationary-patch (CMFSP) network is proposed by combining a pre-trained large language model (LLM) as a text encoder with three improved attention mechanisms (AMs) for spatiotemporal representation learning. Specifically, a dual-stream design processes time-series data and textual descriptions, where the pre-trained LLM extracts auxiliary semantic and statistical representations from the textual modality, and a cross-modality fusion AM integrates complementary information across modalities. Subsequently, an explicit NS spatiotemporal correlation extractor is designed to suppress short-term NS interference, aggregate historical context, and preserve informative NS patterns for spatial correlation learning. Moreover, a frequency-constrained weighted decay loss is introduced to regulate prediction residuals in the frequency domain with step-dependent weighting. Experiments are conducted on three offshore wind farm datasets that differ in installed capacity and local statistical characteristics, and CMFSP is compared with state-of-the-art forecasting models under short- and medium-term forecasting horizons. The results show that CMFSP achieves favorable forecasting performance.

Due to the harsh operating environments, the incipient fault characteristics of wind turbine yaw bearings are often obscured by intense background noise. Fault detection for wind turbine yaw bearings is always considered a challenging task. To address this issue, an innovative characteristic extraction method named cyclostationarity mode extraction (CME) is proposed to extract the underlying fault components from the vibration signals of yaw bearings. Essentially, the proposed CME designs an adaptive filter bank to achieve characteristic extraction without requiring prior information about the fault period. Firstly, a bandpass filter bank initialized by the Hamming window is constructed to divide the frequency band of the vibration signal into a series of temporary filtered signals. Subsequently, the enhanced autocorrelation strategy (EAC) is developed to estimate the cyclic frequencies concealed within strong noise interference. With the assistance of the EAC strategy, the fault information is accurately captured, and its characteristics are enhanced with the filter coefficient update. Meanwhile, redundant modes are eliminated during the optimal mode selection process based on the minimum Euclidean distance criterion. Additionally, the influence of key parameters in CME is analyzed through simulation signals. A novel effectiveness indicator, the crest factor of β -power envelope spectrum, is defined and further incorporates the multiple grid search algorithm to realize the automatic selection of the filter length as well as the decomposition level. The reliability of CME is confirmed by analyzing the simulation and the experiment signals, and CME outperforms the compared methods in fault detection of the wind turbine yaw bearings.

Wind turbine yaw bearings play an important role in power generation operation, and their damages will cause security risks. To identify the damages of wind turbine yaw bearings effectively, a novel deconvolution method named maximum average Gini index deconvolution (MAGID) is proposed. In this method, two key parts of deconvolution are designed for extracting damage component. On the one hand, a dedicated phase shift optimization strategy is designed to isolate individual damage cycles and get the optimal phase shift from the raw signal. The signal is first segmented in the angle domain using the damage characteristic order to isolate individual damage cycles. Subsequently, the optimal phase shift is determined by maximizing the average Gini index (AGI) to ensure the best alignment of damage features. On the other hand, the core deconvolution procedure is used to solve iteratively for the optimal deconvolution filter by maximizing the AGI criterion. The filter optimization is efficiently performed by solving a numerical optimization problem. At each iteration, the filter coefficients are updated through eigenvalue decomposition. Based on identification results from both simulation and experimental signals, this method accurately identifies local damage in yaw bearings. The effectiveness demonstrates clear reference value for engineering applications.

Existing cross-domain fault diagnostic algorithms often presume prior knowledge of the target domain’s failure patterns and train models on labeled data from the source domain. However, in practical engineering, new domains may include new failure modes, making it difficult to acquire labeled fault samples. This paper proposes a multi-classifier residual universal domain adaptation network (MUDAN) for cross-domain rotor fault diagnosis without making explicit assumptions about fault modes. The proposed model effectively extracts feature information from source vibration data and migrates model parameters to the target domain. In target domain training, an open-set classifier is trained using open-set entropy minimization to identify unknown faults, while a closed-set classifier is trained by weighting classification loss with nearest neighbor labels. Simultaneously, pseudo-labels are refined using nearest-neighbor knowledge to reduce pseudo-label noise. To lessen reliance on experimental data, the dynamics of the Jeffcott rotor with crack faults are modeled, and the system’s vibration response serves as a simulation dataset for migration tests. Three migration experiments are conducted using a simulated dataset and two test bench datasets. Results show that the MUDAN approach achieves higher diagnostic accuracy in cross-domain tasks compared to other methods, providing new reference value for cross-device rotor fault transfer diagnosis when target domain labels are unknown.