Artificial Intelligence Applications
in Geotechnical Engineering
Geotechnical Systems Group,
School of Engineering,
University of Durham, Durham, DH1 3LE, UK
e-mail Dr Toll
INVITED PAPER
|
|
Artificial Intelligence Applications
Geotechnical Systems Group, |
|
|
INVITED PAPER |
Abstract
The paper reviews artificial intelligence (AI) systems that have been developed for geotechnical applications. It covers knowledge-based ('expert') systems and neural network approaches. A significant number of systems have been developed for site characterisation, classification of soils and rocks, foundations, earth retaining structures, slopes, tunnels and underground openings, mining, liquefaction, ground improvement, geotextiles, ground water/dams, roads and earthworks. It is suggested that AI systems should be developed as support tools, rather than attempting to replace human expertise. It is also recognised that AI techniques are good for some aspects of solving engineering problems, but that other approaches are still valid for many applications. Therefore, the way forward will be the development of hybrid systems which mix different artificial intelligence techniques and conventional programming.
KEYWORDS: knowledge-based systems, expert systems, neural network systems, site characterisation, soil classification, foundations, earth retaining structures, slopes, tunnels, underground openings, mining, liquefaction, ground improvement, geotextiles, ground water, dams, earthworks
INTRODUCTION
Geotechnical engineering is known as an 'imprecise' area of engineering due to the fact that we are dealing with a material produced by nature (the ground). In many circumstances, our fundamental understanding of soil and rock behaviour still falls short of being able to predict how the ground will behave. Under these circumstances, expert judgement plays an important role, and empirical approaches to design are widely used. Since artificicial intelligence (AI) techniques can make use of heuristic knowledge (rules of thumb) or pattern matching techniques, as opposed to solving a set of mathematical equations, they should be ideally suited for application in the field of geotechnical engineering.
Moula et al (1995) reviewed knowledge-based (expert) systems developed up to 1993, concentrating mainly on soil engineering applications. The conclusion was that the systems developed showed the potential for applying knowledge-based system technology in geotechnical engineering. It was thought likely that, in the next decade, more systems would develop to the stage where they will be used commercially. Since 1993 there has been a significant number of new systems developed in all aspects of geotechnical engineering. This paper brings that earlier review up-to-date, and also includes more rock engineering applications. Since neural network approachs are also coming to prominence, they have also been included in the review. There has been little, if any, use of other A.I. techniques (eg genetic algorithms) in geotechnical engineering.
SITE CHARACTERIZATION
This section reviews artificial intelligence techniques for site characterization. Systems have been developed for planning site investigations, interpreting site investigation data to generate a model of the ground conditions, classification of soil and rock, and the interpretation of geotechnical parameters.
Site Investigation Planning
SOILCON (Wharry and Ashley, 1986; Siller, 1987) was one of the earliest KBSs to address the problem of determining the required level of geotechnical investigation. This is based on the requirements of a proposed structure and the level of information known about the site. The aim is to reduce the risk involved with the subsurface to an acceptable level. Probabilistic analysis was used by Halimet al (1991) in a KBS to assist engineers in making site exploration decisions. The system generates an inference of prior estimates of soil and anomaly characteristics (such as lenses or pockets of soft soils within the regular soil deposit) and uses probabilistic analysis in the selection of the most appropriate exploration program. The system is also intended for evaluating geotechnical design for shallow foundations or slope stability.
Smith and Oliphant (1991) describe a KBS to assist with the planning stages of a site investigation. The system provides suggestions as to the next stage of the site investigation (e.g. desk study, site reconnaissance, ground investigation etc). The information obtained from the subsoil exploration stage is also used to create a 2-D visual representation of the soil layers. This has been developed further by Oliphant et al (1996) as part of the ASSIST system which is described later. CESSOL (Magnan, 1992) is a more advanced system that can give qualitative advice on the type of investigation needed, and what sort of testing would be required. It can also give quantitative advice on the number of boreholes and piezometers and amount of testing required.
The system for site investigation of trunk road projects described by Thomas et al (1992) and Winter and Matheson (1992) is similar to the approach used by Smith and Oliphant (1991). In this case an activity log of an investigation is produced for comparison against a list of mandatory and advisory procedures contained within the system. The system can therefore be used to highlight omissions in the way that an investigation has been carried out.
