A Machine Learning-Integrated Electronic Nose Framework for Real-Time Seafood Freshness Evaluation and Spoilage Prediction

Abstract

The freshness of seafood is crucial for food safety, quality assurance and consumer confidence. Spoilage causes ill health and financial losses. The work is a portable real-time e-nose system that using ML together with volatile chemical sensor information automatically assess the freshness of seafood. The analysis is conducted using Fish Freshness Evaluation data set from a Kaggle data sets. This data set encompasses five different kinds of seafood (Cod, Crab, Tuna, Salmon and Shrimp), and sensor data over a period of time for various levels of freshness. To select the features which considerably influence the categorisation of freshness and prediction of spoiling, PPS approach is used. Several ML techniques are implemented such as AdaBoost, RF, XGBoost, DT, NB, Gradient Boosting, and KNN for classification, and DT, XGBoost, AdaBoost, and KNN for regression. The model is assessed based on accuracy, precision, recall, F1-score, RMSE, R2 score and cross validation accuracy. The whole data is preprocessed using SMOTE, StandardScaler. Later, ensemble extensions are applied on RF, DT, NB, Voting Classifier (RF + DT soft voting), and Voting Regressor. XAI techniques such as LIME, SHAP are used to explain the model behaviour. The real-time prediction is done with Flask based interface along with SQLite. The CV Accuracy of 100% for the Voting Classifier and RF, and CV MSE of 0.0106 for the Voting Regressor, demonstrate excellent precision, resilience and generalisation in the prediction of seafood freshness.