Conducting Research: A Personal Technical Experience
-Part VII: At Two Institutions in Germany
By JIRARA
© JIRARA, September 2022, Published by JIRARA on matrubharti.com
All rights reserved. No parts of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, for any commercial purposes without the prior permission of the author and/or publisher.
Disclaimer: This is exactly the same as in the previous parts of the story.
**
The first part of the story is dedicated to Prof. Dr. S. M. Hamel, the then director of the IFM, Dr. K. F. Doherr, the group head, and Dr. R. V. Jategaonkar (RVJ), senior scientist with the group, for their technical and moral support during my visit to their institute, and for their sponsorship; and for their hospitality.
**
The second part of the story is dedicated to Prof. Florin Ionescu, professor and director of the institute of mechatronics, University of Applied Sciences (UAS), Konstanz, for inviting me on the DFG (Germany Research Organisation) fellowship for a period of three months, and for his (and his family’s) hospitality.
***
1. At the Institute of Flight Mechanics (IFM), DLR, Braunschweig; German Aerospace Institute (April-December 1994).
During a joint workshop-cum-seminar under the DLR cooperative program between IFM and FMCD, that was held in (November?) 1993 at FMCD-NAL (-CSIR, India), my group had presented three papers in the area of system identification and parameter estimation, and one of these was on parameter estimation using recurrent neural networks. This impressed the group coordinator from IFM, and he invited me to go there for 9 months on the existing cooperative exchange program to conduct advanced research on artificial neural networks and their applications for aircraft parameter estimation, and I had accepted this offer.
***
One scientist and the group leader from IFM had come to pick me at the Hannover airport, since I had flown from Bangalore to Frankfurt and then to Hannover.
I was put in a two-bedroom apartment in the first floor of a building in the area called Hondelage. They had bought some bread, apple and butter for me, since I was new to the place, and it was I think Friday or Saturday that I had reached there; and I was also supposed to immediately go to a nearby post-office to get myself insured for health, because if something happened to me in the next few days, it would be a very costly affair; I would have to pay lot of Deutsch marks.
Discussion of the plan for the research work:
On Monday, I went to the institute with the coordinating scientist with whom I was supposed to carry out the research, and we had a meeting with the group leader and this scientist; and luckily I had already thought out what I would do; perhaps this was also partially chalked out at the time of their visit to FMCD-NAL a few months ago, so it was just refinement of the work-time schedule, and I had 9 months to complete the planned work: a) study the theory of artificial neural networks (ANNs), b) study the possibility of applications of these to estimate the parameters of dynamic systems, especially for aircraft parameter estimation, c) develop the necessary algorithms, d) validate these techniques using simulated data, and e) publish a report or a paper as the case may be.
Actual research work:
I started studying the relevant research papers that I had brough with me, and decided to work out my own plan in more details to carry out the research. Since, I had to validate the algorithms, I needed the MATLAB software tool, and a desktop PC; these both were bought for me, since I was supposed to use these for next nine months.
**
In the process of my studies, I had found that the problem was more difficult than I had thought of; but before coming to IFM, I had done some preliminary work, as well as I had also submitted one paper in this area, which was under the review process with an international journal.
***
I had divided the total work in two areas, study of: a) various aspects of recurrent neural networks (RNNs), and b) feedforward neural networks (FFNNs); and also develop algorithms for their applications to aircraft parameter estimation.
****
a) RNNs: After extensive study of several papers, and doing the detailed analysis of the methods described in the open literature, I figured out that the RNN structure is highly amenable to be cast in the format of the so-called equation error method (EEM, which I had studied earlier in NAL), and then one can develop algorithms to estimate parameters of any dynamic system from its measured data; in fact, one needs the measurements of the states and their derivatives, which can be easily gathered.
**
I validated the algorithm with the MATLAB based realistically simulated data, and I found very encouraging results. I also figured out similarities and distinctions between 3 to 4 prevailing RNN-structures in the literature; then, I developed the equations for the trajectory estimation of dynamic systems using each structure of the RNNs; this was very revealing thing for me.
**
Meanwhile, my paper that was submitted to an international journal came back with the reviewers’ comments (and a note that the journal had been discontinued because one of its editors had expired). So, I had revised and enhanced that paper, and also had written two more papers from my work on RNN and its validation for parameter estimation; and I had submitted these three papers at different times to the IEE proceedings of Control Theory and Application, and all the three papers were accepted with only minor comments; since I had thoroughly checked the contents, equations, grammar and spellings.
**
Then, the RNN-based parameter estimation algorithm was used by RVJ for the real flight data and a paper of comparative results with his own filter error method (FEM) was subsequently accepted for an AIAA conference, USA; and later on he had gone to present the paper, by that time I had already returned to NAL.
b) FFNNs: I found that this structure can be used for senor fault detection, so I generated some simulated data and validated it and published a paper in the journal of aeronautical society of India, Bangalore.
Note: I had developed all the algorithms on my own, and I had also scripted the codes in MATLAB, and validated these for parameter estimation of dynamic systems; even though by that time MATLAB based neural network tool box was available.
**
Then, combining all the results of my study of these two types of ANNs, I wrote one brief report for the institute.