A KBS for assisting in the selection of appropriate field tests is presented by Moula (1993) and Moula & Toll (1993). The system can be used to advise which test can provide a required parameter (with a specified reliability) and how applicable it is to the specified ground conditions. Fung and Kay (1996) have used a probablistic approach to the selection of an appropriate field testing strategy.
Interpreting Ground Conditions
One of the earliest geotechnical knowledge-based systems was SITECHAR (Norkin, 1985; Rehak et al, 1985). This is a KBS which uses geometrical reasoning to develop inferences about the depositional patterns of the subsurface materials and their physical properties. It uses field and laboratory data but also takes into account existing experience of geology and geomorphology at a specific site or at similar ones. The same rule-based approach was further developed as LOGS (Lok, 1987; Adams et al, 1989). LOGS treats information from several boring logs and provides the user with two dimensional subsurface profiles. The system tries to identify marker beds, lenses (wedge-shaped deposits) and lentils (strata with boundaries within the confines of the site).
The approach used by Carpaneto and Cremonini (1991) for geotechnical site characterisation uses the concept of a 'site pattern', i.e. a simplified general soil profile. The depths and field description of soil layers, results of laboratory and insitu penetration tests, are compared against the 'site pattern' and a measure of certainty calculated. SAGITAIRE (Vergobbi et al, 1992) can also be used to merge data from soil descriptions, classification data from laboratory testing and results from insitu tests to form a final borehole log.
Another system which uses the integration of different types of data is described by Kovalevsky & Kharchenko (1992). Their system is used for classifying seabed soils based on an integration of geophysical and geotechnical data, e.g. compressional- and shear-wave velocity sections, and borehole profiles.
Toll et al (1992), Toll (1994), Toll (1995) describe SIGMA, a KBS for interpreting ground conditions from borehole logs. It can also assist with the derivation of design parameters from laboratory or field test results. The approach used for correlating soil layers between boreholes (based on soil descriptions) is described by Vaptismas & Toll (1993). A similar approach was also used by Oliphant et al (1996). Their system (ASSIST) can also generate graphical representations of the ground conditions.
Kinnicutt et al (1994) describe a system called NOMAD which can be used for three dimensional stratigraphic characterization. NOMAD can use the functionality of KRIBS (Kinnicutt, 1995) to create ground profiles from borehole data. This is done by combining geostatistical and knowledge-based approaches. The geostatistical interpretation can be combined with subjective data entered by the user. Inazaki (1994) combines a dynamic depth warping algorithm for borehole correlation with a KBS for combining layers into geotechnical zones.
Adams & Bosscher (1995) describe the integration of geographical information systems (GIS) and knowledge-based systems for subsurface characterization. Thomaz & Altschaeffl (1994), in their sketchy outline of GeoSYS, also suggest that a combination of tools is necessary to support the site investigation process. These developments are the logical extension of the idea of a 'geotechnical site characterisation workbench' suggested by Rehak et al (1985).
Al-Garni (1995) describes the use of a KBS in the desk study phase of a site investigation. This system uses data from aerial and space imagery to assist in evaluating potential sites. The KBS uses features abstracted from the images (topography, drainage, gullies, tone, land cover and land use) to assess a site.
SITECLAS (Wong et al, 1989) is a KBS used to classify a site (according to the Australian Standard AS2870.1) rather than to generate an interpretation of the ground conditions. The site is identified as one of six classes based on information about the soils found at the site (using Australian soil groups), presence of fill, footing type, whether the site is subjected to landslip, erosion, collapse etc.
The methods so far described for analysis and interpretation of geotechnical site investigation data make use of either geometrical reasoning or statistical techniques. Zhou & Wu (1994) describe the use of neural networks for this purpose. Their neural network system is used to characterize the spatial distribution of rockhead elevations. Similar applications relevant to ground water characterization are described by Rizzo & Dougherty (1994) and Basheer et al (1996). Basheer et al descibe how neural networks can be used to map the variation of permeability in order to identify boundaries of a landfill.
