Implicit constitutive modelling for viscoplasticity using neural networks

Up to now, a number of models have been proposed and discussed to describe a wide range of inelastic behaviours of materials. The fatal problem of using such models is however the existence of model errors, and the problem remains inevitably as far as a material model is written explicitly. In this paper, the authors define the implicit constitutive model and propose an implicit viscoplastic constitutive model using neural networks. In their modelling, inelastic material behaviours are generalized in a state-space representation and the state-space form is constructed by a neural network using input–output data sets. A technique to extract the input–output data from experimental data is also described. The proposed model was first generated from pseudo-experimental data created by one of the widely used constitutive models and was found to replace the model well. Then, having been tested with the actual experimental data, the proposed model resulted in a negligible amount of model errors indicating its superiority to all the existing explicit models in accuracy. © 1998 John Wiley & Sons, Ltd.     

Autoprogressive training of neural network constitutive models

A new method, termed autoprogressive training, for training neural networks to learn complex stress–strain behaviour of materials using global load–deflection response measured in a structural test is described. The richness of the constitutive information that is generally implicitly contained in the results of structural tests may in many cases make it possible to train a neural network material model from only a small number of such tests, thus overcoming one of the perceived limitations of a neural network approach to modelling of material behaviour; namely, that a voluminous amount of material test data is required. The method uses the partially-trained neural network in a central way in an iterative non-linear finite element analysis of the test specimen in order to extract approximate, but gradually improving, stress–strain information with which to train the neural network. An example is presented in which a simple neural network constitutive model of a T300/976 graphite/epoxy unidirectional lamina is trained, using the load–deflection response recorded during a destructive compressive test of a [(±45)₆]_S laminated structural plate containing an open hole. The results of a subsequent forward analysis are also presented, in which the trained material model is used to simulate the response of a compressively loaded [(±30)₆]_S structural laminate containing an open hole. Avenues for further improvement of the neural network model are also suggested. The proposed autoprogressive algorithm appears to have wide application in the general area of Non-Destructive Evaluation (NDE) and damage detection. Most NDE experiments can be viewed as structural tests and the proposed methodology can be used to determine certain damage indices, similar to the way in which constitutive models are determined. © 1998 John Wiley & Sons, Ltd.     

Fractional derivatives and recurrent neural networks in rheological modelling – part I: theory

The aim of this paper was to develop a general approach based on fractional time derivatives and recurrent neural networks to model the rheological behaviour of asphalt materials. The paper focuses on elastic and viscoelastic material characteristics. It consists of two parts. In this first part, the theoretical aspects of modelling are discussed. A brief introduction into the theory of rheological elements based on fractional time derivatives is provided. The fractional differential equation of a general rheological element (base element) is developed from which a huge variety of other rheological elements can be derived, e.g. fractional Newton, Kelvin and standard solid elements. A new approach is presented for solving the fractional differential equations. Artificial neural networks are developed to compute the stress–strain–time behaviour of fractional rheological elements in a numerical efficient way. The approach is tested and verified. The second part of this work will appear later. It will be focused on applications of the new theoretical work to pavement engineering problems.     

神经网络七十年: 回顾与展望

边界条件的施加是求解偏微分方程定解问题的重要步骤。神经网络方法求解偏微分方程定解问题时,将原问题转化为对应的构造变分问题后,损失函数是包含控制方程与边界条件的泛函。采用经典罚函数法及其改进方法施加边界条件时,罚因子的取值直接影响计算精度和求解效率;直接采用Lagrange乘子法施加边界条件,计算结果可能偏离原问题最优解。为破解此局限性,使用广义乘子法施加边界条件。基于神经网络获得原问题的预测解,再使用广义乘子法构建神经网络的损失函数并计算损失值,利用梯度下降法进行参数寻优,判断损失值是否满足要求;不满足则更新罚因子与乘子后再进行求解直至损失满足要求。数值算例的计算结果表明：与采用经典罚函数法、L₁精确罚函数法和Lagrange乘子法施加边界条件构造的神经网络相比,该文提出的方法具有更好的数值精度和更高的求解效率,且求解过程更加稳定。     

Nonlinear constitutive models for FRP composites using artificial neural networks

This paper presents a new approach to generate nonlinear and multi-axial constitutive models for fiber reinforced polymeric (FRP) composites using artificial neural networks (ANNs). The new nonlinear ANN constitutive models are complete and have been integrated with displacement-based FE software for the nonlinear analysis of composite structures. The proposed ANN constitutive models are trained with experimental data obtained from off-axis tension/compression and pure shear (Arcan) tests. The proposed ANN constitutive model is generated for plane–stress states with assumed functional response in some parts of the multi-axial stress space with no experimental data. The ability of the trained ANN models to predict material response is examined directly and through FE analysis of a notched composite plate. The experimental part of this study involved coupon testing of thick-section pultruded FRP E-glass/polyester material. Nonlinear response was pronounced including in the fiber direction due to the relatively low overall fiber volume fraction (FVF). Notched composite plates were also tested to verify the FE, with ANN material models, to predict general non-homogeneous responses at the structural level.     

