Sistem ve Kontrol Mühendisliği Anabilim Dalı

Çevresel algılama ve harıtalama ıle ınsan takıbı

This thesis is concerned with human following by a mobile robot that also needs to navigate in a socially compliant manner. The importance of this problem is due to the increasing usage of service robots in human-populated areas. This is a challenging task because the robot needs to consider proxemics while navigating. At the same time, it needs to keep the human in sight. In this thesis, this problem is studied from four aspects. First, an extensive evaluation of the previously proposed social navigation method, Social APF-RL, has been conducted to verify that the robot's navigation is socially compliant. Following, in order to better keep the human target within its field of view, a social robot head design has been realized based on a pan-tilt mechanism. Thirdly, human following in a socially compliant manner has been considered using the developed robot head, including body-head coordination while navigating. For this, a deep learning network based on reinforcement learning has been developed. Finally, recovery in case of human target loss is considered through human tracking and environmental sensing of doors using a specially trained YOLOv8 network. All the proposed methods are tested extensively in the Gazebo simulation environment.

Giyilebilir sensörler ile kas yükü tahmini

Assessment of workload intensity and muscle load measurement is important to understand the quality of the intended motion. There have been numerous attempts to quantify workload levels for different motion for different parts of the body using wearable sensors. This thesis investigates human locomotion, particularly walking and running, which are fundamental activities of daily life. Electromyography (EMG) is a technique to quantify the level of muscle activity. In this study, the aim is to lay foundation for feature engineering and to build a machine learning model to predict EMG sensor output with minimal number of sensors during walking. Two pressure insole sensors and a single EMG - IMU combined sensor have been used during the experiments. The process starts with building a machine learning model to predict EMG sensor output using only raw input signals, which provided a poor accuracy. In order to improve the model, lagged features, derivatives and window-based statistics like moving averages and standard deviations are added to feature set and accuracy is improved. Then the focus is put on hyperparameter tuning and regularization is introduced to decrease overfitting. Although these steps led to incremental improvements, the test performance eventually plateaued with overfitting. In order to overcome this issue, data is shuffled before splitting it into training and test sets. After shuffling during training and test data split, the model delivered its best results—both training and test accuracy improved, and overfitting was significantly reduced.

A data mining approach to predict failures in banking sector

A survey on systems architecture development and a helicopter communication system conceptual design

A system is commonly defined to be a collection of hardware, software, people, and procedures organized to accomplish some common objectives. These objectives are required by the stakeholders of the system. Systems are not developed at a point in time. The system development process to bring a system into being and into operational use from user requirements, requires a systems development life cycle approach that includes analysis, design, implementation, integration, maintenance and retirement. To obtain efficient systems, in the design process of the system, system?s architecture is built to manage to prevent design conflicts and undesired solutions. System?s architecting contributes to the development of a system from its initial concept until its retirement from use in this life cycle process.This thesis mainly focuses on system?s architecture design context and systems architecture design methodologies. In the first chapter of the study, system development life cycle models and system?s architecting design process is introduced. In addition, the context of the systems architecure is explained. In the next chapter, Structured Architecture Methodology and Object Oriented Architecture Methodology are introduced and explained. In the last chapter the study is concluded by architecting a communication system of an attack helicopter by Structured Systems Architecture Methodology.Key Terms: Systems Architecture Context, Systems Architecture Methodologies, Structured Architecture Methodology, Object-Oriented Architecture Methodology

A surveillance algorithm for fall detection and initiation of an e-mail

Demographic patterns are demonstrating that the world is a maturing society. Advances in medical treatments extend people's life spans. Elderly care is a burden on families' and states' budget. Moreover, elder people want to age in their place without the breach of their privacy and losing independency. Statistically unexpected falls happen to one third of individuals in excess of 65. This thesis focuses on fall problem on elderly, who age at their place. Smart assisting for elderly is an essential need for health care and emergency response when needed. Since sleep assistance is a complex subject, this project covers only the part that, if they fall while standing or sitting they would get help. Thus this thesis aims to help people age in their place by providing fall detection via image processing. The approach can be described as data analysis and programming an algorithm in the area of fall detection. The system analyzes minimum bounding rectangle of a moving object, considers aspect ratio, centroid and diagonal angle. In the literature fall algorithms use computationally expensive algorithms to distinguish the focused person in the image. In this study distinguishing an inactive person is not included considering that a fall necessarily contains motion. The person of interest is distinguished with image differencing. However this technique amplifies the noise in the binarized image. This issue is eliminated using Gaussian filtering on differenced image. Due to experiment constraints an adequate amount of statistics was not possible to collect. However human movements can be imitated with computer software literally. For academic purposes open access to this software can make statistics available for fall algorithms. A series of scenarios of fall is presented in section 2.7.4 and for each category a sample was recorded. Scenarios are a total of 20 with 4 recoveries and 5 fall-like cases. Among them only 11 are real fall cases and they are all detected as true positives besides one. The rest 9 cases are either with recovery or fall like cases and only one of them gives a false positive alarm. Hence the true positive percentage is 91 per cent while the false positive ratio is 11, 1 per cent.

