Integrated application model for visual detection of welding quality based on visual neuron under edge-end collaboration

Abstract

ABSTRACT The traditional ‘end-to-end’ integrated machine-vision application model is limited by high economic costs, as well as by computing power and storage capacity when solving the tasks with high time delay sensitivity and complex inference processes in automated welding quality detection of whole vehicles. To address this problem, a visual neuron was defined as an edge node, and an integrated application model was proposed that incorporated the idea of edge-end collaboration. First, to provide the visual neurons with the edge capability to solve the welding quality visual detection tasks, a multi-level visual knowledge hypergraph was constructed, which decouples and abstracts the knowledge units for the tasks. An object-oriented knowledge representation was used to establish the knowledge hyperedge. Secondly, Petri net technology was employed to analyze the task data flow in the integrated application pattern and establish the solving model N-COPN of ‘control-inference-decision’. Moreover, edge-end data transfer was implemented based on the message queue telemetry transfer protocol and RESTful architecture. Finally, the quality of auto component gluing in the welding process of the vehicle was detected using this model. The experimental results showed that the model performed better in terms of fast-integrated inference and concurrent task processing.