Classification and Parameter Assessement (Soil)
CONE (Mullarkey, 1986; Mullarkey and Fenves, 1986) is a KBS that interprets raw data from the cone penetrometer (CPT) in order to check the validity of the raw data and to classify the soil types (to generate a profiling). It represents an early use of fuzzy sets in geotechnical engineering. The classification is used to infer values for the shear strength of sands and clays.
Alim and Munro (1987) present a very simple prototype KBS for soil identification that uses rather simplistic textbook knowledge. It provides judgement concerning the most likely foundation type under given soil and loading conditions, based on visual and physical observation of soil characteristics.
A KBS was developed by Davey-Wilson (1991a) for soil shear strength analysis. The system uses soil descriptions as input in order to infer values for friction angle (f). Similarly, Gillette (1991) describes CASS (Computerised Adviser on Soil Strength), a KBS to assist in the selection of shear strength parameters (c and f) for use in stability analysis. Agrawal etal (1994) use a neural network approach for predicting c'and f' for silty clay from dry density and water content.
Davey-Wilson's earlier work has been much extended by Davey-Wilson and Mistry (1995) which uses a case-based approach to the estimation of geotechnical parameters. An object-oriented approach to the same problem is described by Toll & Giolas (1995). The KBS makes use of a knowledge base which is structured to represent the ground at different levels of detail. It can advise on the typical range of values for a number of geotechnical design parameters.
A neural network approach to soil classification is described by Cal (1995) who uses three main factors (plastic index, liquid limit and clay content) to generate a quantitative soil classification. Goh (1995c) has used neural networks for modelling soil correlations.
The stress-strain behaviour of soils has also been modelled using neural networks. Penumadu et al (1994) have attempted to model the stress-strain behaviour of clays, incorporating rate dependant behaviour. Ellis et al (1995) have used grain size distribution and stress history as input parameters in order to simulate stress-strain relationships for sand. This can be compared with similar attempts by Zhang et al (1991), Millar & Clarici (1994) and Millar & Calderbank (1995) to use neural networks to model stress-strain behaviour of rocks.
Classification and Parameter Assessement (Rock)
Rock Mass Classification systems make use of a set of reasonably well defined rules and are therefore ideally suited for implementation as knowledge-based systems. A number of systems have been developed, some of which have been reviewed by Coulthard (1995). These include Zhang et al (1988) based on Gu's qualitative classification scheme; Juang & Lee (1989) and Madhu et al (1995) based mainly on Bieniawski's rock mass rating (RMR) system, and using fuzzy logic; Butler & Franklin (1990) using Barton's Q and Bieniawski's RMR systems.
Classification systems that have been developed for specific purposes have also been implemented as knowledge-based systems. Bearmanet al (1990) describe the development of KBS for predicting crushing requirements based on a comminution index. Koczanowskiet al (1991) describe a KBS for rock rippability assessment.
Neural networks have been used for rock classification. Millar & Hudson (1994) describe the use of neural network methods for performance monitoring of rock masses for mining geomechanics. They describe an application for the collection of data relating to the condition and subsequent classification of rock masses. They have also used neural networks to predict of the likely future performance of rock masses, particularly when they have been perturbedfrom their natural condition by mining engineering activity. Cai(1995) has used neural networks to classify rocks for the purposesof blast design and Yi & Lindqvist (1995) have used neuralnetwork mdels for predicting rock quality parameters.
Zhang et al (1991), Millar & Clarici (1994) and Millar &Calderbank (1995) have used neural networks for modelling rockdeformability behaviour. Input parameters include mineralogy,particle size and shape, porosity, grain compressibility etc.
FOUNDATIONS
Artificial intelligence techniques have been used in a range of systems relating to foundations. Developers have particularlyfocused on conceptual design, ie selecting appropriate foundationtypes. However, there are also systems for detailed design, forfoundations problems and construction. The prediction of pilecapacity from pile driving data is an area which has been foundto be well suited to neural network approaches.
Conceptual Design of Foundations
A number of KBS have been developed to address the problem of selecting an appropriate type of foundation. (Shukla,1988; Adams et al, 1989) is such a system for building foundations. It provides a list of all feasible foundation alternatives, based on soil conditions, water table location, depth of bedrock and the imposed loading conditions from the structure. It was extended by Meyer (1992) to use preliminary soil data and the potential configuration of the building in order to produce a set of feasible foundation solutions.