Personalized product recommendation on second hand platforms

With the advent of online marketplaces which millions of people worldwide visit and make purchase every second, the shopping experience and competition between these platforms have been significantly changed and recommendation systems have become a more critical part of these platforms and gained popularity in the literature. One of these online marketplaces, in which the recommendation system plays a key role, is second hand platforms. In addition to general recommendation problems, these platforms have several problems which are specific to this domain such as compromising extremely unique item sets that makes the problem difficult with respect to other domains. In this study, we propose two staged model pipelines using state-of-the-art NLP techniques word2vec and paragraph2vec to address these problems with high quality personalized product recommendation in a scalable architecture. The model performance is evaluated on both offline experiments which are conducted on historical user clickstream dataset that is gathered from a popular second hand platform and A/B test on a production system. As a consequence of these experiments, the proposed model outperforms the baseline collaborative filtering-based models with respect to selected metrics, in addition, provides significant uplifts on several business metrics in the product system.

Design of a social robot and safe social navigation with deep reinforcement learning

This thesis is concerned with the design and development of a social robot that can navigate around in a socially compliant manner. The importance of this problem is due to the growing demand of using robots in human-populated environments. In this thesis, this problem is addressed in two concurrent parts. The first part has focused on the physical design and development of a social robot - named as SempRob. SempRob is aimed to have a sympathetic appearance while also having a design in which its visual sensors are located appropriately for environmental sensing. In the second part, the social navigation capability of the social robot is developed. First, a novel navigation method referred to as artificial potential function with reinforcement learning (APF-RL) method. In addition, an ellipse-based representation of obstacles is developed for efficient obstacle representation. Furthermore, environmental complexity measures are defined in order to ensure that learning scenarios incorporate a range of maneuvering difficulties. Both simulation and experimental results with SempRob demonstrate that APF-RL method enables the robot to move safely and efficiently in complex environments. Following, APF-RL method is extended to Social APF-RL method so that the robot additionally respects the comfort zones of the humans while navigating. This requires the robot to detect the humans in its surroundings and to track them spatially. A deep learning based human detection algorithm is combined with a Kalman filter for this purpose. Finally, Social APF-RL method is modified to be applicable in human following as well. All the proposed methods are tested on the developed robot successfully.

Deep learning based text regression

Most financial analysis methods and portfolio management techniques are based on risk classification and risk prediction. Stock return volatility is a solid indicator of the financial risk of a company. Therefore, forecasting stock return volatility successfully creates an invaluable advantage in financial analysis and portfolio management. While most of the studies are focusing on historical data and financial statements when predicting financial volatility of a company, some studies introduce new fields of information by analyzing soft information which is embedded in textual sources. Forecasting financial volatility of a publicly-traded company from its annual reports has been previously defined as a text regression problem. Recent studies use a manually labeled lexicon to filter the annual reports by keeping sentiment words only. In order to remove the lexicon dependency without decreasing the performance, we replace bag-of-words model word features by word embedding vectors. Using word vectors increases the number of parameters. Considering the increase in number of parameters and excessive lengths of annual reports, a convolutional neural network model is proposed and transfer learning is applied. Experimental results show that the convolutional neural network model provides more accurate volatility predictions than lexicon based models.

Nonlinear model predictive control based fuel-efficient adaptive vehicle spacing strategy for heavy-duty vehicle platooning

