Reference Number: INTAS-97-0606
INTAS logo Development of an Intelligent Sensing Instrumentation Structure I.S.I.S.
Using Neural Networks in ISIS
Home | The main features | Contradictions | Simulating of drifts | Combination of drifts | Error correction
The main contradiction at using of neural networks for sensor drift prediction
It is known that the quality of neural networks training depends on volume of data used for training in strong degree. It causes a main contradiction of neural networks using for sensors drift correction (see Figure 2). The high-quality neural network training allows sharply reducing the prediction error. It allows increasing of an intertesting interval that obtained by testing or calibration volume of data will appear insufficient for high-quality neural networks training.

Figure 2. The main contradiction at using of neural networks for sensor drift prediction
For solution of this contradiction, that is for artificial increasing of the training points number for predicting neural network it is proposed using two methods: (i) method of two neural networks with various properties and (ii) method of historical data integrating using special neural network. The historical data are the data of drift of the same type sensors in the similar exploitation conditions. Using proposed methods allows increasing data volume for neural network training without calibration or testing number increasing. The experimental researches using Integrating Historical Data neural network (IHDNN), Approximating neural network (ANN) and Predicting neural network (PNN) are considered below.