• Journal of the Korean Institute of Intelligent Systems
• /
• v.15 no.1
• /
• pp.87-91
• /
• 2005

The cascade-correlation (CC) learning algorithm of Fahlman and Lebiere is one of the most influential constructive algorithm in a neural network. Cascading the hidden neurons results in a network that can represent very strong nonlinearities. Although this power is in principle useful, it can be a disadvantage if such strong nonlinearity is not required to solve the problem. 3 models are presented and compared empirically. All of them are based on valiants of the cascade architecture and output neurons weights training of the CC algorithm. Empirical results indicate the followings： (1) In the pattern classification, the model that train only new hidden neuron to output layer connection weights shows the best predictive ability; (2) In the function approximation, the model that removed input-output connection and used sigmoid-linear activation function is better predictability than CasCor algorithm.

• Steel and Composite Structures
• /
• v.18 no.3
• /
• pp.547-563
• /
• 2015

Deflection in a beam of a composite frame is a serviceability design criterion. This paper presents a methodology for rapid prediction of long-term mid-span deflections of beams in composite frames subjected to service load. Neural networks have been developed to predict the inelastic mid-span deflections in beams of frames (typically for 20 years, considering cracking, and time effects, i.e., creep and shrinkage in concrete) from the elastic moments and elastic mid-span deflections (neglecting cracking, and time effects). These models can be used for frames with any number of bays and stories. The training, validating, and testing data sets for the neural networks are generated using a hybrid analytical-numerical procedure of analysis. Multilayered feed-forward networks have been developed using sigmoid function as an activation function and the back propagation-learning algorithm for training. The proposed neural networks are validated for an example frame of different number of spans and stories and the errors are shown to be small. Sensitivity studies are carried out using the developed neural networks. These studies show the influence of variations of input parameters on the output parameter. The neural networks can be used in every day design as they enable rapid prediction of inelastic mid-span deflections with reasonable accuracy for practical purposes and require computational effort which is a fraction of that required for the available methods.

• Journal of Korea Society of Industrial Information Systems
• /
• v.24 no.4
• /
• pp.1-8
• /
• 2019

Plant phenomics is a technique for observing and analyzing morphological features in order to select plant varieties of excellent traits. The conventional methods is difficult to apply to the phenomics system. because the color threshold value must be manually changed according to the detection target. In this paper, we propose the convolution neural network (CNN) structure that can automatically segment plants from the background for the phenomics system. The LeafNet consists of nine convolution layers and a sigmoid activation function for determining the presence of plants. As a result of the learning using the LeafNet, we obtained a precision of 98.0% and a recall rate of 90.3% for the plant seedlings images. This confirms the applicability of the phenomics system.

• Journal of The Korean Society of Agricultural Engineers
• /
• v.54 no.4
• /
• pp.65-71
• /
• 2012

The objective of this study was to estimate surface runoff from rice paddy plots using an artificial neural network (ANN). A field experiment with three treatment levels was conducted in the NICS saemangum experimental field located in Iksan, Korea. The ANN model with the optimal network architectures, named Paddy1901 with 19 input nodes, 1 hidden layer with 16 neurons nodes, and 1 output node, was adopted to predict surface runoff from the plots. The model consisted of 7 parameters of precipitation, irrigation rate, ponding depth, average temperature, relative humidity, wind speed, and solar radiation on the daily basis. Daily runoff, as the target simulation value, was computed using a water balance equation. The field data collected in 2011 were used for training and validation of the model. The model was trained based on the error back propagation algorithm with sigmoid activation function. Simulation results for the independent training and testing data series showed that the model can perform well in simulating surface runoff from the study plots. The developed model has a main advantage that there is no requirement for any prior assumptions regarding the processes involved. ANN model thus can be a good tool to predict surface runoff from rice paddy fields.

• Journal of the Korea Society of Computer and Information
• /
• v.7 no.1
• /
• pp.174-183
• /
• 2002

In this paper a control chart pattern recognition methodology based on the back propagation algorithm and Multi layer perceptron, a neural computing theory, is presented. This pattern recognition algorithm, suitable for real time statistical process control. evaluates observations routinely collected for control charting to determine whether a Pattern, such as a cycle. trend or shift, which is exists in the data. This approach is promising because of its flexible training and high speed computation with low-end workstation. The artificial neural network methodology is developed utilizing the delta learning rule, sigmoid activation function with two hidden layers. In a computer integrated manufacturing environment, the operator need not routinely monitor the control chart but, rather, can be alerted to patterns by a computer signal generated by the proposed system.

