Abstract
This study presents a hybrid framework that integrates the finite element method (FEM) with a convolutional neural network‐long short‐term memory model (CNN‐LSTM) for faster prediction of maximum strain and fault‐deviation assessment in wind turbine gears under seasonal dynamic loading. FEM was first used to evaluate maximum principal strain, a failure‐sensitive indicator of material deformation, for healthy, cracked and pitted gear conditions under winter, spring, summer and fall torque profiles. The FEM results showed that maximum strain values ranged from 0.0062 to 0.0121 for healthy gears, 0.0074–0.0124 for cracked gears, and 0.0146–0.0236 for pitted gears, with summer producing the highest strain response due to increased loading. Compared with healthy conditions, cracks increased the average peak strain by 9.30%, while pitting produced a much larger increase of 103.49%, confirming the stronger influence of surface defects located near high contact‐stress regions. To reduce the computational cost of repeated FEM simulations, four recurrent/deep‐learning models, namely recurrent neural network (RNN), gated recurrent unit (GRU), LSTM and CNN‐LSTM, were developed and compared for predicting healthy maximum strain. The models were evaluated using mean absolute deviation (MAD), root mean square error (RMSE), mean absolute percentage error (MAPE), training time and prediction time. Among the investigated models, CNN‐LSTM achieved the best overall performance, providing the lowest average prediction errors while requiring only 36.38 s for combined training and prediction. The CNN‐LSTM model also achieved strong peak‐strain prediction accuracy, reaching 99.9% in winter and 96%–99% across the other seasons for the healthy condition. The predicted healthy strain response was then used as a baseline for comparison with FEM‐simulated responses under cracked and pitted conditions. The deviation analysis showed that crack‐ and pit‐induced strain changes could be distinguished from the healthy baseline, with pitting producing the largest deviation. The proposed FEM‐CNN‐LSTM framework provides a computationally efficient approach for estimating healthy gear strain responses and assessing fault‐induced deviations in wind turbine gears.