A Self-Attention Enhanced Deep CNN-LSTM-Based Irregular Surface Recognition Approach for Integration Into Lower Limb Prosthesis Systems to Ensure Safety Through Predictive Walking

Advancements in instrumentation and control systems for lower limb prostheses have substantially improved mobility for amputees. However, significant challenges persist when users encounter irregular surfaces, as most prosthetic systems lack the capability to dynamically adapt to surface variations. This limitation restricts mobility, compromises safety, and diminishes user confidence and security during walking. To address these challenges, integrating inertial measurement units (IMUs) with artificial intelligence (AI) techniques, particularly deep learning (DL) methods, has emerged as a promising solution for surface classification and safety enhancement. This study proposes a self-attention enhanced deep CNN-LSTM model to automatically classify walking surfaces as regular or irregular, utilizing IMU acceleration data collected from prosthetic limbs. The model employs the strengths of convolutional and recurrent neural networks combined with a self-attention mechanism to enhance feature representation and improve classification accuracy. Experimental evaluations reveal that the proposed method achieves a classification accuracy of 99.94%, outperforming existing approaches. This result underscores the model’s potential to serve as the basis for AI-driven IMU-based systems, enabling real-time surface recognition and safety alerts in prosthetic devices. By enhancing walking safety and user confidence, this method represents a significant advancement for lower limb prosthesis systems.