No description available.

No description available.

No description available.

The prediction results demonstrated that the method using six frames of variables as the input vectors for the BPNN model could improve the model prediction accuracy. Also, the number of nodes used in the hidden layer had a significant impact on the performance of the BPNN model. The results indicated that the best prediction accuracies in advance of a driver’s actual driving behavior with a lead time of 1s, 1.5s, and 1.8s were at 89.6%, 84.9%, 78.8% for merge events, and for non-merge events were at 92.2%, 87.5%, 81.1% respectively. / Recently, the applications of some driver assistance systems on vehicles have reduced vehicle accidents. However, studies have shown that the number of vehicle accidents caused by improper lane-changing behavior remains at a high level. Therefore, research has been focusing on developing a lane-changing assistance system to increase the safety level of driving in traffic. Many researchers have attempted to predict lane-changing behavior, and a general trend in the study of predicting driving behavior is the greater application of computational artificial intelligence. Artificial Neural Network (ANN) is one of the artificial intelligence methods, and it is well-known for its high reliability in a variety of applications. An ANN model can mimic human thinking and behavior due to its ability to capture the complex relationship among different variables in an environment of uncertainty. In this thesis, a BP (back-propagation) Neural Network model established by two methods was developed to predict a driver’s mandatory lane-changing decisions (merge or non-merge) at an early stage by considering driving environment features as the input vectors. Vehicle trajectory data from the Next Generation Simulation (NGSIM) dataset was used for training and testing the model. The results of the proposed model indicated that the prediction accuracies in advance of a driver’s actual driving behavior with a lead time of 1s, 1.5s, and 1.8s were at 89.6%, 84.9%, 78.8% for merge events, and for non-merge events were at 92.2%, 87.5%, 81.1% respectively. / Thesis / Master of Applied Science (MASc) / Lane-changing behavior at freeway on-ramps has a significant effect on driving safety and the stability of traffic flow. During the lane-changing process, the information processed by drivers is more complicated than that processed while remaining in a lane. If drivers fail to accurately judge the appropriate lane-changing time or the relative movement characteristics of related vehicles, vehicles accidents may occur. Thus, accurate prediction of lane-changing behavior is essential for a driving assistance system to ensure driver safety.

A hierarchical approach is typically followed in design of consumer products. First, a manufacturer sets performance targets for the whole system according to customer surveys and benchmarking of competitors' products. Then, designers cascade these targets to the subsystems or the components using a very simplified model of the overall system. Then, they try to design the components so that they meet these targets. It is important to have efficient tools that check if a set of performance targets for a component corresponds to a feasible design and determine the dimensions and mass of this design. This dissertation presents a methodology for developing two tools that link performance targets for a design to design variables that specify the geometry of the design. The first tool (called translator A) predicts the stiffness and mass of an automotive joint, whose geometry is specified, almost instantaneously. The second tool (called translator B) finds the most efficient, feasible design whose performance characteristics are close to given performance targets. The development of the two translators involves the following steps. First, an automotive joint is parameterized. A set of physical parameters are identified that can completely describe the geometry of the joint. These parameters should be easily understood by designers. Then, a parametric model is created using a CAD program, such as Pro/Engineer or I-Deas. The parametric model can account for different types of construction, and includes relations for styling, packaging, and manufacturing constraints. A database is created for each joint using the results from finite element analysis of hundreds or thousands of joint designs. The elements of the database serve as examples for developing Translator A. Response surface polynomials and neural networks are used to develop translator A. Stepwise regression is used in this study to rank the design variables in terms of importance and to obtain the best regression model. Translator B uses optimization to find the most efficient design. It analyzes a large number of designs efficiently using Translator A. The modified feasible direction method and sequential linear programming are used in developing translator B. The objective of translator B is to minimize the mass of the joint and the difference of the stiffness from a given target while satisfying styling, manufacturing and packaging constraints. The methodologies for Translators A and B are applied to the B-pillar to rocker and A-pillar to roof rail joints. Translator B is demonstrated by redesigning two joints of actual cars. Translator B is validated by checking the performance and mass of the optimum designs using finite element analysis. This study also compares neural networks and response surface polynomials. It shows that they are almost equally accurate when they are used in both analysis and design of joints. / Ph. D.

Musculoskeletal pain and injury continue to be prevalent sources of disability for thousands of workers in the U.S. every year. Proactive approaches to the reduction of this incidence attempt to prevent the injury by effecting task design so that human capabilities and limitations are driving factors in the task design and analysis process. Knowledge about the posture and kinematics that might be employed by an individual in performing a task is an important element of these proactive approaches to task design and analysis, especially for manual materials handling (i.e., lifting) exertions. In turn, accurate models that predict posture and kinematics can reduce the need for empirical postural and kinematic data in this task development process. Artificial neural networks were used in this investigation to achieve these predictions. As input, these networks received information about lift characteristics (e.g. target location, movement duration) and returned a predicted set of joint angles. Two types of networks were created, one to predict static posture based on target position, the second to predict the time histories of several joint angles (i.e., kinematics) as an object is lifted or lowered. Initial networks were created for sagittally symmetric lifts (two dimensions), but the final set of networks was expanded to make predictions for symmetric and asymmetric lifts in three dimensions. Networks were trained and verified with an empirical set of non-cyclic lifting motions. Notably, the within-subject variability in these motions was similar in magnitude to the associated between-subjects variability. In general, the networks were able to assimilate the data relatively well, especially in predicting kinematics, where root mean square errors were typically smaller than 20 degrees. These errors were similar in magnitude to the levels of within-subject variability observed in the dataset. Network performance also compared favorably to other existing models, typically resulting in smaller prediction errors than these other approaches. In addition, the internal connections of trained networks were examined to infer hypothetical motor control strategies. Results of this examination showed that feedback was an important component in providing kinematic predictions, whereas posture prediction benefited greatly from knowledge about individual anthropometry. Finally, potential improvements to increase prediction accuracy are discussed. Overall, these results support the use of artificial neural network models to predict posture and kinematics for lifting tasks. / Ph. D.

