1. Introduction
Structural analysis, structural design and structural appraisal are three important constituents of structural engineering. Though, the former two topics have been researched well and practiced with greater emphasis, the later has largely been ignored by the profession. Inspite of recorded evidence of major structural failures world-wide, no serious attempts have been made towards investigation on the topic of structural damage assessment, in par with the areas of analysis and design. Only recently the importance and complexity of damage assessment is being discussed openly and professionally.2. Main contributions
The main issues warranting investigation and identified in the literature review along with current state of knowledge are synthesized to form a general damage assessment paradigm. However, the issues particularly addressed in the thesis are highlighted in the following sections.2.1 Damage assessment engineering
The concept of sensitivity and gradient computation has been used extensively in structural optimization and there are well established techniques available in literature to carry out sensitivity studies. Structural response sensitivities are derivatives of structural response such as displacement, strain, stress, eigenvalue and amplitude of derived response (vibration frequency or buckling load factor) with respect to a structural parameters. In the present investigation the role of sensitivity has been explored in several different ways in the field of damage evaluation.2.1.1 Structural identification and sensitivity studies
There are many problems of structural engineering where one has to evaluate the changes in local and global structural response characteristics due to changes in structural parameters. Sensitivity information has been adopted by some investigators in the algorithmic identification of structural components using static and dynamic response. The suitability and limitation of the algorithmic approach of identification is examined in the present context.2.1.2 Sensitivity based structural importance factor
The importance factor adopted in damage assessment has been either assumed heuristically, statistically, or by the process of opinion surveys, based on intuition or experience of the inspector/experts. One of the shortcomings of these methods, particularly in using the structural importance/weightages based on the experts' personal judgment is that it has a expert's bias and subjectivity. These could lead to an unrealistic damage index. Hence, a new concept of damage-sensitivity-based-importance factor has been introduced in the thesis.2.1.3 Sensitivity as a measure of redundancy
The degree of redundancy has often been considered in structural integrity assessment. However, there is no consensus in the literature on quantification of redundancy. A new generalized definition of redundancy applicable to both continuum and discrete structures in terms of structural response sensitivity is proposed.2.1.4 Sensitivity based damage functions
A number of damage functions have been proposed in the literature related to safety, reliability and structural damage assessment. However, no unique definition of damage index is available and hence the search for a universal definition is still on. New damage functions based on structural response sensitivities are proposed. As a matter of general applicability, the proposed damage functions would be equally applicable to both discrete and continuum structures.2.1.5 Sensitivity in instrumentation-decision-making for in-situ measurement of structural response
The response of a structure at a certain stage of life is captured in-situ by measurement through proper instrumentation. Usually, measurement is made on the surface, at the boundaries, at loading points and some times inside the structure. However, the decision on selection of location(s) for instrumentation, its range of measurement and accuracy is based mostly on experience and judgment. The use of structural response sensitivity is proposed as a logical tool for instrumentation decision.2.2 Structural identification paradigms using ANN
When a structure undergoes various degrees of damage, certain characteristics have been found to undergo changes. In order to identify those changes, during inspection a sequence of static/dynamic tests are conducted and the resulting data such as load, displacements, strains, mechanical properties, such as, stiffness/strength, and dynamic characteristic, such as natural frequency, and damping can be estimated with system identification techniques. Structural identification can be done under static and dynamic conditions. These techniques have been demonstrated in the past in structural damage detection using conventional algorithmic techniques. The algorithm adopted is generally complex and is not suited to the situations where measured data is imprecise or inadequate. The recent emergence of ANN can be explored as an alternative tool for identification exercise in such situations. For inverse problems like structural identification of large civil engineering structures such as bridges and buildings where the in-situ measured data is expected to be imprecise and often incomplete, the ANN holds greater promise. This thesis presents novel applications of the multilayer perceptron in identification of damaged members of bridge truss structures in the presence of limited in-situ measurement data.2.2.1 Performance Study of backpropagation algorithm
The performance study of backpropagation algorithm in the context of multilayer perceptron simulation of damaged bridge truss structures has been carried out. The training paradigm which generally affects the performance of the network is examined in great detail. Finally, guidelines for developing ANN architecture for structural damage detection have been proposed.2.2.2 Structural identification using ANN with static measurements data
The generalized delta rule in ANN is implemented for identification of the static test data of steel bridge structures. It is concluded that measurement at only a few locations in the structure is needed as input in the identification process. This observation is of great engineering significance as it can reduce the difficulties in in-situ measurement.2.2.3 Vibration signature analysis using ANN