Neural network constitutive modelling for non‐linear characterization of anisotropic materials

This paper presents a new technique of neural network constitutive modelling for non‐linear characterization of anisotropic materials. The proposed technique, based on a recently developed energy‐based characterization framework, derives the variations of the external work applied to and the strain energy induced in a specimen. The error between the variations of the energies is subsequently applied to correct the neural network properties by using a modified backpropagation algorithm. Unlike the conventional techniques for neural network constitutive modelling, the proposed technique develops models by quantifying the deformation of the specimen on a continuum basis. This allows the neural network constitutive models to be constructed from a single load test of one specimen. Numerical examples first examine the effect of specimen geometries and loading conditions. The effect of noise in the experimental measurements is subsequently investigated while having the applicability for non‐linear constitutive behaviour shown thereafter. The application for anisotropic materials is finally demonstrated by modelling a unidirectional lamina based on the measurements of a biaxial load test on a balanced laminate. Copyright © 2010 John Wiley & Sons, Ltd.     

Fault detection via nonlinear profile monitoring using artificial neural networks

Fault detection is the characterization of a normal behavior of a system using a response function or profile of interest and the identification of any deviation from such normal behavior. As system complexity grows, predicting the underlying structure or form of response function becomes challenging if not impossible. This article presents a data‐driven approach for fault detection of complex systems using multivariate statistical process control based on artificial neural network (ANN) characterization. In this approach, the quality of a system is characterized where one explanatory variable is adequately explained as a function of the other variables using an ANN model. The vector of weights and biases of the ANN model is monitored by using Hotelling T ² through control charts. The proposed method is tested and compared with existing methods such as polynomial and sum of sine function regression for 3 cases from the literature. Moreover, it is applied to a 4‐story reinforced concrete building that uses continuous monitoring to avoid potentially catastrophic failures. The proposed ANN approach outperforms the existing methods for small shifts (deviations) from healthy states. For large and medium shifts, it provides comparable results that are on the conservative side.     

Parameter estimation of an aeroelastic aircraft using neural networks

Application of neural networks to the problem of aerodynamic modelling and parameter estimation for aeroelastic aircraft is addressed. A neural model capable of predicting generalized force and moment coefficients using measured motion and control variables only, without any need for conventional normal elastic variables or their time derivatives, is proposed. Furthermore, it is shown that such a neural model can be used to extract equivalent stability and control derivatives of a flexible aircraft. Results are presented for aircraft with different levels of flexibility to demonstrate the utility of the neural approach for both modelling and estimation of parameters.     

基于CNN的点云图像融合目标检测

针对自动驾驶场景中目标检测存在尺度变化、光照变化和缺少距离信息等问题,提出一种极具鲁棒性的多模态数据融合目标检测方法,其主要思想是利用激光雷达提供的深度信息作为附加的特征来训练卷积神经网络(CNN)。首先利用滑动窗对输入数据进行切分匹配网络输入,然后采用两个CNN特征提取器提取RGB图像和点云深度图的特征,将其级联得到融合后的特征图,送入目标检测网络进行候选框的位置回归与分类,最后进行非极大值抑制(NMS)处理输出检测结果,包含目标的位置、类别、置信度和距离信息。在KITTI数据集上的实验结果表明,本文方法通过多模态数据的优势互补提高了在不同光照场景下的检测鲁棒性,附加滑动窗处理改善了小目标的检测效果。对比其他多种检测方法,本文方法具有检测精度与检测速度上的综合优势。     

Use of neural networks for fitting of FE probabilistic scaling model parameters

The probabilistic crack approach, based on the Monte Carlo method, was recently developed for finite element analysis of concrete cracking and related size effects. In this approach the heterogeneity of the material is taken into account by considering the material properties (tensile strength, Young modulus, etc.) to vary spatially following a normal distribution. N samples of the vector of random variables are generated from a specific probability density function, and the N samples corresponding to a simulation are functions of the mean value and of the standard deviation that define the Gauss density function. The problem is that these statistical moments are not known, a priori, for the characteristic volume of the finite elements used in the analysis. The paper proposes an inverse finite element analysis using neural networks for the determination of the statistical distribution parameters (e.g., for a normal distribution, the mean and the standard deviation) from a given response of the structure (for instance, an average load-displacement curve). From FE-analysis of 4-point bending beam tests, it is shown that the backanalysis technique developed in this paper is a powerful tool to determine the probabilistic distribution functions at the material level from structural tests for material volumes which are generally not accessible to direct testing. This revised version was published online in July 2006 with corrections to the Cover Date.     