This thesis presents an approach for enhancing fuel efficiency in heavy-duty vehicle platooning through the implementation of an adaptive spacing strategy. The optimization design incorporates two crucial components, namely the nonlinear fuel consumption model of a diesel engine and the nonlinear air drag model. By integrating these elements into the overall cost function, a nonlinear model predictive controller is devised to calculate an adaptive time headway strategy. The primary focus is minimizing fuel consumption by adjusting the time headway which also affects the air drag coefficient. However, simply reducing the time headway and the air drag coefficient may not always be the most fuel-efficient strategy. In some scenarios, keeping up to the minimum set time headway can lead to excessive control effort, resulting in higher fuel consumption because the vehicle has to operate its engine within the inefficient fuel map region. The proposed dynamic strategy allows modifying the intervehicular distance within certain boundaries in order to optimize the potential benefits of aerodynamic drag reduction while respecting the engine's fuel map. To assess the effectiveness of the control design and validate the expected outcomes, extensive closed-loop simulations are conducted. A benchmark truck model is utilized, and various road topography conditions involving uphill and downhill slopes are considered. The simulation results underscore the efficacy of the adaptive time headway strategy in reducing fuel consumption for heavy-duty trucks across different scenarios. When compared to the lead vehicle, fuel consumption is reduced by up to 8%, and compared to a constant time headway approach, reductions of up to 3% are observed.

Human-like coordination of body-assisted arm movements for object manipulation

Manipulation is an integral capability for service robots. The goal of this thesis is to design and develop an approach that enables a mobile robot to mimic human manipulation abilities. We consider a differential type of mobile robot that is endowed with an arm and gripper. The robot is assumed to have visual sensing so that it can determine the relative position of the object of interest. First, it is observed that humans exhibit various basic modes of interaction with an object of interest, including extension, flexion, gripping, release and translation. As such, the robot can be programmed to have similar capabilities through establishing the correspondence between the robot and a human with respect to the underlying manipulation mechanisms. More complex behaviors such as putting, pulling, pushing, and shaking are defined using a sequential composition of basic operations. Second, humans are observed to achieve these tasks through the coordination of their body and arm movements. For this, a control approach in which the movements of the robot body and manipulator are coupled temporally and spatially is proposed. As such, if the object of interest is within the robot's reach, then only arm movements are made. If this is not the case, the robot starts moving its body. Depending on the vicinity of the object, this may be accompanied by arm motion or not. The control algorithm results in the robot's body and arm movements to be done in a coupled manner. The proposed approach is evaluated through an extensive set of experiments involving various manipulation tasks.

A reinforcement learning based controller to minimize forces on the crutches of a lower-limb exoskeleton

The majority of the metabolic energy consumption of a lower-limb exoskeleton user comes from the upper body effort, since the lower body can be considered to be passive. However, the upper body effort of lower limb exoskeleton users is ignored during motion controller development process in the literature. In this thesis study, deep reinforcement learning is used to develop a locomotion controller that minimizes the ground reaction forces (GRF) on crutches. The rationale for minimizing the ground reaction forces is to minimize the upper body effort of the user. A model of the human-exoskeleton system with crutches is created in URDF and XML formats. Reward functions that encourage the forward displacement of the center of mass of the exoskeleton-human system without falling and extreme joint torques are shaped. The state-of-the-art methods, Twin Delayed Deep Deterministic Policy Gradient (TD3) and Proximal Policy Optimization (PPO), are employed with the RaiSim and MuJoCo physics simulators and with different algorithm specific parameters in multiple training trials. The employed networks generate the joint torques based on the joint angle and velocities along with the ground reaction forces on feet and crutch tips. These generated joint torques are directly sent to the exoskeleton model and a new state is observed after implementing the action that the deep RL framework provides. Policies trained by the TD3 and PPO methods on RaiSim are observed to fail to generate proper control commands for a stable and natural looking gait. In general, it is observed that the PPO method generated higher rewards than the TD3 method on RaiSim. After failing to develop a desired policy with RaiSim, MuJoCo is employed as the simulator. Eventually, a policy that can generate a reasonable gait with a desired crutch usage and with 35% minimization in GRFs with respect to the baseline policy is developed.

Modular safety-critical control of legged robots

With recent performance improvements, legged robots will soon enter our lives to stay. Various control algorithms are already employed in deploying existing robots, and many more algorithms are still in the making. Safety concerns during the operation of legged robots must be addressed to enable their widespread use and ease their development. Especially machine learning-based control methods would benefit from model-based constraints to improve their safety. This thesis presents a modular safety filter to improve safety, i.e., reduce the chance of a fall of a legged robot. The availability of a robot capable of locomotion is assumed, i.e., a nominal controller exists. During locomotion, terrain properties around the robot are estimated through machine learning which uses a minimal set of proprioceptive signals. A novel deep-learning model utilizing an efficient transformer architecture is used for terrain estimation. A quadratic program combines the terrain estimations with inverse dynamics and a novel control barrier function constraint to filter and certify nominal control signals. The result is an optimal controller that acts as a filter and the filtered control signal allows the safe locomotion of the robot. The resulting approach is general and could be transferred with low effort to any other legged system.