ایجاد قوانین پیش بینی برای مایع سازی از طریق داده کاوی

Prediction of liquefaction is an important subject in geotechnical engineering. Prediction of liquefaction is also a complex problem as it depends on many different physical factors, and the relations between these factors are highly non-linear and complex. Several approaches have been proposed in the literature for modeling and prediction of liquefaction. Most of these approaches are based on classical statistical approaches and neural networks. In this paper a new approach which is based on classification data mining is proposed first time in the literature for liquefaction prediction. The proposed approach is based on extracting accurate classification rules from neural networks via ant colony optimization. The extracted classification rules are in the form of IF–THEN rules which can be easily understood by human. The proposed algorithm is also compared with several other data mining algorithms. It is shown that the proposed algorithm is very effective and accurate in prediction of liquefaction

In soil deposits in undrained condition, earthquakes cause cyclic shear stresses, which may lead to liquefaction. Liquefaction indicates a condition where a soil will subject to a deformation, because of the build-up and maintenance of high pore water pressure. In a soil deposit subjected to dynamic load the pore water generation increases, leading to a reduction in its strength and then liquefaction of the soil. Field evidence of liquefaction usually consisted of observed sand boils, ground fissures or lateral spreads. Liquefaction in loose and saturated sands and silty sands is one of the causes of tilting or subsiding in civil engineering structures. In the literature, there are several methods developed to determine the liquefaction potential of soil deposits using the in situ tests results, such as; standard penetration test (SPT), and cone penetration test (CPT). Simplified methods have been used by scientists to evaluate nonlinear liquefaction potential of soil. Derived from several laboratory and field tests, various simplified procedures (e.g., strain-based, stress-based, Chinese criteria) have been developed by using case studies and undisturbed soil specimens. Over the past 35 years, a stress-based procedure, termed as ‘simplified procedure’, has evolved for evaluating the liquefaction resistance of soils. Following the earthquakes in Alaska and Niigata, Japan in 1964, Seed and Idriss (1971) developed this method. The simplified procedure was developed from evaluations of experimental data and field observation. The number of cycles and the shear stress level are the major requirements in this method. Equivalent stress intensity and the number of cycles need to be determined to correlate actual motion of an earthquake to laboratory harmonic loading conditions. This equivalent shear stress was selected as 65% of the maximum shear stress induced in an earth structure in the study by Seed, Idriss, Makdisi, and Banerjee (1975). However, Ishihara and Yasuda (1975) said that the equivalent shear stress is 57% for 20 cycles of loading. Although the procedure has been corrected and augmented periodically since then with the studies by Seed, 1979, Seed and Idriss, 1982 and Seed et al., 1985, the uncertainty concerning random loading still exists (Baziar & Jafarian, 2007). The results of research by Dobry, Ladd, Yokel, Chung, and Powell (1982) were employed to assess the liquefaction potential of soils which is known as strain-based approach. According to the strain-based approach, pore water pressure initiates to develop when shear strain surpass a threshold shear strain level. Dobry et al. (1982) concluded that the threshold shear strain is approximately 0.001%, and independent of relative density, sand type, initial effective confining pressure and sample preparation technique. The same uncertainty is available for this approach as well, since the Dobry et al. (1982) thought the same equivalent number of earthquake loading as the stress-based approach. Over last three decades, various researchers have studied on an energy-based liquefaction assessment approaches. The amount of frictional energy loss needed to liquefy a soil is dependent on contact density, confining stress, and frictional characteristics of the soil. The cumulative energy loss up to liquefaction has been identified as a useful index for liquefaction potential of soils (Berrill and Davis, 1985, Nemat-Nasser and Shokooh, 1979, Okada and Nemat-Nasser, 1994, Thevanayagam et al., 2000 and Trifunac, 1995). This idea has also come up with the development of energy based liquefaction potential evaluation procedures (Green, 2001 and Kayen and Mitchell, 1997) based on penetration resistances and arias intensity. Researches have proposed energy-based models relating pore water pressure increment ratio, to dissipated strain energy density, loading parameters (cyclic stress ratio, strain level), initial void ratio, or relative density, initial effective confining pressure, and some calibration parameters from curve fitting of experimental data. Parameters used in this approach can be directly related to seismological parameters (Green, 2001). The main advantage of the energy-based approach is that the shear energy required to liquefy a soil deposit is not dependent on the stress history. Geotechnical engineers often solve complex problems having a serious of interacting parameters. However, in some problems, these parameters are not well defined and it is not possible to identify a