Research on Lightweight Industrial Large Models for Fault Diagnosis on Edge Devices

Industrial large models can extract features from industrial data, transfer learned knowledge and adapt to different fault diagnosis tasks. However, their massive parameter counts, high memory usage, and expensive computational costs make direct deployment on resource-constrained edge devices challenging. This paper reviews the current applications and deployment challenges of industrial large models in fault diagnosis and provides an in-depth discussion of six major lightweight techniques: model pruning, quantization, knowledge distillation, lightweight architecture design, parameter-efficient fine-tuning, and hybrid optimization. The paper analyzes the advantages, limitations, and applicability to different hardware of these techniques from dimensions such as diagnostic accuracy, model size, inference latency, memory usage, and energy consumption. In addition, based on the computing capabilities of various types of industrial equipment, this paper proposes a resource-aware deployment framework: for low-power terminals, models with high compression ratios are recommended; for medium-power devices, models that are moderately optimized and combined with an edge-to-cloud collaboration mechanism are suitable; while for complex diagnostic tasks, larger-scale models can be deployed on high-performance edge servers or cloud platforms. Overall, this combination may improve the accuracy and efficiency of edge-based industrial fault diagnosis.