Rock-mechanics parameters such as Young’s modulus and Poisson’s ratio are critical to geomechanical analysis and resource exploration. Because these parameters come from laboratory measurement, they present some characteristics such as insufficient samples and contamination of outliers. In this paper, a novel semi-supervised support vector machine soft sensor is devised considering the characteristics of the parameters. First, it takes into account data similarity and selects labeled data set that are most similar to the continuous unlabeled data set at each iteration to improve estimation performance. Meanwhile, an outlier deletion algorithm is developed for a better similarity comparison. After that, a semi-supervised approach is presented for the estimation of rock-mechanics parameters, it can leverage continuous unlabeled data to train the model dynamically. Finally, the verification of our method is carried out on data set from UCI (University of California, Irvine) and several drilling sites. The results demonstrate that our method outperforms eight well-known methods in estimation accuracy.

杨氏模量和泊松比等岩石力学参数对地质力学分析和资源勘探至关重要。由于这些参数来自实验室测量，因此它们呈现出一些特征，例如样本不足和异常值的污染。在本文中，考虑到参数的特性，设计了一种新颖的半监督支持向量机软传感器。首先，它考虑了数据的相似性，并在每次迭代时选择与连续未标记数据集最相似的标记数据集，以提高估计性能。同时，为了更好的相似性比较，开发了异常值删除算法。之后，提出了一种估计岩石力学参数的半监督方法，它可以利用连续的未标记数据来动态训练模型。最后，我们的方法的验证是在来自 UCI（加州大学欧文分校）和几个钻井地点的数据集上进行的。结果表明，我们的方法在估计精度方面优于八种众所周知的方法。