CN-121980541-A - Underground ore grade spectrum fusion detection method based on deep learning

CN121980541ACN 121980541 ACN121980541 ACN 121980541ACN-121980541-A

Abstract

The invention discloses a deep learning-based underground ore grade spectrum fusion detection method, which relates to the technical field of ore spectrum detection and comprises the steps of performing model training by taking a convolutional neural network as a basic framework and taking a spectral line conflict relation set as a constraint condition, obtaining a spectral line conflict recognition model, performing conflict discrimination and self-adaptive digestion on a multi-mode spectrum sample set by utilizing the spectral line conflict recognition model to obtain a purified multi-source spectrum feature vector set, performing quality evaluation and uncertainty estimation on the purified multi-source spectrum feature vector set to obtain a feature reliability weight vector, and performing weighted fusion on the basis of the feature reliability weight vector to generate a spectrum fusion feature vector. According to the invention, the multi-layer feedforward neural network is utilized to carry out regression prediction on the spectrum fusion feature vector, so that deep learning driving detection and prediction reliability quantification of the ore grade are realized, and the stability and practical value of the underground ore grade detection result are improved.

Inventors

LI XIANYING
LIU XU
ZHANG YUNCHI
WANG XUANYI
Xie xingshan
MENG LINGHAO

Assignees

长春黄金设计院有限公司

Dates

Publication Date: 20260505
Application Date: 20260209

Claims (10)

