Design of a novel adaptive residual fusion technique for SCADA-based wind turbine anomaly detection

Wind turbines are critical for sustainable energy, yet operational failures cause costly disruptions and maintenance challenges. Traditional fault diagnosis methods often lack adaptability and real-time performance. This paper proposes a hybrid fault diagnosis framework integrating wavelet-based feature extraction, sparse auto-encoders (SAE), and a novel Adaptive Residual Fusion Block (ARFB) to detect anomalies in supervisory control and data acquisition (SCADA) wind turbine operations. Unlike static ensemble or fusion methods, the ARFB dynamically learns adaptive feature weights and corrects residual errors, enhancing noise robustness and fault detection across diverse conditions. The model processes SCADA sensor data to achieve accurate, noise-robust anomaly detection. Empirical evaluation on the Turkey SCADA 2018 Wind Turbine Dataset shows the proposed model achieves 95% accuracy, 92% precision, and a 6% false-positive rate, outperforming baseline auto-encoders by 9% in F1-score and reducing false positives by 50% compared to traditional vibration analysis. These results demonstrate superior adaptability and robustness, making the framework suitable for real-world predictive maintenance.