I ' f - 9 6 b 0 625d C W F - Yb 0 769- 7 / AN EXPERT SYSTEM FRAMEWORK FOR NONDESTRUCIWE WASTE ASSAY G.K. Becker Idaho National Engineering Laboratory Lockheed Martin Idaho Technologies Company Idaho Falls, Idaho USA ABsrRAcr The management and disposition of transuranic (TRU) waste forms necessitates a determination of entrained TRU and associated radioactive material quantities as per National T R U Waste Characteriza- tion Program requirements. The technical justifica- tion and demonstration of a given nondestructive assay (NDA) method used to determine TRU mass and uncertainty in accordance with Program quality assurance objectives is a difficult task for many waste forms. Difficulties are typically founded in waste NDA methods that employ standards compensation techniques and/or the employment of simplifying assumptions regarding waste form configurations. The capability to determine and justify TRU mass and mass uncertainty can be enhanced through appropriate integration of waste container data/ information using expert system and empirical data- driven techniques in conjunction with conventional data acquisition and analysis methods. Presented is a preliminary expert system framework that integrates the waste form data base, algorithmic techniques, statistical analyses, expert domain knowledge bases, and empirical artificial intelligence modules into a cohesive system. T h e framework design and bases in addition to module development activities are discussed. BACKGROUND The ability t o manage and disposition TRU waste is predicated o n demonstrating compliance with appli- cable characterization requirements. Non-destructive waste assay methods are used in the National TRU Waste Characterization Program t o determine the mas of waste-entrained radionuclides. The capabil- ity and performance of waste NDA systems employed in the Program must comply with requirements as set forth in the TRU Waste Characterization Program Quality Assurance Program Plan (QAPP).' Parameters for which waste NDA system perform- ance must be demonstrated and technically justified include total bias and total uncertainty as defined per the QAPP. Within the instrument response envelope of commonly employed NDA techniques, certain standard compensation and bias quantifica- tion techniques can be derived to account for a number of prominent measurement bias sources. These techniques rely o n the knowledge of the bias and precision elements and the fact that there is a unique instrument response relation for each particular bias element. In reality, the functional form of a number of bias sources and their interrelationships are not necessarily known and understood for many waste forms. Additionally, each waste NDA technique has limitations where the response is non-unique, interfered with, or there is no measure of a particular waste form attribute from which bias arises. An example of an assay modality and waste form attribute where there is no measure is active thermal neutron interrogation and ffisile material clumping. This commonr& employed measurement modality has n o viable response indicator of the fissile material configuration for which an account can be made of the induced bias. Hence, alternate approaches to the interpretation/ integration of NDA data and supportive information enhancing the characterization capability base is of interest. In consideration of Program quality assurance objectives for total bias and total uncertainty and the reality of difficult-to-account-for waste form bias sources, it is logical to look to other mechanisms for acquiring and processing data/information to a m v e at a defensible assay solution. T h e solution may reside in the acquisition of additional waste form configuration information and/or deducing that one or more different measurement modalities would yield necessary data. Regardless of the approach, additional or reformatted information will in all probability be of value in completing the solution in DISCLAIMER Portions of this document may be illegible in electronic image products. Images are produced from the best available original document. a defensible manner. Another view to take in lieu of additional data/information is to ask if all presently available information and data confirm that the waste form configuration attributes are within the capability bounds of the employed NDA technique. This perspective or concept is most applicable in those cases for which alternate measurement modes and/or information are not available. In either case, consideration of the assay routine is required to EXPERT SYSTEM OVERVIEW either substantiate the implemented technique o r acquire and utilize additional datalinformation to Expert systems are computer programs that emulate justify and demonstrate that a solution with known the reasoning process of a human expert using uncertainty bounds has been achieved. previously established rules for a well-defined domain. They combine knowledge bases of rules and For most waste characterization programs, it is not domain-specific facts with information representing feasible to maintain technical experts in waste NDA specific instances acquired from the environment of to analyze the instrument response acquired on each interest. Functionally expert systems infer and draw waste container to conclude whether the technique conclusions from available data and information. A is applicable or whether additional information is particularly powerful feature is the interactive needed. Therefore, the ability is desired to assess acquisition of new and relevant datahnformation via instrument response and all pertinent datal questions. The expert system can also tell and justify information in an on-line fashion yielding a tangible how a given conclusion has been reached via an assay confidence factor related to compliance explanation facility, potentially giving the user insight criteria. A technology ideally suited to this task is o n the problem at hand. Expert systems are modular that of expert systems technology. In accordance in structure, typically consisting of the following with the