Utilizing Multimodal Logic Fusion to Identify the Types of Food Waste Sources.

Area of Science:

• Industrial Waste Management
• Machine Learning
• Sensor Fusion

Background:

• Identifying food waste sources in industrial settings is challenging due to variable lighting affecting visual recognition models.
• Existing methods struggle with performance degradation under suboptimal illumination.

Purpose of the Study:

• To propose and validate a robust, interpretable, and adaptive multimodal logic fusion method for food waste source identification.
• To dynamically assign sensor dominance based on real-time illuminance intensity for improved accuracy.

Main Methods:

• A multimodal approach combining a MobileNetV3 + EMA image recognition model with an audio model (Fast Fourier Transform + Support Vector Machine).
• Implementation of environment-aware conditional logic for dynamic sensor fusion based on illuminance.
• MobileNetV3 + EMA for image recognition and FFT + SVM for audio classification.

Main Results:

• The image model achieved 99.46% accuracy in optimal lighting (120-240 cd m^-2) but degraded significantly outside this range.
• The audio model demonstrated illumination-independent performance with 0.80 accuracy, 0.78 recall, and 0.80 F1 score.
• The fusion method achieved an overall accuracy of 90.25% on an independent test set, prioritizing audio recognition below 84 cd m^-2.

Conclusions:

• The proposed adaptive multimodal logic fusion method ensures robust food waste source identification under variable industrial lighting conditions.
• Dynamic sensor dominance assignment prevents model failure and enhances classification accuracy in challenging environments.
• The system offers an interpretable and resilient solution for real-world industrial applications.