SACON: A Knowledge-Based Consultant for Structural Analysis

Introduction

We describe an application of artificial intelligence methods to structural analysis, In particular, the development and (partial) Implementation of an "automated consultant" to advise engineers In the use of a general-purpose structural analys1s program". The analysis program numerically simulates . the behavior of a physical structure subjected to vanous mechanical loading conditions. The automated consultant, called SACON (Structural Analysis CONsultant), was constructed using the EMYCIN system, the domain­ Independent version of the MYCIN program [1). Ongmally developed to advise physicians 111 the. d1agnom and treatment of Infectious diseases, the domam-spec1f1c med1cal knowledge tn M YCIN Is represented as production rules, and is kept Independent of the "inference engine" that uses the rules . . By substituting structural engineering knowledge for the med1c_al knowledge, the program was converted easily f'rom the domaan of infectiou~ diseases to the domain of structural analysu.

The purpose of a SACON consultation is to provide advice to a structural engineer regarding the use of a structural analysis program called MARC [2). The MARC program uses finite--element analysis techniques to simulate the mechanacal behavior of objects. The engineer typically knows what he wants the MARC program to do--e.g., ex a mane the behavior of a specific structure under expected loading conditions--but does not know hMtJ the simulation program should be set up to do 11. The MARC program offers a large (and, to the novice, bewildering) choice of analysis methods, materaal properties, and geometries that may be u~ed to ·model the structure of interest. From these options the user must learn to selt>ct . an appropriate subset that will simulate the correct phys1cal behavior, preserve the desired accuracy, and manam1ze the (typically large) computational cost. A year of experaence with the program is the typical time required to learn how to use all of MARC's options proficiently. The goal of the automated consultant Is to bridge the "What-to-How" gap, by recommending an analysis stratP.gy. This advice can then be used to direct the MARC user in the choice of spec1flc anput data--e.g., numerical methods and material properties. Typical structures that cetn be analyzed by both SACON and MARC Include aircraft wings, reactor pressure vessels, rocket motor casings, bridges, and buildings.

· This research was supported by . the Defense Advanced R ese.-.rch Projects Agency (ARPA Order No. 2i9i Contract No. DAHCIS-73-C-0435) and the Air Force Fhght Dynam1cs Laboratory. RSE's present address: DARPA.

In this report we describe the general structure of the structural analysis knowledge base, including the major concepts used and the inferences drawn by the consultant. We conclude by making some observations, In light of the SACON knowledge base, about the EMYCIN system as . a representational vehicle and about the process of acqumng knowledge for rule-based systems. Further details of this application can be found in [3].

2 The SACON Knowledge Base

The obj·ctive of a SACON consultation is to identify an analysis strategy for a particular structural analys1s problem. The engineer can then implement this strategy, usmg the MARC program, to simulate the behavior of the structure. This section Introduces the mathematical and physical concepts used by the consultant when characterizing the structure and

recommending an analysis strategy. .

An analysis strategy consists of a number of analysu classes and their associated analysis recommendations. Analysis classes characterize the complexity of modelling the structure and the ability to analyze the material behav1ors of the structure. Currently, 38 analysis classes are considered; among them, Nonlinear geometry crack grorvth, Bifurcation, Mat erial inJtabiltty, /nelaJtlc Jtiffnm dtgradation, Linear analyJIJ, and No analysis. The analysis recommendations advise the engineer on specific features of the MARC program that should be activated when performing the actual structural analysis. The example consultation concludes with 9 such recommendations (see below).

To determine the appropriate analysis strategy, SACON infers the critical material stress and deflection behaviors of a structure under a number of loading conditions. Among the material stress behaviors Inferred by SACON are Yielding collapse. Cracking potential, Fatigut, and Material instabilities; material deflection behaviors are Excessive deflection, Flexibility changts, Incremental strain failure, Duckling, and Load path bifurcation.

Using SA CON, the engineer decomposes the structure into one or more substructures and provides the system the data describing the materials, the general geometries, and the bound;uy conditions for each of these substructures. A substructure is a geometrically contiguous region of the structure composed of a single material such as high-strength aluminum or structural steel and having a specified set of kinematic boundary conditions. A structure may be subdivided by the structural engineer in a number of different ways; the decomposition is chosen which best reveals the worst-case

material behaviors of the structure.

For each substructure, SACON estimates a numeric total loAding from one or more loadings. Each loading applied to a substructure represents one of the typical mechanical forces on the substructure during its working life. These might, for example, include loadings experienced during various maneuvers such as braking, banking, etc., for planes or, for buildings, loadings caused by natural phenomena such as earthquakes or wind-storms. Each loading is In turn composed of a number of point or distributed load components.

