大数据挖掘与机器学习算法应用已成为成矿预测研究的重要趋势,但如何使非结构化地质数据进行机器可读表达成为需要克服的难点。本研究针对粤西庞西垌矿集区开展地层、岩性、断裂等非结构化地质信息面向机器可读的转换处理,并进而应用机器学习算法对水系沉积物地球化学测试数据和构造、地层等综合地质信息进行挖掘,提取致矿异常特征,最终实现智能圈定致矿异常找矿靶区。独热编码算法与空间加权主成分分析中的权重变量方法组合应用,实现了地层、岩性和断裂构造等非结构化地质信息的结构化转化,并最大限度地保留其所包含的地质信息。单分类支持向量机和自编码网络异常检测算法的应用,解决了研究区已知矿点与非矿点数据不平衡问题。对多源地质数据的集成和综合生成的预测结果与研究区铅锌矿床的空间分布和实际的地质构造情况相对一致,表明上述算法能够有效识别找矿潜力区并寻找潜在的矿床。与传统的勘查地球化学方法相比,本研究中的分析方法能够处理和集成多源的地质致矿信息,可应用于尚未发现矿床的找矿潜力区,提高发现矿床的可能性和找矿工作的效率。

The application of big data mining and machine learning algorithms in mineralization prediction has become an important research trend, but unstructured geological data cannot be directly mined—first they need to be converted to machine-readable expressions. In this study of the Pangxidong ore district in western Guangdong Province, the unstructured geological information such as stratigraphy, lithology, faults are converted into machine-readable format, and two machine learning algorithms, namely, One-Class Support Vector Machine and Auto-Encoder Network, are applied to mine the geochemical test data of stream sediments as well as the comprehensive geological information on faults, stratigraphy, etc. to extract the features of mineralization anomalies and ultimately achieve intelligent delineation of the anomaly areas. Through combined application of One-Hot Encoder and the weighted variable method for spatially weighted principal component analysis, the structural transformation of the unstructured geological information is realized, and geological information is maximally preserved for data mining. It is demonstrated that the application of One-Class Support Vector Machine and Auto-Encoder Network can effectively solve the problem of data imbalance, as the numbers of ore and non-ore spots in the study area are seriously unbalanced. The prediction results generated using the integrated, synthesized multi-source geological data are relatively consistent with the observed spatial distribution of Pb-Zn deposits and the actual geological structure in the study area, indicating the two algorithms can effectively identify potential prospecting targets and ore deposits. Compared with traditional geochemical prospecting methods, the intelligent prediction method can process and integrate multi-source geological information about the ore-forming processes and identify mineralization anomaly areas. This method is applicable in prospecting areas without prior ore discovery, thereby improving the efficiency of ore prospecting and increasing the possibility of finding ore deposits.