Wine discrimination based on multi-sensor fusion of GASF and Mel spectrogram features using an enhanced...

Area of Science:

• Food Science
• Analytical Chemistry
• Machine Learning

Background:

• Accurate classification of wine raw materials is crucial for quality control and authenticity.
• Traditional methods may struggle with subtle differences in sensory profiles.
• Developing objective, high-throughput methods for wine analysis is an ongoing challenge.

Purpose of the Study:

• To develop and validate a novel multi-sensor fusion strategy for discriminating wines based on their raw materials.
• To enhance feature extraction from complex sensor data using advanced transformation techniques.
• To leverage deep learning for accurate classification of wine origins.

Main Methods:

• Collected aroma and taste data using an electronic nose and noble metal electrodes.
• Transformed one-dimensional time-series sensor data into two-dimensional RGB images via Gramian Angular Summation Field (GASF) and Mel spectrogram algorithms.
• Utilized an Enhanced-EfficientNet-B0 deep learning model for image-based classification.

Main Results:

• The multi-sensor fusion strategy successfully discriminated wines from eight different raw materials.
• The Enhanced-EfficientNet-B0 model achieved a validation accuracy of 92.17%.
• The image transformation techniques effectively preserved temporal and spectral information from the sensor data.

Conclusions:

• The integration of multi-sensor data, advanced signal transformation, and deep learning offers an accurate and efficient method for wine raw material classification.
• This approach holds potential for improving wine authentication and quality assessment.
• The study demonstrates the efficacy of combining sensory data with sophisticated computational techniques in food analysis.