The system descibed by Stuckrath and Grivas (1990) focused on the selection of bridge foundations. The system presents preliminary design options including shallow and deep foundations and ground improvement. Similarly the main function of BABE (Bridge and Building Evaluation) developed by Zheng et al (1989) is to aid in the selection of the most appropriate type of foundation for a specific bridge superstructure and a set of site conditions. Cadogan et al (1996) also outline a system for bridge foundation design.
FOUNDCON (Rashad et al (1991) also provides a preliminary design module for selection of the most appropriate foundation system. In addition it has a detailed design module for performing the final design. CONFOUND (Toll & Barr, 1995) is a computer-aided learning package for preliminary (conceptual) foundation design. The system offers a range of choices for foundation types. The KBS then provides a critique of the option selected, based on site information provided by the user. Boissier & Henry (1995) also describe a KBS approach to the generation of foundation solutions.
A number of KBSs have been developed specifically for pile selection. PILE (Santamarina and Chameau, 1987) provides a list of the most promising alternatives based on technical constraints. It is then up to the user to consider additional factors (e.g. economical), in order to reach a final decision. PILEX (Elton and Brown, 1991) considers timber, concrete and steel piles and takes into account geotechnical, geological, structural, and environmental factors that influence the pile selection. SUPILE (Wong et al,1991) evaluates of suitability of different types of piles and can estimate the required pile size and length. The selection of a pile type is performed by generating a suitability score depending on the number of problems that would exist if that pile type was used.
Detailed Design
Yehia and El-Hajj, 1987) is a KBS to assist in the selection and design of spread footings. It uses a database of previous designs and tries to match a new problem to one of the existing cases in the database. Its main purpose is the structural design of a foundation and no real geotechnical design is included. Rowlinson (1989) briefly describes the early stages of development of Geotech, a KBS to assist in foundation design in Hong Kong. The system is intended to ensure that all design is constrained by the relevant regulations. A KBS with the same name, GEOTECH, has been developed by Parikh and Kameswara Rao, (1991) as an aid in shallow foundation design by calculating bearing capacity and settlement. The output is in the form of a list of the most promising alternatives with corresponding confidence factors.
Kulhawy & Trautman (1991), Trautman & Kulhawy (1995) have included some KBS features into their CUFAD+ (Compression and Uplift Foundation Analysis & Design) programme. The KBS component assists the user with selecting realistic values for the input data. Flavigny et al (1992), Izadi et al (1995) also describe a KBS for the design of shallow foundations based on the pressuremeter, commensurate with French practice.
Pile Driving
Chow et al (1995) present a neural network approach to the prediction of pile capacity. A stress-wave matching technique is used which makes it feasible to determine the static pile capacity in real time in the field. Chan et al (1995) have used neural networks as an alternative to pile driving formulas. Similarly, Goh (1995b, 1996b) and Lee & Lee (1996) have both used neural networks to estimate the load capacity of driven piles based on the hammer characteristics, the properties of the pile and soil, and the pile set. All these authors suggest that neural networks provide better predictions than conventional pile driving formulas. Jongmans (1996) describes a KBS to predict the level of ground vibration caused by pile driving based on the geometry and the dynamic characteristics of the ground.
Foundation Construction
Kato et al (1995) outline a KBS for the planning and progress of foundation work. It selects an appropriate method of construction (for pile or slab) and incorporates these into the construction plan. Yeh et al (1991) describe a diagnostic knowledge-based system PCPILE (prestress concrete pile) for diagnosing the damage to a pile during the construction process. Fisher et al(1993), Fisher et al (1995) describe a decision support system called DS2 which can suggest an appropriate construction methods for constructing a drilled shaft based on geological information. It can also prepare a preliminary cost estimate, and suggest key specification items.
Foundation Problems
Hadipriono et al (1991) described a KBS which was under development for determining the causes of foundation failures. The system contains knowledge on possible causes for foundation failure such as soil settlement, expansive soil, soil erosion, bearing capacity failure, slope instability and foundation corrosion.
Wiseman et al (1992) describe a KBS for foundations on expansive soil, extending their system for heave prediction (Wiseman et al, 1987). The system requires input about the soil and profile, the building environment (changes in drainage, vegetation) and details of existing and proposed buildings and attempts to quantify the amount of heave expected.
