Water source identification in mines combining LIF technology and ResNet

Abstract

The problem of mine water source has always been an important hidden danger in mine safety production. The water source under the mine working face may lead to geological disasters, such as mine collapse and water disaster. The research background of mine water source identification involves many fields such as mining production, environmental protection, resource utilization and technological progress. It is a comprehensive and interdisciplinary subject, which helps to improve the safety and sustainability of mine production. Therefore, timely and accurate identification and control of mine water source is very important to ensure mine production safety. Laser-Induced Fluorescence (LIF) technology, characterized by high sensitivity, specificity, and spatial resolution, overcomes the time-consuming nature of traditional chemical methods. In this experiment, sandstone water and old air water were collected from the Huainan mining area as original samples. Five types of mixed water samples were prepared by varying their proportions, in addition to the two original water samples, resulting in a total of seven different water samples for testing. Four preprocessing methods, namely, MinMaxScaler, StandardScaler, Standard Normal Variate (SNV) transformation, and Centering Transformation (CT), were applied to preprocess the original spectral data to reduce noise and interference. CT was determined as the optimal preprocessing method based on class discrimination, data distribution, and data range. To maintain the original data features while reducing the data dimension, including the original spectral data, five sets of data were subjected to Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) dimensionality reduction. Through comparing the clustering effect and Fisher’s ratio of the first three dimensions, PCA was identified as the optimal dimensionality reduction method. Finally, two neural network models, CT+PCA+CNN and CT+PCA+ResNet, were constructed by combining Convolutional Neural Networks (CNN) and Residual Neural Networks (ResNet), respectively. When selecting the neural network models, the training time, number of iterative parameters, accuracy, and cross-entropy loss function in the classification problem were compared to determine the model best suited for water source data. The results indicated that CT+PCA+ResNet was the optimal approach for water source identification in this study.