concept of evaluating available datal parts: (1) a knowledge base, (2) an inference engine, information, a n expert system can be configured to (3) a global o r working memory, and (4) a user classify measurement data and associated waste form interface. The knowledge base contains the expert configuration information as either consistent or domain knowledge for use in problem solving. The inconsistent with respect to all available knowledge. working memory is used as a scratch pad and to Consistency leads to confidence in parameters of store interim reasoning data and infomation importance, Le,, fissile material mass and alpha provided to the system from the environment of activity and a quantifiable indicator o f quality interest. The inference engine uses the domain control. Inconsistency can take many forms as a knowledge together with acquired information to function of which particular sets of information and provide an expert solution. High-level ruies are data provide confirmatory evidence and which do implemented in the inference engine to avoid blind not. Regardless, significantly more is known about searches of the solution space. The user interface is the assay process in that the measurement results the link to the outside world and provides an support all other knowledge and vice versa o r that explanation facility to illustrate to the user how a inconsistencies a r e present, prompting an investiga- particular decision was reached. tion into technique applicability. There are many ways to represent knowledge in an An expert system can b e constructed with capabilities expert system. The three most popular representa- expanded beyond basic consistency testing, thereby tion schemes are rules, semantic networks, and employing additionaI evaluation tools and data from frames. The rule scheme is the primary method used other measurement modalities. Relevant analysis in the subject system. Rules a r e used to build a resources include exploratory statistical routines and domain-specific knowledge base represented via empirical artificial intelligence techniques such as domain facts and heuristics that specify a set of neural networks and fuzzy logic, hypothesis testing actions to be performed for a given situation. A rule twb, adaptive system learning routines, and waste is composed of a n antecedent and a consequent of form generation information bases. An expanded the general form: IF antecedent THEN consequent. expert system of this nature will necessarily be The antecedent of a ruIe is a set of conditions (or derived from experience gained with the construction conditional elements) that must be satisfied for the testing and validation of a system based on the rule to be applicable. The consequent of a rule is simpler consistency verification concept. interim goal of the data analysis component of the Waste Assay Measurement Information System project2 is to develop an initial phase expert system that addresses the concept of assay consistency with quantifiable compliance capabilities. the set of actions to be executed when the rule is applicable. Confidence measures are associated with the various rules to capture the probabalistic nature of the rules often called certainties. Several methods (such as tables of probability-related information) exist to define and represent rule certainties. An important advantage of expert systems is the ease with which knowledge bases can be modified as new rules and facts become known. An inference mechanism, or engine, using logical reasoning is necessary to determine the most appropriate response when the knowledge base is consulted. T h e inference engine is the driver program for the expert system using the knowledge base t o reach a particular conclusion. It is responsible for ensuring that questioning is done in a concise, logical manner, determining when to search the knowledge base for the information and scheduling other necessary actions. It will take action as indicated via a knowledge state found to be true based on the current facts presented to the expert system. The inference engine is, in effect, the intelligence that allows the expert system to make conclusions based on the expertise or knowledge stored in the knowledge base. Inference engine reasoning methods primarily operate in o n e of two ways. They may be data- driven, known a s forward chaining, or they may work backwards from conclusions or hypotheses, known as backward chaining. A forward chaining system begins with problem domain input data and moves down the inference chain until a conclusion is reached. Goal-directed reasoning is termed backward chaining where the system starts with a hypothesis or goal and works backward through the facts until it reaches a final node or conclusion. The actual reasoning process consists of constructing new rules or sentences from existing ones and ehsuring that the new ones represent facts that actually follow from the facts of previous sentences represented. Hence, the inference procedure derives logical sentences from t h e knowledge base that represent facts, which follow from facts embodied in the knowledge base. NONDESTRUCTIVE WASTlE ASSAY FRAMEWORK Waste NDA has many features suitable for the application of expert system technology. The measurement data set acquired from a given waste container may or may not embody infomation sufficient to extract the parameters of importance and the associated uncertainty. Use of an expert system allows one to utilize available knowledge concerning a given waste container in addition to the capabilities of the applied measurement technique to subsrantiate a viable solution. For instance, waste form generation process knowledge can bound and/or yield probable radioactive material composi- tions, radioactive material age, radioactive material c h e m i c a l c o m p o u n d , p r o b a b l e e l e m e n t a l compositions, probable density distributions, waste form configuration, etc. Oftentimes, the generator has previously acquired assay data in the same o r differing packaging configurations, which can be evaluated for consistency. Data