The research work that has been carried out is presented here in the form of brief scientific description of the main results. These are conclusions and comments of the detailed analysis and studies which are presented in the publications of this project.
The main ISIS requirements are formulated as follows:
high accuracy (small deviation of result of physical quantity measurement from its real value);
reliability (maintenance of a given measurement accuracy during long-time ISIS exploitation in industrial conditions);
adaptability (initial adaptation of ISIS structure to the requirements of measuring task and adaptation of processing algorithms of measurement results to the operation conditions).
A detailed analysis of Digital Acquisition systems (DAQ), has shown that:
their errors are normalized without errors of sensors of physical quantities;
initial error of the majority of new sensors (for example, see standards DIN IEC 751, IEC 584-2 and the advertisements of leading firms) exceed errors of DAQ systems from 5 : up to 200 times;
changing of conversion characteristics of the majority of sensors during exploitation (drift) is commensurable with initial sensor error and it is characterized by considerable individual deviations;
finally, the measuring error of physical quantity basically is determined by sensor errors instead of errors of DAQ systems.
The analysis of experimental research on sensor drift has shown that sensor drift prediction by known mathematical methods does not provide satisfactory results. It is depended on personal nature of drift and random error of real drift value definition during calibration.
Capability of neural networks (NN) after training to data generalization and data prolongation justifies their application for sensor drift prediction. The improvement of prediction quality for intercalibration interval increasing is possible at the expense of high-quality NN training. However, increasing of interesting interval reduces quantity of data for NN training and makes worse prediction quality
[10, 13]
. On the initial stage of system exploitation this contradiction becomes apparent for each sensor. The overcoming of such contradiction demands special development of NN technique
[9, 31, 38]
.
A new method of sensor drift prediction is proposed using an ensemble from two NN with different properties: approximating NN (ANN) and predicting NN (PNN)
[10, 22]
. ANN allows receiving sufficient sampling for training of PNN on the basis of 5-6 calibrations only. No more than 11% of percentage error of prediction has been achieved by six intercalibration intervals
[17, 38]
.
A new method of sensor drift prediction is proposed by integration of historical data about sensor drift (data about real sensor drift, which are exploited earlier in similar exploitation conditions)
[17, 22, 24, 36, 38, 44]
. The set of integrating historical data neural networks (IHDNN) is trained using such data, which allows receiving sufficient sampling for ANN training on the basis of two calibrations.
For the implementation of the proposed methods of sensor drift prediction the different NN architectures were investigated
[21, 26]
. A new algorithm of multilayer NN training is developed
[7, 33, 34]
, which has certain desired features over conventional algorithms: new approach to adaptive initializing of NN weight coefficients
[6, 10, 12]
andmodification of calculation of adaptive learning rate for linear and non-linear neurons
[2, 3, 11]
. The algorithms are implemented as software routines on C++.
The use of neural-based methods into ISIS has allowed proposing three-level structure, which is optimum for this application
[1, 4, 5, 10]
. Thus the training of neural networks is executed on the higher ISIS level and use of neural networks is limited to the middle ISIS level. Is it shown, that each ISIS level should execute the tasks in its own real time scale: for the lower, middle, higher level. Consideration of time scales has allowed us to distribute computational power to ISIS levels
[27, 28]
and to formulate the requirements of hardware and software components of all ISIS levels, taking into account the main requirements to ISIS as a whole
[5]
.
The following data acquisition units are developed in order to provide using heterogeneous and polytypic sensors on the lower ISIS level
[5, 9, 22, 25, 37, 41]
:
HSADC - 8-channel 12-digit analog-to-digital converter of digit-by-digit balancing on the basis of MC68HC05 microcontroller,
IADC - 8-channel 16-digit analog-to-digital converter with two-cycle weight integration on the basis of AT89C51 microcontroller,
SDAC - 8-channel 12-digit digital-to-analog converter on the basis of AT89C51 microcontroller with software selection of executed functions.
For providing of adaptability and universality, the controller on the middle ISIS level is developed on the basis of AT89C51/251 microcontroller with on-line remote reprogramming capability
[1, 4, 5, 16, 32, 35, 37, 42]
. The software of both controllers is developed and implemented using C language.
The software of the higher ISIS level is developed using client-server technology with the purpose of maintenance of real time scale of mathematical models replacement
[5, 14]
. The functional analysis has shown, that it is expedient to develop the following software routines: the supervisor, NN manager, accuracy manager, ISIS database
[14, 46]
. The software routines implemented on Delphi v.6.0 with usage of server-modules on Microsoft C ++ v.6.0. The main software functions are:
presentation of measurement results to the user;
communication with the middle level (loading of the current working routines to the middle level controllers, receiving the measurement results, replacement of the mathematical models of correction factors);
the control of adequacy of mathematical models of correction factors;
the NN training according to developed algorithms of sensor drift prediction. The machine-user interaction is implemented using Windows - interface.
Neural Network training on the higher ISIS level is analyzed with the purpose of decreasing of temporary complexity of iterative training algorithms. It is shown, that the sequential training of neural networks allows reducing total training time on 30% in comparison with multitasking Windows mode
[5]
. The parallel algorithms of NN training are developed
[29]
. The four schemes of parallelisation using MPI technology are implemented
[45]
. The high-performance multiprocessor computer Origin2000
[29]
and the segment of personal computers were used for the experiments. The different NN architectures at different training samples were researched. The results of researches on the computer Origin2000 have shown increasing of NN training speed in 3-4 times for predicting neural networks
[45]
and in 26-150 times for integrating historical data neural networks
[29]
.
The implemented ISIS prototype has high flexibility in application and can be easily adapted (by software) to the requirements of measuring task
[37]
. The testing of ISIS prototype was conducted using fast-response temperature sensors (30K5A1 thermistors)
[4, 22, 23]
. A testing method of sensor accuracy improvement using RTD Pt100 was used. The high speed of sensors drift was set by their operation near to high bound of permissible temperature. The experimental results have shown, that the application of the ensemble from three NN (set IHDNN+ANN+PNN) leads to an increase of the intercalibration interval 6-12 times, with simultaneous increase of measurement accuracy about 2-3 times and with drift influence decreasing in 3-5 times
[27, 28]
.
The method of scaled-down simulation was applied for ISIS verification. The mathematical model of sensors drift was constructed on the first phase (on the basis of experimental researches). The mathematical model accounts: (i) non-uniformity of drift in time; (ii) systematic and random errors of definition of real drift value
[46]
. A D/A converter was used as a test bench for the developed mathematical model of drift, which was included in the measuring circuit of sensor on the second phase. The evaluation has shown, that the application of developed methods allows to increase an intercalibration interval in 4-12 times at simultaneous decreasing drift influence in 3-5 times depending on kind of drift ("with saturation", "with acceleration", combination of drift).
Developed ISIS can be referred to intelligent systems because it has good adaptive capabilities at the expense of training. ISIS provides continuous increasing of measurement accuracy of physical quantities during exploitation at usage of non-accurate sensors due to data storage about sensors drift and periodic updating of individual mathematical models of sensors drift.
