Classification of auditory signals from a combine harvester based on Mel-frequency Cepstral coefficients and machine learning.

Abstract

As agricultural machinery moves into the digital era, significant developments in available technology will likely make autonomous farm vehicles more feasible, affordable, and desirable. One of the challenges of effective autonomous vehicle control specific to agriculture is the ability of the vehicle to interpret and adapt to constantly changing conditions. Auditory information is a primary indicator of changing conditions to an in-cab operator, particularly in situations such as detecting mechanical overload in a combine. This paper explores the potential for auditory information to be used in autonomous vehicle control. The sound was recorded at a sampling rate of 48 kHz near the straw chopper of a combine for three different operating modes during the same harvest day. Samples from each clip were segmented and analyzed to extract 31 audio features. Six different feature selection methods ranked the importance of each of the 31 features to identify the features that lead to accurate classification with a minimal number of calculations. These six rankings were assessed by Fagin’s algorithm to yield two features (both mel-frequency cepstral coefficients). Twenty-five distinct machine learning classification methods were evaluated using these two features. Three of these classification methods reached 100% accuracy, and 9 classifiers exceeded an individual success rate of more than 99% using those same features. These feature extraction and classification steps took less than 1 s, assuring that such a classification system could be implemented in real-time.