bases containing previous measurement records can be consulted to see if certain statistical parameters fit historical distributions. Algorithmic assessment routines may be implemented to determine and/or evaluate specific parameter values per the acquired data and associated knowledge. Empirical artificial intelli- gence (AI) techniques can be utilized to look for data patterns and features that should be or not be present. Statistical hypothesis testing can be performed at various points in the problem domain to support t h e decision-making process of the system. The system capability of adaptive learning based on cumulative system experience could be exploited to enhance: the knowledge of t h e system. Finally and most importantly, the expert knowledge base allows the implementation of domain expert rules based in theory as well as heuristics. This is considerably more information that can be used to define and bound a solution than can be determined from the response of an NDA instrument calibrated in accordance with a specific regimen. An illustration of a preliminary expert system framework delineating applicable knowledge and dara base components is shown in Figure 1. Each component is defined based o n a particular type of knowledge or data/information source. These com- ponents or modules support the reasoning processes residing in the expert domain knowledge base. It is notable that such a modular arrangement supports the documentation, validation, and modification of the various knowledge and data modules. NDAFRAMEWORK COMPONEN?s T h e following is a brief description of the function of each module in the preliminary expert system. The . .- - __..,__i_. learning computation NDA System Measurement Data INPUT FIGURE 1. PRELIMINARY WASTE NDA EXPERT S Y S T E M FRAMEWORK modules are defined based on a logical partitioning of the knowledge and data/information space perti- nent to the problem. The collective set of modules is intended to contain all available data, information, and knowledge a n NDA expert would utilize to produce a best estimate o n the fissile material mass and associated uncertainty for a given container. The structure of the data and information contained in the module set to be used by the expert system is that of an object-oriented database? Master Expert System - The master expert system oversees operation of the system as negotiated by the expert domain knowledge based expert system. For exampIe, the master expert system interacts with the user, handIes resuft conflicts, and manages the adaptive learning controller based on the accumula- tion of data and information in the system over time. A part of the adaptive learning component would be a statistical analysis routine that continually updates parameters and correlations of interest to the expert system. Non-Destructive Waste Assay Measurement Data (System Input) - This input data vector contains pertinent raw acquired NDA measurement data in addition to the reduced data and associated parameters of significance. It constitutes the funda- mental data source from which the forward chaining inference procedure starts. T h e input data set format is such that differing fields can be extracted by other moduIes for various operations supporting the expert system analysis, e.g., empirical AI routines looking for patterns. Data preprocessing can also be implemented a t this stage, providing a format more amenable to the various components of the expert system. Expert Domain Knowledge Base Module (Expert System) - This knowledge base contains the funda- mental principles of the waste NDA problem domain in the form of rules and heuristics. This module contains the conditionaI questions and search direction data. Waste Generation Process Knowledge Base (Expert System) - This knowledge base contains previously established waste form attributes by generation process. Such attributes are the waste form genera- tion method and the resultant characteristics of the waste form. The knowledge base will be consulted in terms of the certain attributes variable range. The expert domain module will, as appropriate, consult this knowledge base regarding expected radionuclide compositions, waste form configurations, etc. Characteristics and attributes of the waste form must be consistent with measured parameters. Statistical Module - T h e employment of statistical models in the context of the expert system can take the form of a classical, hypothesis-driven o r confirm- atory data analysis approach o r a data-driven, exploratory data analysis approach. Applied statisti- cal methods in this module range through descriptive statistics, statistical inference procedures, hypothesis testing, goodness-of-fit tests, regression and correlation analysis, analysis of variance techniques, principal component analysis, classification via cluster analysis techniques, conditional expectation evaluations, linear discriminant functions, etc., for parameters of interest per predefined populations. Algorithmic Processing Module - This module contains first-principle algorithms for the evaluation of specific parameters. Examples include a figure of merit for quantifying the alpha,n component, the comparison of measured parameters to those predicted by MCNP models, and the execution of bounding computations to support validity determin- ations. Empirical AI Module - Empirical data exploration and processing techniques generally classified as artificial intelligence techniques are addressed in this module. The focus of these techniques is to perform classification functions using pattern recognition techniques, supervised and unsupervised learning and clustering techniques, and data-driven empirical methods. Examples include the class of neural networks, learning vector quantization, genetic algorithms, K-Nearest Neighbor regression, adaptive kernel methodsflocal basis function methods, fuzzy inference systems, fuzzy neural systems, generalized memory-based learning, and constrained topological mapping. O n e implementation has been the use of the Fuzzy A R W neural network for waste form matrix