Recognition and classification of single and concurrent unnatural patterns in control charts via neural networks and fitted line of samples

The correct, prompt recognition and analysis of unnatural and significant patterns in Schewhart’s control charts are very important since they remind out-of-control conditions. In fact, pattern extraction increases the sensitivity of charts when identifying out of control conditions. Artificial neural networks have been used to identify unnatural patterns in many research studies due to their high efficiency in pattern recognition. In most of such studies, there is a significant risk of misclassification of highly sensitive patterns. To put it more clearly, the proposed models offered for the recognition of patterns with low parametric coefficients are mistaken. This study, offers a model for the recognition and analysis of basic patterns in process control charts using LVQ and MLP networks along with a fitted line analysis. In this model, not only does risk of misclassification at different levels of sensitivity decrease remarkably, but there will also be the possibility for recognition and analysis when basic pattern occur simultaneously. The efficiency and effectiveness of the model are shown by conducting tests based on simulation.     

Determination of S–N curves with the application of artificial neural networks

The present paper describes the application of artificial neural networks for estimating the finite-life fatigue strength and fatigue limit. A comprehensive database with results of single-stage tests on specimens which simulate structural components is evaluated and prepared for processing with the use of neural networks. The available data are subdivided into different classes. A total of six different data classes are specified. The results of the prediction by means of neural networks are superior to those obtained with conventional methods for calculating the fatigue strength. The experimental results are estimated with high accuracy.     

Deep learning for photoacoustic tomography from sparse data

The development of fast and accurate image reconstruction algorithms is a central aspect of computed tomography. In this paper, we investigate this issue for the sparse data problem in photoacoustic tomography (PAT). We develop a direct and highly efficient reconstruction algorithm based on deep learning. In our approach, image reconstruction is performed with a deep convolutional neural network (CNN), whose weights are adjusted prior to the actual image reconstruction based on a set of training data. The proposed reconstruction approach can be interpreted as a network that uses the PAT filtered backprojection algorithm for the first layer, followed by the U-net architecture for the remaining layers. Actual image reconstruction with deep learning consists in one evaluation of the trained CNN, which does not require time-consuming solution of the forward and adjoint problems. At the same time, our numerical results demonstrate that the proposed deep learning approach reconstructs images with a quality comparable to state of the art iterative approaches for PAT from sparse data.     

基于SSM/I数据的神经网络方法反演海面风速

研究了单参数神经网络(SANN)模型、多参数神经网络(MANN)模型及复合多参数神经网络(CMANN)模型对使用SSM/I数据反演海面风速精度的影响,并对增加85.5GHz垂直和水平极化亮温作为输入项对反演精度的提高进行了验证.重点发展了一种新型的CMANN算法,并分析了它在不同风速范围下的反演效果.随着风速的增加,反演风速的精度也有提高,高风速(≥15m/s)较低风速有更小的风速误差.比较表明,此方法的反演效果优于以往的各种SSM/I反演风速算法.反演风速的范围为0～25m/s,在晴天和云天混合状况下反演风速与实测风速的均方根误差为1.61m/s,晴天则达到1.46m/s.     

Teeth category classification via seven-layer deep convolutional neural network with max pooling and global average pooling

Accurately classify teeth category is important in further dental diagnosis. Analyzing huge dental data, that is, identifying the teeth category, is often a hard task. Current automatic methods are based on computer vision and deep learning approaches. In this study, we aimed to classify the teeth category into four classes: incisor, canine, premolar, and molar. Cone beam computed tomography was used to collect the data. We proposed a seven-layer deep convolutional neural network with global average pooling to identify teeth category. Data augmentation method was used to enlarge the size of training dataset. The results showed the sensitivities of incisor, canine, premolar, and molar teeth are 88%, 86%, 84%, and 90%, respectively. The average sensitivity is 87.0%. We validated max pooling gives better results than average pooling. Our method is better than three state-of-the-art approaches.     

Three stage cervical cancer classifier based on hybrid ensemble learning with modified binary PSO using pretrained neural networks

ABSTRACT

Cervical cancer is one of the major challenges in developing nations like India.In recent years, a lot of research has been done todetect cervical cancer at an early stage through the pap-smear test, human papillomavirus test (HPV), etc. In this study, we have proposed athree-stage cervical cancer classifier to classify cervical cells among normal and abnormal cells using a hybrid ensemble classifier based onfeatures extracted using pre-trained neural networks. Furthermore, this work extends to classify the cells among different levels of dysplastic mainly mild, moderate and severe. The accuracy achieved for 2-class classification among normal and abnormal cells is up to 100% while for 4-class classification among normal, mild, moderate and severe dysplastic cells is up to 98.91% and 99.12% for new and old Herlev university hospital datasets respectively.