Given the descriptions of the component substructures and the descriptions of the loadings applied to each substructure, the consultant estimates stresses and deflections for each substructure using a number of simple algebraic equations. The behaviors of the complete structure are found by determining the sum of the peak relative stress and deflection behaviors of all the substructures. Based on these peak responses (essentially the worst-case behaviors exhibited by the structure), its knowledge of available analysis types, and the tolerable analysis error, SACON recommends an analysis strategy. Figure I Illustrates the basic inferences drawn by SACON during a consultation.

Analysis Strategy of the Structure

1

Worst-case Stress and Deflection Behaviors of the Structure

1

Symbolic Stress and Deflection Behaviors of each Sub-structure

1

Composite Numeric Stress and Deflection Estimates for each Loading

t

Numeric Stress and Deflection Magnitudes for each Load Component

Fig. I Inference structure during the consultation.

Knowledge about the structural analysis task is represented In the form of production rules. An example rule, which provides the transition from simple numeric estimates of stress magnitudes to symbolic characterizations of stress behaviors for a substructure, Is Illustrated below. For details on this representational scheme, see [i).

RULEOSO

• If: I) The material composing the sub-structure is one of: the metals, and
• The analysis error (in percent) that is tolerable is between 5 and 30, and
• The non-dimensional stress of the sub-structure Is greater than .9, and
• The number of cycles the loading is to be applied is between 1000 and 10000

Then: It Is definite ( 1.0) that fatigue is one of the stress behavior phenomena in the sub-structure

The current consultant is able to select from among 36 nonlinear analysis strategies. If nonlinear analysis is not Indicated by low response estimates, the col}sultant recommends either linear analysis or no analysis. The knowledge base consists of 170 rules and about liO consultation parameters. A typical consultation (2 substructures, 3 loadings, 3 load components) requires about 25 minutes at an interactive terminal.

To bring the SACON program to its present level of performance, we estimate that two man-months of the experts' time were required to explicate the consultation task and formulate the knowledge base, and about the same amount of time to implement and test the rules. For the first 170 rules in the knowledge base, the average cost of formulating and implementing a rule was about i hours. The marginal cost for a new rule is about 2 hours. The following transcript illustrates SACONs interactions with an engineer and SACON's recommendation of an analysis strategy.

--------5 T R UCTUR E-28--------

1) What is the name of STRUCTURE-28?

** BOEING 7i7 WING

1. What Is the analysis error (In percent) that is tolerable? ** 10
2. Do you want to examine the responses of the boeing 7i7 wing, to evaluate its Instability, or both?

** EXAMINE-RESPONS~S

1. Does the boeing 7i7 wing have time dependent terms in its equations of equilibrium?

**NO

5) Is heating or cooling of interest in determining structure responses?

**NO

6) Are there any sub-structures?

**YES

--------5 U B-S TR UCTU R E-1--------

7) What is the name of SUB-STRUCTURE-I?

**TOTAL WING

8) What Is the material composing the total wing?

**HIGH-STRENGTH-ALUMINIUM

9) What is the length of the total wing?

** li20 INCHES

10) What is the weight of the total wing?

** i6000 POUNDS

II) Is the total wing solid or thin-walled?

**THIN-WALLED

12) What is the configuration of support for the total wing? **ONE-SIDE

I!\) What is the modelling dimensionality of the total wing? ** 2

The first loading applied to the total wing wiU be called: --------LOA DING-1--------

16) What Is the name of LOADING-I?

**FLIGHT

17) Enter the number of loading cycles to be applied:

** 20000

The first load component of LOA DING-I will be called:

--------LOAD-COMPONENT -1--------

18) What Is the distribution of LOAD-COMPONENT- I?

**DISTRIBUTED

HI) What is tht surface to which it acts normal?

**WIDTH-LENGTH

26) Art the support conditions of the structure nonlinear? **NO

your structure: General-inelastic

The following are specific analysis recommendations you should follow when performing the structure analysis: Activate incremental stress- incremental strain analysis. Model nonlinear stress-strain relation of the material. Solution will be based on a mix of gradient and Newton methods.

Logic to scan peak stress at each step and evaluate fatigue Integrity should be used.

Logic to scan stresses, smooth, and compare with allowable stresses (with appropriate safety factors) should be used.

Logic to scan deflections, calculate relative values, and

comp:ue with code limits, should be called upon.

Cumulative strain damage should be calcuated.

Analysis should include two or more load cycles (if cyclic) with extrapolation for strain accumulation.