Damage detection by measuring and analyzing Vibration Signals in a machine component is an established procedure in mechanical and aerospace engineering. Hence, as a novel application, multilayer perceptrons with the backpropagation algorithm have been adopted for network training using vibration signals for simulated damaged states in steel bridge structures. The training patterns in terms of vibration signature have been generated analytically for a moving load traveling on a trussed bridge structure at a constant speed to simulate the inspection vehicle. The issues related to the performance of the network are examined. For the cases illustrated, the performance of the network trained even with single-point vibration signature measurement at a suitably chosen location is found to be reasonably good as compared to that of three-point and five-point vibration signature measurements. This has a very important engineering significance as the measurement only at single node is good enough for identification purpose.2.2.4 Time-delay neural network in vibration signature analysis
Limitations of Traditional Neural Networks (TNN) in dealing with patterns that may vary in time domain has given birth to Time- Delay Neural Networks (TDNN). The Traditional Neural Networks and the Time-Delay Neural Networks are implemented in detecting the damage in bridge structures using vibration signature analysis. A comparative study has been carried out for the various cases of complete as well as incomplete measurement data. It is observed that Time-Delay Neural Networks have performed better than Traditional Neural Networks in the structural identification application for the cases illustrated.2.3 Damage assessment paradigms using Artificial Intelligence (AI)
Knowledge Based Expert System (KBES)technique, a branch of artificial intelligence (AI), offers the possibility to formulate a system to encode and utilize the knowledge of a domain expert or a system which combines the background of many experts into a single, expert system for assessing the structural damage. If maintained and updated over time, such a system could become a repository for corporate knowledge, so that experience would not be lost when engineers retire or resign. In the development of expert system, three main issues are knowledge acquisition, knowledge manipulation and knowledge representation. Various knowledge engineering issues related to damage assessment are discussed in the thesis and outlined in the following sections.2.3.1 Knowledge engineering issues
Acquisition of domain knowledge is a major difficulty in the development of expert system in the field of structural engineering. Though, expertise related to design has been largely codified, "Enquiry" is suggested as a way of elicitation of engineering knowledge required in the context of damage assessment. It is generally observed that the raw knowledge obtained from experts/sensors may contain some noise and hence needs to be processed using knowledge engineering methodologies. The knowledge sensors available for obtaining bridge inspection data can be identified as visual, photographic, and NDE instrumentation, etc. However, the raw knowledge obtained from such sensors needs to be examined by experts and processed for imprecision, uncertainty, vagueness and other forms of noise. The topic of knowledge acquisition, knowledge processing and knowledge representation is still an active area of AI research. However, some of the relevant issues have been addressed here.2.3.2 Uncertainty in visual data
Bridge inspection is a very complex procedure. It covers visual inspection of each bridge components (e.g. the deck, superstructure, substructure etc.) and subcomponents (e.g. joints, expansion bearings etc.). During inspection, the bridge inspector usually rates the condition of components and subcomponents by assigning qualitative or quantitative values. The rating thus assigned relates to the effectiveness of the components, subcomponents or condition of a bridge structure in performing the function which was intended. Usually, decision making about damageability based on information gathered at various stages by visual inspection is generally imprecise and needs to be processed using knowledge engineering techniques. The information can be separated into objective and subjective components. The objective components comprises, numerical, countable, or quantitative information. Whereas the subjective components include intangible or qualitative information, which obviously involves the wisdom, judgment and experience of the expert. The procedure to handle such information has been developed using the recent developments in fuzzy sets.2.3.3 Role of fuzzy sets and Fuzzy weighted average
The development of expert systems for structural damage assessment often involves handling of uncertain and vague information as explained earlier. The complexity arises from the use of subjective opinion and imprecise numerical data. Very often the information from the experts is available either in an interval or in linguistic terms. Fuzzy sets are adopted to express such linguistic variables. The processing of fuzzy data is carried out using fuzzy arithmetic. However, procedures to evaluate fuzzy functions are yet to be fully established. The governing expressions such as the Fuzzy Weighted Average which which has been adopted for assessing the degree of damageability in structures has severe complexity in computation. An efficient computational procedure for evaluating fuzzy functions, interval based functions and their combination has been implemented.2.3.4 Sensitivity based weighted average
