Abstract
With the rapid advancement of industrial intelligence, ensuring occupational safety has become an increasingly critical concern. Among the essential personal protective equipment (PPE), safety helmets play a vital role in preventing head injuries. There is a growing demand for real-time detection of helmet chinstrap wearing status during industrial operations. However, existing detection methods often encounter limitations such as user discomfort or potential privacy invasion. To overcome these challenges, this study proposes a non-intrusive approach for detecting the wearing state of helmet chinstraps, inspired by the mechanosensory hair arrays found on spider legs. The proposed method utilizes multiple MEMS inertial sensors to emulate the sensory functionality of spider leg hairs, thereby enabling efficient acquisition and analysis of helmet wearing states. Unlike conventional vibration-based detection techniques, posture signals reflect spatial structural characteristics; however, their integration from multiple sensors introduces increased signal complexity and background noise. To address this issue, an improved adaptive convolutional neural network (ICNN) integrated with a long short-term memory (LSTM) network is employed to classify the tightness levels of the helmet chinstrap using both single-sensor and multi-sensor data. Experimental validation was conducted based on data collected from 20 participants performing wall-climbing robot operation tasks. The results demonstrate that the proposed method achieves a high recognition accuracy of 96%. This research offers a practical, privacy-preserving, and highly effective solution for helmet-wearing status monitoring in industrial environments.