基于WMT-CNN与双阶段光谱匹配的岩矿识别模型研究

Rock and Ore Identification Based on WMT-CNN and Two-Stage Spectral Matching Study

摘要: 针对当前矿物识别中特征工程复杂、多元素同步预测能力不足的问题，本研究提出了一种多模态岩矿智能识别框架。该框架创新性地融合了激光诱导击穿光谱(LIBS)与可见光–短波红外(VIS-SWIR)光谱数据。首先，构建加权多任务卷积神经网络(WMT-CNN)模型，实现对Si、Al、Fe等9种元素的高精度同步定量预测，R²达0.8933。继而，设计基于KPCA-KNN与孪生网络的双阶段光谱匹配算法筛选矿物。最终，通过融合元素定量结果、光谱匹配排序与基团信息，建立“元素–光谱–基团”协同约束机制，实现岩矿大类的智能识别。实验表明，该集成框架对金属矿石和岩浆岩的识别准确率分别达到95.8%和88.9%，显著提升了复杂矿物组分分析的自动化程度与可靠性，为野外高效勘查提供了关键技术方案。

Abstract: To address the challenges in current mineral identification—namely the complexity of feature engineering and the insufficient capability for synchronous prediction of multiple elements—this study proposes a multimodal intelligent rock-ore identification framework. The framework innovatively integrates laser-induced breakdown spectroscopy (LIBS) with visible to shortwave infrared (VIS-SWIR) spectral data. First, a weighted multi-task convolutional neural network (WMT-CNN) model is developed to achieve high-accuracy synchronous quantitative prediction of nine elements, including Si, Al, and Fe, with an R² of 0.8933. Next, a two-stage spectral matching algorithm based on KPCA-KNN and a Siamese network is designed to screen minerals. Finally, by fusing quantitative elemental results, spectral matching rankings, and functional group information, an “element-spectrum-group” collaborative constraint mechanism is established to enable intelligent identification of broad rock–ore categories. Experimental results show that the proposed integrated framework achieves identification accuracies of 95.8% and 88.9% for metallic ores and igneous rocks, respectively, significantly improving the automation and reliability of complex mineral composition analysis and providing a key technical solution for efficient field exploration.