Instrumentation & Measurement

Application Category: Instrumentation & Measurement

Keywords: feedforward neural networks, microwave measurement, microwave six-port device, calibration

Company Name: Commonwealth Scientific and Industrial Research Organization (CSIRO) of Australia

Department: Division of Minerals

Main Investigator: Dr Yi Liu, Principal Research Scientist - Signal Processing & Process Control

Related Web Site: http://www.minerals.csiro.au

Contact Email: Y.Liu@minerals.csiro.au

Application Summary:

Six-ports have been gaining the popularity in many industrial on-line microwave measurement instruments. Many industrial applications of the on-line microwave measurements have been developed, such as on-line moisture measurement for coal and food, on-line carbon-in-fly-ash measurement for coal power stations, and multi-phase mass flow measurement for oil industry, etc. However, calibration of a six-port device remains a difficult task, particularly if a large dynamic range of the measurement is required. Firstly, accurate calibrations (usually involving non-linear fittings) are needed for each individual power detector (diode) over the dynamic range at the measurement frequencies. Then complex non-linear calibrations have to be carried out for both phase and attenuation.

The development of the artificial neural network (ANN) has provided a powerful tool for non-linear approximations (a black-box approach). A multi-layered ANN with enough neurons can approximate almost any non-linear input-output mapping at any required accuracy. An ANN calibration technique has been developed to calibrate six-ports for industrial on-line microwave measurement instruments using multi-layered neural networks.

We used a programmable vector modulator and an automatic network analyzer as well as some fixed attenuators, to set up the multiple standards for the calibration. The attenuations were set at about 2 dB steps from -10 dB to -60 dB (50 dB dynamic range) and the phases were set at about 10 degree steps from -180 degrees to 180 degrees. All true attenuation and phase values at each step were measured by the automatic network analyzer. A total of 1042 standards was set up and their attenuation and phase values were measured. Then the programmable vector modulator was connected to the six-port in the same way as to the network analyzer and ran through the same settings. All readings from the four power detectors were measured and recorded by the gauge computer. After detecting and omitting saturated power readings, we had a total of 990 data points.

We then randomly selected two thirds of the data (660 samples) as the training data set and the rest of the data as the cross-validation data set (330 samples). The ANN chosen for calibrating the attenuation measurement of the six-port had two hidden layers and one output layer. It had four inputs fed by the four power meter readings of the six-port. The first hidden layer had 10 neurons with hyperbolic tangent transfer functions. The second layer had 8 neurons with the same nonlinear transfer functions. The third layer is the output layer of the ANN with one linear node which predicted the attenuations (in dB). The ANN used for calibrating phase had a similar structure.

The ANN training was carried out on a 486/66 MHz computer using the Marquardt method instead of the standard back propagation algorithm. This algorithm offers very fast training performance compared with the standard back propagation and its variants. However, the memory requirement of the Marquardt algorithm is huge and this can make the algorithm impractical for very large neural networks or very large training data sets. Usual training time for the above ANNs was between 1 to 5 hours. The software package NeuroSolutions from NeuroDimension Inc. was later used to fine tuning the trained ANNs and then to generate C++ codes of the final ANNs for the implementation of the calibration in the gauge computers.

The technique simplifies the procedure to calibrate a six-port into a one-step process. The power measurements from four diodes are directly calibrated against the attenuation and phase shift measured by a network analyzer via multi-layered neural networks. Over the dynamic range of attenuation 50 dB and phase from -180 to 180 degrees, the ANN calibrations have achieved the accuracy of standard errors of 0.26 dB for attenuation and 2.76 degrees for phase for the calibration (training) data set, and standard errors of 0.31 dB and 3.05 degrees for the cross-validation data set (not used in the calibration). The accuracy achieved by the six-port with the ANN calibration are almost equal to the accuracy of the network analyzer used in calibration.

Clearly, the ANN technique for six-ports calibrations offers some advantages over the conventional calibration methods. It can achieve higher accuracy over a wider dynamic range. It has less stringent requirements on diodes’ linearity and uniformity. Although the ANN calibration process itself requires considerable computing power, it is however very simple to implement the calibration (the trained neural network) and only requires very little computing power (comparable with the linear calibrations).

Related Publications:

1. Y. Liu, "Calibrating An Industrial Microwave Six-Port Instrument Using Artificial Neural Network Technique", IEEE Trans IM, Vol. 45, No. 2, pp.651-656, 1996.
2. Y. Liu, "Industrial Microwave Six-Port Calibration: A Comparison of Two Nonlinear Techniques", in Proceeding of ICEMI’95, Shanghai, China, pp.139-142, Oct., 1995.

Download this Application Summary in Word97 format.
If you don't have Word97....Download the Word Viewer.