Jul 2026· PHM Society European Conference· Vol 9· 0 citations· 37 references
Abstract
Efficient health monitoring is indispensable for the reliable operation of wind turbines. Damage to wind turbine blades, such as cracks and holes, typically generates whistle-like sounds during rotation. This study proposes a two-stage edge-server collaborative system for detecting blade damage using acoustic signals captured by arrays built from commodity microphones. The first stage employs a lightweight attention-based convolutional neural network to run on edge devices for the real-time binary classification to determine whether anomalous sounds are present. Suspicious time segments are stored for further analysis. The second stage uses a time-frequency sound event detection model that employs a detection transformer with an audio spectrogram transformer backbone to identify the time and frequency ranges of sound events via bounding boxes in the spectrograms. Owing to its high computational demand, this in-depth analysis is performed on a server. To validate the proposed system, acoustic signals were recorded intermittently for more than a year using micro-electromechanical system (MEMS) microphones externally attached to wind turbine towers. The models were trained and evaluated on a manually annotated dataset comprising 4,210 audio clips (15 s each) containing 14,420 sound events. The experimental results demonstrated that the binary classification model achieved an area under the receiver operating characteristic curve (AUC) of 0.920, whereas the sound event detection model attained an average precision at a 50% intersection-over-union threshold (AP50) of 0.510. Furthermore, evaluations on test data under unseen conditions, comprising 496 clips with 135 sound events recorded by handheld recorders at different locations, yielded an AUC of 0.867 and an AP50 of 0.440. The results highlight the robustness of the proposed system to variations in microphone types, recording locations, and environmental noise, demonstrating its strong potential for practical continuous automatic damage detection in wind power infrastructure.
Accurate identification of operating conditions in rotating machinery is essential for performance monitoring and early detection of abnormal behavior. Rotating systems generate both structural vibration and airborne acoustic emissions that reflect changes in speed and mechanical imbalance.
In this study, synchronized vibration and acoustic measurements were collected from a laboratory table fan using one triaxial accelerometer and two microphones. The fan was evaluated under four operating conditions: low speed, high speed, low speed with added mass on one blade, and high speed
with added mass on one blade. Time-domain signals were segmented into overlapping 2-second windows and transformed into the frequency domain using Fast Fourier Transform (FFT). Short-Time Fourier Transform (STFT) was applied to generate spectrograms representing the time-frequency structure
of both vibration and acoustic signals. A supervised Convolutional Neural Network (CNN) was trained on these multi-channel spectrograms to classify the operating condition. The dataset was divided into training, validation, and testing subsets to evaluate generalization performance. The trained
model successfully distinguished the four operating states, achieving high classification accuracy across multiple sensor configurations, including single-sensor and fused measurements. Multi-sensor configurations incorporating acoustic data demonstrated faster convergence while maintaining
perfect classification performance.
Lucky Islam Pataky, Feruza A. Amirkulova, Sabrina Hsu et al.· INTER-NOISE and NOISE-CON Co...· 0 citations
Pipeline vandalism and crude theft remain the primary threats to the resilience of Nigeria's energy infrastructure and environment. Conventional pressure monitoring systems often generate high rates of false alarms or detect breaches only after containment is lost. This paper evaluates a lightweight, hybrid software architecture combining Distributed Acoustic Sensing (DAS) data with hydraulic pressure transient analysis to identify intrusion attempts before a spill occurs. Due to the classified nature of empirical breach data, the methodology involved generating a high-fidelity synthetic training dataset modeling manual digging, mechanical drilling, and standard operational background noise. A Multi-Input Convolutional Neural Network (CNN) was developed to perform feature-level sensor fusion. The architecture transformed 1D acoustic time-series data into 2D Mel-spectrograms, fusing these spatial features with 1D temporal pressure wave data. The hybrid model achieved a 92% accuracy rate in distinguishing between theft attempts and benign operational vibrations, significantly outperforming standalone pressure monitoring systems. Crucially, the system demonstrated a geolocation accuracy of within 20 meters, enabling rapid security response. The study concludes that advanced, computationally efficient sensor fusion democratizes digital transformation for marginal field operators, successfully shifting the focus from reactive leak detection to proactive, cost-effective interference prevention.
