Our Portfolio: Characterization Projects

CMST Index | Characterization | TFA

Neural Network Raman Cone Penetrometer Signal Extraction and Enhancement Information
Resources

Technology
Summary Sheet

Technology
Summary Book

Tech ID: 242
Overview
A hybrid neural network (NN) system was developed for real-time data and analysis of chemical constituents in underground storage tanks (UST). This was coupled with an advanced LLNL fiber optic Raman spectroscopic probe for in situ cone penetrometer deployment. This supports the in situ Raman cone penetrometer development for chemical characterization of the Hanford underground storage tanks (USTs). The NN is utilized to identify and measure tank chemical constituents identified as safety and operational concerns to satisfy DQO that have been defined for UST waste retrieval and remediation efforts.
Technology Description
 A plug-in hardware module that enhances the capabilities of the Raman technique by performing signal extraction, automatic signal analysis, and feature identification on-line and in real time (one second or less per spectrum) was developed. The NN hardware is developed collaboratively with POC as a continuation of a project funded in FY94 by the Strategic Environmental Research and Development Program (SERDP) for the identification and enhancement of the Raman spectral signatures of chlorinated hydrocarbon solvents. POC’s NN system for the on-line analysis of tank waste component Raman spectral features is a pattern recognition system that combines conventional image processing and feature extraction methods with a proprietary hybrid NN. The neural network draws upon algorithms supplied by chemometrics, principal component analysis, cross correlation, acoustics, sonar, image processing, and oil exploration for a front end preprocessing package. These preprocessing algorithms are used to extract information from the complex raw spectral data, reducing large data sets to information on shapes, locations, intensities, ratios, and slopes of spectral features. The condensed features are fed into the input neurons of the NN for nonlinear processing and algorithms and NN architecture for identifying and measuring tank waste materials. LLNL is providing POC with the necessary Raman training spectral data for prepro-cessing algorithm and NN architecture development and optimization. These data consist of individual chemical components in concentrations ranging from 100 to 0.1 weight percent in both solid and aqueous matrices. These concentration ranges cover the concentrations of interest, as required by the waste tank retrieval and remediation DQOs. The tank constituents that are used for training the NN system for real time identification have been selected from the tank DQO lists and include, but are not limited to, sodium nitrate, sodium nitrite, bismuth phosphate, sodium carbonate, sodium sulfate, uranyl nitrate, sodium ferrocyanide, sodium nickel ferrocyanide, edta, tributylphosphate, acetone, sodium formate, dibutylphosphate, ammonium nitrate, sodium chromate, sodium dichromate, butanol, dimethylamine hydrochloride, formamide, sodium aluminate, sodium cyanate, and kerosene. POC is refining the NN package as needed to provide detection limits of tank constituents to necessary limits, greater than or equal to 0.1 weight percent.

Return to Top