Artificial Neural Networks Approach in Microwave Filter Tuning

Abstract

This paper presents a novel method of cavity filter tuning with the usage of an artificial neural network (ANN). The proposed method does not require information on the filter topology, and the filter is treated as a black box. In order to illustrate the concept, a feed-forward, multi-layer, non-linear artificial neural network with back propagation is applied. The method for preparing, learning and testing vectors consisting of sampled detuned scattering characteristics and corresponding tuning screw deviations is proposed. To collect the training vectors, the machine, an intelligent automatic filter tuning tool integrated with a vector network analyzer, has been built. The ANN was trained on the basis of samples obtained from a properly tuned filter. It has been proved that the usage of multidimensional approximation ability of an ANN makes it possible to map the characteristic of a detuned filter reflection in individual screw errors. Finally, after the ANN learning process, the tuning experiment on 6 and 11-cavity filters has been preformed, proving a very high efficiency of the presented method.

Filter Tuning Based on Linear Decomposition of Scattering Characteristics

“…The described coarse set P C = {∆z, s C } has a very big advantage over the set P F = {∆z, s F }, collected randomly in the whole N R space considered in [4]. The random set P F has maximum (2K+1) R pairs, so it grows very fast with the filter tuning elements R and we are never certain that we have collected "appropriate" pairs to model the scattering characteristic optimally.…”

“…To model scattering characteristic s the only pieces of information that must be collected are these that stem from detuning a filter on a single element only while the others are set to their proper positions. It would allow us to reduce significantly the number of {∆z, s} pairs which need to be collected comparing to methods presented in [4][5][6][7].…”

“…On the other hand it can deal with many types of filters including very complex ones as shown in a tuning experiment, Section 5. An important advantage is a reduction of time needed to prepare a model for new filter type compared to methods [4][5][6][7]. A comparison of mentioned methods is depicted in Figure 1.…”

“…This mapping is performed with the use of artificial neural network (ANN) multidimensional approximator [4,5], neurofuzzy system [6] or linear matrix operator [7]. Especially [4,5] methods based on ANN are very well examined and proved to be very fast and able to deal with any type of filter (high complexity, many crosscouplings).…”

“…Similarly as in methods [4][5][6][7], method [8] considers direct correspondence between filter characteristics and tuning elements. Elaboration [8] is motivated by the fact that in methods [4,6,7] an inconvenient process of controlled, random filter detuning must be done for each filter type to train, e.g., Artificial Neural Network. That process takes a long time because it makes it necessary to collect a lot of random {∆z, s} pairs containing tuning element deviations ∆z and corresponding scattering parameters s. In [8] an approximator A: s→ ∆z is proposed, which requires a very coarse set of {∆z, s} pairs.…”