An optimization strategy for HMI panel recognition of CNC machines using a CNN deep-learning network. (March 2021)

• Record Type: Journal Article
• Authors: Guo, Bo; Lee, Fu-Shin; Lin, Chen-I; Lin, Yuan-Jun
• Other Names: Vijayakumar K guest-editor.
• Is Part Of: Concurrent engineering, research and applications. Volume 29:Number 1(2021)
• Page Start: 35
• Page End: 48
• Publication Date: 2021-03
• Subjects: training optimization -- image recognition -- hyperparameter -- cloud server -- CNC -- robot -- control panel -- and invariant feature extraction; Production engineering -- Periodicals; Concurrent engineering -- Periodicals; 621.39
• DOI: 10.1177/1063293X21998083
• Languages: English
• ISSNs: 1063-293X

Abstract

This paper suggests an optimization strategy to train a CNN deep-learning network, which successfully recognizing working status on the HMI panels of CNC machines. To verify the developed strategy, the research experiments using a prototype that consists of a CNC milling machine and an industrial robot. In the optimization strategy, the research first defines a length-varying hyperparameter list for the deep-learning network, and the entities in the list adjust themselves to optimize the model scales. During the optimization process, this paper adopts a two-stage training scheme that gradually augments image datasets to improve HMI control-panel recognition performances, such as recognition accuracy and recognition speed to identify the CNC machine working status. Using an open-source PyTorch platform, this research establishes a cloud-based distributed architecture to build training codes for the deep-learning network, in which an applicable optimization model is deployed to recognize the CNC control-panel working status. The optimization strategy employs minimal codes to rebuild the architecture and the least efforts to reform the manufacturing system. The optimally trained model provides up to a 99.34% CNC panel-message recognition accuracy and a high-speed recognition of 100 images in 0.6 s. Moreover, the developed optimization strategy enables the prediction of necessitated dataset augmentation to training a practically implemented CNN network.