Alexander Hosovsky

Fuzzy adaptive control for pneumatic muscle actuator with simulated annealing tuning

Pneumatic artificial muscles - based robotic systems usually necessitate the use various nonlinear control techniques in order to improve their performance. Moreover, their robustness to parameter variation, which is generally hardly predictable, should also be tested. Here a fast hybrid adaptive control is proposed, where a conventional PD controller is placed into the feedforward branch and a fuzzy controller is placed into the adaptation branch. The fuzzy controller compensates for the actions of PD controller under conditions of inertia moment variation. The design of fuzzy controller is based on the results of optimization using simulated annealing algorithm. The results confirm fast action of the control scheme as well as its robustness to changes in inertia moment variation.

Improved modeling of pneumatic muscle actuator using recurrent neural network

Derivation of models of complex nonlinear systems usually incorporates a number of simplifications in modeled phenomena with the level of these simplifications being dictated primarily by its intended purpose. If the overall model accuracy is insufficient, it might be helpful to use the powerful approximation capabilities of universal approximators like neural networks which are capable of approximating certain types of functions to arbitrary degree of accuracy. On the other hand, using black-box modeling techniques can impair the resulting extrapolation qualities of the model as well as eliminate its physical interpretation. Here an improved dynamic modeling of one-DOF pneumatic muscle actuator using recurrent neural network is proposed. The proposed method preserves the physical meaning of the model while improving its accuracy compared to the original analytic model. System and model responses are compared in closed-loop (using conventional PD controller) and all unmodeled dynamics is treated as disturbance which is identified using Elman neural network It is shown that the resulting model is applicable for model-based control system design with greater precision.

Implementation of wearable sensors for fall detection into smart household

The proposed paper describes implementation of wearable sensors into smart household environment. Wearable sensor prototype is based on MEMS inertial measurement unit, containing accelerometer and gyroscope. Fall detection using wearable sensors has been determined as suitable function in several implementations. Fall detection algorithm, considering information about acceleration, tilt and angular velocity was developed and tested. Communication of IMU with PC is described as well as experimental verification of the algorithm.

Monitoring of influence of significant parameters during anodizing of aluminium

The paper deals with the possibilities of control the technological process of aluminium anodic oxidation using the Design of Experiments (DoE) and the higher order neural unit to monitor the influence of the significant parameters on the resulting AAO (anodic aluminium oxide) film thickness. It also compares the relationship between individual inputs factors and their mutual interactions on the AAO thickness at monitored current density of 1.00 A·dm-2 and 6.00 A·dm-2. The developed predicted model describes the influence of input factors on the final AAO thickness by cubic function and its reliability is 99.37 % at current density of 1 A·dm-2 and 99.47% at current density of 6 A·dm-2. The electrolyte temperature and the size of an applied voltage had the most important influence.

Recognition of Assembly Parts by Convolutional Neural Networks

The paper describes the experiments with the use of deep neural networks (CNN) for robust identification of assembly parts (screws, nuts) and assembly features (holes), to speed up any assembly process with augmented reality application. The simple image processing tasks with static camera and recognized parts can be handled by standard image processing algorithms (threshold, Hough line/circle detection and contour detection), but the augmented reality devices require dynamic recognition of features detected in various distances and angles. The problem can be solved by deep learning CNN which is robust to orientation, scale and in cases when element is not fully visible. We tested two pretrained CNN models Mobilenet V1 and SSD Fast RCNN Inception V2 SSD extension have been tested to detect exact position. The results obtained were very promising in comparison to standard image processing techniques.

DESIGN OF ADJUSTABLE SMART VISION SYSTEM BASEDON ARTIFICIAL MUSCLE ACTUATORS

The paper describes a proposal of positional adjustable camera system based on artificial muscles for rapidly changing products with different height and sizes on a conveyor belt. Measuring accuracy by camera system is heavily dependent on the distance of the camera system from the captured part. The ideal situation is that component occupying a largest area of the processed image. Current positioning technologies on the principle of ball screws with an electric motor do not reach sufficient speed to dynamically adjust the position of the camera system. Traditional technologies significantly increase the risk of vibration of the whole structure in the repeated resizing captured parts. Artificial muscles due to its flexibility significantly reduce the risk of vibration and noise of designed mechanism. The proposed smart camera system can be used to measure the dimensions or fault finding on manufactured parts in automated lines.