Hybrid 1D-CNN and attention-based Bi-GRU neural networks for predicting moisture content of sand gravel using NIR spectroscopy

• Published in: Construction & building materials 2022-10, Vol.350, p.128799, Article 128799
• Main Authors: Yuan, Quan, Wang, Jiajun, Zheng, Mingwei, Wang, Xiaoling
• Format: Article
• Language: English
• Subjects: Characteristic wavelength selection; Convolutional neural network; Gated recurrent unit neural network; Moisture content prediction; Multi-head self-attention mechanism; Near-infrared spectroscopy; Sand gravel
• ISSN: 0950-0618; 1879-0526
• DOI: 10.1016/j.conbuildmat.2022.128799

•A hybrid 1D-CNN and attention-based Bi-GRU model is developed for moisture content prediction.•The proposed model can simultaneously extract spectral local abstract information and position information.•The proposed model shows superior performance on both LUCAS and sand gravel spectral datasets.•A moisture content characteristic wavelength (CW) screening process is established.•The top ten CW points are calculated to help realize low-cost discrete NIR spectrometer. A non-destructive and rapid moisture content detection method of sand gravel material is required in loose material dams. The near-infrared (NIR) spectrum of sand materials is closely related to its moisture content. Recently, there is a growing need for fully using spectral information when establishing calibration models for sand gravel moisture content detection. To address these issues, a hybrid one dimensional-convolutional neural network (1D-CNN) and attention-based bidirectional gated recurrent unit (Bi-GRU) neural network was proposed to detect sand gravel moisture content with NIR spectrum. Two learners, namely, 1D-CNN and Bi-GRU, were constructed to extract local abstract information and sequence position information from the spectrum, respectively. In the 1D-CNN learner, multiple kernels CNN layers and one dimensional-separable convolution layers were conjunct to improve model accuracy and reduce network parameters. In the Bi-GRU learner, a multi-head self-attention mechanism was appended to evaluate the weights of the output features extracted by Bi-GRU layers. The proposed model achieved the best prediction results in LUCAS dataset (R2 greater than 0.75, RPD greater than 2.0) and our sand gravel spectral dataset (R2 = 0.96, RPD = 5.06) compared to other deep learning and conventional spectroscopy analysis methods. In addition, the top ten characteristic wavelength points of sand gravel were identified. These can be used to choose a discrete spectrum measuring instrument, which has a relatively low cost.