TL;DR
This paper introduces a novel self-powered microsystem capable of continuous vibration sensing and target recognition for unattended applications. By integrating a composite energy source, optimized low-power modules, and a lightweight SVM algorithm, it achieves long-term operation with minimal maintenance, paving the way for more pervasive surveillance networks.
What Problem Does This Paper Address?
The Internet of Things (IoT) is increasingly relying on unattended microsystems for tasks like target detection and recognition. However, a significant challenge is their power supply. Traditional battery-powered microsystems have limited lifespans, necessitating frequent maintenance. While self-powering solutions exist, they often compromise on size, power consumption, sensing capabilities, or recognition accuracy. Furthermore, many existing systems require data to be sent to a central server for processing, leading to high latency and communication burdens. This paper tackles the need for a compact, long-lasting, and autonomous microsystem that can perform both sensing and recognition locally, without external power sources, for applications like covert, large-scale, and long-term monitoring.
Methods
The core of this work is a microsystem designed for low-power consumption and self-powering. Key methodological aspects include:
- Composite Energy Source: A combination of a lithium battery and photovoltaic panels provides a self-sustaining power supply.
- Optimized Hardware Modules: The microsystem integrates a low-power MCU (STM32WB55), a MEMS tri-axis accelerometer (LSM6DSM) for vibration sensing, a MEMS vibration trigger switch for event wake-up, and an RF antenna. These components are compactly integrated onto a single PCB.
- Event Wake-up Mechanism: A vibration trigger switch filters out non-target vibrations, activating the system only when significant events occur, thus conserving power.
- Low-Power Software Workflow: The system operates in two states: a low-power sleep mode and a high-power active mode. It uses an event-driven approach, waking up for target detection and then entering active mode for processing.
- Lightweight Recognition Algorithm: A Support Vector Machine (SVM) algorithm, specifically a one-versus-rest noise-target classifier followed by triple one-to-one target classifiers, is employed for target recognition. This algorithm is chosen for its low computational complexity, making it suitable for the microsystem's limited processing power.
- Feature Extraction: Five vibration signal features are extracted: sum of acceleration amplitudes (asum), peak frequency (fpeak), zero-crossing number (Nz), maximum adjacent zero-crossing time interval (Δtmax), and first-to-last zero-crossing time interval (Δt₁). These features are selected for their low computational complexity and ability to distinguish between different target vibrations.
- Preprocessing: Mean filtering and autocovariance methods are used to enhance the signal-to-noise ratio of the raw vibration data.
Datasets
The paper describes data acquisition and training using vibration data collected from three target types: pedestrians, wheeled vehicles, and tracked vehicles.
- Data Acquisition: Vibration data was collected on a 150m long, 5m wide flat soil pavement. Five high-accuracy vibration sensors were placed at varying distances from the pavement edge. Targets moved at specific speeds (pedestrian: 4-8 km/h, wheeled vehicle: 10-30 km/h, tracked vehicle: 3-11 km/h).
- Training Data: For SVM training, 60 groups of tracked vehicle data, 50 groups of wheeled vehicle data, 60 groups of pedestrian data, and 50 groups of environmental signals were used. This data was split into a 7:3 training-to-test set ratio.
- Benchmark: While no explicitly new benchmark dataset is introduced, the paper uses collected real-world vibration data to train and evaluate its microsystem, simulating both indoor and outdoor scenarios.
Key Insights and Ideas
- Synergistic Design is Crucial: The paper emphasizes that optimizing hardware modules and software algorithms as a whole is key to achieving the desired characteristics of small size, low-power consumption, and self-supply.
- Vibration as a Rich Signal: Different targets produce distinct vibration signatures, making vibration sensing a viable and effective method for target recognition, even over long distances.
- Trade-offs in Recognition: Achieving high recognition accuracy often comes at the cost of increased computational power and energy consumption. The paper demonstrates a successful compromise by using a lightweight SVM and accepting slightly lower accuracy in challenging outdoor conditions.
- Self-Powering Enables Long-Term Autonomy: The composite energy harvesting and storage system, combined with the low-power sleep mode, allows the microsystem to operate for extended periods (24 hours of sleep mode with just 11 minutes of charging), making it truly "long-life."
- Event Wake-up is Key to Efficiency: The vibration trigger switch effectively filters out noise and unnecessary activations, significantly reducing power consumption by only waking the system for relevant events.
Key Results
- Microsystem Size and Weight: The implemented microsystem is compact, measuring 2.5 cm x 2.5 cm x 2.4 cm, with a volume of 15 cm³ and a weight of 35 g.
- Power Consumption:
- Sleep Mode: 0.18 mW
- Active Mode: 36 mW
- Self-Powering Performance: Under midday sunlight, the microsystem achieved self-powering of up to 22.7 mW. This is approximately 126 times the power consumption in sleep mode, enabling 24 hours of operation in sleep mode with just 11 minutes of charging.
- Recognition Accuracy:
- Indoor: Higher than 84% for all three targets (tracked vehicle: 92.4%, wheeled vehicle: 84.0%, pedestrian: 86.7%).
- Outdoor: Higher than 65% for all three targets (tracked vehicle: 90.0%, wheeled vehicle: 65.0%, pedestrian: 65.0%).
Limitations and Future Work
- Outdoor Recognition Accuracy: The paper acknowledges that outdoor recognition accuracy is lower than indoor accuracy. This is attributed to:
- Poor coupling of vibrations to the accelerometer due to the soil surface and contact interface.
- Larger differences in signal quality for pedestrians and wheeled vehicles compared to tracked vehicles.
- Future Work Suggestions:
- Increasing Weight/Burying: While effective for improving vibration coupling, this contradicts the microsystem's advantages of small size and easy deployment.
- More Training Samples: Collecting more data can improve recognition algorithms but increases memory and processing demands.
- Feature Selection: Identifying more efficient features for recognition can optimize algorithms. The paper suggests that
fpeakis crucial for distinguishing pedestrians from wheeled vehicles, andasumfor distinguishing tracked vehicles from others. Future work can focus on optimizing feature selection. - Further Performance Improvements: The authors aim to further improve system performance in terms of smaller size, lower power consumption, more robust self-power supply, and better recognition capabilities.
Summary updated