relationship between the parameters, or the problem could be too complex to be described in a mathematical function. One of the most common approaches to solve these problems uses experimental data to develop empirical models that relate the variables in the system. Modern techniques such as fuzzy systems and neural networks have been considered to develop models based on data owing to their ability to learn and recognize trends in data patterns (Goh & Goh, 2007). For example, artificially intelligence (AI) applications have been used in several applications in civil engineering. Artificial neural network (ANN), one of the AI approaches, has the capability to mimic the learning abilities of human brain by processing the data. ANN models were recently adopted by researchers in the area of geotechnical engineering (e.g., Abu Keifa, 1998, Baziar and Jafarian, 2007, Chen et al., 2005, Goh, 1994, Goh, 1995, Hanna et al., 2004, Hanna et al., 2007a, Hanna et al., 2007b, Teh et al., 1997 and Wang and Rahman, 2002). Such as, Baziar and Jafarian (2007) developed an ANN model to correlate some of the soil parameters with the strain energy required liquefaction triggering. They carried out parametric study and confirmed the results using previously published cyclic triaxial, simple shear and torsional shear test results. Chen et al. (2005) assessed the liquefaction probability using an energy-based method with back-propagation neural network. The method proposed by them has a capability in evaluating the probability of soil liquefaction. Hanna et al., 2007a and Hanna et al., 2007b proposed an alternative general regression neural network model for the two major earthquakes in Turkey and Taiwan. The proposed model was found to be a viable tool to assess the seismic conditions susceptible to liquefaction. Khozaghi and Choobbasti (2007) made a prediction for the liquefaction potential in south-east of Tehran city in Iran using neural network approach. Comparing their results with the results of Seed (1979) method, they found that the method they proposed using neural network made a prediction with 92% accuracy for the studied area. Apart from the recent studies, Wang and Rahman (2002) made one of the considerations that can be particularly connected with the neural network model for horizontal ground displacement caused by liquefaction. They had indicated that the neural network model serves as a reliable and simple predictive tool for the amount of ground displacement. As it can be seen from the literature review, ANNs have found application potential in prediction of liquefaction. Although, ANN can achieve high prediction accuracies, an important drawback of them is their very limited explanation capability. ANNs are generally considered as black boxes as it is very difficult to understand how they learn and solve a given problem. In order to overcome this drawback several approaches have been proposed in data mining/pattern recognition research area. In these approaches, trained ANN is processed by another algorithm (which is generally a metaheuristic algorithm) in order to extract accurate classification rules in the form of IF–THEN rules which can be easily understood by human decision makers. In the present paper, a data mining algorithm which is recently proposed by the authors is used to predict liquefaction first time in the literature. The proposed data mining algorithm works on trained ANN to extract accurate prediction (classification) rules by making use of an ant colony optimization algorithm. The algorithm is named as TACO-miner ( Özbakır, Baykasoğlu, & Kulluk, in press). The details of the proposed ant colony algorithm are given in the following sections of this paper. Several other classical data mining algorithms along with a powerful genetic programming based technique (MEPAR-miner, which is also recently proposed by authors; Baykasoğlu & Özbakır, 2007) are also applied to liquefaction prediction problem in order to compare the performance of the proposed algorithm and to show how classification data mining techniques can be used.

Prediction of liquefaction is very important in geotechnical engineering. Prediction of liquefaction is a hard problem as it depends on many different physical factors, and the relations between these factors are highly non-linear and complex. In this paper a new approach which is based on classification data mining is proposed first time in the literature for liquefaction prediction. The proposed algorithm is based on ant colony optimization and neural networks. The algorithm tries to discover the hidden knowledge within the connection weights of trained neural networks. Although neural networks are determined as very accurate classifiers, the black-box nature of them does not allow them to produce understandable classification rules. The present algorithm tries to optimize the activation function values by exploring the search space and then translates this information into a comprehensible classification rules. It uses the advantages of both neural network structure and ant colony optimizer at the same time. The proposed algorithm is experimentally evaluated to evaluate its performance in terms of prediction accuracy on the liquefaction dataset. It is shown that the proposed algorithm is able to generate very accurate prediction rules for liquefaction. The proposed algorithm also outperforms most of the well known rule based classifiers such as C4.5, DecisionTable, NBTree, PART and a powerful genetic programming based algorithm MEPAR-miner from the recent literature.