The mining industry faces challenges such as hazardous working conditions, labor shortages, and accidents. Autonomous mining, using AI and digitalization techniques, offers a solution to improve productivity, safety, and efficiency. While partial automation exists, fully autonomous mining remains challenging due to the complex and scalable nature of field environments. This research addresses the limitations of existing approaches by proposing a novel framework based on parallel intelligence, a methodology that leverages the power of multiple computational approaches to create a more robust and adaptable system. This framework is designed to overcome the inherent complexities of autonomous mining in real-world settings, thereby achieving a significant advancement in the field. The increasing global demand for energy necessitates enhanced mineral resource productivity, making the development and implementation of autonomous mining systems a critical endeavor.

Previous research has focused on partially automating specific mining processes, such as autonomous truck development, remotely controlled excavation, automated drilling and blasting, and mining rescue robots. However, these efforts have primarily been confined to simplified and idealized scenarios, lacking the ability to scale up to large-scale, heterogeneous, and complex real-world mining operations where collaborative efforts among multiple machines are essential. The existing challenges include generalization issues in field environments (uneven terrain, poor visibility), complex collaboration among diverse mining devices, and the need for cost-effective and comprehensive pre-validation of operations. Current simulation approaches struggle to bridge the sim-to-real gap, especially for large-scale integrated systems.

The proposed autonomous mining framework, "Parallel Mining," utilizes parallel intelligence and consists of three main components: 1) Model-driven virtual mine simulation: This involves creating a detailed virtual model of the mine environment, including sensors and machinery dynamics. 2) Data-driven digital twins: This component collects real-world data from on-site equipment (robots and sensors) to replicate real-world scenarios in the virtual mine, enhancing the simulation's accuracy. 3) Parallel computing experiments: This uses a scenario library (real-world, hybrid, and generated virtual scenarios) to evaluate and optimize mining operations in the virtual mine, generating prescriptive intelligence that is applied to real-world operations. The system's design prioritizes a hierarchical structure to enhance stability. This includes an autonomous driving module, a scheduling and cooperative operations module, and a parallel management module. The parallel management module is critical, acting as the central hub for data collection, processing, analysis, testing, and optimization. It also monitors the hardware's operational status and provides early warnings. This system architecture ensures that critical tasks can continue even if one module fails. The virtual simulation engine is crucial for testing and validation, providing a safe and cost-effective environment for algorithm development and refinement. The engine is capable of highly realistic simulations that mirror real-world conditions and allows for the creation and testing of a wide variety of scenarios, including those that are rare or dangerous in the real world. The generation of scenarios is achieved through a combination of model-driven and data-driven approaches, leading to a versatile and reliable system. Specifically, the descriptive learning component uses real-world data to supervise and refine the virtual sensor models. Predictive learning involves constructing the mapping relationship between mining operations and entities' state transformation through deep neural networks. Prescriptive learning uses the simulation system to generate massive amounts of data, ultimately generating the algorithms that direct specific mining operations. The process involves task decomposition, initialization algorithms, iterative system operation using data-driven and reinforcement learning, and curriculum learning-based model optimization for scheduling.

The "YuGong" autonomous mining system, developed based on the Parallel Mining framework, was deployed and extensively tested at the Yimin Open-pit Mine in Inner Mongolia, China. The system demonstrated stable and efficient autonomous operations over an extended period. Testing included scenarios focusing on autonomous driving (fundamental self-driving, obstacle avoidance, rough and winding roads), cooperative operations (inter-truck cooperation, cooperative loading, cooperative unloading), and adverse operational conditions (snow, high dust, low light). YuGong exhibited superior performance to human operators in several key areas. In autonomous driving, the system achieved accurate perception (Table 1), effectively avoided obstacles, and navigated challenging terrain. The cooperative operations demonstrated efficient coordination among multiple vehicles, improving loading and unloading times (Table 2). The system successfully operated under adverse weather conditions, although with a reduction in speed for safety. Long-term testing revealed a significant efficiency improvement compared to human-operated systems. YuGong achieved a 10% increase in loading and unloading efficiency, 33% faster speeds for empty trucks, and a 35% increase in daily transportation shifts. Furthermore, it resulted in 12% fuel savings per ton of minerals produced. Beyond Yimin, YuGong has been successfully deployed in more than 30 mines, covering over 4.1 million kilometers of operation without major accidents. The high-definition map created by the system was also highlighted, showing its ability to capture the detailed features of the mining environment. This virtual map allows the system to plan efficient routes and ensure safe operations. The ability to accurately perceive the environment and plan accordingly is critical to the success of the system and is showcased in the detailed figures showing perception results and maps.

The results demonstrate the feasibility and effectiveness of the Parallel Mining framework for achieving fully autonomous mining operations. The YuGong system successfully addresses the challenges of generalization, complex collaboration, and adequate pre-validation. The system's success in various scenarios, including those with adverse weather conditions, highlights the robustness of the parallel intelligence approach. While the current implementation focuses on open-pit mining, a relatively simpler scenario, the framework's modular design and adaptability suggest potential for extension to more complex underground mining operations. The system has been successfully deployed in multiple mines, suggesting scalability and generalizability. The long-term testing and fuel savings data further validate the system's economic viability and environmental benefits. However, there remains a need for further research in the development of a plug-and-play system that can be easily adapted to new mining sites, possibly using large foundation models tailored for mining. The overall efficiency improvements seen in the system suggest that parallel intelligence is a viable approach for building robust and efficient AI systems for complex tasks in real-world environments.

This research presents a significant advancement in autonomous mining by introducing a novel framework based on parallel intelligence. The YuGong system, implemented based on this framework, demonstrates high efficiency, robustness, and safety in real-world open-pit mining scenarios. The system's success underscores the potential of parallel intelligence for addressing complex challenges in various industrial settings. Future research should focus on extending the framework to more complex mining environments (underground), developing a universal, plug-and-play autonomous mining intelligence, and integrating societal information into the framework for macro-scale resource allocation and global sustainability.

The current study primarily focuses on open-pit mining, which may not fully represent the complexities of underground mining operations. The system's adaptability to different mining sites requires further improvement, currently involving a manual fine-tuning process with regional data. The reliance on specific sub-module algorithms, while currently effective, may limit the system's flexibility. Furthermore, while the data-driven approach enhances the accuracy of the virtual mine model, there is still a potential gap between the simulated and real-world environments, which could affect the performance of the system.