Date Available

5-7-2021

Year of Publication

2021

Document Type

Master's Thesis

Degree Name

Master of Science in Electrical Engineering (MSEE)

College

Engineering

Department/School/Program

Electrical and Computer Engineering

Advisor

Dr. Kevin D. Donohue

Abstract

Within recent years, the demand for organic produce has greatly increased due to many factors, including increasing knowledge about such things as dietary fiber and balanced gastrointestinal bacterial ecosystems. This increase in demand, coupled with the financial penalties for sending invasive species and pests across borders, presents a need for a scalable and accurate system to non-destructively detect infestation. The proposed work addresses this problem by testing the performance of a non-destructive vibro-acoustic method for detecting lava activity in apples. This involved 3 steps; design a mechanical data collection prototype for testing apples, a evaluate a set of features, and test the detection performance using machine learning algorithms. The mechanical data collection prototype aims to solve some of the issues that arose when collecting repeatable vibro-acoustic data from apples. The second piece aims to show the feasibility of a scalable model which takes vibro-acoustic data, performs multi-domain feature extraction, and then utilizes a SVM/ANN backend to detect codling moth infestation in apples. The final piece describes a procedure in which a novel CNN architecture pair is created to assess the quality of results with and without an acoustic reference channel. The data collection prototype produced higher quality data than previous setups. The feature extraction and SVM/ANN showed the ability to characterize patterns and detect infestation. The best of these was an SVM which had 87.34% accuracy on classifying 5 second segments from apples not in the training set, which was run on one iteration of a randomized dataset split. The CNN architectures showed potential for further development, with the noise-inclusive model performing over 8% better. However, both models show limited potential for generalizing to new apples with accuracies of (35.15% without noise, 43.92% with noise). The lower detection rates were limited by the intermittent larval activity rates, since the low accuracy rates were driven primarily by missed detections in the 5 second windows on apples labeled as infested. If the percentage of activity in any five second window is too low, then the “infested” sample will get classified as healthy due to that window having no larval sounds. The other notable issue regarding generalization potential was the sample size: the number of distinct apples used was too small, especially for deep learning applications. A much larger number of apples will be needed for future work.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2021.119

Funding Information

This work was supported by the Kentucky Agricultural Experiment Station and the National Institute of Food and Agriculture, U.S. Department of Agriculture, Foundational and Applied Science Program Grant with Award Number: 2019-67021- 29692, in Summer 2019, Summer 2020, and Spring 2021.

Download

Share

COinS