PSI - Issue 64

1526 Manish Prasad et al. / Procedia Structural Integrity 64 (2024) 1524–1531 Manish Prasad / Structural Integrity Procedia 00 (2019) 000 – 000 3 where is the beam width, d is the effective depth, is the longitudinal reinforcement ratio, and ′ is the compressive strength of concrete. In a subsequent experimental investigation, they observed that the load at which CCS occurred was always higher than the V c . Therefore, to estimate the load causing CCS, V db,end , they introduced a multiplication factor η to the shear capacity of the concrete V c as expressed in Eq.(3) (Smith & Teng, 2002b). (3) In addition, CCS is also influenced by crack spacing at failure (D. Zhang et al., 2012), for which analytical prediction is not very accurate. Besides that, modeling CCS using principles of mechanics of materials involves complex combinations of stresses that usually lead to simplifying hypotheses of the phenomenon, leading to poor prediction of the failure load. To deal with CCS failure, Al-Ghery et al. (Al-Ghrery et al., 2021) proposed a model using Gene Express Programming (GEP), a regression model based on fitting the data with given sets of mathematical functions. The prediction of CCS load with GEP model were in better agreement with the experimental results than the analytical models discussed in their paper. Machine Learning (ML) regression models have been proven very effective in the field of predicting mechanical behavior of concrete elements strengthened with EB FRP such as its shear strength (Abuodeh et al., 2020) or the FRP to concrete bond strength (F. Zhang et al., 2023). However, to the best of authors ’ knowledge, application of ML has not been explored yet for predicting CCS. This paper presents an application of ML to predict CCS failure load in RC beams strengthened with EB CFRP using data from existing experimental tests on beams that failed by CCS. Data on 140 four-point bending tests, from 45 existing experimental programs have been collected and used to train four ML models: K-nearest neighbor, support vector machine, random forest, and extreme gradient boosting. 2. Regression with machine learning Regression is one of the common methods to predict a behavior inherent to a dataset, if there is enough data, and that the response follows a mathematical trend. However, when there are many independent variables influencing a target output, it becomes an intensive task to perform the regression manually. Machine learning (ML) is an iterative procedure which recognizes information about a dataset and makes prediction on new data based on the learnt information. With modern computational power, ML can perform regression with large amounts of data and find the relationships between the input and output variables. In this paper, ML algorithms are trained with a database on existing experimental data from four-point bending tests of beams externally strengthened with CFRP that failed with CCS. In this case, we are dealing with a regression problem, as we are trying to predict a value (CCS failure load). Therefore, we use predictive ML methods for regression. 2.1. Data collection and pre-processing The first step to a ML application is data collection. This data will be used to train and validate the ML model. Both the quality and quantity of the data affect the performance of an ML model. Quality of data can be measured in terms of its distribution and missing values of one or more features. Quantity of data can be measured in terms of its size to properly train an ML model. Data preprocessing involves shaping the raw collected data to make it compatible for ML algorithm of choice. Some common steps are to remove rows with missing data and outliers. Variables, also called features, in a dataset can be of different orders of magnitude according to the nature of their origin (as an example, value for FRP thickness can be 1 mm and that for FRP elastic modulus can be 170000 MPa), or simply due to their units. ML is sensitive to this; therefore, data are usually transformed in a similar scale using scaler or normalization methods (De Amorim et al., 2022). Finally, the pre-processed data is split into two sets, one to be used to train the model and the other to test the fitting of the trained model. , db end c V V  =