Advanced geological prediction method and system based on perception while drilling

An advanced geological prediction method based on perception while drilling, comprising:

• acquiring drilling parameters during an advanced drilling;
• obtaining physical and mechanical parameters of tunnel surrounding rocks by carrying out an inversion of the drilling parameters through a trained prediction model;
• acquiring rock slag or rock powder from a flushing fluid collected during drilling;
• acquiring geochemical characteristic parameters of the rock slag or the rock powder;
• inputting the physical and mechanical parameters of tunnel surrounding rocks and the geochemical characteristic parameters into a pre-trained deep learning model, to analyze engineering geological conditions ahead of a tunnel face of the tunnel;
• and
• carrying out a directional intervention on unfavorable geological conditions recognized and predicted ahead of the tunnel face of the tunnel or adjusting tunnel engineering construction parameters according to analysis results output by the pre-trained deep learning model, to avoid geological disasters in the tunnel;
• wherein, using the pre-trained deep learning model to process the physical and mechanical parameters of tunnel surrounding rocks and the geochemical characteristic parameters, and outputting the analysis results, comprises:
• performing a characteristic extraction on the input physical and mechanical parameters of tunnel surrounding rocks and the geochemical characteristic parameters respectively by using a first-layer structure of the pre-trained deep learning model comprising more than one fully connected layer, then concatenating the characteristics after the extraction;
• performing continuously the characteristic extraction on the characteristics after the concatenation by using a second-layer structure of the pre-trained deep learning model comprising more than one fully connected layer, to obtain a characteristic extraction vector X;
• concatenating the characteristic extraction vector X and a transfer parameter Ht-1, then being multiplied with different weight matrices, and then converted by an activation function, to obtain matrices Y1, Y2, Y3, and Y4 between −1 and 1;
• wherein, the transfer parameter Ht-1 is a parameter obtained by self-learning of the pre-trained deep learning model;
• performing characteristic extraction the matrices Y1, Y2, Y3, and Y4 respectively by using a plurality of residual fully connected blocks, to obtain characteristic vectors Z1, Z2, Z3, and Z4;
• performing point multiplication of the characteristic vector Z1 and a parameter Ct-1 introduced in a previous analysis of the pre-trained deep learning model, to obtain an attention matrix A1;
• performing point multiplication of the characteristic vectors Z2 and Z3, to obtain an attention matrix A2;
• obtaining a transfer parameter Ht for a next analysis of the pre-trained deep learning model by adding the attention matrices A1 and A2 and activating them through a first activation function;
• obtaining an attention matrix A3 by adding the attention matrices A1 and A2 and activating them through a second activation function, and then point multiplication with the characteristic vector Z4;
• and obtaining a parameter Ct for the next analysis of the pre-trained deep learning model by activating a result of the point multiplication with the characteristic vector Z4 through the first activation function;
• and
• outputting the analysis results obtained by calculating the attention matrix A3 by using a cross entropy function.