joshi@wuche2.wustl.edu (Amol Joshi) (04/30/91)
The following tech report is available. This work was presented at the AIChE Spring Meeting, April 1991, Houston, TX. The departmental policy requires that requests for the tech report be official (e.g. on the company letterhead). The cost of the tech report is US$5.00 (check payable to `Department of Chemical Engineering'). A. Joshi (ps: We would like to express our gratitude to *Prof. Eugene Norris* of West Virginia U. whose backprop simulator, BPS, was used in this work.) ___________________________________________________________________ Process Trend Analysis Using the Frazier-Jawerth Transform and Neural Networks A. Joshi and R. Motard Department of Chemical Engineering Washington University in St. Louis St. Louis, MO 63130. Keywords: abnormality detection, multiresolution descriptions, pattern recognition, signal interpretation, time-frequency analysis ABSTRACT This work employs a novel signal processing technique, the Frazier-Jawerth transform (FJT) [1,2], in conjunction with an artificial feedforward neural network (FFN) to detect trends in process operational data. Two trained neural networks are employed --- one to assign a qualitative descriptive term to each signal and the other to detect process abnormalities by associating patterns among multiple signals. These tasks are critical to many applications of the AI technology to process control. A time-domain description obscures frequency characteristics of a signal whereas the information about its evolution in time is hidden in its Fourier domain representation. FJT coefficients provide a *joint* time-frequency description and thus make both time and frequency domain features in signals explicit. The paper argues that generating concise, explicit input data representations is an important step in improving the generalization properties of FFNs. Less importantly, the convergence properties of the FFNs are also improved. In this work, transformed signals are fed to a trained FFN that employs the conventional back-propagation algorithm. The FJT is closely related to the wavelet transform which is receiving much attention in the engineering community lately [3]. The advantage of using the FJT is that it is rather easy to construct the FJ analyzing functions such that they have small essential supports in both time and frequency domains. FJT can be used on-line and yields *multiresolution* descriptions [4]. An appropriate resolution can then be chosen to examine the trends. This is crucial in situations where signals have different characteristic time constants (e.g. signals from distillation columns). The FJT and neural nets together provide a promising paradigm for extracting information from sensor data for on-line applications. REFERENCES: [1] Frazier, M. and Jawerth, B., "Decomposition of Besov Spaces," Indiana U. Math J., Volume 34, No. 3, pp. 777-799, 1985. [2] Kumar, A., Fuhrmann, D., Frazier, M. and Jawerth, B., "A New Transform for Time-Frequency Analysis," submitted to IEEE Trans. Acoustics, Speech, and Signal Processing. [3] "A New Wave in Applied Mathematics," Science, Volume 249, pp. 858-859, August 1990. [4] Joshi, A., Kumar, A. and Motard, R.,"The Frazier-Jawerth Transform and its On-line Implementation," submitted to Computers and Chemical Engineering. -- ------------------------------------------------------------ Amol Joshi | joshi@wuche2.wustl.edu Department of Chemical Engineering | Washington University in St. Louis.|
enorris@gmuvax2.gmu.edu (Gene Norris) (05/01/91)
In article <1991Apr29.201633.13203@wuche2.wustl.edu> joshi@wuche2.wustl.edu (Amol Joshi) writes: > > >(ps: We would like to express our gratitude to *Prof. Eugene Norris* of >West Virginia U. whose backprop simulator, BPS, was used in this work.) Reports concerning Prof Norris' affiliation with West Virginia University are somewhat obsolete. Though he was a member of the WVU faculty from 1969 until 1972, he has been at George Mason University since 1980. Bps is available via ftp from gmuvax2.gmu.edu (129.174.1.8). Connect via anonymous ftp, cd to nn, read and take what you find. Executables for SUN 3, Sparc 1.0, MS-DOS, Ultrix, and Macintosh (old version) are there. Source licenses are available from the undersigned, who should be contacted for details. Prof. Eugene M. Norris CS Dept George Mason University Fairfax, VA 22032 (703)323-2713 enorris@gmuvax2.gmu.edu FAX: 703 323 2630