The use of Support Vector Machine Learning for the Early Detection of Kickback in Chainsaws

By Drew Arnold

Candidate for Master of Science in Mechanical Engineering

Abstract

Chainsaw kickback is the most dangerous phenomenon facing chainsaw operators. Kickback occurs when the chain at the tip of the chainsaw bar is suddenly stopped, transferring the kinetic energy of the cutting chain, to gross motion of the saw body, causing the saw to rotate backward toward the operator rapidly. The small body of published research on the topic of chainsaw kickback was conducted in development of standardized testing for consumer chainsaws. Modern chainsaws are equipped with safety measures such as low-kickback cutting chains and hand-guard braking mechanisms. These mechanisms have greatly improved the safety of chainsaws, but their inherent mechanical simplicity leaves room for improvement.

Phase 1 of this work utilized two accelerometers and a gyroscope to determine if distinguishing a kickback event from normal cutting operations was possible. The three sensors were combined through simple addition, and a weighting coefficient was applied to each sensor. The weighting coefficients were varied to obtain the greatest margin between normal cutting and kickback. The result of this study showed that using only a gyroscope was the most effective at detecting kickback. Phase 2 focused on detecting kickback as early as possible. Three analysis methods were applied: Signal Differentiation, a Simplified Bag of Words method, and a Support Vector Machine with selective under-sampling and a stack of classifier vectors. Signal differentiation, while detecting the kickback events earlier, also suffered from many false positives. The Bag of Words method was unsuccessful in creating results different than the method used in Phase 1. The Support Vector Machine classification was able to detect kickback an average of 19.4 ms before the simple threshold method with no occurrence of either false positives or false negatives. This method was the most reliable and provided the greatest likelihood of early kickback detection.

Tuesday, November 27, 2012

9:00 am, Covell 117

School of Mechanical, Industrial, and Manufacturing Engineering