# An Adaptive Compression Method for Lightweight AI Models of Edge Nodes in Customized Production

## TL;DR

This paper introduces a new method to compress AI models for edge devices in manufacturing, adapting to changing tasks and hardware constraints.

## Contribution

The novel adaptive compression method dynamically adjusts AI models for edge nodes using reinforcement learning and Bayesian optimization.

## Key findings

- The proposed method improves inference efficiency while maintaining high accuracy in object recognition tasks.
- It outperforms conventional fixed compression strategies in adaptability and edge-deployment performance.

## Abstract

In customized production environments featuring multi-task parallelism, the efficient adaptability of edge intelligent models is essential for ensuring the stable operation of production lines. However, rapidly generating deployable lightweight models under conditions of frequent task changes and constrained hardware resources remains a major challenge for current edge intelligence applications. This paper proposes an adaptive lightweight artificial intelligence (AI) model compression method for edge nodes in customized production lines to overcome the limited transferability and insufficient flexibility of traditional static compression approaches. First, a task requirement analysis model is constructed based on accuracy, latency, and power-consumption demands associated with different production tasks. Then, the hardware information of edge nodes is structurally characterized. Subsequently, a compression-strategy candidate pool is established, and an adaptive decision engine integrating ensemble reinforcement learning (RL) and Bayesian optimization (BO) is introduced. Finally, through an iterative optimization mechanism, compression ratios are dynamically adjusted using real-time feedback of inference latency, memory usage, and recognition accuracy, thereby continuously enhancing model performance in edge environments. Experimental results demonstrate that, in typical object-recognition tasks, the lightweight models generated by the proposed method significantly improve inference efficiency while maintaining high accuracy, outperforming conventional fixed compression strategies and validating the effectiveness of the proposed approach in adaptive capability and edge-deployment performance.