{"title":"An artificial neural network for digital image correlation dynamic subset selection based on speckle pattern quality metrics","authors":"D. Atkinson, M. van Rooyen, T. H. Becker","doi":"10.1111/str.12471","DOIUrl":null,"url":null,"abstract":"Precise and accurate digital image correlation computed displacement data requires sufficient noise suppression and spatial resolution, which improve and diminish, respectively, with increased subset size. Furthermore, spatially varying speckle pattern quality and displacement field complexity ideally necessitate a location‐specific optimal subset size to obtain a favourable compromise between noise suppression and spatial resolution. Although dynamic subset selection (DSS) methods have been proposed based on speckle pattern quality metrics (SPQMs), they do not ensure such a favourable compromise.This work investigates using an artificial neural network (ANN) for DSS. An ANN is trained to predict the displacement error standard deviation of a subset from multiple SPQMs and the standard deviation of image noise, such that the smallest subset offering sufficient noise suppression, dictated by a displacement error standard deviation threshold, is appointed.Validation, both within and outside the domain of the training images, shows that the smallest subset providing sufficient noise suppression offers a favourable compromise for up to moderate displacement gradients. Additionally, the proposed method is shown to perform with greater consistency and reliability relative to existing SPQM‐based DSS methods.The novel proposition lies in utilising an ANN as an error prediction tool, based on multiple SPQMs, and hence, is an attractive alternative for DSS.","PeriodicalId":51176,"journal":{"name":"Strain","volume":null,"pages":null},"PeriodicalIF":1.8000,"publicationDate":"2023-12-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Strain","FirstCategoryId":"88","ListUrlMain":"https://doi.org/10.1111/str.12471","RegionNum":3,"RegionCategory":"材料科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"MATERIALS SCIENCE, CHARACTERIZATION & TESTING","Score":null,"Total":0}
引用次数: 0
Abstract
Precise and accurate digital image correlation computed displacement data requires sufficient noise suppression and spatial resolution, which improve and diminish, respectively, with increased subset size. Furthermore, spatially varying speckle pattern quality and displacement field complexity ideally necessitate a location‐specific optimal subset size to obtain a favourable compromise between noise suppression and spatial resolution. Although dynamic subset selection (DSS) methods have been proposed based on speckle pattern quality metrics (SPQMs), they do not ensure such a favourable compromise.This work investigates using an artificial neural network (ANN) for DSS. An ANN is trained to predict the displacement error standard deviation of a subset from multiple SPQMs and the standard deviation of image noise, such that the smallest subset offering sufficient noise suppression, dictated by a displacement error standard deviation threshold, is appointed.Validation, both within and outside the domain of the training images, shows that the smallest subset providing sufficient noise suppression offers a favourable compromise for up to moderate displacement gradients. Additionally, the proposed method is shown to perform with greater consistency and reliability relative to existing SPQM‐based DSS methods.The novel proposition lies in utilising an ANN as an error prediction tool, based on multiple SPQMs, and hence, is an attractive alternative for DSS.
期刊介绍:
Strain is an international journal that contains contributions from leading-edge research on the measurement of the mechanical behaviour of structures and systems. Strain only accepts contributions with sufficient novelty in the design, implementation, and/or validation of experimental methodologies to characterize materials, structures, and systems; i.e. contributions that are limited to the application of established methodologies are outside of the scope of the journal. The journal includes papers from all engineering disciplines that deal with material behaviour and degradation under load, structural design and measurement techniques. Although the thrust of the journal is experimental, numerical simulations and validation are included in the coverage.
Strain welcomes papers that deal with novel work in the following areas:
experimental techniques
non-destructive evaluation techniques
numerical analysis, simulation and validation
residual stress measurement techniques
design of composite structures and components
impact behaviour of materials and structures
signal and image processing
transducer and sensor design
structural health monitoring
biomechanics
extreme environment
micro- and nano-scale testing method.