Deployment of edge-lightweight AI vertical models in multimodal anomaly detection and energy optimization of tech-long production line

Abstract

Aiming at the problems that the abnormal state of high-speed liquid packaging production lines is difficult to identify in time, the computing power of edge controller is limited, and the energy consumption regulation and condition monitoring are separated from each other, a multimodal lightweight artificial intelligence (AI) system embedded in Tech-Long production controller is constructed; this realizes the collaborative optimization of abnormal operation window identification and energy consumption of the whole line. First, this study collects vibration, acoustic, and thermal signals from the bottle blowing, filling, and packaging sections; it also constructs cross-sectional anomaly characterization through time synchronization, feature-level alignment, and hierarchical fusion. Subsequently, Feather Convolutional Neural Network (FeatherCNN) is reconstructed using kernel fusion, channel compression, and 4-bit Integer (INT4) quantization to reduce model computation and storage load; thus, the network can adapt to resource-constrained Advanced RISC Machine (ARM) edge controllers. Finally, a dual-layer control chain is built, composed of continuous basic energy-saving and anomaly risk-weighted correction, so that anomaly scores influence the updates of temperature zones, valve openings, and heat-sealing parameters via dynamic weights and adjustment gains. The experimental results demonstrate that the anomaly window recognition accuracy of multimodal FeatherCNN reaches 0.931, which is superior to traditional machine learning models, deep learning baselines and various single-modal structures. The lightweight processing reduces Floating-Point Operations (FLOPs) from 3.42×10⁹ to 0.58×10⁹, a decrease of approximately 83%. The inference latency of the complete model on the ARM controller is shortened from 18.63 milliseconds (ms) to 7.41 ms. Under the controlled conditions of fixed production line speed, material specifications, and statistical windows, the dual-layer collaborative control has reduced the overall energy consumption of the whole line by 26%. The results show that multimodal anomaly identification, model lightweighting, and energy-saving parameter tuning can work together in the same edge decision chain. The framework verifies the feasibility of resource-constrained controllers supporting abnormal monitoring and energy consumption regulation of the entire packaging line; meanwhile, it provides an engineering reference for edge intelligent deployment of similar high-speed liquid packaging lines.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