EARTH RETAINING STRUCTURES
Artificial intelligence systems have been developed for design of retaining structures, for predicting movements and analysing failures.
Hutchinson et al (1987) present RETWALL, a KBS for the selection and preliminary design of earth retaining structures. The system evaluates the applicability of of the nine wall types that are included in its knowledge base (brick wall, blockwork wall, crib wall, gabions, gravity wall, railway sleeper wall, reinforced earth, reinforced concrete wall, sheet piling). Oliphant and Blockley (1989) developed a KBS with a knowledge base comprising three parts, the construction process, the design process and environmental impact. The system includes 11 case studies of retaining structures and provides a narrative of the history of each one in terms of why it was selected or considered as an alternative, allowing the user to compare these with a proposed retaining wall.
A KBS for retaining wall selection and design is presented by Arockiasamy et al (1991). The system has knowledge about ten wall types including concrete gravity, cantilever, counterfort, gabions, reinforced-earth, crib, slurry, sheet-pile, tieback, and soil nailed walls. Amer & Abdel Rahman (1994) describe a KBS for sheet pile selection, with links to programs for detailed design.
WADI (Chahine and Janson, 1987) is a KBS developed for the preliminary diagnosis of retaining wall failures. WADI is applicable to two types of retaining walls: cantilever reinforced concrete wall sand gravity concrete or rubble walls, having a maximum height of 8 metres. RETAIN (Adams et al, 1989) is a KBS that allows categorisation and organisation of knowledge relating to failure and rehabilitation of earth retaining walls. Upon solving the failure diagnosis, a table of wall failure modes with associated certainties is produced.
Goh et al (1995) have shown how neural networks can be used to estimate lateral wall movements in braced excavations. The neural network was used to synthesize data derived from finite element studies on braced excavations in clays.
SLOPES
Soil slopes
Grivas & Reagan (1988) describe a KBS (STABCON) for evaluating slope instability and recommending appropriate types of treatment for soil slopes. It is linked to analytical methods for calculating slope stability. Hirokane et al (1993) also describe a KBS for deciding on appropriate slope treatment. It includes 44 different types of slope protection ranging from vegetation and seeding through to concrete crib-work and retaining walls.
SISYPHE and XPENT (Asté et al, 1995) are two KBSs for slope instability which have been developed in parallel. XPENT (Faure et al, 1988; Faure et al, 1991; Mascarelli et al, 1992; Faure et al, 1995) is a KBS for assisting in slope stability analysis. It assists in diagnosing the type of landslide on the basis of information about the geology, vegetation, geomorphology (large and small scale), and hydrogeology. It canal so advise on methods of stabilisation based on the size of the slide, the material, accessibility of the site, etc. SISYPHE (Asté, 1992) is a KBS for investigating slope instabilities. It can be used in diagnosis of a landslide as well as for hazard evaluation. For diagnosis purposes, SISYPHE provides the ability to develop three dimensional representations of the ground surface, piezometric surfaces and the slip surface itself.
Wang et al (1994) describe a KBS for investigating potential landslides. It contains knowledge bases relating to the spatial distribution of an unstable zone, the geotechnical properties, methods of assessing stability, and methods of treatment.
Wislocki and Bentley (1991) describe the development of a KBS for the determination of planning applications with respect to landslide hazard in South Wales. The system attempts to assess the landslide hazard that may affect proposed development sites and it produces output in the form of planning response options.
Rock Slopes
Expert Slope Design System (ESDS) presented by Kizil & Denby (1990) and Denby and Kizil (1991) is a KBS to assist geotechnical engineers in the assessment of proposed slope designs in opencast coal operations in the UK. Ozgenoglu & Ocal (1994) describe SEVDUR a KBS for slope stability analysis relating to mining operations.
Hao & Zhang (1994) describe a KBS for stability analysis of rock slopes. This uses fuzzy sets for representation of joint sets. Zhou (1994) uses a probabalistic approach in a KBS for the prediction of slope stability. An approach called MAQEGP - Mechanism Analysis and Quantitative Evaluation through Geological Processesis used. Moon et al (1995) have also used a neural network integrated with a KBS for preliminary design of slopes.
