TL;DR
This thesis presents a low-power, battery-operated device designed to detect if a cargo package has been dropped during transit. By utilizing an accelerometer and gyroscope, coupled with a machine learning model embedded in a microcontroller, the system aims to ensure the safe delivery of sensitive goods in the growing e-commerce landscape.
What Problem Does This Paper Address?
The rapid growth of e-commerce, accelerated by recent global events, has led to an increase in the transportation of high-value and fragile items. A significant challenge in current logistics is ensuring these sensitive goods are handled safely and arrive undamaged. Traditional fall detection systems often focus on personal safety (e.g., for the elderly), but this paper addresses the need for a solution specifically for cargo, aiming to provide a mechanism to verify the integrity of packages during their journey from sender to receiver.
Methods
The core of the solution is a low-power device featuring a Renesas RL78/G13 microcontroller and a Bosch BMI160 inertial measurement unit (IMU), which combines an accelerometer and gyroscope. The system is designed for minimal power consumption, allowing for extended battery life.
The detection mechanism relies on analyzing data from the accelerometer and gyroscope. Specifically, the Signal Vector Magnitude (SVM) from the accelerometer and the angular velocity from the gyroscope are calculated. These values are then fed into a Logistic Regression machine learning model. This model was trained using collected data and then embedded into the microcontroller's firmware. The device operates in a low-power "sleep mode," waking up only when the IMU detects significant motion or vibration exceeding a predefined threshold. Upon waking, it collects sensor data for a short period and uses the embedded model to determine if a drop has occurred.
Datasets
The paper describes the creation of a custom dataset for training and testing the fall detection model. This dataset consists of recorded accelerometer and gyroscope data from various simulated drop and non-drop events. The dataset is presented in Table 2, detailing sample numbers, filtered peak values of accelerometer and gyroscope, a "Drop or Not" label, and a description of the event. The first 40 samples were used for training, and the remaining for testing.
Key Insights and Ideas
- Low-Power Design is Crucial: The emphasis on a low-power device is paramount for a battery-operated system that needs to function throughout the shipping process without frequent battery replacements.
- IMU for Motion Analysis: Combining accelerometer and gyroscope data provides a robust way to capture the complex motion patterns associated with a package drop.
- Embedded ML for Real-time Decision: Embedding a machine learning model directly onto the microcontroller allows for immediate, on-device decision-making without the need for constant cloud connectivity, further saving power.
- Sleep Mode Optimization: The strategy of keeping the microcontroller in a deep sleep state and only waking upon significant motion is key to achieving the reported low current consumption.
Key Results
- Train and Test Accuracy: The Logistic Regression model achieved a training accuracy of 90.00% and a test accuracy of 88.00%.
- Low Power Consumption:
- In sleep mode, the device consumes approximately 0.23 µA (microamperes).
- In run mode (active detection), the current consumption is around 2.6 milliamperes.
- The accelerometer alone consumes 20 µA in sleep mode.
- Future Deep Sleep Mode: The authors propose a "deep sleep mode" that would further reduce current consumption to 5 µA, potentially quadrupling battery life when no acceleration is detected.
Limitations and Future Work
- Limited Dataset Size: While a dataset was created, the paper implies that more extensive data collection and analysis could improve model robustness.
- Future ML Algorithms: The authors plan to explore alternative machine learning algorithms to further refine the detection accuracy and reduce erroneous drop detections, particularly for events occurring near the decision boundary.
- Enhanced Battery Life: Future work will focus on software development to further extend battery life, including the implementation of the proposed "deep sleep mode."
- Real-world Testing: While the study focuses on a designed device and collected data, extensive real-world testing in various shipping conditions would be beneficial.
Summary updated