Structural health monitoring (SHM) is critical for ensuring the integrity and safety of composite structures subjected to dynamic loading conditions. Traditional damage detection methods are often limited in real-time applications, necessitating advanced data-driven approaches. This study investigates the use of Long Short-Term Memory (LSTM) networks for damage detection in composite panels under varying boundary conditions (fixed and pinned) and applied weights to assess their impact on vibrational characteristics. Time-series vibration data, including velocity, acceleration, and displacement, were collected under controlled conditions using a shaker-induced excitation. The LSTM model was trained to classify panels as intact or damaged, leveraging its ability to capture long-term dependencies in sequential data. The results show that the LSTM model effectively distinguishes between damaged and undamaged states, achieving high classification accuracy while adapting to variations in loading and boundary constraints. The model performed better under fixed boundary conditions, suggesting that structural stiffness enhances classification effectiveness by producing clearer vibrational patterns. To further evaluate performance, Precision-Recall curves and confusion matrices were analyzed, demonstrating the model’s ability to balance false positives and false negatives. Additionally, a robustness analysis was conducted by introducing Gaussian noise (0.1, 0.2, and 0.3) into the dataset, confirming that the model maintains stable accuracy despite increasing noise levels, making it reliable for real-world applications. These findings highlight the potential of LSTM-based SHM systems for real-time damage detection in composite structures, with promising applications in aerospace, marine, and civil engineering.