Neural Network Techniques for the Control and Identification of Acceleration Sensors

Abstract

This thesis reports the research carried out by the author to determine the appropriateness and feasibility of using neural network techniques to facilitate improved in-service performance of micro-engineered acceleration measuring devices. As a result of the original research work performed by the author, the feasibility of a sensor-neural network approach has been positively established and several original sensor control/identification configurations based on neural networks have been proposed.
The study has been performed both in simulation and experimentally. Whilst the research has been based on proprietary capacitive, bulk-micro machined acceleration sensing elements, the findings are generic and applicable to any capacitive accelerometers.
The motivation for the work stems from the fact that, in spite of advances in micromachining technologies, acceleration sensors exhibit imperfections such as offset, drift, non-linearity and noise, and the magnitude of these imperfections is found to vary. Moreover, fundamental characteristics of the sensor, e .g. sensitivity, .may be subject to manufacturing tolerances, varying material properties and ambient effects. In service, the capacitive sensors considered here exhibit another serious drawback, namely, an irreversible latch-up condition which may occur for input accelerations larger than the specified dynamic range of the sensor and for shocks in acceleration. The aim of the work therefore is to compensate for departures from ideal behaviour of such systems by using neural network based, nonlinear control techniques.
Solutions for eliminating latch-up conditions are also considered.
Based on the above, the system-level requirements for "smart" transducers are considered and an evaluation of existing open-loop transducers and a closed-loop linearly controlled prototype has been made. The main contributions in this thesis are the design and evaluation of two novel transducers and the identification of the sensing element.
The first design was of an open-loop transducer, in which a direct-inverse control strategy was adopted. The design procedure was based on measured data and validated initially in simulation for two acceleration sensors, of the same type. Further, a hardware neural, open loop transducer prototype was produced. The developed system is aimed at static/low frequency acceleration measurement applications. Although not able to avoid latch-up conditions, when compared with the "off-the-shelf', uncompensated sensors, the performance of the neural prototype was increased significantly.
The second design was of a closed-loop neural transducer, in which a non-linear gam feedback control strategy was adopted. A mathematical model of the sensing element has been used as a basis for the design work. Electrostatic forces were used for actuation. Neural networks are employed for both providing a linear feedback and for controlling the sensor.
The existing linearly controlled prototype is guaranteed to be stable only over the dynamic range of the transducer. In contrast, the neural system proposed here offers the prospect of overcoming this crucial deficiency by providing stability over double the dynamic range.
Further possible application-oriented modifications to the novel design are suggested in order to increase the dynamic range or to increase accuracy, as required.
A further contribution is that of exploiting the self-learning capabilities of neural networks in the context of sensor identification. In the development of non-linear, model-based control strategies the choice of the process/sensor model is of paramount importance.
Accurate mechanistic models (which would include manufacturing tolerances and faults) are difficult to generate for micro machined devices. The neural network approach, however, is a potentially valuable generic modelling technique which could provide for rapid non-linear model formulation, whilst also allowing the capture of essential process/system characteristics.
Although the starting point of the modelling process, in this work, was an available mathematical model of the sensor, it is shown that the identification procedure developed can be readily applied to physical sensors.

Date of Award	Apr 2000
Original language	English
Awarding Institution	Coventry University
Supervisor	Nigel Steele (Supervisor), Richard Rider (Supervisor) & Keith Burnham (Supervisor)