classification? Non-Destructive Examination Module - Important NDA parameters can be extracted from NDE data. Such parameters of use include the container fill height and the effective matrix density and its distribution. The availability of such information, appropriately formatted, is a useful input to the expert system. HistoricaVGenerator Data Module - Historical generator data include a great deal of information regarding the time of waste production and packag- ing and shipment, assay values acquired prior to shipment, waste form and configuration data, etc. All of this information can be used when bounding the uncertainty in the assay value. Ancillary Characterization Dataflnformation - This module contains information and data derived from specific characterization projects. An example would be a sludge drum coring project and the debris characterization project. In these efforts, drums were opened and sampled to determine and veriq their contents and radionucIidic compositions. Such data provide confirmatory evidence of actual waste form characteristics and serve as a benchmark when evaluating related containers. It is important to note that this is a comprehensive system that may take a considerable effort to estab- lish and validate. It is quite plausible to build the expert system in discrete subsets and apply each as developed. For example, t h e first stage of knowledge base development could be directed at evaluating the acquired measurement data with respect to algorith- mic bounding parameters and waste form generation data. Once established, the benefits of this expert system component could be realized whiIe other more-time-consuming components are in develop- ment and testing phases, an advantage of the modular nature of expert systems. Human experts do not proclaim to be exact in their reasoning and will factor a measure of uncertainty into such processes and final decisions. Likewise, expert systems have mechanisms for representing and managing uncertain knowledge entailed in the set of facts and heuristics as well as environmental data input to the system. There are a number of ways to represent uncertain knowledge, the most common being the use of probabalistic techniques, certainty factors, and fuzzy logic. Bayesian networks and fu logic techniques are currently under investigatio36 for application in the subject waste NDA expert system. Bayesian or belief networks can be used to represent uncertainty through the production of a compl model of a domain that is computation Fuzzy logic is currently being used as a classification tool and as a method to represent the degree of certainty in the classification. After the preliminary classification studies, investigation into the utility of a fuzzy expert system through the assignment of rules, heuristics, and W t e m input to fuzzy sets will be undertaken. In such a system, the inference engine provides a fuzzy output, which may need to be defuzzified depending on the application domain. The management of uncertainty may also be treated via evidential reasoning. Evidential reasoning is based o n the Dempster-Shafer theory (DST) and is an effective method to represent ignorance, incom- plete information, o r inexact rules in expert systems. DST also provides a mechanism to handle conflicting data and rules and is ideal to integrate knowledge from different sources. CONCLUSIONS Waste NDA techniques in many cases do not take advantage of all available knowledge about the problem domain, which can lead to difficulty in technically justifj4ng adherence to compliance criteria. Experts system technology has been identi- fied and investigated as a viable means to accomrno- date a11 available domain-specific knowIedge into a cohesive system. A system for t h e accumulation of available knowledge for use by t h e expert system has been developed, the Waste Assay Measurement and Integration System. The identification of waste NDA data/information sources and means to represent and evaluate such knowledge and t h e associated uncertainty are under study. The expert system is compatible with compliance demonstration activities, and the solution technique is tractable in that the explanation facility output details those factors used to determine the soIution and how uncertainty was treated in each particular case. REFERENCES 1. Transuranic Waste Characreniation @ality Assurance Program Plan, Revision 0, CAO-94-1010, U.S. Department of Energy, Carlsbad Area Office, National TRU Program Office, April 30, 1995. 2. T. J. Roney, G. K. Becker, and K, Mousseau, Methods to Improve the Interpretation of Data porn NDA and NDE of Waste Containers, A Summary of the Waste Assay Measurement Integration System, INEL-96/233, Idaho National Engineering Laboratory, July 1996. 3. G. K Becker et al., "Utility of Neural Networks in Nondestructive Waste Assay Measurement Methods," Proceedings of the 4th Nondestructive A s s q and Nondesmtctive Examination Waste Characterization Conference, October 24-26,1995, Salt Lake City, UT. 4. J. Hodges, S. Bridges, S. Yie, Y. Johnson, and M. Gentry, The Design of an Object-Oriented Radiological Waste Database, MSU-960617, Department of Computer Science, Mississippi State Universiry, MS, June 17, 1996. 5. J. Hodges, S. Bridges, and S. Yie, Preliminary Results in the Use of Fuzty Logic for a RadioIogicaI Waste Characterization Expert System, MSU-960626, Department of Computer Science, Mississippi State University, MS, June 26, 1996. 6. S. Bridges, J. Hodges, and B. \yooley, Preliminary ResuIts in the Use of Bayesian Network for a Radiological Waste Characteriration Expert System, MSU-, June 17, 1996, Department of Computer Science, Mississippi State University, MS. This work was supported by the U.S. Department of Energy under contract DE-AC07-941D 13223. DISCLAIMER This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsi- bility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Refer- ence herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise. does not necessarily constitute or imply its endorsement, recom- mendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. -