F. M. Kelechi, A. Aribisala, M. Evwerhamre· SPE Nigeria Annual Internati...· 0 citations
In 2025, wind energy was the primary source of renewable electricity in the EU. As wind energy deployment continues, automatic inspection of wind turbines, especially rotor blades, has gained increasing interest for the economically competitive wind energy market. Therefore, this study introduces a three-stage framework for combined damage detection and localization in wind turbine rotor blades based on acoustic event detection. Bridging the gap between established SHM techniques analyzing either low-frequency vibrations or ultrasonic structure-borne sound waves, our methodology is based on airborne sound in the audible frequeny range. First, we apply an adaptive thresholding technique to extract short-term percussive sound signals, which are subsequently localized using Time Differences of Arrival. Finally, a spatio-temporal accumulation of the emitted acoustic energy is introduced as a damage indicator. Our evaluations on two large-scale rotor blade fatigue tests demonstrate that a configuration of only two microphones is sufficient to detect and coarsely localize structurally relevant damage within a 12-m-long blade segment. To the best of the authors' knowledge, such a favorable trade-off between sensor spacing and localization accuracy remains unprecedented among established structure-borne SHM techniques.
Alexander Lange, Tomas Verardo, Reemt Hinrichs et al.· e-Journal of Nondestructive...· 0 citations
The detection of water pipe leaks using machine learning to identify ground-based acoustic signals has long been a hot topic in the field of water supply safety. However, achieving reliable leak detection in complex real scenarios remains a significant challenge. Based on an acoustic signal dataset collected from an actual pipeline network, this study proposes a sliding window traversal method with intra-window normalization (SWT-IWN) for data preprocessing, which effectively enhances the separability of acoustic features between leakage and non-leakage pipe segments. On this basis, a dataset fused with spatial position information of sampling points is constructed. Furthermore, a leakage identification model named CST-Net is designed, which uses a convolutional neural network (CNN) to extract time-frequency features and a Swin-Transformer to model the positional sequence correlation of sampling points, thereby realizing collaborative representation of the two feature types. Experimental results show that CST-Net accurately identifies leakage pipe segments of 2.5 m in length, with an identification accuracy of 92.35%. The model also demonstrates strong identification performance and stability under diverse sampling conditions, well meeting the requirements of practical engineering applications.
Yonggang Shen, Haoli Wang, Jianxin Shen et al.· Journal of Hydroinformatics· 0 citations
To see and comprehend their surroundings, autonomous and sophisticated advance driver assistance systems require a very large count of cameras and LiDAR sensors in the automotive systems. However, due to some of situations like bad weather, poor lighting, and obstructions that can interfere with its dependability, leading to safety risks and perceptual malfunctions which lead to a risky and deadly outcome. Emergency sirens and other acoustic cues usually precede visual signals and continue to be perceived even in the event that these vision sensors fail or malfunction. Despite this benefit, due to the energy usage of the battery-operated automotives and their computational limitations, audio-based perception is rarely used. The hybrid Artificial Neural Network–Spiking Neural Network (ANN–SNN) framework developed in this study enables the identification of emergency acoustic incidents in cars in real time environments. Utilizing spectral traits extracted from actual audio data, a lightweight ANN is trained. Poisson, temporal, and delta encoding techniques are then used to transform these extracted spectral characteristics into spike representations. In order to provide rapid and energy-efficient inference, leaky integrate-and-fire neurons are used to process the spiking signals. The suggested system greatly decreases latency and the energy consumption while achieving accuracy on par with ANN baselines, according to experimental results. A browser-based real-time implementation further supports the practicality of this suggested approach for in-vehicle deployment.
Kiran Yadav V., V. S, K. E.· International Conference Com...· 0 citations
Wind turbines (WT) need to perform well to meet the ever-growing demand of green energy. However, the performance of a wind turbine can be jeopardized due to occurrence of damage within its components. Acoustic emission (AE) is highly sensitive to occurrence of damage and can be used for damage identification. However, the localization of an AE source is challenging using a time domain method for structures that involve material anisotropy and complex geometry (e.g., curved surfaces and variable thickness). With this view, this work presents an unsupervised framework that operates on narrow frequency bands (NFBs) for artificial AE source determination in composite wind turbine blades. Unlike the conventional time-domain methods that require a network of sensors and the time of arrival information, the new method requires only one sensor and the information contained in the reduced frequency bands within each AE signal to discern its source. The unsupervised framework is found to be highly efficient in finding pattern in the multi-dimensional frequency domain dataset extracted from AE signals and can easily cluster the AE signals as per their zone of occurrence in a WT blade. This method can be easily implemented in real and complex structures as it requires only one sensor. Due to the very high degree of accuracy, the method can be applied to real complex systems and structures where time domain methods are not feasible. The method can also be included in a digital twin model for accurate prediction of an AE source.
Deba Datta Mandal, S. Sikdar, Rakesh Mishra· e-Journal of Nondestructive...· 0 citations