Nigel Steele

Neuro-fuzzy control of a mobile robot

Abstract Fuzzy systems are able to treat uncertain and imprecise information; they make use of knowledge in the form of linguistic rules. Their drawbacks are caused mainly by the difficulty of defining accurate membership functions and lack of a systematic procedure for the transformation of expert knowledge into the rule base. Neural networks have the ability to learn but with some neural networks, knowledge representation and extraction are difficult. First, we consider a rule-based fuzzy controller and a learning procedure based on the stochastic approximation method. The radial basis function neural network is then considered and it is shown that a modified form of this network is identical with the fuzzy controller, which may thus be considered as a neuro-fuzzy controller. Numerical examples are presented to demonstrate the validity of the approach and it is shown that such an adaptive neuro-fuzzy system is successful in the control of a mobile robot.

Neuro-Fuzzy Control for Basic Mobile Robot Behaviours

The work described here was originally conceived in conjunction with employing a robotic system for cleaning the interior of railway carriages, although the ideas clearly extend to other industrial operations.

Closed-loop neural network controlled accelerometer

Abstract The purpose of this paper is to present aspects of an integrated micromachined sensor-neural network transducer development. Micromachined sensors exhibit particular problems such as non-linear characteristics, manufacturing tolerances and the need for complex electronic circuitry. The novel transducer design described here, based on a mathematical model of the micromachined sensor, is aimed at improving in-service performance and facilitating design and manufacture over conventional transducers. The proposed closed-loop transducer structure incorporates two modular artificial neural networks: a compensating neural network, which performs a static mapping, and a feedback neural network, which both linearizes and demodulates the feedback signal. Simulation results to date show an excellent linearity, wide dynamic range and robustness to shocks for the proposed system. The design was approached from a control engineering perspective due to the closed-loop structure of the transducer.

Radial basis function artificial neural networks for the inference process in fuzzy logic based control

This paper illustrates the fuzzy logic based approach to the control of a plant or a system, and discusses some of the possible shortcomings of the usual inference mechanisms. Radial basis function artificial neural networks have been shown to be effective in a number of applications, and have the advantage that network training is a very rapid process due to their structure. In fact, this is usually accomplished by the solution of a system of linear equations, a process for which fast and reliable algorithms are available. Radial basis function networks are shown to provide a means of constructing an ‘inference engine’ capable of handling a rule base in which plant state and control actions are specified in terms of fuzzy sets. The resulting inference mechanism is shown to avoid the phenomena of ‘rule overlap’ which can be a feature of fuzzy control algorithms. It is interesting to note that in this application of radial basis function networks, the usual problems on the number and location of centres do not arise. The paper concludes with a brief discussion of some experimental results achieved.ZusammenfassungDieser Artikel illustriert den auf Fuzzy-Logik basierenden Ansatz zur Steuerung eines Systems and diskutiert einige der möglichen Nachteile der üblichen Inferenzmechanismen. Es wurde gezeigt, daß neuronale Netzwerke mit radialen Basisfunktionen für viele Anwendungen sehr effektiv sind und den Vorteil haben, daß das Trainieren des Netzwerkes wegen seiner Struktur ein sehr schneller Prozeß ist. Praktisch wird dies für gewöhnlich durch das Lösen eines linearen Gleichungssystems durchgeführt, ein Schritt für den schnelle und robuste Verfahren verfügbar sind. Es wird gezeigt, daß neuronale Netzwerke mit radialen Basisfunktionen ein Mittel zur Konstruktion einer Inferenzmaschine sind, wobei eine Regel-Basis zur Steuerung von Systemzuständen und Systemänderungen durch Fuzzy-Logik dargestellt wird. Der resultierende Inferenzmechanismus vermeidet das Phänomen der Regelüberlappung, das bei Fuzzy-Logic-Steueralgorithmen auftreten kann. Es ist interessant, daß bei dieser Anwendung von neuronalen Netzwerken mit radialen Basisfunktionen die üblichen Probleme bei der Anzahl und Plazierung der Zentren nicht auftreten. Der Artikel schließt mit einer zusammenfassenden Diskussion und einigen experimentellen Ergebnissen.

Developing smart micromachined transducers using feedforward neural networks: a system identification and control perspective

Describes some possible applications of feedforward neural networks in the sensorial field. The subject of the research was a micromachined acceleration sensor, with a capacitive type of pick-off. Static sensor identification (based on measurement results) and dynamic identification (based on the mechanical model of the sensor) was performed with a view to develop, neural, open- and closed-loop transducers with improved performance characteristics. Measurement results are presented for the open loop, neural transducer, which was implemented in hardware. Two closed-loop structures were proposed which used static and/or dynamic networks. The performance of these transducers was assessed based on simulation results. All neural network controlled transducers showed an extended measurement range compared to the off-the-shelf sensors and, in the closed loop designs, the latch-up condition was eliminated.

Direct Inverse Control of Sensors by Neural Networks for Static/Low Frequency Applications

This paper addresses the issue of direct inverse control for two types of nonlinear transducer systems characterised by: piecewise linear input-output transfer function; hysteresis occurring in the input-output transfer function; with the aim of establishing whether some relationship exists between the severity of different nonlinearities and the complexity of the network required to control such nonlinearities in static/low-frequency sensor applications.

On parity problems and the functional-link artificial neural network

This paper contains the proof of a theorem on the capability of functional-link artificial neural networks both to represent and to learn the n-dimensional parity problem. The result is obtained by an embedding of the problem into a space of dimension 2n — 1. It is shown that the Volterra expansion of the data in n-dimensions provides the necessary transformation. By computing the parity function, it is shown that a suitable set of neural network weights can be deduced. Finally, it is demonstrated that 2n — 1 is the minimum embedding dimension for the problem.

Neural network based compensation of micromachined accelerometers for static and low frequency applications

In this work, a single-shot direct inverse compensation procedure based on neural networks is proposed, with application to micromachined accelerometers. Compensation was first considered from an empirical viewpoint to determine whether or not some kind of relationship exists between the severity of different nonlinearities and the complexity of the network required to control such nonlinearities. The procedure was then validated by applying direct inverse control to the measured static characteristic of a micromachined acceleration sensing element.

Modelling Non-Numeric Linguistic Variables

We consider how non-numeric linguistic variables may take their values from a pre-ordered set of vaguely defined linguistic terms. The mathematical structures that arise from the assumption that linguistic terms are pair-wise tolerant are considered. A homomorphism between tolerance spaces, filter bases and fuzzy numbers is shown. A proposal for modelling non-numeric linguistic variables with an ordered set of fuzzy numbers is introduced.

Development of a Prototype Neural Network Controlled Accelerometer

Abstract In this paper, sensors that perform the task of measuring the physical quantity of acceleration are discussed These sensors are of special significance since by integrating their output signal, accelerometers can additionally provide a measure of velocity and position. The system-level requirements for smart accelerometers are considered in this paper and a novel neural network based transducer design is presented which aims to satisfy these requirements. The design is based on a micromachined sensing element with capacitive pick-off. Nonlinear open-loop compensation of the sensor is performed based on direct inverse control. A prototype of the transducer has been built.