In today's world, there is a high pressure to change lifestyle, which is increasing the incidence of neurological diseases, such as Parkinson's disease. To assess motor dysfunction in these patients, approaches based on markerless motion capture (MMC) technology have been tested in recent years. Despite the high sampling rate and accuracy of commercial depth sensors such as the Leap Motion Controller (LMC), their versatile use is limited due to irregular sensing or processing errors. These affect their reliability and question clinically meaningful data. To mitigate the impact of errors during measurements, we introduce visual feedback for the specialist physician in the form of a real-time display of the measurement data recorded by the LMC. In this proof-of-concept study, we evaluate data from 10 patients with Parkinson's disease and 12 healthy subjects during the finger tapping test (FTT). To verify the suitability of using the LMC sensor for this purpose, we validate the results by simultaneous measurement with digital camera and two contact sensors: an accelerometer and two three-axis gyroscopes placed on the fingertips. The preliminary results confirmed the effectiveness of introducing visual feedback when performing FTT by reducing the impact of LMC sensor failure by 4.3%. Additionally, we used machine learning techniques to determine the clinical relevance of the measured and extracted features, achieving an average classification accuracy of 90.41%.
1 February 2025 in Journal of Electrical Engineering
The proposed article discusses the principles of implementing artificial intelligence (AI) methods in industrial control systems focusing on the deployment of neural networks (NNs) within programmable logic controllers (PLCs). Recent advancements in AI have led to significant improvements in modeling, control, quality control and predictive maintenance. This progress is further supported by the development of newer CPU architectures in PLCs, which offer enhanced processing speeds and capabilities. These advanced CPUs facilitate the implementation of more complex AI algorithms enabling systems to perform real-time data analysis. By leveraging the power of AI and improved hardware, industries can achieve higher levels of automation and better decision-making.
The paper deals with a soft computing state control method for multi input - multi output (MIMO) non-linear dynamic model of a robot. Soft methods based on neural networks and genetic algorithms have proven their effectiveness for this application. They are based on quite simple principles, but take advantage of their mathematical nature: non-linear iterative computation solutions. One way of controlling such nonlinear systems is to use of neural networks as effective controllers. In this paper a new methodology is proposed, where neural controller structure and parameters are computed by a genetic algorithm (GA). The proposed approach is represented by a direct neural controller using a multilayer perceptron (MLP) network in the feedback control loop. The training method using GA allows finding optimal adjustment of neural network weights so that high performance is achieved. The proposed control method is realized in Matlab/Simulink and demonstrated on a typical non-linear system with two inputs and two outputs (two-link robot).
1 October 2025 in Journal of Electrical Engineering
Nowadays, mobile robots require accurate and reliable control algorithms. The parameters of these algorithms are strongly influenced by the characteristics of the terrain over which the robot moves. The type of terrain can be identified using machine learning approaches. One promising approach is classification using measurements from inertial sensors. This method provides effective classification without the need for complex models. This paper deals with the analysis of different machine learning models using the vibration of inertial sensors. We experimented with various manual and automatic feature extraction techniques, each combined with different classification algorithms. At the conclusion of the evaluation phase, the most successful method from each feature extraction category will be selected based on performance metrics. These selected models were further validated on a custom dataset to assess their generalization capabilities in real-world conditions.
This paper explores extensions and restrictions of shallow convolutional neural networks with fixed kernels trained with a limited number of training samples. We extend the work recently done in research on Receptive Field Neural Networks (RFNN) and show their behaviour using different bases and step-by-step changes within the network architecture. To ensure the reproducibility of the results, we simplified the baseline RFNN architecture to a single-layer CNN network and introduced a deterministic methodology for RFNN training and evaluation. This methodology enabled us to evaluate the significance of changes using the (recently widely used in neural networks) Bayesian comparison. The results indicate that a change in the base may have less of an effect on the results than re-training using another seed. We show that the simplified network with tested bases has similar performance to the chosen baseline RFNN architecture. The data also show the positive impact of energy normalization of used filters, which improves the classification accuracy, even when using randomly initialized filters.