Online noise-adaptive Kalman filter integrated novel autoencoder for multi-fault detection and early warning of wind turbines

Abstract

For wind turbines (WTs), It is essential to implement proactive maintenance strategies that predict and minimise potential failures, thereby ensuring the reliable operation of wind turbines. Supervisory Control and Data Acquisition System (SCADA) data and intricate spatio-temporal dynamics impede the timely and precise diagnosis of various fault anomalies. There is a need in the literature to effectively incorporate online learning mechanisms or to work on studies that dynamically adapt to sensor noise and uncertainty during real-time operation. To address this gap, this study introduces an Online Noise-Adaptive Kalman Filter (ONAKF)-based hybrid Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Attention Layer (AL) autoencoder (AE) model designed for fault detection utilising SCADA data, facilitating multi-fault early warning capabilities. The proposed model demonstrates superior performance compared to existing methods, achieving across five distinct fault types, with R2 values ranging from 0.9333 to 1.0000. Across all five fault models, the Mean Absolute Error (MAE) ranges from 1.41 × 10^–5 to 1.23 × 10^–3, the Mean Squared Error (MSE) ranges from 8.84 × 10^–10 to 2.17 × 10^–3, and the Root Mean Squared Error (RMSE) ranges from 2.97 × 10^–5 to 4.66 × 10^–2. Additionally, the model demonstrates notable detection efficiency, attaining precision (1.0), recall (1.0), and F1 scores (1.0) across all five fault categories. The model provides initial alerts 10 to 167 h prior to the occurrence of five specific issues. The findings demonstrate the model’s efficacy in enhancing maintenance schedules and monitoring the conditions of wind turbines.