An evaluation of recurrent neural network modelling for the prediction of damage evolution during forming

Xiong, Yi; Withers, Philip J.

doi:10.1016/j.jmatprotec.2005.01.016

Cited by 7 publications

(4 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Furthermore, once deconvoluted, the expected values (predictions) from the model are in general closer to the underlying smooth type-I strain profile, f(x), than one would measure with a completely faithful infinitesimally fine sensor, which would pick up the microstructural fluctuations. Similar behaviour has been observed in the dynamic model for predicting deformation microstructures (Xiong & Withers, 2005). As one would expect, its performance in recovering the profile is better for the sensors with the narrower response functions.…”

Section: Test Example Isupporting

confidence: 77%

A deconvolution method for the reconstruction of underlying profiles measured using large sampling volumes

Xiong

Withers

2006

J Appl Cryst

View full text Add to dashboard Cite

A deconvolution method for diffraction measurements based on a statistical learning technique is presented. The radial-basis function network is used to model the underlying function. A full probabilistic description of the measurement is introduced, incorporating a Bayesian algorithm based on an evidence framework. This method allows predictions of both the convolution and the underlying function from noisy measurements. In addition, the method can provide an estimation of the prediction uncertainty, i.e. error-bars. In order to assess the capability of the method, the model was tested first on synthetic data of controllable quality and sparsity; it is shown that the method works very well, even for inaccurately measured (noisy) data. Subsequently, the deconvolution method was applied to real data sets typical of neutron and synchrotron residual stress (strain) data, recovering features not immediately evident in the large-gauge-volume measurements themselves. Finally, the extent to which short-period components are lost as a function of the measurement gauge dimensions is discussed. The results seem to indicate that for a triangular sensor-sensitivity function, measurements are best made with a gauge of a width approximately equal to the wavelength of the expected strain variation, but with a significant level of overlap ($80%) between successive points; this is contrary to current practice for neutron strain measurements.

show abstract

Section: Test Example Isupporting

confidence: 77%

A deconvolution method for the reconstruction of underlying profiles measured using large sampling volumes

Xiong

Withers

2006

J Appl Cryst

View full text Add to dashboard Cite

show abstract

“…Saxen et al 33 describe some issues related to time series modelling of hot metal silicon content in a blast furnace, where approximation error and the number of weights for ANNs are minimised (optimised) simultaneously by a GA approach. Xiong et al 34 examined the efficiency and capability of a recurrent neural network model used to predict damage evolution during hot non-uniform, non-isothermal forging on the basis of a limited number of snapshots during the process. A Bayesian algorithm was introduced to optimise the hyperparameters in the ANN related to the noise level (uncertainty in inputs and output predictions) and to the weight decay (the model fitting process).…”

Section: Data Miningmentioning

confidence: 99%

Invited review: Adaptive numerical modelling and hybrid physically based ANM approaches in materials engineering – a survey

