Multivariate Gas Sensor E-Nose System with PARAFAC and Machine Learning Modeling for Quantifying and Classifying the Impact of Fishing Gears

Abstract

Highlights: This research addresses critical gaps in the real-time quality assurance of the hydrobiological industry by developing a novel monitoring methodology. Accordingly, our findings are structured around the following key questions: What are the main findings? A novel methodology integrating MOSR sensors, PARAFAC, and Machine Learning was implemented for real-time quality monitoring. The L2 Norm metric successfully quantified metabolic stress deviation, establishing a data-driven distinction between various fishing gears. What are the implications of the main findings? The methodology provides an objective tool for seafood traceability and quality differentiation. Generating gear-specific stress profiles can be utilized as a robust standard to support modern sustainability and food quality certification schemes. The quality of seafood is intrinsically linked to the accumulated history of stress, feeding, handling, and physical damage imposed by the fishing gear employed. This study proposes an innovative methodology using an E-nose sensor. The study species was Sparus aurata. Eight fishing gears were studied. The methodology integrates Parallel Factor Analysis (PARAFAC) for impact quantification and Machine Learning (ML) for classifying the fishing gear of origin. Longline was established as the method with the lowest deviation. The impact hierarchy, from highest to lowest deviation, is as follows: Aquaculture 50.61% (95% CI: 34%, 68%), Purse seine 37.92% (95% CI: 22%, 54%), Trawl 35.92% (95% CI: 21%, 51%), Gillnet (three panels) 27.69% (95% CI: 14%, 41%), Gillnet (single panel) 24.63% (95% CI: 9%, 40%), Gillnet (two panels) 18.12% (95% CI: 4%, 31%) and Hook and line 1.36% (95% CI: −10%, 13%). For the classification task, 33 ML models were evaluated. Subspace KNN model yielded the best results with an accuracy of 97.14% in the validation and 98.08% in the testing, using 35 variables. Using 10, 15, 20, 25, and 30 variables, an accuracy higher than 85% was achieved. These results demonstrate the high precision in fish traceability by exploiting the sensor response profile left by each fishing gear.