A team of researchers from Beihang University, Jiangsu University of Technology and the University of Alberta has cast new light on the integration of large language models (LLMs) with reliability systems engineering (RSE), revealing both the opportunities and challenges it presents.
Reliability systems engineering is a sophisticated discipline that focuses on the entire product life cycle, aiming to reduce failure rates, enhance system reliability and extend product lifespans. The introduction of LLMs into this field has the potential to significantly boost industrial production efficiency and flexibility. The new study provides a comprehensive overview of the current development of LLMs in RSE, analysing key application scenarios and technologies, and identifying existing challenges.
Advertisement
The research team systematically collected literature from the Web of Science database, using keywords related to LLMs and RSE. Their analysis revealed that LLMs can be applied across various dimensions of the RSE V-model, including requirements, design, manufacturing, verification and maintenance. These models are capable of supporting traditional tasks such as design modelling and requirements analysis, as well as more advanced applications such as perceptual enhancement and intelligent manufacturing. For instance, fine-tuning LLMs with domain-specific data has shown high accuracy in detecting vulnerabilities in Ethereum smart contracts and improving fault diagnosis accuracy in bridge cranes.
However, the integration of LLMs into RSE isn’t without challenges. One significant issue is the systemic deficiencies in domain data.
Advertisement
Complex engineering tasks often involve imbalanced and incomplete data, particularly regarding failures, which can lead to biases in failure prediction. This not only hampers the summarisation of design experiences but also limits the training effectiveness and inference capabilities of LLMs. Additionally, the complexity of RSE analysis, characterised by the phased and diverse structure of the V-model, requires LLMs to integrate physical laws, statistical patterns and domain-specific knowledge.
The ‘black-box’ nature of LLMs further complicates matters, as their probabilistic inference outputs often lack transparency and explainability, making it difficult to meet the high standards of complex engineering tasks.
Advertisement
Looking ahead, the researchers suggest several future development directions for LLMs in RSE. One key area is the exploration of pre-trained industrial LLMs capable of addressing complex engineering issues. This involves optimising knowledge extraction and retrieval methods, integrating expert knowledge and generating high-quality datasets to enhance training effectiveness.
Another direction is improving the adaptability of LLMs for solving complex analysis processes by integrating them with probabilistic models, Bayesian inference and reinforcement learning. This would enable LLMs to better manage uncertainty and make decisions in real-world scenarios. Finally, the organic integration of human, machine and intelligence is seen as crucial for coordinated development, especially in safety-critical domains such as aerospace and healthcare.
The integration of LLMs into reliability systems engineering holds significant promise, yet several critical areas require further development. Despite the unique advantages that LLMs offer in knowledge extraction and application within RSE, they still face limitations in data resources, performance and explainability. Future efforts should focus on exploring more practical engineering application scenarios, promoting collaborative interaction with other intelligent models and developing systematic engineering frameworks. By doing so, LLMs can enhance the intelligence of design and production processes, ultimately empowering RSE with greater efficiency and reliability.
Advertisement
The research has been published in Engineering.