Beehive bioacoustics noise reduction using liquid neural networks
Loading...
Supplementary material
Other Title
Authors
Malhotra, Raveesh
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2024
Supervisors
Ardekani, Iman
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
New Zealand
honeybees (Apis mellifera)
beehive monitoring
audio data processing systems
noise removal algorithms
Artificial Neural Network (ANN)
Long Short-Term Memory (LSTM)
Convolutional Neural Network (CNN)
apiculture industry
honeybees (Apis mellifera)
beehive monitoring
audio data processing systems
noise removal algorithms
Artificial Neural Network (ANN)
Long Short-Term Memory (LSTM)
Convolutional Neural Network (CNN)
apiculture industry
ANZSRC Field of Research Code (2020)
Citation
Malhotra, R. (2024). Beehive bioacoustics noise reduction using liquid neural networks (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology https://hdl.handle.net/10652/6492
Abstract
Monitoring the health and activity of bee colonies has become increasingly crucial in recent years, owing to bees' critical role in pollination and their worrisome drop in population. Beehive sound analysis is a promising tool for non-invasively measuring beehive health. However, these auditory waves are noisy, making precise assessment difficult. This thesis investigates the application of advanced artificial neural network architectures for denoising beehive sounds, with a focus on the comparative performance of Convolutional Neural Networks, Long Short-Term Memory networks, and Liquid Time-Constant networks, with a particular emphasis on Liquid Time-Constant Networks, which have been recently proposed by researchers. Liquid Neural Networks, exemplified by Liquid Time-Constant neural networks, are inspired by the brain's continuous-time processing and have demonstrated significant adaptability and robustness.
The work begins with the collection and preprocessing of beehive audio data, which is then used to develop and train Convolutional Neural Networks, Long Short-Term Memory networks, and Liquid Neural Networks models. To improve denoising capabilities, each model goes through a rigorous hyperparameter tuning procedure that includes learning rate optimization and batch normalization. The major goal is to assess the real-world applicability of Liquid Time-Constant networks by comparing model performance in terms of noise reduction effectiveness and computing efficiency. According to experimental results, while Convolutional Neural Networks have basic denoising capabilities, they are outperformed by both Long Short-Term Memory networks and Liquid Time-Constant networks. Long Short-Term Memory networks perform better in addressing temporal dependencies in audio signals, resulting in the best overall denoising outcomes. However, Liquid Neural Networks with Liquid Time- Constant networks perform similarly to Long Short-Term Memory networks and greatly outperform Convolutional Neural Networks. The Liquid Neural Networks show special potential due to their dynamic adaptability capabilities and robustness in dealing with shifting noise levels.
This study demonstrates the viability of Liquid Neural Networks as a competitive and economical solution for real-world audio denoising applications, particularly in the context of beehive health monitoring. While Liquid Time-Constant networks fall short of Long Short-Term Memory networks in terms of regression performance, they outperformed Convolutional Neural Networks. The models' performance was largely measured using Mean Squared Error and Signal-to-Noise Ratio, two key metrics for audio denoising tasks. They have significant advantages in model simplicity and adaptability, making them an important tool in environmental monitoring and advanced auditory signal processing. This thesis contributes to the developing field of neural network applications in environmental monitoring and establishes the framework for future research in this area by offering a breakthrough in adaptive, real-time audio denoising using novel Liquid Time Constants for crucial applications like beehive health assessment by surpassing traditional models in both adaptability and efficiency.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved