Tiny Machine Learning Sensor Platform for Local Sensor Data Fusion and Evaluation

Abstract:
Cunently prevalent methods for analyzing sensor data rely on centralized processing. The transmission of collected sensor data leads to increased latency, and applications could additionally suffer from bandwidth constraints and privacy concerns. On-device data analysis is crucial for many future edge computing applications. Embedded machine learning is an increasingly important approach to enable local sensor data analysis on resource constrained edge devices. This paper demonstrates a concrete implementation of the embedded machine learning life cycle on the developed Tiny Machine Learning Sensor Platform to enable local sensor data fusion and evaluation for generic applications. The featured Tiny Machine Learning Sensor Platform exhibits cormnon microcontroller interfaces to attach generic sensors suitable for domain-specific applications in a wide range of use cases. It is intended for time series domain specific task such as anomaly detection, classification and prediction. The platform runs machine learning models and neural networks that are specifically optimized for resource constrained embedded microcontrollers. The advantage of on-device data analysis is that the sensor platform itself can decide locally how to proceed when the deployed machine learning model detects certain events. Data only needs to be transmitted to a centralized service when an event has occuned. The device does not need to constantly send its monitored data when there is little or no new information reducing the amount of data transmitted significantly and allowing extended run times of battery-powered systems. The platfonn can react autonomously to specific events without waiting for a server response, reducing the latency for real-time, safety critical applications. This paper describes the applied hardware, software, and the data science pipeline used to develop machine learning enhanced applications on embedded hardware. A first version of the Tiny Machine Learning Sensor Platform is already deployed in several applications. It is planned to adapt the Tiny Machine Learning Sensor Platform to additional use cases and further develop identified challenges. In summary, the practicability of tiny machine learning in real world time series data applications is demonstrated.