Convergence of improved particle swarm optimization based ensemble model and explainable AI for the accurate detection of food adulteration in red chilli powder

Food adulteration involves the practice of adding or mixing inferior substances to food products, which undermines quality and safety. Adulteration of red chilli with brick powder is a significant food safety issue as it poses serious health risks to consumers. Accurate identification of the adulteration presents a significant challenge, particularly when adulteration is present in minuscule amounts. Existing methods aimed at identifying such micro levels of food adulteration are less accurate and lack interpretability. This study aims to address the research gaps in food adulteration by developing a robust model that integrates machine learning and explainable artificial intelligence methods. The key contributions of the proposed work are a deep convolutional generative adversarial network to enhance the model performance in limited data scenarios; improved particle swarm optimization as a promising metaheuristic optimization method to select the robust and highly discriminative features and to address premature convergence; explainable artificial intelligence methods (SHAP and LIME) to enhance the ensemble stacking model transparency and interpretability. A custom dataset is generated in the work, and it comprises 250 natural samples distributed among 5 categories (50 samples per category), ranging from no adulteration to adulteration in varying concentrations of 1 %, 2 %, 3 %, and 4 %, respectively. The proposed work is implemented on the synthetic data (200 samples per category) generated by the deep convolutional generative adversarial network. The potent combination of improved particle swarm optimization and explainable artificial intelligence enhances the accuracy, interpretability, and transparency of the proposed model by providing deeper insights which in turn bolsters confidence in distinguishing between pure and adulterated red chilli powder samples, thus contributing to improved food safety measures. The proposed model has shown a remarkable accuracy of 92.42 % on the synthetic data.