In factory automation, efforts are being made to increase productivity while maintaining high-quality products. In this study, a CNN network structure was designed to quickly and accurately recognize a cigarette located in the opposite direction or a cigarette with a loose end in an automated facility rotating at high speed for cigarette production. Tobacco inspection requires a simple network structure and fast processing time and performance. The proposed network has an excellent accuracy of 96.33% and a short processing time of 0.527 msec, showing excellent performance in learning time and performance compared to other CNN networks, confirming its practicality. In addition, it was confirmed that efficient learning is possible by increasing a small number of image data through a rotation conversion method.
Future autonomous vehicles need to recognize the ego lanes required for lane change and the side left and right lanes differently. Therefore, multi-lane recognition is needed. In this study, using the YOLO network, mainly used for object recognition, the proposed method recognizes the ego, left and right side lanes as different objects and identifies the correct lanes. As a result of the performance evaluation on the TuSimple test data, the proposed method recognized the ego lanes and the left and right side lanes differently. It showed very stable lane recognition results. And by detecting lanes that do not exist in the ground truth of TuSimple data, the proposed method is very robust in lanes detection. Nevertheless, studies related to learning data reinforcement in which lanes are located in the center or at the left and right edges of the image and accurate network learning for lanes are needed.
In this study, the multi-lane detection problem is expressed as a CNN-based regression problem, and the lane boundary coordinates are selected as outputs. In addition, we described lanes as fifth-order polynomials and distinguished the ego lane and the side lanes so that we could make the prediction lanes accurately. By eliminating the network branch arrangement and the lane boundary coordinate vector outside the image proposed by Chougule’s method, it was possible to eradicate meaningless data learning in CNN and increase the fast training and performance speed. And we confirmed that the average prediction error was small in the performance evaluation even though the proposed method compared with Chougule’s method under harsher conditions. In addition, even in a specific image with many errors, the predicted lanes did not deviate significantly, meaningful results were derived, and we confirmed robust performance.
위치 기반 AR 게임에서 모바일 기기의 위치를 정확하게 추정하는 것은 중요한 요소 중 하나이다. 모바일 기 기에 내장되어 있는 위치 추정 시스템이 얼마나 정확하게 위치정보를 추정하는 지가 증강되는 AR 콘텐츠의 정확도를 결정한다. 하지만 도시 환경에서는 건물, 건축물, 광고판, 표지판 등 지형지물 및 장애물에 의해서 위치 추정을 위해 필요한 신호가 반사, 굴절, 회절, 차단 등이 발생하게 되고 그로 인하여 모바일 기기의 위 치 추정 시스템으로부터 추정되는 위치에 오차가 생기게 된다. 본 논문에서는 도시 환경에서 모바일 기기의 위치 오차가 증가하는 현상에 대해서 상용 스트리트 뷰와 문자 태그를 활용하여 심플하면서도 신뢰성 있는 위치 보정 방안을 제안한다. 제안하는 위치 보정 방안은 쿼리 이미지와 스트리트 뷰 파노라마 이미지로부터 생성된 문자 태그를 대조하는 것을 통해서 매칭 스코어를 계산한고 매칭 스코어에 따라 쿼리 이미지를 촬영 한 실제 위치와 가까운 스트리트 뷰를 검색하는 것을 통해서 모바일 기기의 위치를 보정한다. 제안하는 위치 보정 방안은 위치 오차가 43.71m인 위치를 위치 오차가 4.09m인 위치로 보정하였으며 낮과 밤에 관계없이 위치를 보정할 수 있다는 장점이 있다.
We need data such as the number of lanes for lane change on the road as well as environmental and object recognition of the road for the autonomous vehicle of the future. This study proposed an algorithm that recognizes the left and right lanes and the center lane while driving differently from the black box image taken from a car. In general, deep learning does not recognize lanes individually but recognizes all lanes as only one lane. Therefore, using YOLO's object recognition function, the left and right lanes and the center lane were detected as different lanes, and a heuristic method was applied to recognize multi-lanes as more correct lanes. As a result of the performance evaluation, we confirmed that the proposed method detects the lane more accurately than Fast R-CNN and only YOLOv2.
