The Internet of Things (IoT) and 5G networks promise ultrafast data transmission, enabling numerous interconnected electronic devices in smart buildings. However, camera-based surveillance raises privacy concerns. Alternative approaches like laser beam scanning have limitations in cost, power consumption, and information acquisition. This research proposes using floor-embedded sensors as a frequently interactive interface to gather rich sensory data about human movement, including position, activity, and identity. This data is valuable for applications like fall detection in elderly care, home automation, and security monitoring. Existing floor sensors often use resistive, capacitive, piezoelectric, or triboelectric mechanisms. Piezoelectric and triboelectric sensors offer advantages in reduced power consumption and potential self-sustainability. However, previous floor sensors lacked scalability and were costly to implement at a large scale. The challenges include increasing the number of sensing pixels and electrodes for large areas, leading to complex electrode layouts and signal processing. This research aims to address these challenges by combining the low-cost triboelectric sensing mechanism with large-scale printing techniques, along with advanced artificial intelligence (AI) for data analysis to extract comprehensive sensory information including identity information.

Triboelectric sensors, based on contact electrification and electrostatic induction, offer advantages in simple configuration, manufacturing compatibility, scalability, and low cost. Printing techniques (roll-to-roll, inkjet, screen printing) are suitable for large-scale device fabrication. Previous research explored using four edge electrodes and output ratios to determine contact position, but this approach is unsuitable for large areas due to small induced outputs. Connecting electrodes with distinct patterns in parallel has been proposed to reduce electrode numbers, but large-area applications like floor sensing were not demonstrated. Most sensor functionality relies on time-domain data analysis (signal magnitude, frequency), potentially losing important features like identity information. AI and machine learning (ML) can enhance monitoring systems by extracting complete sensory data for personalized authentication and object/intention identification, improving position monitoring, home automation, and healthcare.

This study developed deep learning-enabled smart mats (DLES-mats) using the triboelectric mechanism for an intelligent, low-cost, scalable floor monitoring system. The system integrates a minimal-electrode-output triboelectric floor mat array with deep learning (DL)-based data analytics. DLES-mats were fabricated using screen printing, offering cost-effectiveness, scalability, and self-sustainability. Each mat has a unique electrode pattern with varying coverage rates, mimicking QR codes for identification. Parallel connection in an interval scheme ensures minimal two-electrode outputs. The mats are distinguished based on the relative magnitudes of output signals, allowing indoor positioning and activity monitoring. Deep learning, using a convolutional neural network (CNN), extracts identity information from walking gait patterns. The minimal two-electrode outputs reduce the computational cost compared to traditional image-based or massive-channel processing, enabling real-time applications. The researchers first characterized the individual DLES-mats, measuring output voltages on 1 MΩ and 100 MΩ external loads with repeated stepping motions in different directions. They analyzed the maximum, minimum, and peak-to-peak voltages to determine the relationship between electrode coverage rate and output magnitude. They then investigated the effect of different users and contact materials on output performance. Next, they evaluated the performance of a parallel connection of 12 DLES-mats (two sets of 0–100%) in a one-dimensional arrangement. Due to overlapping voltage pulses from simultaneous stepping motions, they introduced an interval parallel connection scheme to improve signal stability. A 3x4 DLES-mat array with an interval parallel connection was constructed for position sensing. Walking tests were conducted to demonstrate the system's capability for position and trajectory detection. The system's ability to monitor activity (slow walking, normal walking, fast walking, running, jumping) was also assessed. The output voltage and power characteristics with varying external resistances were measured to evaluate energy harvesting potential. Finally, they integrated deep learning into the system for user recognition. A CNN model was trained using data from 10 users, each with 100 samples (80% for training, 20% for testing). The performance was evaluated using the prediction accuracy. A virtual corridor environment was used to demonstrate real-time position sensing and identity recognition, with the system controlling lights and door access based on the detected position and identified user.

The study successfully demonstrated a scalable, low-cost smart floor monitoring system using triboelectric DLES-mats and deep learning. The key findings include: 1. **Scalable and cost-effective fabrication:** Screen printing enabled the creation of large-area, low-cost DLES-mats with unique electrode patterns for easy identification and parallel connection, reducing system complexity and the number of required electrodes significantly. 2. **Effective signal differentiation:** The unique electrode patterns and an interval parallel connection scheme allowed clear differentiation of signals from individual mats, providing accurate position sensing even in dense arrays. 3. **Accurate position sensing and activity monitoring:** The system accurately determined walking position and trajectory in real-time, and it also successfully differentiated various activity levels such as slow walking, normal walking, running, and jumping based on the output signal magnitudes and frequencies. 4. **High-accuracy user recognition:** The deep learning model achieved a high average recognition accuracy of 96% for distinguishing between 10 different users based on their unique walking gait patterns. This recognition remained robust even when users moved at different speeds (normal walking, fast walking, running). The system could distinguish 12 different classes with an accuracy of 89.17% (different user with different speeds), and an accuracy of 91.47% if the walking speed is ignored. 5. **Energy harvesting capability:** The system demonstrated energy harvesting capability. While the maximum output power of the 3x4 DLES-mat array was 8.57 µW, a single larger mat (40cm x 40cm) was able to provide 169.46 µW, and even a smaller mat (30cm x 12cm) achieved 800.84 µW, suggesting the potential of powering low-power IoT devices through energy harvesting. 6. **Privacy-preserving user recognition:** The system could be adapted for privacy-preserving recognition; instead of identifying specific individuals, it focuses on distinguishing valid users from invalid users. 7. **Real-time demonstration:** The integration of the DLES-mat array with deep learning allowed for a real-time demonstration of position sensing and identity recognition in a virtual corridor environment, simulating practical applications in smart building control and security.

The findings address the research question of developing a scalable, low-cost, and privacy-preserving floor monitoring system. The system successfully integrates triboelectric sensing, large-area printing, and deep learning for high-accuracy position sensing, activity monitoring, and user identification. The use of unique electrode patterns and parallel connection minimizes the number of electrodes, reducing system complexity and cost. The deep learning model efficiently extracts features from walking gait patterns for accurate user recognition, providing a privacy-preserving alternative to camera-based systems. The demonstrated energy harvesting capability offers the potential for self-powered operation. The results are highly relevant to the field of smart buildings and homes, offering a promising approach to improve automation, healthcare, and security applications. Future work could focus on enhancing energy harvesting efficiency, exploring different materials and electrode designs, and implementing more complex deep learning models for more sophisticated activity recognition and user authentication.

This research successfully developed a smart floor monitoring system using triboelectric DLES-mats and deep learning. The system offers a low-cost, scalable, and privacy-preserving solution for indoor positioning, activity monitoring, and user identification. High accuracy was achieved in both position sensing and user recognition. The energy harvesting capabilities demonstrate potential for self-powered operation. Future work could explore improved energy harvesting, advanced materials, and more complex deep learning algorithms for enhanced functionality.

The current system's accuracy is dependent on the consistency of user walking patterns. Variations in walking style, footwear, or carrying items could affect recognition accuracy. The energy harvesting capacity may be limited for high-power applications. The testing was conducted in a controlled environment; the system’s performance in real-world scenarios with more varied conditions and interference needs further evaluation.