Addressing Gap between Training Data and Deployed Environment by On-Device Learning (2203.01077v4)

Abstract: The accuracy of tinyML applications is often affected by various environmental factors, such as noises, location/calibration of sensors, and time-related changes. This article introduces a neural network based on-device learning (ODL) approach to address this issue by retraining in deployed environments. Our approach relies on semi-supervised sequential training of multiple neural networks tailored for low-end edge devices. This article introduces its algorithm and implementation on wireless sensor nodes consisting of a Raspberry Pi Pico and low-power wireless module. Experiments using vibration patterns of rotating machines demonstrate that retraining by ODL improves anomaly detection accuracy compared with a prediction-only deep neural network in a noisy environment. The results also show that the ODL approach can save communication cost and energy consumption for battery-powered Internet of Things devices.

• The paper presents a neural network-based on-device learning (ODL) methodology to address the gap between training data and deployed environments for improved edge AI anomaly detection.
• Experiments using vibration data demonstrate that ODL significantly enhances anomaly detection accuracy, particularly in noisy conditions, and reduces execution time and energy consumption on low-end IoT devices.
• The proposed ODL framework is practical for resource-constrained edge AI applications, offering benefits like continuous adaptation to environmental changes, improved energy efficiency, and enhanced data privacy.

Overview of On-Device Learning for Edge AI

The paper "Addressing Gap between Training Data and Deployed Environment by On-Device Learning" presents a significant improvement in the accuracy and efficiency of anomaly detection in edge AI applications through an innovative neural network-based on-device learning (ODL) methodology. This approach specifically addresses the discrepancies that arise between training data and the conditions present in deployed environments, which are often dynamic and unpredictable.

Core Contributions

The primary contribution of this research is the development of a semi-supervised, sequential training algorithm involving multiple neural network models designed to operate effectively on low-end internet of things (IoT) devices. The authors meticulously designed and implemented this methodology on wireless sensor nodes powered by Raspberry Pi Pico, achieving a structured system that performs retraining in its deployed environment. This is crucial for applications where prediction accuracy is affected by variable environmental factors such as noise, calibration inconsistencies, and temporal changes.

Experimental Evaluation and Results

A noteworthy aspect of the paper is its experimental validation using vibration pattern data from rotating machines to assess anomaly detection capability. The results robustly demonstrated that the ODL approach substantially enhances anomaly detection accuracy compared to conventional prediction-only deep neural networks, especially in environments with higher noise interference.

The execution time and energy consumption data further validate the practical benefits of the ODL approach. The paper details a comparative analysis between several cases, illustrating that ODL reduces communication costs and conserves energy in battery-powered IoT devices. For instance, in tasks operating at a frequency of per-second anomaly detection, ODL managed execution times and energy consumptions that were sustainable for long-term sensor node operations, something not feasible with conventional prediction-only approaches.

Practical and Theoretical Implications

The ODL framework holds significant promise for advancing edge AI technology, particularly in resource-constrained and battery-operated devices. It effectively addresses the challenge of concept drift by allowing continuous learning and adaptation in the field without requiring extensive cloud resources. The minimization of data transmission to the cloud also bolsters energy efficiency—a critical factor for IoT devices with limited power sources—and enhances data privacy by avoiding unnecessary data uploads.

From a theoretical perspective, the research offers a deeper understanding of applying online sequential learning algorithms like OS-ELM in real-world, non-static conditions. The combination of autoencoders with OS-ELM in the ODL framework for semi-supervised anomaly detection indicates a promising direction for building more efficient, adaptable models for edge computing scenarios.

Future Research Directions

The findings in this paper lay the groundwork for further exploration into more complex anomaly detection systems, potentially expanding beyond simple vibration analysis to other sensory modalities. Moreover, the development of dedicated hardware optimized for ODL could yield performance benefits that allow even higher efficiency in terms of speed and power consumption. Additionally, research could probe into integrating more sophisticated automatic switching between prediction and training modes using machine learning-based concept drift detection methods.

In conclusion, the paper presents a meticulously engineered approach to bridging the gap between training and deployment environments, offering both immediate practical benefits and rich avenues for future inquiry in on-device learning and edge AI applications.