Estimating cement compressive strength from microstructure images using convolutional neural network

Abstract

The estimation of cement compressive strength (CCS) plays an important role in the quality inspection of cement. Physical experiment using cement compressive strength testing machine measures cement strength exactly, but is destructive for the cement specimen, and is unable to realize the estimation of one specimen repeatedly and continuously. Therefore, various computational intelligence methods are proposed to estimate cement strength, in order to achieve non-destructive and continuous estimation. In these methods, a large amount of experimental data and micro-components need to be measured and the error is also unavoidable. However, different gray values in the cement microstructure images represent different substances during the hydration and include the structure features of micro-components. Therefore, microstructure images reflect the relation between micro-components and CCS, and using these images avoid the measurement of micro-components. In addition, convolutional neural network (CNN) has recently shown a powerful advantage in classification and recognition of images. This study proposes a method to estimate CCS from microstructure images directly using CNN. The method extracts the abstract features of cement image is helpful in reducing measurement error of parameters and accomplishing non-destructive estimation. Furthermore, it provides a solution for simulated microstructure to estimate its strength. Experimental results show that the proposed method has favorable estimation accuracy.