TUNNELS AND UNDERGROUND OPENINGS
Many of the systems for rock mass classification described above have an application in design of tunnelling support (Zhang et al, 1988; Juang & Lee, 1989; Madhu et al, 1995; Butler & Franklin, 1990). In addition, Fairhurst & Lin (1985) have discussed the use of a fuzzy methodology in the design of tunnel support systems. Feng & Lin (1992) also present a KBS (OSDES) for tunnel support design. The system considers rock mass classification; groundwater; type, span, and service time of opening; depth of overburden; dynamic, swelling, and rheological properties of the rock.
Ghosh et al (1987) describe a KBS for deciding on rock bolt length and spacing for supporting coal mine roofs. Zhang et al (1991) discuss an early application of a neural network to coal mine support. Similarly, Deb et al (1994) describe a neural network approach to roof stability in long wall mining. This is intended to provide real-time monitoring of leg pressure on shields in order to provide early warning of possible collapse. King & Signer (1994) also describe a neural network approach to selection of roof supports in mining. The neural network was used to identify patterns of discontinuities in coal mine roofs. Zhang et al (1995) describe use of a neural network for forecasting rock deformation in Chinese colliery roadways.
Lee & Sterling (1992) describe a neural network for identification of probable failure modes for underground openings from prior case history information. The neural network forms part of a KBS for assisting with tunnel design (Sterling & Lee, 1992). The neural network is used to identify similar cases to that being designed. The case histories can then be extracted for the the user to examine. Moon et al (1995) have also used a neural network approach integrated with a KBS for preliminary design of tunnels.
Gökay (1993) has made use of Hudson's (1992) systems approach to rock engineering to develop a KBS to assist in rock engineering decisions relating to mine excavation. The system deals with rock mass type and structure; in situ stress state; hydro-geology; mining methods and assits with excavation stability, location, and orientation.
SIMSECTION (Halabe & Einstein, 1994) is a KBS that acts as the user interface for DAT (Decision Aids for Tunnelling). The KBS assists the user with the definition of the problem and provides consistency checking before performing an analysis. Coulthard & Ciesielski (1991) describe SAGA, a KBS to assist with the selection of a stress analysis program for rock excavation design. This can assist with choosing between ten different stress analysis packages.
Zhang et al (1993) present a KBS for prediction of potential disaster due to excavation of tunnels or underground structures within carbonate rock areas. It is based on the knowledge of Chinese experts in karst science and in underground engineering.
Although most AI applications in tunneling have been developed for rock engineering applications, Mi and Jieliang (1989) report on a KBS for soft ground tunnelling. It has been developed to predict the value of surface settlement and the degree of damage to corresponding buildings caused by shield-driven tunnelling. As well as estimating settlement, the system can also propose prevention and strengthening measures. Russell & Alhammad (1993) describe a KBS framework for selection of appropriate construction methods. The approach is illustrated using a prototype KBS, called CMSA (Construction Methods Selection Assistant), to select a shoring system for cut-and-cover tunnelling. A KBS has also been developed for providing assistance for the planning of support for trenches (Konkoly, 1986; Siller, 1987). The system is based on two soil classification systems developed by the US National Bureau of Standards in order to increase the safety of this type of excavation.
MINING
Some of the A.I. systems described in the section on slopes are relevant to opencast mining operations (Kizil & Denby, 1990; Denby and Kizil, 1991; Ozgenoglu & Ocal, 1994). Similarly, some of the systems in the section on underground openings relate to deep mining operations (Ghosh et al, 1987; Zhang et al, 1991; Gökay, 1993; Deb et al, 1994; King & Signer, 1994; Zhang et al, 1995). This section describes mining systems that do not fit into either of these previous categories.
Yao et al (1992), Reddish et al (1994), Reddish et al (1995) present ESDAS (Expert Structural Damage Assessment System). This was developed to evaluate damage due to mining subsidence. The system uses a risk-assessment technique based on certainty factors to predict the likely damage to a particular structure that is subject to mining subsidence.
Yu & Vongpaisal (1996) describe a new blast damage criteria that has been developed with special reference to mining operations. It can be used for assessing damage by incorporating the vibration level, rock properties, site characteristics and the effects of ground support systems. The approach has been used within a ground control KBS module.