This paper proposes a model predictive controller of robot manipulators using a genetic algorithm to secure the best performance by performing parameter optimization with the genetic algorithm. Genetic algorithm is a natural evolutionary process modeled as a computer algorithm and has excellent performance in global optimization, so it is useful for tuning control parameters. The sliding mode controller and inverse dynamics controller are included in the lower part of the model prediction controller to minimize the problems caused by non-linearity and uncertainty of the robot manipulator. The performance superiority of the proposed method as described above has been confirmed in detail through a simulation study.
The temperature distributions were numerically calculated for the two-dimensional transient conduction heat transfer problem of a square plate. The obtained temperature distributions were converted into colors to create images, and they were provided as learning and test data of CNN. Classification and regression networks were constructed to predict representative wall temperatures through CNN analysis. As results, the classification networks predicted the representative wall temperatures with an accuracy of 99.91% by erroneously predicting only 1 out of 1100 images. The regression networks predicted the representative wall temperatures within errors of C. From this fact, it was confirmed that the deep learning techniques are applicable to the transient conduction heat transfer problems.
To apply CNN to a fluid problem, we need a method to effectively convert the physical quantities of fluid into an image. The performance of CNN was evaluated using the image transformation method using the minimum and maximum values of the pressure distribution data and the image transformation methods using the normal distribution of the pressure distribution data. Through the performance evaluation of the learned CNN, the image transformation methods of Method 4 and Method 5, which applied the normal distribution of representative pressure distribution data, were very effective. In particular, Method 5 includes the initial and final pressure distribution data to include overall pressure distribution data, thereby improving the resolution of the color map to improve classification performance.
The numerical analysis of two-dimensional transient flow around the obstacle with rotated square cross sections was carried out. The obtained velocity distributions for each time step and each rotation angle were imaged to provide data for CNN(convolutional neural network). Both classification and regression neural networks were used for prediction of rotation angle. As results The classification method incorrectly predicted the rotation angle in only 2 of the 470 images. The regression method predicted the rotation angle errors within except 2 out of 470 images. From these facts, it could be concluded that both methods can be sufficiently applicable to the flow analysis.
This paper deals with the dynamic control of redundant robot manipulator. Traditionally, the kinematic control schemes for redundant robot manipulator were developed from the point of speed and used under the assumption that the dynamic control of manipulator is perfect. However, in reality, the precise control of redundant robot manipulator is very difficult due to their dynamics. Therefore, the kinematic controllers for redundant robot manipulator were employed in the acceleration dimension and may be combined with the computed torque method to achieve the accurate control performance. But their control performance is limited by the accuracy of the manipulator parameters such as the link mass, length, moment of inertia and varying payload. Hence in this paper, the proportional and derivative control gains of the computed torque controller are optimized by the genetic algorithm on the typical payloads, and the neural network is applied to obtain the proper control gains for arbitrary loads. The simulation results show that the proposed control method has better performance than the conventional control method for redundant robot manipulator.
The flow analysis of two dimensional transient flow over the obstacles with rectangular cross sections was performed. And 190 velocity distributions for each aspect ratio were imaged to provide input data for convolutional neural network learning. The classification and regression methods were used in estimating the aspect ratio from given velocity distributions. As a result the classification method was more exact than the regression method. But both the classification and regression methods gave relatively accurate prediction of the defined aspect ratio judging from the imaged velocity distributions. This confirms that the deep learning technique is applicable to the flow analysis.
Robot manipulators are highly nonlinear system with multi-inputs multi-outputs, and various control methods for the robot manipulators have been developed to acquire good trajectory tracking performance and improve the system stability lately. The computed torque controller has nonlinear feedforward control elements and so it is very effective to control robot manipulators. If the control gains of the computed torque controller is adjusted according the payload, then more precise control performance is attained. This paper extends the conventional computed torque controller in the joint space to the Cartesian space, and optimize the control gains for some specified payloads in both joint and Cartesian spaces using genetic algorithms. Also a neural network is employed to have proper control gains for arbitrary payloads using generalization properties of the neural network. Computer simulation results show that the proposed control system for robot manipulators has excellent performance in various conditions.
