Göksu Taş

This study examines traditional and advanced Kalman filter methods for signal filtering and estimation. Synthetic sinusoidal data were first used, followed by human activity signals. Initially, the filtering performances of the standard Kalman Filter (KF), Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), and Bayesian-Optimized Kalman Filter (Bayes-KF) with automatically tuned parameters were investigated on synthetic signals. Human activity signals were estimated using acceleration data from the UCI's Human Activity Recognition (HAR) dataset. In terms of prediction accuracy, the Bayesian-KF approach outperformed the traditional KF, EKF, and UKF methods by improving the filter parameters (Q and R). Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) indicators were used to evaluate the performance of each approach. The Bayesian-KF method produced the lowest prediction error, with a value of 5.0996 x 10-5 based on the MSE metric. The comparative study demonstrates that Bayesian-KF is a reliable method for both synthetic and real-world human activity signals. Therefore, it is particularly useful in applications requiring motion analysis and sensor data processing.

The emergency stop moment is of critical importance for robots to meet legal and industrial safety requirements, protect human health and equipment, and respond quickly in times of crisis. In this study, the emergency stop time of the UR3 robot arm, which is widely used in the industry, was determined according to the axis speed, current value, and temperature values. The emergency stop moment of the robot was determined successfully with the state-of-the-art deep learning models. The performance of the state-of-the-art deep learning models was compared using the most common classification metrics. In addition, the most important hyperparameter values of the Bidirectional Recurrent Neural Network method were determined using the Egret Swarm Optimization Algorithm, which is new in literature. Among all models, the ESOA-BiRNN model was the most successful method with a value of 0.946124 according to the ROC AUC metric. The success of the proposed method is given together with other classification metrics.

In this study, the capacity estimation of the lithium-ion battery was successfully proposed using the genetic algorithm-Kalman filter method. During the design phase of electrical devices, lithium-based batteries have begun to be preferred instead of wired electricity transmission due to flexibility, freedom of movement, and portability problems. In addition, from an environmental perspective, the use of electric vehicles has become more important than the use of internal combustion engine vehicles. In this study, the capacity estimation of the 18650 lithium-ion battery, which is the most preferred in electric vehicles, was made quickly and accurately. The performances of the Standard Kalman Filter and the Kalman Filter, whose parameters are determined by the Genetic Algorithm, were compared by estimating the battery capacity. By giving the results obtained by the Genetic Algorithm during the parameter search process, the most appropriate values and the important parameters of the Kalman Filter have been determined. The success of the proposed method is given by the experimental results. In the performance comparison, the success of the proposed method is given using RMSE, MSE, and R2 metrics. When the average of all experiments was calculated using the R2 metric, the Genetic Algorithm-Kalman Filter approach achieved the best results in estimating the capacity of the 18650 lithium-ion battery, with a value of 0.999874.

In this study, the discharge energy of the lithium-ion battery was estimated by using the FOX- Bidirectional Recurrent Neural Network (Bi-RNN) method for the Dynamic Stress Test (DST) driving cycle method and different temperatures. For lithium-based batteries, discharge energy estimation is critical for long-term use, while problems such as overheating are major problems. For this reason, in this study, the discharge energy of lithium-ion batteries under different temperature conditions was estimated using bidirectional-based deep learning methods. In addition, the hyperparameter values of the BiRNN method were determined with FOX optimization, and the FOX-BiRNN method was proposed. The discharge energy estimations of FOX-BiRNN, BiRNN, Bidirectional Gated Recurrent Unit (Bi-GRU), and Bidirectional Long-short term (Bi-LSTM) methods were compared. The obtained estimation results were compared using the most commonly used battery parameter estimation metrics in the literature for performance comparison. The estimation success of the proposed method was presented using many comparison metrics and graphics. The FOX-BiRNN method was the most successful method for discharge energy estimation by obtaining values of %99.4186 at 0 0C according to the R2 metric, %99.6080 at 25 0C according to the R2 metric, and %99.4148 at 45 0C according to the R2 metric.

In this study, the temperature estimation of lithium-ion batteries is proposed by Chaos Particle Swarm Algorithm-Unscented Kalman Filter (UKF). 18650-type lithium-ion batteries are widely used in electric vehicles due to their compact design and long life. The accurate estimation of the temperature parameter of these batteries is critical for reasons such as balancing the performance and predicting chemical degradation. Therefore, in this study, the temperature parameter estimation of an 18650-type lithium-ion battery is made by UKF-based methods. Due to the intensive and mathematical processing load of the UKF method, the parameter values are determined by Chaos Particle Swarm Optimization (PSO) methods, and their estimation performances are compared. The parameter values such as alpha, kappa, and R matrix of the UKF method are determined by Particle Swarm Optimization (PSO), Chaos Particle Swarm Optimization (CPSO), Comprehensive Learning Particle Swarm Optimization (CLPSO), and hyperparameter values determined by trial and error. The hyperparameter values obtained from these four different methods were applied to the UKF method separately and their estimation performances were compared. The CPSO-UKF method became the most successful method by reaching an accuracy of 99.99228% in estimation according to the R2 metric. The success of the proposed method is also given with other regression metrics.