Embedded decision support system for ultrasound nondestructive evaluation based on extreme learning machines

• Published in: Computers & electrical engineering Vol. 90; p. 106891
• Main Authors: Silva, Lucas C., Simas Filho, Eduardo F., Albuquerque, Maria C.S., Silva, Ivan C., Farias, Claudia T.T.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 01.03.2021; Elsevier BV
• Subjects: Artificial neural networks; Autoencoder; Carbon steels; Data acquisition; Decision support systems; Defects; Embedding; Extreme learning machine; Machine learning; Microcontroller; Microcontrollers; Nondestructive evaluation; Nondestructive testing; Personal computers; Principal component analysis; Principal components analysis; Probability density functions; Reduction; Steel plates; Systems analysis; Training; Ultrasonic imaging; Ultrasound; Nondestructive evaluation; Extreme learning machine; Autoencoder; Microcontroller; Principal component analysis
• ISSN: 0045-7906; 1879-0755
• DOI: 10.1016/j.compeleceng.2020.106891

•ELM-based Embedded decision support system for ultrasound nondestructive evaluation.•Comparison between PCA and ELM Autoencoders for ultrasonic feature set reduction.•Evaluation of ELM performance regarding four probability density functions.•ELM’s random weights quantization achieved good performance with small bit count.•ELM networks trained in microcontroller showed good time and accuracy performances. [Display omitted] Decision support systems for nondestructive evaluation of defects are often trained on a personal computer environment, which may slow the assessment task. This work investigates the feasibility of embedding extreme learning machines in microcontrollers, in order to execute in situ training for an ultrasound nondestructive evaluation. Principal component analysis and autoencoders were evaluated as dimensionality reduction methods in a data set acquired from welded SAE 1020 carbon steel plates containing four types of defects. A binary-encoded random weight approach is proposed, whose results were tested over four different probability density functions used to generate the random weights. The obtained experimental results matched the accuracy of a previous work, run in computer environment for higher dimensionality data. In this work a dimensionality reduction of 75% and elapsed time for system training of less than 0.5 ms in microcontroller were achieved, indicating the suitability of such embedded networks for the studied case.