To quantify the global damageability of a structure the weighted average function as mentioned above, has generally been used in the literature. The weighted average function is defined in terms of damage condition and the importance or weightage of structural elements. However, both these aspects are judged by domain experts. Especially, the judgment of structural importance of a structural element is difficult task. Based on sound structural principle, using structural response sensitivity, the computation of Weighted Average function is proposed. The weighted average thus computed would be more realistic index for integrity assessment. Examples of weighted average computation comprising fuzzy sets and importance factors obtained from normalized damage sensitivity have been demonstrated.2.3.5 Integration of expert opinion
The topic of aggregation of expert opinion is relevant in the domain of structural damage assessment as the opinions expressed by domain experts are usually in linguistic terminology. The issue of consensus is important not only from engineering point of view, but also in various decision making situations from legal, financial and social aspects. A methodology based on nested weighted average function is proposed to integrate interval based or linguistic expert opinions. The aggregation methodology is demonstrated with examples in structural assessment.2.3.6 Inexact inference using Petri nets
As stated earlier, in structural damage assessment, the data available from various knowledge sources are imprecise and often interpreted by experts in linguistic terms. In this context, as a novel application, the use of Fuzzy Petri net for making inference on damageability using imprecise data is demonstrated. The Petri net, is an abstract, formal tool for representing information flow. The basic elements of Petri nets are places (P), transitions (T) and tokens that reside in the places. The linguistic knowledge is presented by generalized fuzzy production rules and the inference is made with composite rules formed using AND (MIN) and OR (MAX) nodes (places) with fuzzy tokens in the Petri net. A new method of firing the transitions with fuzzy marking with a probabilistic threshold is also proposed. The dynamics of the Petri net is illustrated with a general decision model in the context of structural damage assessment.2.3.7 Codal provisions
Damage assessment of structures (at local or global level), involving detailed investigation, is nothing more than a process of determining compliances with the requirements of the codes, standards and regulations at the stage of integrity assessment. The codal provision have been incorporated in a prototype knowledge base for design verification of steel structures with existing attributes. The domain knowledge from design codes is implemented in using hierarchical combination of limited-entry and extended-entry decision tables.2.4 The BLACKBOARD and integration of various paradigms
2.4.1 Blackboard architecture
The blackboard architecture is a special case of opportunistic problem solving. Originally, its basic approach has been thought to partition a problem into loosely coupled subtasks which were thought of as areas of specialization. Knowledge sources change the contents of the blackboard as a result of their reasoning, and all communication between knowledge sources passes through the blackboard. The blackboard architecture contains three basic components namely, Blackboard, Knowledge sources and a Control. A structured global database is analogous to and also named the blackboard through which the knowledge sources interacts and in which the state of the problem solving process is recorded. The knowledge sources are analogous to experts, comprising the knowledge necessary for solving the assigned task. The control determines the sequence of problem solving activity. The blackboard architecture is flexible and powerful, both conceptually and computationally. It has been used in many diversified application areas.2.4.2 Knowledge source
All communication between knowledge sources are routed through the blackboard, and thus it facilitates the development of isolated knowledge sources. This advantage allows the construction of a complete system to be partitioned into independent modules. Thus, the problem-solving knowledge required by the system can be partitioned cleanly into independent specialities. Moreover, depending on the type of knowledge, the knowledge sources are implemented in procedural and declarative languages, allowing them to be executed at reasonable speed and sequence as dectated by the control.2.4.3 Control issues
Control-information are contained within the knowledge sources, on the blackboard as well as in a separate modules as necessary (there is no actual control unit specified as part of a blackboard system). The control knowledge monitors the changes to the blackboard and determines what the immediate focus of attention should be in solving problems.2.4.4 Implementation issues
Using the concepts and paradigms developed in the sections 2.1 to 2.3 and keeping in view the developmemnt of a knowledge based expert system (KBES) for damage assessment of steel railway bridges, a blackboard architecture has been proposed. The various knowledge sources (e.g. visual inspection data integration, expert opinion integration, computation of damage index, static and dynamic test data evaluation etc.) have been implemented independently. Though, some of the modules have been developed in various environments, the coupling is possible through the control module. Here, the modular implementation of knowledge sources represents the most important aspects of the development of a prototype knowledge based expert system.2.4.5 Applications and Illustration
The application of the various paradigms have been demonstrated in the context of Indian railway steel bridges.3. Conclusions
This investigation has contributed in a very important but somewhat neglected field of structural engineering, where in modern computational mechanics has been integrated with wisdom of the profession using recent computing tools, AI and ANN. It is believed that the work presented here is of practical relevance, particularly for performance monitoring of important structures such as railway steel bridges. The paradigms presented are of general nature and would be equally applicable to any other type of structure with relevant domain information coded in the knowledge sources of the BLACKBOARD.