1. The underground ore grade spectrum fusion detection method based on deep learning is characterized by comprising the following steps of, Collecting a multi-source spectrum data set and aligning according to the time stamp and the position code to form a multi-mode spectrum sample set; Performing spectrum preprocessing and spectrum candidate interval extraction on the multi-mode spectrum sample set, and constructing a spectrum conflict relation set based on spectrum overlapping relations among different spectrum modes; Taking a convolutional neural network as a basic framework, taking a spectral line conflict relation set as a constraint condition to carry out model training, obtaining a spectral line conflict recognition model, and carrying out conflict discrimination and self-adaptive resolution on a multi-mode spectrum sample set by using the spectral line conflict recognition model to obtain a purified multi-source spectrum feature vector set; performing quality evaluation and uncertainty estimation on the purified multi-source spectrum feature vector set to obtain feature reliability weight vectors, and performing weighted fusion based on the feature reliability weight vectors to generate spectrum fusion feature vectors; and carrying out regression prediction on the spectrum fusion feature vector through a multi-layer feedforward neural network, and outputting an ore grade detection result and corresponding prediction reliability.
2. The method for detecting the grade spectrum fusion of the underground ore based on deep learning according to claim 1, wherein the steps of forming the multi-modal spectrum sample set are as follows, The multi-source spectrum data set comprises laser-induced breakdown spectrum data, X-ray fluorescence spectrum data and working condition information data; Extracting time stamps and position codes in the multi-source spectrum data set, performing clock synchronous calibration and space coordinate mapping, and outputting a multi-source spectrum data sequence; mapping the multi-source spectrum data sequence to a unified three-dimensional grid coordinate point, and splicing and organizing spectrum data of the same space position according to modes to form a multi-mode spectrum sample set.
3. The method for detecting the grade spectrum fusion of the underground ore based on deep learning according to claim 1, wherein the steps of performing spectrum preprocessing and spectral line candidate interval extraction on the multi-mode spectrum sample set are as follows, The spectral preprocessing comprises denoising, baseline correction, normalization and wavelength calibration; and (3) performing peak detection operation and spectral line interval demarcation on the multimode spectrum sample set subjected to spectrum pretreatment by adopting a peak detection algorithm, and obtaining a candidate characteristic spectral line set.
4. The method for detecting the grade spectrum fusion of the underground ore based on the deep learning of claim 1, wherein the method is characterized in that a spectrum conflict relation set is constructed based on spectrum overlapping relations among different spectrum modes, and comprises the following steps of, Calculating the spectrum similarity among different spectrum modes in the candidate characteristic spectrum line set, and identifying an overlapped spectrum line pair set; and carrying out conflict label assignment on each spectral line pair in the overlapping spectral line pair set to generate a spectral line conflict relation set.
5. The method for detecting the underground ore grade spectrum fusion based on deep learning according to claim 1, wherein the method is characterized in that a convolutional neural network is used as an infrastructure, a spectrum conflict relation set is used as a constraint condition for model training, a spectrum conflict recognition model is obtained, the steps are as follows, Extracting spectral line candidate interval data corresponding to the candidate characteristic spectral line set from the multi-mode spectrum sample set, and pairing with the conflict label to generate a training sample set; and calculating a training loss value based on the training sample set, and carrying out iterative updating on model parameters of the convolutional neural network through a back propagation algorithm to obtain a spectral line conflict recognition model.
6. The method for detecting the grade spectrum fusion of the underground ore based on deep learning of claim 1, wherein the method for detecting the grade spectrum fusion of the underground ore based on deep learning is characterized by comprising the following steps of performing conflict discrimination and self-adaptive resolution on a multi-mode spectrum sample set by utilizing a spectral line conflict recognition model to obtain a purified multi-source spectrum characteristic vector set, Inputting the multi-mode spectrum sample set into a spectral line conflict recognition model for forward reasoning, and outputting a conflict discrimination result set; Based on the conflict judging result set, performing self-adaptive weighted suppression on the spectral line features judged to be conflicts, and generating intermediate spectral features; and fusing and recalibrating the intermediate spectral features and spectral line features which are not judged to be conflicting to obtain a purified multi-source spectral feature vector set.
7. The method for detecting the grade spectrum fusion of the underground ore based on deep learning according to claim 6, wherein the adaptive weighted suppression of spectral line features which are judged to be conflicts is carried out by the following steps, Calculating a corresponding weight coefficient through a parameter mapping relation of a spectral line conflict recognition model based on the conflict strength of spectral line features which are judged to be conflicts in the conflict judgment result set; and performing feature-by-feature self-adaptive weighting adjustment on the spectral line features judged to be in conflict by using the weight coefficient to generate intermediate spectral features.
8. The method for detecting the grade spectrum fusion of the underground ore based on deep learning of claim 1, wherein the method for detecting the grade spectrum fusion of the underground ore based on deep learning is characterized by comprising the following steps of performing quality evaluation and uncertainty estimation on a purified multi-source spectrum feature vector set to obtain a feature reliability weight vector, Calculating the signal-to-noise ratio and variance stability index of each spectrum characteristic dimension based on the purified multi-source spectrum characteristic vector set to generate a characteristic quality score vector; Based on the purified multi-source spectrum feature vector set, calculating a confidence interval and a variation coefficient of each spectrum feature dimension estimation value by adopting a Bootstrap resampling method, and generating a feature uncertainty vector; And performing feature-by-feature reliability weighting calculation and normalization on the feature quality score vector and the feature uncertainty vector, and generating a feature reliability weight vector.
9. The method for deep learning based downhole ore grade spectrum fusion detection of claim 1, wherein the spectrum fusion feature vector is generated by performing a weighting operation on a set of purified multi-source spectrum feature vectors according to a feature reliability weight vector.
10. The method for detecting the spectrum fusion of the underground ore grade based on the deep learning of claim 1, wherein the method for detecting the underground ore grade based on the deep learning of the underground ore grade is characterized by carrying out regression prediction on the spectrum fusion feature vector through a multilayer feedforward neural network and outputting an ore grade detection result and corresponding prediction reliability, and comprises the following steps of, Inputting the spectrum fusion feature vector into a multi-layer feedforward neural network, carrying out deep feature learning and abstraction through layer-by-layer nonlinear transformation, and outputting a high-dimensional abstract feature vector; Performing linear regression on the high-dimensional abstract feature vector through a regression output layer of the multi-layer feedforward neural network to output an ore grade predicted value; in the reasoning process of the multi-layer feedforward neural network, a plurality of times of random forward propagation is carried out on the spectrum fusion feature vector, and the statistical variance is calculated according to the ore grade predicted value output for a plurality of times and is used as the prediction credibility.

