Volume 11 (2021): Issue 3 (July 2021)

This paper addresses the issue of data-driven smoothing parameter (bandwidth) selection in the context of nonparametric system identification of dynamic systems. In particular, we examine the identification problem of the block-oriented Hammerstein cascade system. A class of kernel-type Generalized Regression Neural Networks (GRNN) is employed as the identification algorithm. The statistical accuracy of the kernel GRNN estimate is critically influenced by the choice of the bandwidth. Given the need of data-driven bandwidth specification we propose several automatic selection methods that are compared by means of simulation studies. Our experiments reveal that the method referred to as the partitioned cross-validation algorithm can be recommended as the practical procedure for the bandwidth choice for the kernel GRNN estimate in terms of its statistical accuracy and implementation aspects.

Let a class of proper curves is specified by positive examples only. We aim to propose a learning novelty detection algorithm that decides whether a new curve is outside this class or not. In opposite to the majority of the literature, two sources of a curve variability are present, namely, the one inherent to curves from the proper class and observations errors’. Therefore, firstly a decision function is trained on historical data, and then, descriptors of each curve to be classified are learned from noisy observations.When the intrinsic variability is Gaussian, a decision threshold can be established from T ² Hotelling distribution and tuned to more general cases. Expansion coefficients in a selected orthogonal series are taken as descriptors and an algorithm for their learning is proposed that follows nonparametric curve fitting approaches. Its fast version is derived for descriptors that are based on the cosine series. Additionally, the asymptotic normality of learned descriptors and the bound for the probability of their large deviations are proved. The influence of this bound on the decision threshold is also discussed.The proposed approach covers curves described as functional data projected onto a finite-dimensional subspace of a Hilbert space as well a shape sensitive description of curves, known as square-root velocity (SRV). It was tested both on synthetic data and on real-life observations of the COVID-19 growth curves.

In the paper we develop an algorithm based on the Parzen kernel estimate for detection of sudden changes in 3-dimensional shapes which happen along the edge curves. Such problems commonly arise in various areas of computer vision, e.g., in edge detection, bioinformatics and processing of satellite imagery. In many engineering problems abrupt change detection may help in fault protection e.g. the jump detection in functions describing the static and dynamic properties of the objects in mechanical systems. We developed an algorithm for detecting abrupt changes which is nonparametric in nature and utilizes Parzen regression estimates of multivariate functions and their derivatives. In tests we apply this method, particularly but not exclusively, to the functions of two variables.

Handwritten text recognition systems interpret the scanned script images as text composed of letters. In this paper, efficient offline methods using fuzzy degrees, as well as interval fuzzy degrees of type-2, are proposed to recognize letters beforehand decomposed into strokes. For such strokes, the first stage methods are used to create a set of hypotheses as to whether a group of strokes matches letter or digit patterns. Subsequently, the second-stage methods are employed to select the most promising set of hypotheses with the use of fuzzy degrees. In a primary version of the second-stage system, standard fuzzy memberships are used to measure compatibility between strokes and character patterns. As an extension of the system thus created, interval type-2 fuzzy degrees are employed to perform a selection of hypotheses that fit multiple handwriting typefaces.

Over the last several decades, neuro-fuzzy systems (NFS) have been widely analyzed and described in the literature because of their many advantages. They can model the uncertainty characteristic of human reasoning and the possibility of a universal approximation. These properties allow, for example, for the implementation of nonlinear control and modeling systems of better quality than would be possible with the use of classical methods. However, according to the authors, the number of NFS applications deployed so far is not large enough. This is because the implementation of NFS on typical digital platforms, such as, for example, microcontrollers, has not led to sufficiently high performance. On the other hand, the world literature describes many cases of NFS hardware implementation in programmable gate arrays (FPGAs) offering sufficiently high performance. Unfortunately, the complexity and cost of such systems were so high that the solutions were not very successful. This paper proposes a method of the hardware implementation of MRBF-TS systems. Such systems are created by modifying a subclass of Takagi-Sugeno (TS) fuzzy-neural structures, i.e. the NFS group functionally equivalent to networks with radial basis functions (RBF). The structure of the MRBF-TS is designed to be well suited to the implementation on an FPGA. Thanks to this, it is possible to obtain both very high computing efficiency and high accuracy with relatively low consumption of hardware resources. This paper describes both, the method of implementing MRBFTS type structures on the FPGA and the method of designing such structures based on the population algorithm. The described solution allows for the implementation of control or modeling systems, the implementation of which was impossible so far due to technical or economic reasons.