Concrete classification using Wi-Fi channel state information and convolutional neural networks

• Published in: Construction & building materials 2024-08, Vol.439, p.137280, Article 137280
• Main Authors: Gacem, Mohamed A., Zakaria, Amer, Ismail, Mahmoud H., Yehia, Sherif, Tariq, Usman
• Format: Article
• Language: English
• Subjects: Channel State Information (CSI); Classification; Concrete; Convolutional Neural Networks (CNN); Non-destructive testing; Wi-Fi sensing
• ISSN: 0950-0618; 1879-0526
• DOI: 10.1016/j.conbuildmat.2024.137280

Concrete is one of the most versatile and commonly used composite materials in construction, with its properties being influenced by the type and proportion of materials used in its mixing process. Ensuring accurate mixing is crucial for a structure’s durability and safety. Therefore, assessing concrete mixture quality, homogeneity, and composition post-hydration is vital to meet application requirements. This study proposes a novel non-destructive approach for concrete classification, focusing on homogeneity, conductivity, and composition analysis in dry concrete mixtures. The proposed technique is based on analyzing changes to Wi-Fi signals that are radiated through concrete specimens. These variations are in the form of changes to the magnitude and phase of the Wi-Fi’s Channel State Information (CSI). These data are then utilized to train Convolutional Neural Network (CNN) classifiers for concrete mixture categorization. The proposed method successfully estimates homogeneity across concrete samples with varied particle sizes and achieves a 100% validation accuracy in estimating conductivity and composition. Additionally, distinct CSI responses are observed for mixtures with differing compositions. The validation experiments with new concrete samples demonstrate the models’ robustness and generalization capability, achieving a 78.18% correct prediction rate despite variations in dimensions and compositions. While the models exhibit high accuracy in detecting certain materials, such as silica fume and fly ash, NW aggregates, and steel fiber, challenges persist in detecting materials like carbon powder due to potential confusion with other conductive materials. Overall, the trained CNN models demonstrate strong generalization ability and robustness, highlighting their potential for accurate concrete property detection in various scenarios. •Commodity Wi-Fi is utilized to estimate the properties of concrete non-destructive.•System estimates three different concrete properties using the same collected measurement.•Trained convolutional neural networks are used to estimate the properties of concrete.•Channel state information (CSI) measurements of the Wi-Fi signal are used to train CNN and to estimate the concrete properties.