Acoustic emission (AE) technology is suitable for monitoring wind turbine blades because of its high-frequency response and sensitivity. However, existing damage localization approaches rely on multi-sensor arrays, which limit their engineering applicability. Moreover, deep learning models for single-sensor localization remain constrained by insufficient time-frequency representation and limited interpretability. To address these issues, a novel single-sensor AE source localization model, the Large-kernel dual wavelet attention network, was proposed. First, pencil lead break tests on real blade structures were conducted to investigate AE signal propagation, revealing strong nonlinear behavior and multi-scale attenuation characteristics. On this basis, a Large-kernel dual wavelet convolution layer was developed using Morlet and Mexican hat wavelets. With a mechanism-guided large wavelet kernel design, the arbitrariness in hyperparameter selection and the Gibbs phenomenon associated with incomplete truncation were mitigated, enabling the effective extraction of multi-scale time-frequency features. A time-grouped frequency attention mechanism module was introduced to adaptively recalibrate convolutional kernel weights through the fine-grained perception of localized time-frequency features, thereby improving the discrimination of critical features. Experiments on a real blade specimen demonstrated that the proposed method outperformed existing models in localization accuracy, achieving a minimum mean distance error of 20.29 mm. Furthermore, the model exhibited advantages in terms of feature interpretability and noise robustness. These results provide a promising and interpretable framework, laying the foundation for single-sensor damage localization in structural health monitoring of wind turbine blades.
This article introduces a novel physics-informed large-margin learning framework for diagnosing rotating component faults, addressing the limitations of traditional methods and “black-box” artificial intelligence models in handling complex, noisy vibration data. The methodology integrates feature mode decomposition and short-time Fourier transform to generate physics-informed time–frequency maps, from which interpretable features are extracted. A large-margin learning model then utilizes these features, with its five critical hyperparameters adaptively optimized using the slime mould algorithm to ensure robustness and optimal performance. Evaluated on the Case Western Reserve University (CWRU) and intelligent maintenance systems (IMS) bearing datasets, the proposed framework achieves a diagnostic accuracy of 88.75% on CWRU and 96.67% on IMS, outperforming seven state-of-the-art classification baselines implemented under identical experimental conditions. This approach enhances interpretability, diagnostic accuracy, and generalization, offering a more reliable solution for machine health monitoring.
The composite fault data of aero-engine gas paths generally exhibit significant multiple correlations and couplings while confronting practical challenges, including sample scarcity, imbalanced distribution, and intense noise interference. Existing methods often fail to fully capture the deep-level feature information of raw data, leading to generated data that not only deviates from the raw data but also retains the inherent coupling characteristics, thereby limiting the accuracy of fault diagnosis models trained on such data. To address these issues, a dual-stream Wasserstein transformer generative adversarial network model is proposed for composite fault data augmentation. This model effectively extracts local detailed features and global distribution patterns from raw data through an elaborately designed dual-stream architecture and realizes adaptive fusion of these two feature components via a dedicated gating mechanism. Furthermore, the Wasserstein distance with gradient penalty is introduced into the adversarial training process, which significantly enhances the stability and convergence of model training. Experimental results demonstrate that the data generated by the proposed method are highly consistent with the raw data in terms of distribution characteristics while achieving effective decoupling of composite fault features, thereby significantly enhancing the performance of fault diagnosis models trained on the generated data. In particular, the area under the curve (AUC) of the convolutional neural network (CNN) model reaches 99.07%. Moreover, under conditions of an extremely small sample size, severely imbalanced data distribution, and extremely high noise intensity, the AUC of the CNN model decreases by merely 0.92%, 1.15%, and 0.42%, respectively.
The fault characteristics of planetary gearbox under strong noise conditions show the characteristics of multiscale coupling and weak characteristics, which lead to the problems of insufficient feature extraction and weak antinoise ability in traditional diagnosis methods. In response to this challenge, this article proposes a multiscale Rényi entropy wavelet convolutional network with entropy-guided attention. First, Rényi entropy wavelet convolution is designed. Through different scales of time–frequency feature extraction and feature uncertainty measurement, the feature extraction ability under strong noise is improved. Second, through multiscale decomposition convolution, the features can more accurately locate the distribution characteristics of the fault on the time–frequency plane and enhance the richness of feature expression. Then, the correlation coefficient interaction mechanism is designed, and the correlation coefficient is used to effectively fuse the parallel branch features, so that the information between heterogeneous features can be more comprehensively represented. Finally, the entropy-guided attention mechanism is used to focus on the fault features more accurately. Comparative test verification is carried out on the public gearbox test vibration data set. This method provides an effective solution for intelligent fault diagnosis of planetary gearboxes.
Vibration measurements from civil structures inherently contain fundamental dynamic characteristics and are widely used to monitor structural behavior to inform condition assessment. Machine learning, particularly unsupervised deep learning models, has demonstrated strong potential in automatically learning patterns from baseline vibration data and detecting anomalous structural behaviors. Complex multiscale temporal patterns may arise in structural responses, particularly under rapidly varying external loads, potentially leading to obscured anomalies and reduced anomaly detection accuracy. The primary contribution of the study is the establishment of a benchmark dataset from vibration tests of a laboratory-scale truss bridge. The benchmark vibration dataset contains structural responses under stationary and moving-load excitations, including multilocation and multilevel structural anomaly scenarios. The established dataset can be further used as a benchmark for moving load identification, modal identification, structural model updating, and structural anomaly detection, among others. Additionally, a multiscale pattern learning approach for vibration-based structural anomaly detection by integrating the TimeMixer network with the one-class support vector machine (TimeMixer-OCSVM) is introduced. The methodological contribution lies in establishing a TimeMixer-OCSVM validation baseline, which demonstrates the capability of the benchmark dataset to support anomaly detection studies. The TimeMixer model is adopted to learn multiscale patterns from structural vibration data, improving the anomaly detection performance under varying external excitations. The constructed benchmark dataset is used to validate the proposed method, and the results demonstrate the effectiveness of the proposed method in identifying structural anomalies. Compared to the other typical methods, the proposed method achieves higher structural anomaly detection accuracy.
