Keywords: data mining, k-nearest neighbours, experimental database, ultrasonic pulse velocity, concrete, high temperatures.

The evaluation of the structural stability of a concrete structure that was subject to fire, as well as the consequent decisions that should be taken about its upgrading or demolishing are tasks of high complexity and the responsibility that requires the use of appropriate equipment of inspection, and the application of sophisticated numerical models. As a result of the scarcity of this type of model, as well as material constitutive laws capable of simulating, with enough accuracy, the residual strength of concrete structures submitted to fire, research in this domain is quite necessary and relevant.

The possibilities of using data mining (DM) techniques for the assessment of the compression behaviour of concrete columns after having been subject to temperature exposure were explored. The raw database used in this study was collected from an experimental program of non-destructive tests that was carried out measuring ultrasonic pulse velocity (UPV) through concrete column specimens of various steel reinforcement arrangements and exposed to several levels of high temperature (250°C, 500°C and 750°C). The concept of relative strain (strain divided by the strain at peak stress) was selected as the key variable, and several DM techniques were used taking the ultrasonic pulse velocity (UPV), the exposed temperature (T) and the type of reinforcement arrangement as the known variables of a database.

The comparison of performance measures shows that the technique based on k-nearest neighbours (k-NN) is the best in the prediction of relative strain followed by the non-linear techniques based on support vector machines and neural networks. The top-down hypothesis of predicting the relative strain as a function of the UPV, T and the reinforcement arrangement, allowed a good reliability, particularly by using the k-nearest neighbour technique.

A sensitivity analysis shows that for any of the DM models used to predict the relative axial strain, the UPV, as expected, has the predominant importance (close to 70%) on the value of this parameter. Because of the low values of the importance of the reinforcement arrangement variable (1 to 7%), particularly for the k-NN model, it is recommended that the relative strain is adjusted as a function of UPV and T only. However, by using this procedure a lower reliability is obtained for the k-NN model, with a correlation coefficient of 0.836. If only the input variable UPV is considered the correlation coefficient is 0.619. The highest capability of the k-nearest neighbour technique can be achieved from its capacity of clustering data, based on the T and the reinforcement arrangement variables. On the other hand, the most effective parameter influencing the accuracy of estimated relative strain of fire-damaged concrete elements is identified as the UPV in the concrete.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £78 +P&P)