This thesis describes an automatic CT image interpretation approach that can be used to detect hardwood defects. The goal of this research has been to develop several automatic image interpretation systems for different types of wood, with lower-level processing performed by feed forward artificial neural networks. In the course of this work, five single-species classifiers and seven multiple-species classifiers have been developed for 2-D and 3-D analysis. These classifiers were trained with back-propagation, using training samples of three species of hardwood: cherry, red oak and yellow poplar. These classifiers recognize six classes: heartwood (clear wood), sapwood, knots, bark, split s and decay. This demonstrates the feasibility of developing general classifiers that can be used with different types of hardwood logs. This will help sawmill and veneer mill operators to improve the quality of products and preserve natural resources. / Master of Science

A neural network prediction method has been developed to compute self-noise of airfoils typically used in wind turbines. The neural networks were trained using experimental data corresponding to tests of several different airfoils over a range of flow conditions. The experimental data corresponds to the NACA 0012, Delft DU96, Sandia S831, S822 and S834, Fx63-137, SG6043 and SD-2030 airfoils. The chord of these airfoils range from 0.025 to 0.91 m and they were tested at Reynolds numbers of up to 3.8 million and angle of attack up to 15° depending on the airfoil. Using experimental data corresponding to different airfoils provides to the neural network the capacity to take into account the geometry of the airfoils in the predictions.geometry of the airfoils in the predictions. The input parameters to the network are the flow speed, chord length, effective angle of attack and parameters describing the geometrical shape of the airfoil. In addition, boundary layer displacement thickness was used for some models. The parameters used for taking into account the airfoil's geometry are based on a conformal mapping method or a polynomial approximation. The output of the neural network is given by sound pressure level in 1/3rd octave bands for nine frequencies ranging from 630 to 4000 Hz. The present work constitutes an application of neural networks to aeroacoustics. The main objective was to assess the potential of using neural networks to model airfoil noise. Therefore, this work is focused in the modeling of the problem, and no mathematical analyses about neural networks are intended. To this end, several models were investigated both in terms of the configuration and training approach. The performance of the networks was evaluated for a range of flow conditions. The neural network technique was first investigated for the NACA 0012 airfoil only. For this case, the geometry of the airfoil was not incorporated as input into the model. The neural network approach was then extended to account for airfoils of any geometry by including data from all airfoils in the training. The results show that the neural networks are capable of predicting the airfoils self-noise reasonably well for most of the flow conditions. The broadband noise due to the turbulent boundary layer interacting with the trailing edge is estimated very well. The tonal vortex shedding noise due to laminar boundary layer-trailing edge interaction is not predicted as well, most likely due to the limited data available for this noise source. In summary, the research here demonstrated the potential of the neural network as a tool to predict noise from typical wind turbine airfoils. / Master of Science

Aligning Word2vec word embeddings using a compass in a system of Compass-aligned Distributional Embeddings (CADE) creates stable and accurate temporal word embeddings. This thesis seeks to expand the CADE framework into the area of dynamic topic modeling (DTM), where temporal word2vec embeddings can be used to describe temporally and unsupervised evolving topics. It also seeks to improve upon the CADE framework through a theoretical and experimental exploration of compass parameters, cluster and topic generation techniques, and topic descriptor creation. This method of Temporal Topic Embeddings with a Compass (TTEC) will be compared to other DTM techniques in the ability to create coherent and diverse clusters and will be shown to be competitive compared to traditional and transformer-aided DTM architectures. In addition to a qualitative discussion of results, there will be a political theoretical overview of the nature of this technique and potential use cases, with interviews from political actors of various backgrounds as to how the technique and machine learning as a whole can be used in the organizational setting. / Master of Science / Diachronic word embeddings look at how the context words appear in evolve over time. Dynamic Topic Modeling (DTM) is the ability to computationally discover topics and how they evolve over time. This thesis creates a DTM technique called Temporal Topic Embeddings with a Compass (TTEC) based off diachronic word embeddings, allowing a user to simultaneously look at word and topic evolution over time. There is also an exploration of the use case of TTEC and similar machine learning models within various political organizational settings through interviews.

The proposed study develops a framework that can accurately capture and model input and output variables for multidisciplinary systems to mitigate the computational cost when uncertainties are involved. The dimension of the random input variables is reduced depending on the degree of correlation calculated by relative entropy. Feature extraction methods; namely Principal Component Analysis (PCA), the Auto-Encoder (AE) algorithm are developed when the input variables are highly correlated. The Independent Features Test (IndFeaT) is implemented as the feature selection method if the correlation is low to select a critical subset of model features. Moreover, Artificial Neural Network (ANN) including Probabilistic Neural Network (PNN) is integrated into the framework to correctly capture the complex response behavior of the multidisciplinary system with low computational cost. The efficacy of the proposed method is demonstrated with electro-mechanical engineering examples including a solder joint and stretchable patch antenna examples.