A Novel Condition-Monitoring System Based on a Parametric Correlation Learning Network for Wind Turbine Blade Fault Identification

Abstract

The role of wind turbine blades in determining the reliability and performance of wind energy systems is critical. However, environmental and operational stressors, as well as wind shear, can damage blades. The safe and effective operation, free of failures and with reduced blade maintenance costs, requires the reliable assessment and detection of blade damage. The existing damage detection methodologies, however, fail to recognize operational turbine acoustic emissions due to their complexity, uncharacteristic and nonlinear behaviour, particularly under noisy, varying operational conditions. An attention-enhanced temporal convolutional network (AMTCN) and a bidirectional gated recurrent unit (AMTCN-BiGRU) are proposed in this study for analysing acoustic time-series data. Unlike conventional hybrid TCN-RNN architectures, the proposed framework is designed to hierarchically learn fault-sensitive acoustic correlations by combining attention-guided multiscale temporal feature extraction with bidirectional sequential dependency modelling for improved blade fault discrimination. Improving feature representation and more effectively capturing temporal dependencies is the objective of the integrated attention, temporal convolution and bidirectional sequence processing. This formulation fits the damage identification scenario with limited, noisy datasets, where acoustic signals exhibit substantial variation. The proposed methodology has been benchmarked against other models such as TCN, TCN-LSTM and AMTCN-GRU. The results showed that the AMTCN-BiGRU model outperformed the other models, achieving an RMSE of 0.1369, an MAE of 0.059 and an R² of 0.990 with 97.3% accuracy. The findings show the effectiveness of the proposed method for detecting damage at early stages and for structural health monitoring of wind turbine blades.