Shandong Energy Group Pioneers Intelligent Mining with World's First Commercial Large AI Model
Produkty, Rozwiązania i usługi dla przedsiębiorstw
Shandong Energy Group (SEG) is a leading company in many industries, including mining, high-end chemical industry, electric power, new energy materials, high-end equipment manufacturing, and modern logistics trade. As the third-largest coal producer in China, the company leads in intelligent mining with 9 of its mines chosen as demonstrations on a national level.
The mining industry is experiencing a significant increase in AI, to the point it is involved in national policies. SEG, a world-class coal company, is dedicated to advancing intelligent mining. However, the company's original AI solution was proving to be both costly and slow. This is because one model trained for one mine production scenario had to be retrained when applying it to another one, resulting in additional expenses and delays. Not only this, but the old solution had several other issues, including:
• Low development efficiency: The majority of AI models were developed in silos. This means that for each scenario, a new model had to be created, trained, fine-tuned, and iterated separately, resulting in low efficiency and prolonged development periods.
• High entry barriers: The development of AI involves various stages such as problem definition, data access, data processing, feature extraction, model training, model assessment and deployment, and model management. This demands seasoned AI algorithm specialists to streamline the development pipeline and fine-tune top-notch models.
• Low accuracy and poor generalization: AI models should be able to adapt to the evolving industry standards. However, the original AI models fell short of meeting the actual production demands in terms of accuracy, performance, and scalability when working conditions changed. This lack of adaptability resulted in low accuracy and poor generalization, which made it difficult to adopt these models on a large scale.
SEG has since deployed an AI training and application platform based on Huawei Pangu models, which consists of four foundation models: computer vision (CV), graph neural network (GNN), multimodal, and natural language processing (NLP). By leveraging Pangu models, the company was able to revolutionize its intelligent production modes across six business domains. Details are as follows:
• High sample training efficiency: Pangu models allow for seamless data transfer between the group center cloud and the mining edge cloud through cloud-edge synergy. Compared with other models, they require less data, smaller samples, and less code for training, yet deliver equivalent or superior accuracy, which allows for automatic optimization and continuous learning.
• Massive throughput information processing: The Pangu CV model has undergone pre-training on over 1 billion images and more than 100 terabytes of video data using the unsupervised learning method. This has enabled it to extract and store vast amounts of knowledge within its extensive network, allowing it to represent intricate visual features with ease.
• Strong generalization performance: Pangu models have superior generalization performance compared to small models, because a Pangu model trained for one scenario can be applied to another scenario with a detection accuracy of over 23%. Additionally, these models can be quickly deployed in new mines, eliminating the need for repeated training.
• High data filtering efficiency: Compared with traditional small models, Pangu models are more efficient in screening defect samples in new scenarios, reducing the labor cost for data labeling by over 85%.
• High detection accuracy: The L2 layer of Pangu models is designed to pre-train production, safety observation, and decision-making models based on the principle of marking anything not normal as abnormal. These models require only a small number of samples for training and boast a 10% higher accuracy in detecting abnormalities compared to smaller models.
SEG has taken a huge leap forward by building an AI training center, which marks the first-ever commercial application of Pangu models in the mining industry. The center's primary objective is to explore AI applications for all mining scenarios, including coal mining, drivage, equipment control, transportation, ventilation, and coal preparation. By implementing AI on a large scale, the company has successfully moved most mining workers from underground to above ground, resulting in efficient and safe operations.
• The AI model developed for Xinglongzhuang Coal Mine, which serves as a benchmark for the SEG, will be applied to upwards of 70 coal mines owned by the company.
• The implementation of the Pangu mine model and Pangu CV model in coal mines has not only improved the working conditions for inspection personnel, but also reduced manual inspection frequency from once a day to once a week.
• The Run-up Program offers AI developers advanced training, helping build a skilled talent pool for SEG's digital transformation.
Pangu models will bring the following benefits:
• Improve AI return on investment (ROI): SEG has established a 'four-in-one' AI center to streamline the management of AI assets and prevent redundant investment and isolated development. The center offers various services, including public computing, scenario-based innovation and business conversion, application industrialization, and AI talent development. It has also cut down the average delivery time of scenario-based models from 18 person-days to 12, shortening the time taken to go from investment and application to production.
• Improve production quality and efficiency: In coal preparation and blending scenarios, the input of production process data is complex and cannot be determined entirely by employee experience. To address this concern, SEG has turned to Huawei Pangu models for assistance in accurately predicting and controlling related parameters based on its datasets. For example, in coal blending tasks for coking factories, specialists can use GNN technology to train a coal blending optimization model. This will reduce the time taken for blending from a couple of days to just 1 to 2 minutes. In the coal preparation optimization scenario, specific models can be built based on the Pangu GNN model to automatically predict the coal density and ash content of products. Through cyclone and full-process control parameter optimization, the models quickly adjust working parameters such as suspension density and inlet pressure based on the predicted ash ratio. This ensures stable ash content proportions and improves the yield of refined coal by 0.1%–0.2%.
• Improve innovation capabilities: SEG is dedicated to promoting the integration of AI and smart mine construction, with the goal of building more intelligent exemplary coal mines. Together with Yunding Technology and Huawei, SEG has made 30 national patents (including 20 invention patents and 10 utility model patents), and published 20 papers.
• Foster the industry ecosystem: SEG has developed an industry-specific AI platform based on Pangu models to support the maturing of scenarios, enablement & promotion, and ecosystem operation. By fostering a sci-tech mining ecosystem that encompasses academia, research institutes, and businesses, the models enable upstream and downstream enterprises to harness the potential of AI innovation and application.
• Reduce production safety risks: Pangu models are more effective than traditional small models in detecting safety risks, with 10% higher accuracy. This allows for timely alarms to be sent, reducing the frequency of accidents. For instance, in the Xinglongzhuang Coal Mine Project, the detection accuracy of Pangu models exceeded 90% in scenarios involving unauthorized entry to dangerous areas. Additionally, Pangu models help to standardize the behavior of underground personnel and enhance their safety awareness by warning them of the risks involved when they do not strictly follow protocol.
• Encourage talent: With the help of Pangu models, SEG has improved its internal operations and talent development by cutting down on knowledge acquisition costs. To further advance AI development and operations, the company plans to enhance its training and certification system and expand career development opportunities for AI professionals. This will encourage them to create more innovative solutions and achieve greater scientific research breakthroughs, ultimately contributing to their personal growth and the company's success.