Description

Underground ore grade spectrum fusion detection method based on deep learning Technical Field The invention relates to the technical field of ore spectrum detection, in particular to a downhole ore grade spectrum fusion detection method based on deep learning. Background Under the background of mine intellectualization and digital exploitation, the rapid and accurate detection of the grade of the underground ore has important significance for resource assessment, exploitation decision and flow control, the conventional method mainly utilizes a spectrum analysis means of Laser Induced Breakdown Spectroscopy (LIBS) and X-ray fluorescence spectroscopy (XRF) to invert element components and grade by analyzing characteristic spectrums generated by the excitation of ore elements, and in recent years, a deep learning model is gradually introduced along with the improvement of detection technology and data processing capability to model the nonlinear mapping relation between the characteristics of the light and the grade of the ore, so that the detection automation level and the real-time response capability are improved under the complex working condition. However, when multi-source spectral data is utilized, physical overlapping of spectral lines of different modes (such as LIBS and XRF) can exist at specific wavelengths, the conventional method lacks explicit modeling on mutual interference and contribution competition relationship among the spectral lines, unrecognized conflict information is easy to exist in fusion features, the deep learning-based regression prediction is used for directly modeling the fusion features, and the quality difference and uncertainty of input features are concerned only to a limited extent, so that the stability and credibility of model output results are expressed relatively roughly. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides a deep learning-based underground ore grade spectrum fusion detection method which solves the problems that spectral line overlapping conflict among multiple source spectrum modes is difficult to effectively model and the credibility characterization of an ore grade prediction result is insufficient. In order to solve the technical problems, the invention provides the following technical scheme: The invention provides a deep learning-based underground ore grade spectrum fusion detection method which comprises the steps of collecting a multi-source spectrum data set, aligning according to a time stamp and position codes to form a multi-mode spectrum sample set, carrying out spectrum preprocessing and spectrum candidate interval extraction on the multi-mode spectrum sample set, constructing a spectrum conflict relation set based on spectrum overlapping relations among different spectrum modes, carrying out model training by taking a convolutional neural network as a basic framework and taking the spectrum conflict relation set as a constraint condition to obtain a spectrum conflict recognition model, carrying out conflict discrimination and self-adaptive digestion on the multi-mode spectrum sample set by utilizing the spectrum conflict recognition model to obtain a purified multi-source spectrum feature vector set, carrying out quality evaluation and uncertainty estimation on the purified multi-source spectrum feature vector set to obtain a feature reliability weight vector, carrying out weighted fusion on the feature reliability weight vector to generate a spectrum fusion feature vector, carrying out regression prediction on the spectrum fusion feature vector by using a multi-layer feedforward neural network, and outputting an ore grade detection result and corresponding prediction reliability. As an optimal scheme of the deep learning-based underground ore grade spectrum fusion detection method, the method for forming the multi-mode spectrum sample set comprises the following steps of, The multi-source spectrum data set comprises laser-induced breakdown spectrum data, X-ray fluorescence spectrum data and working condition information data; Extracting time stamps and position codes in the multi-source spectrum data set, performing clock synchronous calibration and space coordinate mapping, and outputting a multi-source spectrum data sequence; mapping the multi-source spectrum data sequence to a unified three-dimensional grid coordinate point, and splicing and organizing spectrum data of the same space position according to modes to form a multi-mode spectrum sample set. As an optimal scheme of the underground ore grade spectrum fusion detection method based on deep learning, the method comprises the steps of performing spectrum pretreatment and spectral line candidate interval extraction on a multi-mode spectrum sample set, The spectral preprocessing comprises denoising, baseline correction, normalization and wavelength calibration; and (3) pe