Application of neural networks to measurement of temperature sensor response time

A neural network, implementing the backpropagation paradigm, has been developed to predict the time constants of resistance temperature detectors (RTDs) from loop current step response (LCSR) test transients. It eliminates the difficulties involved in the LCSR application: complicated computation, specialized equipment, and highly trained personnel.

The neural network consists of three fully connected layers: an input layer, a hidden layer, and an output layer, with the number of input-layer processing elements (PEs) is varied from 20 to 60. The best results are obtained by the network consisting of 60 input- layer PEs, 150 hidden-layer PEs, and 1 output-layer PE.

A series of LCSR tests on 2 RTDs, the type of sensor used in most pressurized water reactors (PWRs) to trip safety systems, generates the response transients of the sensors, the input data of the networks. Plunge tests are used to determined the time constants of the RTDs, the desired output of the neural networks. Neural networks have been trained using these sets of input/output data from one RTD. The trained networks are used to predict the time constant of the other RTD. The time constant predictions of the trained networks produce the average relative error of about 5 percent.

In order to identify the network's sensitivity, tests with imperfect equipment have been performed to generate imprecise iii LCSR data and other tests use the LCSR data to which simulated noise has been added to the LCSR data.

The time constant predictions of the networks using the test sets of imprecise data produce the average relative error of about 6 percent. The average relative error of the time constant predictions of the networks using the test sets of data contaminated with 3 and 5 percent noise is within 8 percent. This indicates that backpropagation networks have been able to overcome contaminated data and equipment imperfections.