Introductory Remarks

Within the Savannah River Site (SRS) and throughout the DOE complex, very large quantities of data have been, and continue to be, collected in programs for the National Security, Clean Energy and Environmental Management missions. Moreover, the data is obtained from measurements of multiple effects, ranging from visual morphology to spectroscopic analyses of material composition, along with thermochemical properties, transient and cyclic behavior, and many more. The value of the data is in its composite evaluation, utilizing the combined measurements, to provide a full understanding of system performance and the efficacy of the processes employed. Unfortunately, because of the size of the data sets, compilation and evaluation is time and cost prohibitive, especially when feature identification depends upon multiple attributes.

At the Savannah River National Laboratory (SRNL), a particular area of interest is the evaluation of image data used to inspect 3013 canisters, used to store Pu-bearing waste, for corrosion. The 3013 canister system consists of three nested canisters: a “convenience can”, containing the Pu-bearing material, that is sealed inside an inner 3013 canister serving as the primary containment, which is, in turn, placed in an outer 3013 canister that serves as a secondary containment. In safety evaluations of the 3013 system, it is required that there is no gas leakage from the inner 3013 can. The top of the inner 3013 canister is sealed by welding, and the weld-affected region is called the Inner Can Closure Weld Region (ICCWR). Over time, halides contained in the Pu-bearing material can be released from the convenience can into the gas space within the inner canister. In simulated exposures using test coupons, exposure to the halides have been found to produce corrosion preferentially in the ICCWR.

The inspection is part of the Material Identification and Surveillance (MIS) program for long term storage of plutonium-bearing materials in the DOE complex. The MIS project, which mandates this effort, has an objective to develop evaluation methodologies and metrics based on Laser Confocal Microscope (LCM) scans of cans having varying levels of corrosion. These scans are being used to analyze statistics for features associated with corrosion. After establishing the statistical protocol, SRNL must work through a 5-year backlog of images while continuing to receive new can images for analysis.

In the inspection process, randomly selected canisters are cut into sections and examined via LCM imaging. The LCM produces binary data files that contain surface height data, RGB color data (both optical and [optical + laser intensity]) and laser intensity data, yielding 8 layers of data per pixel. There are approximately 6000 images produced per can with approximately 7800 pixels per image, with the aforementioned 8 layers of data for each pixel. For evaluation, the binary data must be converted back to physical properties; namely height, RGB color and grayscale image values. Corrosion is associated with surface pitting, but not exclusively. Some pits are artifacts of fabrication, impact or other non-corrosion events. Manual observations identify corrosion via the combined properties of pit depth, area, edge contour, color and clustering. Given the volume of data and its complexity, an effort was under taken to implement numerical data analysis methods and machine learning algorithms. Prior to implementation of these methods, the LCM binary data must be processed to extract and interpret physical properties. Afterwards, feature classification in the context of known association with corrosion events can be used to construct labeled data for training and testing supervised machine learning algorithms.

Data analysis and progress toward implementation of machine learning algorithms for 3013 corrosion surveillance is the result is a collaborative effort by the computer sciences department at the University of South Carolina and the Savannah River National Laboratory. Implementation is carried out as a stepwise process.

LCM data is extracted from large binary files, with software written to convert the data to physical attributes (i.e. height, color and grayscale values; all as functions of location on a projection of the canister surface onto a plane). Since the sample surface is not flat, the software computes the mean surface height as a regression and uses it as a reference to determine the depth of depressions and protrusions. The user interface for the software permits selective downloading of the binary data and interrogation of the attributes. User input thresholds are used to flag attributes of interest. Physical attributes are used to compute features associated with corrosion, such as volumes, shape and cross-sectional areas of depressions, pits, cracks and prominences. The features are accessible to permit further interrogation of the surface data through the user interface, as well as for use in machine learning algorithms. As individual lCM images represent only a small part of the total surface, they must be stitched to give full coverage. The stitching routine accounts for variable overlay by matching relative surface height at the edges of the LCM images. Further, for each image the LCM measures surface height relative to an arbitrary reference. This means that the absolute heights reported by the LCM can differ discontinuously between adjacent images, resulting in a vertical mismatch between the images along the stitched edges. The software corrects for this by using a regression to correct the absolute height along the edges, thus smoothing the discontinuity.

At present, the above capabilities have been developed and tested for sample data. The software has been shown to clearly identify features associated with surface damage and corrosion. A user interface has been developed by USC to facilitate identification of pertinent data and to test the relationship between corrosion events and features. Results of the image analysis, including attribute interrogation, flagging and the computation of surface features will be shown in the presentation.

The next stage of the process will be feature classification, in the context of known association with corrosion events. The associative mapping will be used to generate labeled data sets for training and testing supervised machine learning algorithms for identification of corrosion events. Algorithm selection will be made to reduce training and test errors. Candidate software platforms for machine learning/deep neural network application include TensorFlow, Caffe and RapidMiner.