**
c) It was my inference that the estimation results by using RNN and FEM for the case of real flight data were almost the same or largely similar (AIAA paper); however, the FEM is very complicated and sophisticated software (developed by RVJ, at IFM, DLR), and would take more computational time compared to the RNN-based algorithm; and another notable feature of RNN is that it can be easily parallelised, and FEM is very difficult to parallelise, if at all it could be done; however, the ANNs are inherently amenable to parallelisation and very suitable for concurrent data processing, and hence such algorithms are also very useful for pattern recognition and image processing; and thereby offer several merits for being used for developing the algorithms, software, and computing systems, including embedded algorithms/systems for the design and synthesis of AI-based systems; especially mobile intelligent autonomous systems; in that case FEM can be considered as modern and yet classical approach that is tested and proven, and can be used as a bench mark method for state and parameter estimation.
**
So, in all I had published four journal papers, one AIAA conference paper, and one report from my work for the 9-months’ visit to IFM; and my work on (theory and practice of) parameter estimation using RNNs was original, and subsequently it had been followed and extended by some researchers in USA, and elsewhere. Subsequently, I had included some of the material from my papers in my IET/IEE (London, UK) book published in 2004.
2. At the University of Applied Sciences, Konstanz (March-May 2001).
Professor Florin had visited Bangalore earlier in 1999-2000 to present a technical paper at some conference; for which it seems he had needed a letter of invitation to facilitate his visa procedure, and for which he had written earlier to Dr. A. R. Upadhya, who was a senior scientist in a defence R & D lab in Bangalore; but, being in the defence unit, he could not give the letter so he had requested Dr. SSK, of FMCD-NAL for the letter, and in turn I drafted that letter which was sent to Prof, Florin; hence when he had attended the conference, I had gone to meet him, and then he also visited FMCD, and especially the system identification lab, and I had described our group’s activities, including the one on sensor data fusion; so impressed by it he had invited me to his institute for 3 months, for which after going back he had applied for the fellowship (including return airfare).
Discussion of the plan for the research work:
When my wife and I reached Konstanz (from Stuttgart, where we had gone by flight) by train, Prof. Florin had come to pick us up, and then he took us to the residence which was booked for us earlier.
Then, on Monday we had a small meeting wherein we discussed the plan of my work, since the total period of research work was only three months (12 weeks), he had insisted that I brief him every Friday on the progress of my ongoing work, and hence I had a very tight work-time schedule, and I had prepared a plan for the work for every day; a very difficult thing to do indeed.
Actual research work:
I had already done some initial work before coming to Konstanz, had collected the relevant papers on H-Infinity filters, since my proposal was to initiate research on the application of these filters for sensor data fusion.
**
Here also, I was given a desktop PC with MATLAB and I had started developing the algorithms for some H-Infinity filters, simulated two-sensor system for generating the appropriate data with measurement noise and started validating the algorithms with and without measurement-data loss.
**
I worked on three types of filters: i) Kalman-filter based gain fusion scheme, ii) H-Infinity based sensor data fusion, and iii) Global H-Infinity based fusion filter. I got very good results and wrote three papers which were accepted for three individual conferences in: i) Canada, ii) Japan, and iii) USA; the professor attended the first two; and this happened after I had left Konstanz after completing my work. I had also written one report for the DFG records that included the complete list of all the algorithms/programs developed by me.
**
Interestingly, I had also prepared the PPT-slides for all these papers, since the professor was to attend the conferences and present the results; mostly he had to go since he would get the travel funding easily, and I would not get, and also my term of the scholarship would have been over, being only three months, and the conference dates might be much later on; but I was happy that from my three months’ work I was able to make one report and three papers, which were accepted by the international conferences.
***
This also gave me good confidence that my new idea of application of H-Infinity filters to the problem of sensor data fusion was a viable approach and also the feature of measurement data loss worked very well in these filters. Subsequently, I had included some of the material from my papers into my book on sensor data fusion published by CRC Press, Florida, USA in 2009-10.
****
Some facts, and hiccups:
a) when I implemented the algorithms and ran them, no sensible results would come; every time I ran an algorithm, a different result would come, this perturbed me a lot, I went through some sleepless nights; and then one day I got the idea, that this might have had to do with the noise generation and addition to the measured true data; for generating realistic data it was necessary to add some noise in the data; but every time a different noise was getting added, and the results were varying; so then I fixed the seed numbers for generating the random noise and I started getting good and consistent results. For all the cases, either aircraft parameter estimation or sensor data fusion, it was necessary to try the working of these algorithms by adding some noise with varying intensity (covariance) and getting the consistent results;
**
b) When the new PC and MATLAB were bought and installed for me (with the internet connection), I had complained that my PC was getting corrupted by viruses, so my co-working scientist had requested one computer specialist to install the anti-virus software in my PC, and that was done. But, after a few days the group coordinator had come to see my results on the PC, and he found that I had anti-virus SW, and that since he had not loaded it, he quipped: from where did you steal this SW?; to this I really got annoyed, but since he himself had invited me and was a fund-giver, I did not show my displeasure to him; I had said that his scientist himself had come one day and loaded it; then he did not say anything. I had felt very bad, but gulped this insult, since I was supposed to be there for a long period of nine months; he might have had some history of the kind with someone who might have come on this kind of deputation from my or any other country; but the wrath of that had fallen on me inappropriately.
*****
