Loading...
Construction Materials Classification Using Wi-Fi and Convolutional Neural Networks
Gacem, Mohamed Ait
Gacem, Mohamed Ait
Date
2021-11
Author
Type
Thesis
Degree
Citations
Altmetric:
Description
A Master of Science thesis in Electrical Engineering by Mohamed Ait Gacem entitled, “Construction Materials Classification Using Wi-Fi and Convolutional Neural Networks”, submitted in November 2021. Thesis advisor is Dr. Mahmoud Ibrahim and thesis co-advisors are Dr. Amer Zakaria and Dr. Usman Tariq. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
The process of identifying the properties of a construction material is vital in several industrial and quality assurance applications. Such process normally requires not causing damage to the observed sample and having a high accuracy with low cost of implementation. In this work, a novel non-destructive construction materials classification tool is proposed. The proposed method is based on passing Wi-Fi signals through the observed samples, then analyzing the Channel State Information (CSI) amplitude and phase components. Wi-Fi signals are affected by the channel variations in the form of amplitude attenuation and phase shift. Hence, placing different objects in the channel will yield different CSI responses. The collected CSI data packets are pre-processed by performing an averaging operation. Then the resulting data are divided into training and validation sets to be used to train a Convolutional Neural Network (CNN). The trained CNN models are formulated as classifiers when the data can be sorted into specific classes. Alternatively, the models are formulated as regression models when the data have a continuous nature. The proposed method is used to classify materials through two phases. The main goal of phase 1, is to investigate the proposed method’s potential to perform materials classification using a relatively simple set of samples, which are composed of one type of materials. The experiments are namely, thickness estimation of Plexiglas, estimating the water content value in fine materials and confirming the compaction of individual materials. While in phase 2, the studied samples are heterogeneous, which makes them more challenging since having several materials with different properties in the same mixture generates a more complex response. Nevertheless, such samples are ideal in highlighting the proposed method’s generalization efficacy and possible limits. The experiments in phase 2 include confirming concrete mixtures homogeneity and detecting the presence of a specific construction material within a concrete mixture. The obtained experimental results effectively demonstrate the potential and merits of the proposed method. Overall, the CNN models achieved a 100% validation accuracy and a low validation loss, which confirms that the method is valid and highly accurate.