Enhancing low quality spectral data analytics for food quality control

Abstract

This research delves into the realm of food quality assessment using spectral data from portable spectrometers. By employing a consumer-centric approach, a diverse array of spectral datasets was created encompassing oils, fruits, and powders. The work introduces the Multi Session Testing Methodology (MSTM), highlighting stark performance differences compared to traditional methods and emphasizing the need to address spectra variations originating from portable devices.

A pivotal contribution is the Combined Subclass Linear Discriminant Analysis(CSDA), which aims to overcome the limitations of traditional Linear Discriminant Analysis (LDA) in food quality evaluations. By accounting for subclass clusters, this method discerns features among samples collected across multiple sessions. This was further augmented by the Kernel CSDA (KCDA), offering a kernelized variant of the CSDA to process nonlinear data. The potential of deep learning is crystallized in the CNN Spectra model, a simple yet highly optimized CNN tailored for low-quality spectral data. Contrary to prevailing assumptions, this research underscores that intricate deep network structures aren’t imperative for enhanced performance. Utilizing a hybrid strategy, features from CNN Spectra were harnessed in training a variety of machine learning models, leading to the birth of Deep CSDA.

Key insights emphasize that while portable spectrometers are a beacon of hope for food quality assessment, the challenge of data quality remains critical. Advanced models like CSDA and CNN Spectra offer a fresh perspective, with deep learning paradigms needing more refinement. Furthermore, merging conventional and deep learning approaches can usher in improved results.

Envisioning the future, expanding datasets and investigating shelf-time based evaluations emerge as priorities. The pursuit of distortion-invariant feature extraction can significantly amplify the precision of assessments. With the evolution of AI poised to be embedded in chipsets, we anticipate portable, energy efficient, and privacy-centric solutions. The integration of multimodal sensor data holds promise for a comprehensive food quality evaluation, painting a holistic picture. As progress continues, ethical and regulatory dimensions must be front and center.

Concluding this phase, the research sets the trajectory for a world where consumers lead the charge against food fraud, fortified by state-of-the-art technology. Presented here is not an endpoint but an initiation into a future of fortified food quality and safety.

Date of Award	May 2024
Original language	English
Supervisor	Jun Liu (Supervisor), James Uhomoibhi (Supervisor), Omar Nibouche (Supervisor) & Hui Wang (Supervisor)