基于压缩感知理论的大数据矿产资源评价模型研究与应用示范

Big-data analyzing is the new challenge faced by the modern methodology of mineral resources assessment. Most methods of mineral resources assessment are low efficient for processing big-data. Thus, our research team plans to introduce compressive sensing, which is powerful for analyzing and processing big-data, to the field of mineral resources assessment. We plan to use sparse representation, sparse sampling, and sparse reconstruction, which are the three core techniques for compressive sensing to analyze and process multi-sourced geologic big-data and establish a model system for compressive sensing based mineral resources assessment on big-data. The research content includes: (a) Develop a similar analogy model system for compressive sensing based mineral resources assessment by referring to the research results of compressive sensing based face and handwriting digits recognition and etc.; (b) Develop a dissimilation model system for compressive sensing based mineral resources assessment by referring to the research results of compressive sensing based anomaly event detection in wireless networks and unreconstructed compressive detection of unknown-parameters signals and etc.; and (c)Develop sparse maximum learning machines for mineral resources assessment based on sparse representation and maximum learning machines. We plan to use these three model systems to the mineral resources assessment in the experimental area and set up a typical example of applying compressive sensing to mineral resources assessment on multi-sourced geologic big-data. The research can promote both application of compressive sensing in geosciences and development of the methodology of modern mineral resources assessment. Therefore, it has an important theoretical and practical significance.

大数据分析是当代矿产资源评价方法体系面临的新挑战，绝大多数矿产资源评价方法的大数据处理效率偏低。因此，项目组拟将大数据分析处理功能强大的压缩感知理论引入矿产资源评价领域，用稀疏表示、稀疏采样和稀疏重构三种压缩感知核心技术分析处理多源地学大数据，创建大数据矿产资源评价的压缩感知模型体系。内容包括：①借鉴压缩感知在人脸识别和手写数字识别等方面的研究成果，研制基于压缩感知的“相似类比”矿产资源评价模型体系；②借鉴压缩感知在无线网络异常事件检测和未知信号非重构压缩检测等方面的研究成果，研制基于压缩感知的“求异”矿产资源评价模型体系；③在稀疏表示和极限学习机基础上，研制矿产资源评价稀疏优化极限学习机系列模型。把三种矿产资源评价模型体系应用于实验区，创建大数据矿产资源评价的压缩感知模型体系应用范例。研究工作能够推动压缩感知地学应用研究，促进当代矿产资源评价方法体系的发展，具有重要的理论和实际意义。

压缩感知理论由稀疏采样、稀疏表示和稀疏重构三大核心技术组成。稀疏采样是用测量矩阵把高维信号投影到低维空间；稀疏表示是用过完备字典把原信号变换为稀疏系数；稀疏重构是通过求解欠定方程组把观测值无损或高概率恢复成原始信号。压缩感知能够有效解决高维大数据的模式识别问题，已成功应用于人脸识别、手写数字识别和无线网络分布式异常事件检测等。鉴于压缩感知理论在无线网络分布式异常事件检测方面的优异性能，项目组把矿产资源评价过程划分成地球科学大数据训练过完备字典、借助过完备字典把统计单元变换成稀疏系数向量、根据稀疏系数向量和过完备字典重建统计单元、计算统计单元与其重建之间的欧氏距离四个关键环节。在此框架下，成功研制出了矿产资源远景区预测的字典学习模型，并应用于实验区的矿产资源远景区预测。通过对比字典学习模型、逻辑斯蒂回归模型和单类支持向量机模型的矿产预测性能，发现字典学习模型可与逻辑斯蒂回归模型相媲美，其性能明显优于单类支持向量机模型。可见字典学习模型是一种高性能的矿产资源远景区预测模型。借鉴矿产资源远景区预测的字典学习模型框架，项目组研制出多元地球化学异常识别的字典学习模型并应用于实验区，建立了多元地球化学异常识别的5种字典学习模型，对比了5种字典学习模型、k-最近邻模型(KNN)、组合KNN模型和高斯混合模型(GMM)的多元地球化学异常识别性能，发现5种字典学习模型的多元地球化学异常识别性能明显优于其它三种模型。因此，字典学习模型是高性能的多元地球化学异常识别模型。此外，项目组把邻域主成分分析与字典学习相结合，创建了矿化异常识别的邻域主成分-字典学习混合模型，在实验区应用取得了良好效果。这些重要进展对完善大数据矿产资源评价方法体系和多元地球化学异常识别方法体系具有很好的促进作用。