• Journal of Korean Society on Water Environment
• /
• v.24 no.1
• /
• pp.91-98
• /
• 2008

This work was performed to develop a model possible to predict the influent flow and influent components, which are one of main disturbances causing process problems at the operation of municipal wastewater treatment plant. In this study, artificial neural network (ANN) was used in order to develop a model that was able to predict the influent flow, $COD_{Mn}$, SS, TN 1 day-ahead, 2day-ahead and 3 day ahead. Multi-layer feed-forward back-propagation network was chosen as neural network type, and tanh-sigmoid function was used as activation function to transport signal at the neural network. And Levenberg-Marquart (LM) algorithm was used as learning algorithm to train neural network. Among 420 data sets except missing data, which were collected between 2005 and 2006 at field plant, 210 data sets were used for training, and other 210 data sets were used for validation. As result of it, ANN model for predicting the influent flow and components 1-3day ahead could be developed successfully. It is expected that this developed model can be practically used as follows: Detecting the fault related to effluent concentration that can be happened in the future by combining with other models to predict process performance in advance, and minimization of the process fault through the establishment of various control strategies based on the detection result.

• Journal of radiological science and technology
• /
• v.46 no.4
• /
• pp.277-285
• /
• 2023

In this study, pneumonia identification networks with the small number of layers were constructed by using chest X-ray images. The networks had similar trainable-parameters, and the performance of the trained models was quantitatively evaluated with the modification of the network architectures. A total of 6 networks were constructed: convolutional neural network (CNN), VGGNet, GoogleNet, residual network with identity blocks, ResNet with bottleneck blocks and ResNet with identity and bottleneck blocks. Trainable parameters for the 6 networks were set in a range of 273,921-294,817 by adjusting the output channels of convolution layers. The network training was implemented with binary cross entropy (BCE) loss function, sigmoid activation function, adaptive moment estimation (Adam) optimizer and 100 epochs. The performance of the trained models was evaluated in terms of training time, accuracy, precision, recall, specificity and F1-score. The results showed that the trained models with the small number of layers precisely detect pneumonia from chest X-ray images. In particular, the overall quantitative performance of the trained models based on the ResNets was above 0.9, and the performance levels were similar or superior to those based on the CNN, VGGNet and GoogleNet. Also, the residual blocks affected the performance of the trained models based on the ResNets. Therefore, in this study, we demonstrated that the object detection networks with the small number of layers are suitable for detecting pneumonia using chest X-ray images. And, the trained models based on the ResNets can be optimized by applying appropriate residual-blocks.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.12 no.6
• /
• pp.672-680
• /
• 2019

This paper proposes "A Smart Farm Environment Optimization and Yield Prediction Platform based on IoT and Deep Learning" which gathers bio-sensor data from farms, diagnoses the diseases of growing crops, and predicts the year's harvest. The platform collects all the information currently available such as weather and soil microbes, optimizes the farm environment so that the crops can grow well, diagnoses the crop's diseases by using the leaves of the crops being grown on the farm, and predicts this year's harvest by using all the information on the farm. The result shows that the average accuracy of the AEOM is about 15% higher than that of the RF and about 8% higher than the GBD. Although data increases, the accuracy is reduced less than that of the RF or GBD. The linear regression shows that the slope of accuracy is -3.641E-4 for the ReLU, -4.0710E-4 for the Sigmoid, and -7.4534E-4 for the step function. Therefore, as the amount of test data increases, the ReLU is more accurate than the other two activation functions. This paper is a platform for managing the entire farm and, if introduced to actual farms, will greatly contribute to the development of smart farms in Korea.

• Journal of Biosystems Engineering
• /
• v.43 no.2
• /
• pp.148-159
• /
• 2018

Purpose: Unmanned air vehicle (UAV) remote sensing was applied to test various vegetation indices and make prediction models of protein content of rice for monitoring grain quality and proper management practice. Methods: Image acquisition was carried out by using NIR (Green, Red, NIR), RGB and RE (Blue, Green, Red-edge) camera mounted on UAV. Sampling was done synchronously at the geo-referenced points and GPS locations were recorded. Paddy samples were air-dried to 15% moisture content, and then dehulled and milled to 92% milling yield and measured the protein content by near-infrared spectroscopy. Results: Artificial neural network showed the better performance with $R^2$ (coefficient of determination) of 0.740, NSE (Nash-Sutcliffe model efficiency coefficient) of 0.733 and RMSE (root mean square error) of 0.187% considering all 54 samples than the models developed by PR (polynomial regression), SLR (simple linear regression), and PLSR (partial least square regression). PLSR calibration models showed almost similar result with PR as 0.663 ($R^2$) and 0.169% (RMSE) for cloud-free samples and 0.491 ($R^2$) and 0.217% (RMSE) for cloud-shadowed samples. However, the validation models performed poorly. This study revealed that there is a highly significant correlation between NDVI (normalized difference vegetation index) and protein content in rice. For the cloud-free samples, the SLR models showed $R^2=0.553$ and RMSE = 0.210%, and for cloud-shadowed samples showed 0.479 as $R^2$ and 0.225% as RMSE respectively. Conclusion: There is a significant correlation between spectral bands and grain protein content. Artificial neural networks have the strong advantages to fit the nonlinear problem when a sigmoid activation function is used in the hidden layer. Quantitatively, the neural network model obtained a higher precision result with a mean absolute relative error (MARE) of 2.18% and root mean square error (RMSE) of 0.187%.