LIQUEFACTION
SOLES (Shyu and Hryciw, 1991) is a KBS to assist in the evaluation of liquefaction potential of soil subject to earthquake excitations. SOLES considers four aspects: the earthquake excitation, the soil properties, the analysis results and the overall evaluation. Chouicha et al (1994) describe a KBS called LIQUEFY. This uses five different methods for liquefaction hazard assessment and groups them according to their task.
Goh (1994) has used neural networks to model the complex relationship between seismic and soil parameters in order to investigate liquefaction potential. The network uses the standard penetration test (SPT) value, fines content, grain size, dynamic shear stress, overburden stress, earthquake magnitude, and horizontal acceleration at the ground surface as inputs. Goh (1996a) has also used neural networks to assess liquefaction potential from cone penetration test (CPT)data.
GROUND IMPROVEMENT
IMPROVE (Chameau and Santamarina, 1989) is a KBS designed to assist in the selection of ground improvement techniques. It contains a case-based system that selects case histories that best resemble the project. Similarly, Motamed et al (1991) describe a KBS (ESPGIS) to advise on the selection of ground improvement methods. EPSGIS allows the user to define the problem by specifying (with varying degrees of certainty) the nature of the ground improvement needed, subsurface conditions and other relevant parameters.
Yoon et al (1994) describe what they call a knowledge database for ground improvement technologies. This contains information on the current technologies available, classified by country of use and application. The knowledge base contains information on international/national codes of practice, design methods, state of practice, and case studies.
Kotdawala & Hossain (1994) describe a KBS (PACT) for soil compaction. The system identifies the lift thickness and moulding moisture content to be used in field compaction. It has knowledge of the different types of compaction plant, and particular problems associated with compaction of particular soils. A neural network approach for soil compaction is reported by Basheer & Najjar (1995), Najjar et al (1996). This is intended for predicting optimum moisture content (OMC) and maximum dry density (MDD) based on soil type, grading characteristics, and consistency limits. For natural soils, they have based the prediction on only three variables: liquid and plastic limit, and specific gravity.
GEOTEXTILES
A KBS is described by Maher and Williams (1991) that selects geosynthetic materials and performs detailed designs for different geotechnical applications. The knowledge incorporated in the system contains information about material selection for five different geosynthetic uses, such as stabilisation to reduce erosion, separation of soil layers, reinforcement to improve soil strength, drainage material to remove water, and filtration to reduce cross plain flow of soil particles.
Edge Drain by Expert System (EDxES) has been developed by Dimmicket al (1991) to assist in the design and specificationof the geotextile component of a road pavement edge drain. Thesystem considers commercially available geotextiles and producesa list of the ten thinnest (lightest) candidate products arrangedin ascending order. The system described by Dukes et al(1994) for road design also incorporates the design of a geotextilelayer.
Mannsbart & Resl (1993) describe a KBS for design using Polyfeltgeotextiles. The basis for the design is the design charts inthe technical manual 'Polyfelt TS Design and Practice' and canbe used for the following applications: road construction, hydraulicconstruction, drainage systems, retaining walls, and geomembraneprotection.
GROUND WATER/DAMS
Sieh et al (1988) describe a KBS developed to assist inthe diagnosis of seepage from embankment dams. The system attemptsto define the type of problem (point source seepage, non-pointseepage, sandboils, sinkholes, drainflow), the seriousness ofthe problem and a recommended course of action. EXSEL (Asgianet al, 1988) is another KBS constructed as a diagnostictool for seepage problems associated with dams, such as earth dams,rockfill dams, concrete dams, and roller compacted dams. Ohnishi& Soliman (1995) have used a neural network approach to investigateseepage under a concrete dam founded on rock.
Engel & Beasley (1991) describe a dam site selector (DSS)system. It was developed for use in a graduate-level hydrogologydesign course and can assit with rating potential reservoir sites.
GroundWater eXpert (GWX), presented by Davey-Wilson & May(1989) and Davey-Wilson (1991b), is a KBS that has been developedto advise on appropriate methods for groundwater control in excavations.In its latest version (Davey-Wilson, 1993), the knowledge basecontains information on each of 26 possible methods.
