Researchers embedded a multimodal AI system in a Tech-Long production controller to detect abnormal operation windows and optimize energy use on high-speed liquid packaging lines, where edge computing limits and separate monitoring and energy controls make timely detection difficult. The system collected vibration, acoustic and thermal signals from bottle blowing, filling and packaging sections. It synchronized the signals, aligned features and fused them hierarchically to characterize anomalies across sections.

The researchers adapted Feather Convolutional Neural Network (FeatherCNN) for resource-constrained ARM edge controllers using kernel fusion, channel compression and 4-bit integer quantization. Its anomaly-window recognition accuracy reached 0.931, outperforming traditional machine-learning models, deep-learning baselines and single-modal structures. The lightweight processing cut floating-point operations from 3.42×10⁹ to 0.58×10⁹, a reduction of approximately 83%; inference latency on the ARM controller fell from 18.63 ms to 7.41 ms.

A dual-layer control system combined continuous basic energy saving with anomaly-risk-weighted corrections. Anomaly scores influenced updates to temperature zones, valve openings and heat-sealing parameters through dynamic weights and adjustment gains. In tests with production-line speed, material specifications and statistical windows held fixed, the combined control reduced total line energy consumption by 26%. The researchers said the results demonstrate that anomaly detection, model compression and energy-saving adjustments can operate in one edge decision chain, and offer an engineering reference for similar high-speed liquid packaging lines.

Researchers embedded a multimodal AI system in a Tech-Long production controller to detect abnormal operation windows and optimize energy use on high-speed liquid packaging lines, where edge computing limits and separate monitoring and energy controls make timely detection difficult. The system collected vibration, acoustic and thermal signals from bottle blowing, filling and packaging sections. It synchronized the signals, aligned features and fused them hierarchically to characterize anomalies across sections.

The researchers adapted Feather Convolutional Neural Network (FeatherCNN) for resource-constrained ARM edge controllers using kernel fusion, channel compression and 4-bit integer quantization. Its anomaly-window recognition accuracy reached 0.931, outperforming traditional machine-learning models, deep-learning baselines and single-modal structures. The lightweight processing cut floating-point operations from 3.42×10⁹ to 0.58×10⁹, a reduction of approximately 83%; inference latency on the ARM controller fell from 18.63 ms to 7.41 ms.

A dual-layer control system combined continuous basic energy saving with anomaly-risk-weighted corrections. Anomaly scores influenced updates to temperature zones, valve openings and heat-sealing parameters through dynamic weights and adjustment gains. In tests with production-line speed, material specifications and statistical windows held fixed, the combined control reduced total line energy consumption by 26%. The researchers said the results demonstrate that anomaly detection, model compression and energy-saving adjustments can operate in one edge decision chain, and offer an engineering reference for similar high-speed liquid packaging lines.