Shakedown extrapolation logic should be used.

! EMYCIN as a Knowledge Representation Vehicle

A primary goal of this research was to determine if current "knowledge-engineering" techniques could be usefully applied in the development of a computer-based consultant in structural analysis. Specifically, our research was a test of the applicability of the rule-based formalism of the EMYCIN system. As such, we neither explored the use of other available consultation systems (e.g.. PROSPECTOR, RITA) nor examined with the expert the pros and cons of using the different representation schemes that they provide.

At no time did we find the representation formalism of EMYCIN to be a hindrance to either the formulation of the knowledge by the expert or its eventual implementation in the SACON program. In fact, the simplicity of using and explaining both EMYCIN's rule-based formalism and its b<~ck ward-chaining control structure actually facilitated the rapid development of the knowledge base during the early stages of the consultant's design. Moreover, the control structure, like the rule-based formalism, seemed to impose a salutary discipline on the expert as he formulated the knowledge base.

The development of SA CON represents a major test of the domain-independence of the EMYCIN system. Previous applications using EMYCIN have been primarily medical with the consultations focusing on the diagnosis and prescription of therapy for a patient. Structural analysis, with its emphasis on structures and loadings, allowed us to detect the small number of places where this medical bias had unduly influenced the system design, notably text strings used for prompting and giving advice.

Our expert found that his knowledge was easily cast into the rule-based formalism and that the existing predicate functions and context-tree mechanism provided sufficient expressive power to capture the task of recommending an analy~is strategy. The existing interactive facilities for performing explan<~tion , question-answering, and consultation were found to be well developed and were used directly by our application. None of these features required any significant reprogramming and, for the most part, worked without modification.

One major feature of EMYCIN that was not used in this task was the confidence factor mechanism [i)--i.e., the ability to draw inferences with uncertain knowledge. The present consultation strategy and the <~Ssociated mathematical models were designed to estimate extreme loading conditions, from which SA CON concludes the appropriate analysis class. Consequently, by using a "conservative" model, the rules, though inexact, are sufficiently accurate for predicting response bounds with certainty.

Finally, we note that the development of the knowledge base was greatly facilitated when the knowledge engineering team elicited a well-specified consultation goal for the system and then an Inference structure such as that depleted In Figure I. Without this conceptual structure to give direction to the knowledge explication process, a confused and unusable web of facts issued from the expert. We speculate that the value of these organizational structures is not restricted to the production system methodology. They seem to be employed whenever human experts attempt to formalize their knowledge using some representation formalism. Indeed, when difficulties arise in building a usable knowledge base, we expect that the trouble Is as likely to be because of a poor choice of inference structure than from the use of a particular representation scheme.

The Inference structure Is a form of mtta-knowltdgt, I.e. knowledge about the structure and use of the domain expertise. Our experience shows that this meta-knowledge should be elicited and discussed early in the knowledge acquisition process, in order to Insure that a sufficient knowledge base is acquired to complete a line of reasoning. and to reduce the time and cost of system development.

Currently, the inference structure is not an explicit part of the program (or of any other expert system of which we are aware), and hence is unavailable for the purposes of explanation, tutoring, or further acquisition of the knowledge base. Its critical role in building a successful knowledge engineering application, however, would suggest making such meta-knowledge an explicit part of the consultation system. Following (5), future research on the interactive acquisition of knowledge from experts should benefit from the representation and use of such domain-specific meta-knowledge.

4 Acknowledgements

We wish 10 acknowledge the contributions of Prof. Robert M elosh, who provided the structural analysis expertise, Dr. Lewis Creary, and Dr. Pedro Marcel. William van Melle and Carli Scott provided invaluable technical advice on the care and feeding of the EMYCIN system.

• [I] van Melle, W., A Domain-independent Production-rule System for Consultation Systems, these Proceedings
• MARC User Information Manual, Available from MARC Analysis Research Corporation, 260 Sheridan Ave., Court House Plala, Palo Alto, CA. 9i306, Revised, September 1976.
• (3] Bennett, J.S., Creary, L., Engelmore, R.S., Melosh, R., SA CON: A Knowledge-based Consultant for Structural Analysis, Stanford Computer Science Report CS-78-699, December 1978.
• Shortliffe, E.H., Computtr-Bastd Mtdical Consultations: MYCIN, New York: American Elsevier, 1976.
• Davis, R., Applications of Meta-Level Knowledge to the Construction, Maintenance and Use of Large Knowledge Bases, Ph.D. Thesis in Computer Science, Stanford Computer Science Report CS-76-552, July 1976.