Neural Network Raman Cone Penetrometer Signal Extraction and Enhancement
This task focuses on Raman signal extraction, identification, and limit of detection enhancement using on-line neural network hardware in support of the in situ Raman cone penetrometer for chemical characterization of the Westinghouse Hanford Co. (WHC) underground storage tanks. The neural network will identify key tank constituents targeted as safety and operational concerns to satisfy data quality objectives (DQOs) defined for waste retrieval and remediation efforts.
The neural network will be developed collaboratively with an industrial partner, Physical Optics Corporation (POC). The tasks leading to successful development and deployment of the neural network signal extraction and enhancement package for tank characterization application are: (1) Raman training data collection, (2) neural network design and training, (3) neural network testing and revision, and (3) system integration with a cone penetrometer Raman spectrometer.
Progress
A review of tank waste hot cell characterization reports is in progress. The reports detail the results of characterization events that have occurred over the past 20 years using supernatant grab sampling and drill auger core sampling. The reported data has resulted in an expansion of the list of chemical constituents to be used for neural network training by POC. Additional chemicals include NPH (normal paraffin hydrocarbon) components, which will be used in the screening of tanks for organic and energetic DQOs, such as n-decane, dodecane, tridecane, tetradecane, and pentadecane. Additional organic chelating agents and solvents were identified as resulting from PUREX, B-plant, T-plant, and phosphate fuel reprocessing schemes. These include nitrilotriacetic acid, triphenyl phosphine oxide, acetate, succinate, butanedionate, and acetone among others. Additionally, the reports have provided information about the temperature and moisture content of the tank environment, allowing these variables to be included in training data sets. Chemical samples will be prepared based on the DQO and characterization report data.
A neural network training data set was collected using the cone penetrometer trailer Raman spectrometer and charge coupled device detector during the system's temporary sojourn at Lawrence Livermore National Laboratory. Training sets included spectra of system noise with both the laser on and off, fiber optic silica Raman background reflected off of a silvered and white granular surface from a 60' fiber-optic probe, dry sodium nitrate, and an aqueous sodium hydroxide solution of saturated sodium hydroxide. Blind testing data sets comprising of sodium nitrate at varying weight percent of aqueous sodium hydroxide solution were collected for testing the neural network algorithm for identifying two spectrally similar components in the presence of high noise and silica background. The data were submitted to POC for testing and analysis.
The PI was a presenting participant at the CMST-CP Annual Review Meeting.
PI: Kevin R. Kyle, Lawrence Livermore National Laboratory, (510) 423-3693
February 1996 Report
May 1996 Report
Report Index