Impact force identification using memory-based neural networks: Application to cooling tower fan blades

Accurate impact force identification is vital for effective Structural Health Monitoring (SHM), yet direct measurement is often impractical. Inverse methods offer an alternative but are typically ill-posed. This study introduces a deep learning framework for simultaneous identification of impact location and force time–history on a fiberglass-reinforced plastic (FRP) axial fan blade, a complex target due to its irregular geometry and sparse sensor layout. Four neural network architectures were evaluated: Multi-Layer Perceptron (MLP), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Bidirectional LSTM (Bi-LSTM). While MLP served as a baseline, GRU and LSTM captured temporal dependencies, and Bi-LSTM further improved performance by modeling bidirectional context. Bi-LSTM proved especially robust under limited training data and asynchronous sensor inputs, achieving a reconstruction error as low as 3.48% and a correlation error of 0.01% at one location. In contrast, MLP failed with a reconstruction error of 140.12%, while GRU and LSTM showed moderate improvements. Across multiple impact sites, Bi-LSTM consistently outperformed others, particularly near the blade’s symmetry axis and close to sensors, with reconstruction errors mostly below 1%. Poorer performance at edge locations was linked to torsional vibrations and signal damping, underscoring the importance of sensor placement and structural geometry. By addressing both impact location and force reconstruction in a data-constrained, non-collocated sensor environment, this work demonstrates the practical potential of deep learning, especially Bi-LSTM, for real-world SHM applications.