Time and Memory Efficient Online Piecewise Linear Approximation of Sensor Signals.

Area of Science:

• Data Mining
• Activity Recognition
• Signal Processing
• Embedded Systems

Background:

• Piecewise linear approximation (PLA) is crucial for analyzing sensor signals in Data Mining and Activity Recognition.
• Existing PLA algorithms often require significant memory buffering or have execution times dependent on segment length.
• These limitations are particularly problematic for resource-constrained microcontrollers in wireless sensor nodes.

Purpose of the Study:

• To propose a novel piecewise linear approximation algorithm.
• To achieve constant computational and memory complexity for microcontroller applications.
• To outperform existing state-of-the-art PLA techniques in terms of efficiency and approximation quality.

Main Methods:

• Development of a new PLA algorithm with inherent constant time and memory complexity.
• Experimental evaluation comparing the proposed algorithm against established PLA methods.
• Analysis of approximation quality, parameter space, and latency.

Main Results:

• The proposed algorithm demonstrates one to three orders of magnitude faster worst-case execution time and three to seventy times faster average execution time.
• Achieves constant computational and memory complexity, overcoming limitations of current methods.
• Maintains approximation quality comparable to state-of-the-art techniques while offering a maximum error guarantee per segment.

Conclusions:

• The novel PLA algorithm is highly efficient in terms of time and memory for sensor signal processing on microcontrollers.
• It provides a practical solution for resource-constrained environments without sacrificing approximation accuracy.
• The algorithm's low latency, small parameter space, and guaranteed error bounds make it suitable for real-time applications.