TL;DR
This paper introduces a novel, low-power, wearable sensor system using inverse piezoresistive nanocomposites to accurately identify human sitting postures. By combining a flexible strain gauge with a BP neural network, the system achieves over 98% accuracy in distinguishing between normal, slightly hunched, and severely hunched postures, offering a promising solution for health monitoring.
What Problem Does This Paper Address?
Poor sitting posture is a significant contributor to various health issues, including spinal pain, lumbar degeneration, and even myopia. With modern society increasingly reliant on sedentary work, identifying and correcting poor posture is critical. Existing methods, such as computer vision, are often sensitive to environmental conditions like lighting and are restricted to specific environments. While other sensor-based approaches exist, they can be expensive or lack portability, hindering their widespread adoption for continuous, real-world monitoring. This paper aims to bridge this gap by developing an inexpensive, mobile, and robust system for accurate sitting posture identification.
Methods
The core of this system is an inverse piezoresistive nanocomposite sensor. Unlike traditional piezoresistive sensors where resistance increases with strain, these sensors exhibit a decrease in resistance with increasing strain. This property, combined with a high resting resistance, leads to low power consumption, especially during periods of minimal strain. The sensor itself is composed of a silicone matrix embedded with nickel nanostrands and nickel-coated carbon fibers, optimized through a Design of Experiments (DOE) approach to maximize gauge factor and minimize critical strain.
For data processing, a three-layer Backpropagation (BP) neural network is employed. The raw sensor data undergoes preprocessing, including a moving average filter to reduce noise from body jitter and system measurement. Feature extraction involves calculating the mean, standard deviation, maximum, and minimum of the filtered signal over short time windows. These four features form a vector that is then fed into the BP neural network for classification. The network architecture was empirically optimized, settling on a configuration with an input layer, one hidden layer with four neurons, and an output layer, trained using the Levenberg-Marquardt algorithm.
Datasets
The study utilized data collected from 35 volunteer subjects (17 female, 18 male) with varying demographics. Each subject performed three distinct sitting postures: normal, slight hunchback, and severe hunchback, each maintained for 10 seconds. This resulted in a dataset of 5250 four-dimensional feature vectors. For training and testing the BP neural network, 80% of this data was randomly selected for the training set, and the remaining 20% was used as the test set.
Key Insights and Ideas
- Inverse Piezoresistivity for Low Power: The use of inverse piezoresistive sensors is a key innovation. The significant drop in resistance with strain, coupled with high resting resistance, drastically reduces power consumption, making it ideal for wearable, long-term monitoring applications.
- Nanocomposite Optimization is Crucial: The paper highlights the importance of optimizing the composition of the nanocomposite sensor (nickel nanostrands and nickel-coated carbon fibers) to achieve the desired electro-mechanical response and ensure durability within the expected strain ranges.
- Feature Combination for Robustness: While individual features (mean, std dev, max, min) showed some ability to differentiate postures, combining them into a feature vector significantly improved classification accuracy, particularly in distinguishing between slight and severe hunchbacks where overlap existed with single features.
- BP Neural Network Effectiveness: The BP neural network, with its relatively simple structure and straightforward implementation, proved to be an effective classifier for this complex, non-linear sensor data.
Key Results
The developed system achieved a high overall accuracy of 98.76% in identifying human sitting postures. Specifically, the accuracy for normal posture was 100%, with precision and sensitivity for slight and severe hunchbacks exceeding 98%. The confusion matrix revealed that most misclassifications occurred between slight and severe hunchback postures, with only a small number of samples being confused. The system demonstrated a low power consumption profile due to the inverse piezoresistive nature of the sensors.
Limitations and Future Work
The authors acknowledge that the current implementation has only been evaluated for sitting posture identification. Future work could extend this to other postures like sleeping or running. The device's current design is noted as having some bulk and wiring, and future research aims to reduce its size and integrate it with smart terminals for real-time monitoring and alarms. While the training function was selected based on performance, further optimization of the learning rate and the overall neural network architecture could potentially yield even better results. The authors also suggest that this technology holds promise for broader applications in physiological parameter monitoring, wearable devices, and human-machine interfaces.
Summary updated