Reed¹,

Starink²,

Gunn³

et al. 2009

Materials Science and Technology

View full text Add to dashboard Cite

Many adaptive numerical modelling (ANM) techniques such as artificial neural networks (including multilayer perceptrons), support vector machines and Gaussian processes have now been applied to a wide range of regression and classification problems in materials science. Materials science offers a wide range of industrial applications and hence problem complexity levels from well physically characterised systems (e.g. high value, low volume products) to high volume low cost applications with intrinsic scatter due to commercial manufacturing processes. The authors review a number of recent examples in the literature, with the aim of identifying best practice in the use of these techniques as part of a multistrand modelling approach. The importance of understanding the basic principles of these modelling techniques and how they can link with other modelling strategies is emphasised. In particular the authors wish to identify the importance of hybrid physically based ANM in taking the field forward, which can range from, at the most basic level, careful data selection and data preprocessing to a full integration of physically based models with advanced ANM. A number of case studies are presented to illustrate the main points of the paper. Keywords: Neural networks, Support vector machines, Gaussian processes, Adaptive numerical modelling, Data driven techniques, Hybrid modelling, Applications in materials science IntroductionThe links between processing variables, microstructure evolution, resultant properties and hence mechanical performance are core tenets of materials engineering and a variety of modelling approaches have been applied to these materials modelling challenges. In some cases, a clear physical basis for such models has now been established, but often these can apply only to well defined conditions, which may not be representative of either genuine industrial (scaled up) production processes, or materials performance under realistic service conditions. The advantages of physically based models are:(i) they are robust to interpolation and extrapolation (as long as the underlying physical mechanisms being modelled are in operation) (ii) they should not require large amounts of data, as typically they only have a limited number of fitted parameters; ideally, a physically based model may have no fitted parameters (iii) the mechanistic verifications derived from physically based modelling may also offer insights into novel materials optimisation strategies (iv) a physically based model is relatively transparent, i.e. the mathematical underlying relationships between input and output parameters may be written in a reasonably clear mathematical form: this is part of a loosely defined group of methods which the authors term 'white box'. However, in a truly integrated processing-performance model, with the levels of uncertainty and complexity associated with industrial practice and complex service conditions, expressing these interlinked processes purely on physics based principles is unlikely to be feasible. ...

show abstract

“…Além deste procedimento, as entradas foram aleatorizadas gerando 50 conjuntos de treinamento distintos; de tal forma que o treinamento pudesse ser realizado diversas vezes calculando-se o coeficiente de desempenho (B) e determinando-se a porcentagem de resultados cujo coeficiente situou-se entre 0,9 e 1,0. Este procedimento estatístico deve ser realizado evitando-se "overfitting", ou seja, a memorização dos dados de treinamento ao invés da extração das características gerais que permitiriam a generalização da solução gerada [20]. O desempenho da rede foi avaliado através do cálculo do coeficiente de determinação B que é definido pela equação 11:…”

Section: Cálculo Das Propriedades Dinâmicasunclassified

Redes neurais artificiais aplicadas para a predição do comportamento dinâmico-mecânico de compósitos de matriz epóxi reforçados com fibras de carbono

et al. 2007

View full text Add to dashboard Cite

Neste trabalho, o algoritmo Levenberg-Marquardt foi aplicado para predizer o comportamento dinâmico-mecânico de compósitos de matriz epóxi reforçados com fibras de carbono. Empregou-se o ensaio de vibração amortecida (ASTM E-756) para viga do tipo engastada-livre que forneceu experimentalmente as curvas de amplitude em função do tempo de resposta. O compósito é de uso aeronáutico tendo configuração [0º/45º/90º/0º]S. Uma rede neural do tipo "perceptron" de múltiplas camadas foi empregada, e os resultados mostraram que a aplicação do algoritmo de aprendizado Levenberg-Marquardt conduz a uma elevada qualidade preditiva para compósitos de matriz epóxi, uma vez que o coeficiente de desempenho (B) apresentou-se superior a 0.9 para 64% dos conjuntos testados. Os testes iniciais consideraram uma arquitetura simples 2-[30-30]2-1 resultando em uma baixa qualidade preditiva. Entretanto, o aumento do número de neurônios nas camadas escondidas resultou em uma arquitetura otimizada 2-[100-100]2-1.

show abstract

An evaluation of recurrent neural network modelling for the prediction of damage evolution during forming

Cited by 7 publications

References 14 publications

A deconvolution method for the reconstruction of underlying profiles measured using large sampling volumes

A deconvolution method for the reconstruction of underlying profiles measured using large sampling volumes

Invited review: Adaptive numerical modelling and hybrid physically based ANM approaches in materials engineering – a survey

Redes neurais artificiais aplicadas para a predição do comportamento dinâmico-mecânico de compósitos de matriz epóxi reforçados com fibras de carbono

Contact Info

Product

Resources

About