Incremental Bayesian learning for in‐service analysis of aeronautic composites. Issue 6 (1st November 2013)

• Record Type: Journal Article
• Authors: Cacciola, Matteo; Megali, Giuseppe; Lay‐Ekuakille, Aimé
• Is Part Of: IET science, measurement & technology. Volume 7:Issue 6(2013)
• Page Start: 334
• Page End: 342
• Publication Date: 2013-11-01
• Subjects: aerospace materials -- composite materials -- fault diagnosis -- learning (artificial intelligence) -- neural nets -- nondestructive testing; incremental Bayesian learning -- in‐service analysis -- aeronautic composites -- fault detection -- nondestructive testing -- real‐time monitoring -- human safety -- Bayesian neural network -- VLSI implementation -- re‐programmability -- activation function; Measurement -- Periodicals; Electrical engineering -- Periodicals; Electronics -- Periodicals; Nanotechnology -- Periodicals; Electromagnetism -- Periodicals; Medical instruments and apparatus -- Periodicals; 621.3
• DOI: 10.1049/iet-smt.2012.0151
• Languages: English
• ISSNs: 1751-8822

Abstract

Abstract : Incremental learning could be really useful for fault detection and anticipation in non‐destructive testing and evaluation. The real‐time monitoring could be proficiently exploited when an early warning system is required for the human safety. This is the case of aeronautic transportation of persons and goods. Here, an automated neural‐based system for defect detection in aeronautic composites is proposed. The entire system consists of a stand‐alone defect classifier based on a Bayesian neural network (BNN) combined with advantages of Very Large Scale Integration (VLSI)implementation. Exploiting a parallel implementation is worthwhile when high computational speed, special operating conditions, portability, limited physical size, low‐power dissipation and reliability are required. This study shows how hardware‐based neural network can increase processing speed and defect identification rate. Secondary random access memory‐based field programmable gate arrays represent a suitable platform to realise these models, since their re‐programmability can rapidly change the parameters of the network if a new training is needed. With the hardware‐based BNN, 100% of delamination bottom/top, inclusion bottom/middle/top, porosity and 99.6% of delamination middle were correctly identified. The achieved results highlight the efficient design of the hardware network, obtained also using a new circuit to compute the activation function of neurons.