Browsing by Author "Tas, G"

Now showing 1 - 5 of 5
  • No Thumbnail Available
    Item
    Bloomed or non-bloomed fruit tree classification with transfer learning
    Tas, G; Bal, C
    Agriculture is one of the oldest and most important production sectors in the history of mankind. Agricultural producers suffer losses every year due to the difficulties arising from seasonal conditions. In this study, a deep learning network has been developed to detect bloomed/non-bloomed trees that will support the use of pesticides in order to protect farmers from the damage of frost. For this purpose, a new data set including bloomed/non-bloomed was created over two years. Bloomed or non-bloomed tree detection performances were evaluated using different convolutional neural network (CNN) models with transfer learning on this new data set. Accuracy comparisons are given by including deep learning structures as NASNetMobile, MobileNetV2, ResNet50V2, VGG-16, VGG-19, and InceptionV3, and CNN from Scratch in the evaluation. The results are presented by changing different epochs and optimizer hyperparameter values with different learning rates. In addition, the runtime and parameter values of the examined CNN models were also compared. For the CNN from Scratch model, six different models were obtained by changing the number of convolution blocks and epochs. It was observed that the ResNet50V2 model made the best prediction with 98.65% accuracy after RMSprop optimizer, 10-4 learning rate and 20 epochs of training.
  • No Thumbnail Available
    Item
    A New Lithium Polymer Battery Dataset with Different Discharge Levels: SOC Estimation of Lithium Polymer Batteries with Different Convolutional Neural Network Models
    Tas, G; Uysal, A; Bal, C
    In this study, a new dataset was created for use to estimate the state of charge (SOC) of lithium polymer batteries. A new experimental system was created to obtain the dataset by measuring the current, voltage, and temperature parameters of lithium polymer batteries. A convolutional neural network (CNN)-based deep learning model was used as the SOC prediction method. The effect of both batch size and dense network hyperparameter value on total parameter and deep learning error metric values for CNN-based lithium polymer battery SOC estimation is discussed. The proposed method, unlike deep learning models that require a high processing load in electronic cards, has provided remarkable results by being determined according to four different dense networks and two different batch size values. The proposed model has been obtained by performing experiments on optimizer, learning rate, dense network, and batch size values while determining the appropriate parameters to make successful predictions. The success of the CNN models was compared by conducting deep learning training on a computer with an Nvidia Gtx 1060 graphics card running the Ubuntu operating system. Adadelta optimizer achieved R2 0.977262 prediction success with learning rate 10(-2), batch size 5 x 10(2), dense 10(5) hyperparameter values. According to the results of the experiment, it was concluded that in the CNN deep learning method, large dense layers and small batch size values created less error in SOC estimation.
  • No Thumbnail Available
    Item
    Performance comparison of lithium polymer battery SOC estimation using GWO-BiLSTM and cutting-edge deep learning methods
    Tas, G; Bal, C; Uysal, A
    In this study, the GWO-BiLSTM method has been proposed by successfully estimating the SOC with the BiLSTM deep learning method using the hyper-parameter values determined by the GWO method of the lithium polymer battery. EV, HEV, and robots are used more healthily with successful, reliable, and fast SOC estimation, which has an important place in the Battery Management System. In studies using deep learning methods, it is important to solve the problems of underfitting, overfitting, and estimation error by determining the hyper-parameters appropriately. Thus, this study aims to solve an important problem by investigating the problem of determining the hyperparameter values for the deep learning method with metaheuristic optimization methods. This study was designed to compare the prediction success of the BiLSTM method trained with the optimal hyperparameter values obtained by the GWO method with cutting-edge deep learning methods trained with hyperparameter values obtained by trial and error. The success of the proposed method was verified by comparing the cutting-edge data-based deep learning methods and the BiLSTM method with the SOC estimation MAE, MSE, RMSE, and Runtime(s) metrics. According to the findings obtained during the hyperparameter determination studies, it takes longer time to determine the hyperparameters by trial and error than to determine the hyperparameters by metaheuristic optimization method when estimating lithium battery SOC with the deep learning method. Also, the GWO-BiLSTM method was the most successful method with an RMSE of 0.09244% and an R2 of 0.9987 values according to the average results of SOC estimation made with the lithium polymer battery data set, which was created by experiments performed at different discharge levels and is new in the literature.
  • No Thumbnail Available
    Item
    Estimation of PID parameters of BLDC motor system by using machine learning methods
    Tas, G; Özdamar, M
    Brushless Direct Current (BLDC) motor control in drones and unmanned aerial vehicles is critical for safety, performance, and high precision. In this study, a method based on machine learning rather than traditional methods is proposed to automatically control a system using a BLDC. An experimental system using a brushless direct current motor used in unmanned aerial vehicles was designed and a data set was created with the control studies. For the obtained data set, the Proportional- Integral- Derivative (PID) values were changed at certain intervals and the error values that occurred when applied to the system were recorded. The PID parameters obtained by seven different machine learning methods and the traditional method are compared. The performances of the machine learning methods were evaluated using regression estimation error metrics. According to the results obtained, Kp, Ki, and Kd values were applied to the system. The system response to sine input and step input is compared. When all machine learning experiments were evaluated, the Stochastic Gradient Descent (SGD) method was the most successful method, achieving 99.988% prediction success according to the R2 metric. When the results are analyzed, it is concluded that the system can be successfully controlled automatically using machine learning techniques.
  • No Thumbnail Available
    Item
    Classification of similar electronic components by transfer learning methods
    Tas, G
    Proper selection of electronic components and automated component identification is critical for fast production processes in industry. In addition, for Internet of Things (IoT) systems, accurate and fast selection of similar electronic components is an important problem. In this study, a transfer learning-based method is proposed to classify electronic components that are difficult to select due to their similarity. Eight different convolutional neural network (CNN) models and a novel model developed only in this study were used to classify electronic components. In addition to the transfer learning methods, the parallel CNN method, in which hyperparameter determination is done by trial and error, was developed and used to solve the classification problem. In addition to the transfer learning method, the parameters were tried to be determined by the trial-and-error method for hyperparameter selection. The effect of batch size and learning rate hyperparameter variations on the prediction success of parallel CNN models is analyzed. The effect of two different batch sizes and learning rate values for transfer learning models is also analyzed. Metrics such as confusion matrix, accuracy, and loss were used for evaluation methods. The number of parameters and runtime metrics of the models were also evaluated. All experiments were averaged to obtain a general intuition of success. The success of the proposed method is given by the evaluation metrics. According to the accuracy metric, the Densely Connected Convolutional Networks (DenseNet-121) model was the most successful model with a value of 98.2925%.