Application Adaptive Light-Weight Deep Learning (AppAdapt-LWDL) Framework for Enabling Edge Intelligence in Dairy Processing

The dairy industry is experiencing a surge in data from Edge devices, using spectroscopic techniques for milk quality assessment. Milk spectral data can help understand the species of milk producer and detect inter-species adulteration. Transmitting raw milk spectral data to the cloud for processing faces challenges due to limited network resources such as bandwidth, computational memory, and energy availability. Edge processing offers a solution by training data closer to the source, enhancing efficiency and real-time analysis by providing reduced latency, improved accuracy, resource-aware computation, and real-time customization. However, traditional Deep Learning (DL) methods such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) struggle on resource-constrained Edge devices due to complexity. To address this, we propose an Edge-Centric Application-Adaptive Light-Weight DL approach (AppAdapt-LWDL) for milk species identification and adulteration detection. Our method optimizes DL models via double model optimization, involving low-magnitude pruning and post-training quantization. Our novel application-adaptive algorithm balances speed and accuracy by determining the pruning ratio automatically for the specific application. The chosen model is then quantized for smaller databases, ideal for embedded devices. The AppAdapt-LWDL framework significantly accelerates training, speeds up inferencing, enhances energy efficiency, and maintains accuracy based on application needs.