The industrial large models, which may seem to possess limitless capabilities, actually harbor certain challenges and unseen obstacles. On one hand, issues related to "hallucinations" present threats to the credibility, accuracy, and real-time applicability of these models in industrial scenarios. On the other hand, there are numerous common problems related to data sources, foundational large models, and task adaptations that need to be addressed.

As the integration of these large models with the industrial internet progresses, one must question whether these hidden difficulties are being magnified or effectively resolved. How can the industrial internet provide a nurturing environment for the development of industrial large models, enabling them to adapt more effectively to the complex needs of industrial scenarios?

On November 7, during the 42nd session of the 502 online seminar, representatives from various enterprises gathered to discuss these pressing issues. They shared practical experiences and case studies, delving into how the organic combination of industrial large models and industrial internet can yield greater efficiencies and find new breakthroughs for smart industrial applications.

Current Status and Challenges

Industrial large models are catalysts for enhancing industrial efficiency. The trend toward increasing automation in manufacturing processes is evident, marking an inevitable shift toward Industry 4.0 and beyond into Industry 5.0. Dr. Liu Jing, a researcher at Hebei University of Technology and director of the Hebei Provincial Data-Driven Industrial Intelligence Engineering Research Center, explained that technological advancements are primary drivers of industrial development, guided by the principles of innovation profit as posited by Joseph Schumpeter. While prior phases of industrialization focused on smaller models that produced certain automation results, they also revealed several issues, such as fragmentation of systems, severe information silos, challenges in dealing with unstructured environments, and low efficiency in human-machine interaction. The rise of Industry 5.0 emphasizes human-machine collaboration, underscoring the importance of industrial large models.

Dr. Li Sen, a senior expert in industrial interconnection solutions, noted that large models serve as critical tools for improving efficiency and capabilities in the industrial sector, acting as catalysts in modern industrialization. Historically, the development of Chinese industry has been influenced by Western technologies, leading to impressive advances in equipment innovation and information integration but lacking in fundamental transformations. The integration of large models with industrial internet presents new opportunities, especially given current policy drivers for updating industrial software. Traditional industrial internet systems encounter challenges like inter-company collaboration difficulties, information asymmetry within industrial chains, increasingly heterogeneous production management needs, and poor collaboration among smaller models. The implementation of industrial large models is pivotal in resolving these issues.

Challenges of the "Hallucination" Phenomenon and Common Issues

Guan Lingqian, a product manager at the R&D department of Inspur Cloud's industrial internet Cloud Platform, stated that there are significant differences between small and large models. Small models depend on internal enterprise data for development, training, and application in specific scenarios, continuously optimized through operational data feedback. In contrast, large models rely on vast amounts of internet data and require fine-tuning to adapt to specific industries and scenarios. The inherent probabilistic nature of large model outputs introduces the "hallucination" problem, which poses challenges to credibility, accuracy, real-time performance, and safety in industrial applications.

Dr. Liu also cited common issues faced by industrial large models across various dimensions, such as data sources, foundational models, task adaptation, and industrial applications. For instance, data from different companies often lacks uniformity in format and standards, complicating processing efforts. At the foundational model level, challenges include the presence of heterogeneous multimodal raw data, difficulties in knowledge construction, and low retrieval accuracy, potentially resulting in inaccurate outputs.

Application Practices in Specific Scenarios

The exploration of industrial large model applications encompasses several critical directions. Dr. Liu shared three primary pathways for applying industrial large models.

The first involves the construction of knowledge systems. For example, a training large model tailored for central enterprises' overseas employees can generate personalized materials and courses based on individual circumstances. Other applications include translation models that accommodate multiple complex factors and large models used in EHS (Environment, Health, and Safety) systems to construct safety frameworks.

The second pathway highlights the synergy between small and large models. A case study regarding product weight over-limit issues illustrated how large models can uniquely link to small models to carry out operations such as queries and adjustments.

The third pathway addresses the generation of industrial content. Emphasis is placed on leveraging industry knowledge and data for enhanced training, thus enabling models to better adapt to industrial and product applications. Additionally, while large models can resolve data generation and working condition simulation challenges in the design of process parameters for titanium alloys, issues related to data alignment in multimodal data processing persist.

