Simultaneous localization and mapping (SLAM) is one of the key technologies for coal mine underground operation vehicles to build complex environment maps and positioning and to realize unmanned and autonomous operation. Many domestic and foreign scholars have studied many SLAM algorithms, but the mapping accuracy and real-time performance still need to be further improved. This paper presents a SLAM algorithm integrating scan context and Light weight and Ground-Optimized LiDAR Odometry and Mapping (LeGO-LOAM), LeGO-LOAM-SC. The algorithm uses the global descriptor extracted by scan context for loop detection, adds pose constraints to Georgia Tech Smoothing and Mapping (GTSAM) by Iterative Closest Points (ICP) for graph optimization, and constructs point cloud map and an output estimated pose of the mobile vehicle. The test with KITTI dataset 00 sequence data and the actual test in 2-storey underground parking lots are carried out. The results show that the proposed improved algorithm makes up for the drift of the point cloud map, has a higher mapping accuracy, a better real-time performance, a lower resource occupancy, a higher coincidence between trajectory estimation and real trajectory, smoother loop, and 6% reduction in CPU occupancy, the mean square errors of absolute trajectory error (ATE) and relative pose error (RPE) are reduced by 55.7% and 50.3% respectively; the translation and rotation accuracy are improved by about 5%, and the time consumption is reduced by 2~4%. Accurate map construction and low drift pose estimation can be performed.

Human activity recognition is an imbalance classification problem in essence since various human actions may occur at different frequencies. Traditional ensemble class imbalance learning methods integrate resampling technique with multi-classifier models, which obtain better generalization than single ones. However, the number of base classifiers is often determined in advance, resulting in the redundant structure of a multi-classifier and larger computational cost. Moreover, various combinations of base classifiers may have similar recognition accuracy, forming a multimodel optimization problem. To address the issue, a dual-ensemble class imbalance learning method is presented, in which two ensemble learning models are nested each other. For each sub-dataset, three heterogeneous sub-classifiers compose of internal ensemble learning model, and the one with the highest recognition accuracy is selected as base classifier. In external ensemble learning model, an optimal combination that contains the smallest number of base classifiers and identifies human actions most exactly is found by multimodal evolutionary algorithms. Based on seven imbalanced activity datasets derived from localization data for person activity, statistical experimental results show that the proposed evolutionary dual-ensemble class imbalance learning method provides the simplest ensemble structure with the best accuracy for imbalanced data that outperforms other five widely-used ensemble classification methods on all metrics.

Medical knowledge graph (KG) is sparse KG that contains insufficient information and missing paths. Multi-hop reasoning is an effective approach of medical KG completion, since it offers logical insights of the underlying KG and shows more direct interpretability. However, existing methods based on reinforcement learning focus on the use of historical and current state information but ignore the importance of evaluating the quality of candidate nodes in sparse KGs. Especially, it is difficult for the agent to select the correct search actions in sparse KGs. Occasionally, the agent will be at a dilemma state (i.e., state trap), where few actions can be selected. To address the above issue, we propose an effective relation-based node quality evaluation (RNQE) model for multi-hop reasoning. This model has two merits: (1) it reduces the impact of insufficient information in sparse KGs by synthesizing the reasoning quality information (i.e., the potential reasoning contribution) of candidate nodes; (2) it avoids the state trap by encouraging the agents to explore the path along a set of nodes with more relations. Experiments on both benchmark and real-world medical knowledge graphs demonstrate the promising ability of our proposed method to improve the reasoning performance for KGs.

In knowledge graph construction, a challenging issue is how to extract complex (e.g., overlapping) entities and relationships from a small amount of unstructured historical data. The traditional pipeline methods are to divide the extraction into two separate subtasks, which misses the potential interactio between the two subtasks and may lead to error propagation. In this work, we propose an effective cascade dual-decoder method to extract overlapping relational triples, which includes a text-specific relation decoder and a relation-corresponded entity decoder. Our approach is straightforward and it includes a text-specific relation decoder and a relation-corresponded entity decoder. The text-specific relation decoder detects relations from a sentence at the text level. That is, it does this according to the semantic information of the whole sentence. For each extracted relation, which is with trainable embedding, the relation-corresponded entity decoder detects the corresponding head and tail entities using a span-based tagging scheme. In this way, the overlapping triple problem can be tackled naturally. We conducted experiments on a real-world open-pit mine dataset and two public datasets to verify the method's generalizability. The experimental results demonstrate the effectiveness and competitiveness of our proposed method and achieve better F1 scores under strict evaluation metrics.

The mine landslide risk prediction is a fundamental task for the safety management of the digital mining system, which is dependent on the analysis of the open pit mine exploitation slope stability. Such stability analysis involves a lot of local natural and human factors. Training a learning model using these factors, with the aim of discovering their relationship with the slope stability, is an intuitive way for the mine landslide risk prediction. The key issue is how to explore the complex nonlinear relations among these factors by using a small amount of high-dimensional historical slope data. Traditional factor-led methods for the issue only focus on the impacts of landslide factors, but ignore the correlations between historical slope data, which in fact contain more useful information. In this article, we propose a new mine landslide risk prediction model using knowledge graph. The gradient boosting decision tree is applied to further exploit the crossed features within the historical data, and then a landslide semantic network is constructed using knowledge graph with the consideration of correlation information between historical slope data. In this way, the model is suitable to deal with the small set of high-dimensional data, and it makes a joint use of the features within the landslide factor values and correlations between historical slope data. We conduct a set of experiments on real historical slope data from the real-world open pit mining scene. The experiment results validate the effectiveness and efficiency of the proposed model in predicting landslide risks with a small set of high-dimensional historical data.