Well logging data, as a crucial source of information in petroleum exploration, intuitively reflects the lithological characteristics of geological formations. However, complex correlations exist between different logging curves, and data from different wells with differing reservoir properties invariably exhibit inconsistencies. Consequently, it is difficult to achieve robust lithology identification by directly applying the acquired raw logging data. Traditional lithology identification methods often directly utilize raw well logging data as input features for models, neglecting the potential information inherent within the raw data. This oversight leads to underutilization of the implicit information in well logging data. The key to solving this problem is to obtain correlation construction of information between different wells and invariant feature selection between wells in different reservoirs. Therefore, this paper first introduces a graph representation method to describe the spatial (horizontal) relationships and the temporal (vertical) topological information among different well logging curves. Subsequently, it selects three classic invariant features: the structural tensor (ST), local binary pattern (LBP), and Hu moments (Hu), to robustly represent the local structural information of well logging data across different reservoirs. Finally, using classic machine learning techniques for lithology identification, the study evaluates the integration patterns of heterogeneous features to demonstrate the effectiveness of the robust features extracted in enhancing lithology identification performance. Experiments conducted with multiple actual well logging datasets from the Qijia depression area of the Daqing oilfield indicate that the features selected in this study can resolve issues related to inter-well distribution inconsistency and imbalance in well logging data, thereby improving the accuracy of lithology identification. This approach introduces a novel perspective in well logging data mining, paving the way for more accurate and efficient exploration strategies in the petroleum industry.