Steam tables including superheated, saturated and compressed region were simultaneously modeled using the neural networks. Pressure and temperature were used as two inputs for superheated and compressed region. On the other hand Pressure and dryness fraction were two inputs for saturated region. The outputs were specific volume, specific enthalpy and specific entropy. The neural network model were compared with the linear interpolation model in terms of the percentage relative errors. The criterion of judgement was selected with the percentage relative error of 1%. In conclusion the neural networks showed better results than the interpolation method for all data of superheated and compressed region and specific volume of saturated region, but similar for specific enthalpy and entropy of saturated region.
The model predictive controller performance of the mobile robot is set to an arbitrary value because it is difficult to select an accurate value with respect to the controller parameter. The general model predictive control uses a quadratic cost function to minimize the difference between the reference tracking error and the predicted trajectory error of the actual robot. In this study, we construct a predictive controller by transforming it into a quadratic programming problem considering velocity and acceleration constraints. The control parameters of the predictive controller, which determines the control performance of the mobile robot, are used a simple weighting matrix Q, R without the reference model matrix Ar by applying a quadratic cost function from which the reference tracking error vector is removed. Therefore, we designed the predictive controller 1 and 2 of the mobile robot considering the constraints, and optimized the controller parameters of the predictive controller using a genetic algorithm with excellent optimization capability.
After CNN basic structure was introduced by LeCun in 1989, there has not been a major structure change except for more deep network until recently. The deep network enhances the expression power due to improve the abstraction ability of the network, and can learn complex problems by increasing non linearity. However, the learning of a deep network means that it has vanishing gradient or longer learning time. In this study, we proposes a CNN structure with MLP layer. The proposed CNNs are superior to the general CNN in their classification performance. It is confirmed that classification accuracy is high due to include MLP layer which improves non linearity by experiment. In order to increase the performance without making a deep network, it is confirmed that the performance is improved by increasing the non linearity of the network.
Simultaneous modelling was carried out using the neural networks with three inputs including a distinguishing variable for the steam table. It covered whole steam tables including the compressed, saturated and superheated region of water. And relative errors of the thermodynamic properties such as specific volume, enthalpy, entropy were compared using the neural networks and the linear interpolation method. As a result of the analysis, The neural networks has proven to be powerful in modeling the steam table because it has slightly better results than the interpolation method.
The predictive control system using model-based predictive control is a very effective way to optimize the present inputs considering the states and future errors of the reference trajectory, but it has a drawback in that a control input matrix must be repeatedly calculated with a long calculation time at every sampling for minimizing future errors in a predictive interval. In this study, we applied the neural network simulating the predictive control method for the trajectory tracking control of the mobile robot to reduce complex control method and computation time which are the disadvantage of predictive control. In addition, the neural network showed excellent performance by the generalization even for a different reference trajectory. Therefore, The controller is designed by modeling the model-based predictive control gains for the reference trajectory using a neural networks. Through the computer simulation, the proposed control method showed better performance than the general predictive control method.
The state variables of saturated and superheated region in the steam table were simultaneously modeled using the neural networks. And the results were compared with quadratic spline interpolation and Lagrange interpolation. Two input data without distinguishing parameter were used in the neural networks. For comparison, quadratic spline interpolation method for superheated region and Lagrange interpolation method for saturated region were applied. The overall results revealed that the neural networks were greatly superior to quadratic interpolation method or Lagrange interpolation method.
Zermelo's navigation problem is that the ship reaches a particular target point in the minimum-time when it travels with a constant speed in a region of strong currents and its heading angle is the control variable. Its approximate solution for the minimum-time control may be found using the calculus of variation. However, the accuracy of its approximate solution is low since the solution is based on graph or table form from a complicated nonlinear equations. To improve the accuracy, we use a neural network. Through the computer simulation study we have found that the proposed method is superior to the conventional ones.
The steam table in saturated and superheated region was modeled simultaneously using the neural networks. A variable was introduced to distinguish between the saturation and the superheat. The relative errors were compared with the quadratic spline interpolation method. The relative errors by the neural networks were superior to those by the quadratic spline interpolation method over almost all ranges of temperatures and properties. The overall errors in the saturated region were better than those in the superheated region. From the analysis, it was confirmed that the neural networks could be a very powerful tool for simultaneous modeling of superheated and saturated steam table