Knowledge Consultation and Process Optimization in the Cable Industry

Guan Lingqian described the extensive applications of the Inspur Cloud Knowledge Large Model across various industries. In the cable industry, this model serves as a foundational layer, integrating knowledge databases, algorithm models, and information systems to build applications for knowledge consultation, intelligent equipment maintenance, and production process optimization. This approach effectively addresses customer pain points.

In response to challenges such as data scattering, difficulties in production planning and scheduling, and high equipment maintenance costs, an industry knowledge database is developed to consolidate expertise and facilitate knowledge sharing among industry partners, subsequently training and optimizing the Knowledge Large Model specifically for the cable industry.

Specialized algorithm models are developed to complement large models in dealing with specific issues related to process optimization and equipment maintenance. Here, the large model offers capabilities for understanding demand and planning tasks, while the mechanistic model provides accurate simulations and explanations grounded in physical principles.

Guan highlighted that the application scenarios of Inspur Cloud's industrial large model are realized through the technological capabilities of the Knowledge Service Platform. Firstly, fine-tuning the large model through training enables it to acquire fundamental industry knowledge. Secondly, knowledge graph integration keeps the dynamic knowledge of enterprises connected and updated. Thirdly, orchestration of intelligent agents manages a variety of systems and specialized models for the rapid construction of large model applications, featuring characteristics like multi-source multi-modal capabilities, cloud-edge collaboration, data security, and low-cost deployment.

Boosting Productivity in the Electric Equipment Manufacturing Industry

Dr. Li Sen introduced the launch of the ZhiGong Industrial Large Model on June 4, 2023, marking it as the first large model specifically for the industrial sector in China. Its core capabilities revolve around smart production monitoring and an industrial knowledge engine.

Within the application landscape of the ZhiGong Industrial Large Model, it addresses efficiency and cost issues across the entire value chain for electric equipment manufacturing companies. These capabilities include minimizing material waste during the customization of transformers and enhancing both production operation efficiency and personnel collaboration effectiveness.

Predicting Equipment Failures and Optimization

In the Precision Engineering project, optimizations were made concerning inventory operations and personnel scheduling while predicting equipment failures. In collaboration with Ingersoll Rand, the ZhiGong Industrial Large Model was integrated with edge computing devices, enabling small and medium-sized enterprises to achieve energy optimization and operational efficiency in air compressors.

In the context of China Railway Corporation and aerospace technology sectors, the ZhiGong Industrial Large Model facilitated equipment fault diagnosis and strategy suggestions, allowing for monitoring and early warning of dynamic environment equipment operations.

The ZhiGong Industrial Large Model also excelled in intelligent questioning, invocation, analysis, and optimization. The knowledge engine demonstrates outstanding capabilities in visual interpretation and multi-format support, capable of creating embodied intelligent support across various scenarios.

The knowledge question-and-answer system "ZhiGong · ZhiYu" boasts advantages in products, technologies, and deployment, offering multiple question-and-answer modes, advanced PDF parsing technology, and features for private deployment.

Conclusion

The developments from knowledge system construction to the synergy between small and large models, culminating in the generation of industrial content, illuminate diverse innovative pathways for industrial advancement.

Whether it's the performance of the Inspur Cloud Knowledge Large Model in the cable industry or the successful applications of ZhiGong Industrial Large Model across electric equipment, precision engineering, air compressors, and aerospace sectors, all testify to the formidable potential released following the integration of industrial large models with the industrial internet.

In the future, as the pursuit of industrial intelligence and efficiency continues, ongoing investments in technological improvements for models will be essential to enhance their performance and stability. Furthermore, the expansion of applications must align more closely with the distinct characteristics and needs of diverse industries, ensuring that the advantages of industrial large models extend to a wider array of fields. Comprehensive industrial planning must coordinate resources at a macro level to create a more refined ecological system.

Throughout the event, online participants engaged actively in discussions, posing pertinent questions that were addressed by the experts. The session welcomed valued attendees from companies such as TCL, NIO, COMAC, Zhongtong Service Construction, Zoomlion, State Grid, China Telecom, Baidu, and Huawei. The event concluded amidst a rich exchange of practical experiences and discussions on potential business collaborations.