Gribb & Gribb (1994) and Najjar and Basheer (1996) have bothused neural network approaches for estimating permeability. Najjar& Basheer use thirteen input parameters including classificationtest data (Liquid limit, activity, percent clay etc), density,type of compaction, and weight of compactor in order to predictthe permeability of compacted clay liners.
ROADS AND EARTHWORKS
Pearse et al (1986) describe a KBS being developed forthe evaluation of road corridors taking into account finance,safety, and engineering geological criteria. The system will givea cost for each potential road corridor and a probability of failurewithin its design life, as well as a summary of the main advantagesand disadvantages of each alignment.
Goh (1993) describes a KBS (PAVEDKB) for design of flexible roadpavements. The KBS assists with selection of appropriate soilparameters for the subgrade and also for the properties of thepavement materials. It is linked to algorithmic routines for linearelastic analysis of the pavement structure.
Dukes et al (1994) describe a KBS (ROAD) for design ofprimary and major road highways. It is based on AASHTO designprocedures, and allows the inclusion of a geotextile layer. Itconsiders the mechanical and filtration properties of the geotextilein the design.
Amirkhanian & Baker (1992) describe a KBS for selecting equipmentfor earthmoving operations. The system interprets informationconcerning the soil conditions at a site, operator performanceand required earthmoving operations. PACT (Kotdawala & Hossain,1994) is a KBS that focuses on field compaction. It has knowledgeof the different types of compaction plant, and particular problemsassociated with compaction of particular soils.
DISCUSSION
A significant number of artificial intelligence (AI) systems havebeen developed for geotechnical applications. This is evident from the large number (160) of references listed in this paper.Many of the systems developed are simple prototypes, i.e. they have been developed to show that the techniques could be useful.Relatively few are being used commercially at present. However, as the area develops, and more systems develop beyond the prototypephase, we should see an increase in the use of AI systems in practice.
The majority of the earlier systems used simple rule-based technologies. As time has progressed more complex forms of representation have been adopted, e.g. frame-based, object oriented. We are now seeing a limited number of case-based systems being applied in geotechnical engineering and no doubt this area will develop further. The use of neural networks has developed rapidly and a considerable number of geotechnical applications are now available. Genetic algorithms have yet to be exploited as geotechnical engineering tools but they will certainly have applications in the design synthesis (conceptual design) area.
Knowledge-based system developers identify different intended uses for their systems. Some authors see their knowledge-based systems developing to the stage of becoming 'expert systems' where they would be capable of reasoning at the level of a human expert. They would see the role of their system as replacing human expertise. This is often the case with developers relatively new to the field who do not appreciate the enormous difficulties of acquiring the knowledge required for such a system.
However, as the area of AI has progressed, more seasoned developers now see their systems in a support role, as decision support tools or 'assistants'. It is my belief that we should be developing AI tools in this way, rather than pretending that such systems will replace human expertise. We must also convey this change of philosophy to potential users. There is otherwise a danger that engineers in practice will see AI systems as a threat rather than something that will benefit them (Toll, 1990).
Developers are also finding that AI technology is good for some aspects of solving engineering problems, but that other approaches are still valid for many applications. Therefore, we are increasingly seeing the development of hybrid systems which mix knowledge-based, neural networks, case based, probablistic and algorithmic approaches.
CONCLUSIONS
A large number of artificial intelligence (AI) systems have now been developed for geotechnical applications. Although many of the systems described are simple prototypes, some systems are progressing beyond the developmental prototype phase. Therefore, we should soon start to see an increase in the use of AI systems in geotechnical practice.
It is suggested that AI systems should be developed as decision support tools or 'assistants', rather than pretending that AI systems will develop to the stage where they could replace human expertise. It should also be recognised that AI techniques are good for some aspects of solving engineering problems, but that other approachs are still valid for many applications. Therefore, the way forward will be the development of hybrid systems which mix knowledge-based, neural networks, case based, probablistic and algorithmic approaches.
ACKNOWLEDGEMENTS
In compiling this paper I have drawn heavily on an earlier review by Moula, Toll & Vaptismas (1995). I acknowledge the work of Dr Marina Moula and Dr Nikitas Vaptismas in producing that paper.
REFERENCES
