深部岩体损伤演化与灾变预警中的机理约束数据驱动方法OA

As deep resource exploitation continues to extend into high-stress,high-temperature,and strongly disturbed environments,damage evolution in rock masses is becoming increasingly severe,and disasters such as rockburst,spalling,and surrounding-rock instability are occurring more frequently.In view of the difficulty of traditional empirical criteria and single-parameter threshold methods in meeting the demand for real-time warning under complex geological conditions,recent progress in mechanism-constrained data-driven methods for deep rock-mass damage evolution and disaster early warning was analyzed.The multi-source data basis,task paradigms,and methodological framework were systematically summarized,with emphasis on the applications and limitations of machine learning,deep learning,interpretable learning,uncertainty representation,and physics-informed neural networks in rockburst classification,precursor identification,strength prediction,parameter inversion,and digital-twin warning.Key bottlenecks,including data quality,label reliability,cross-scale representation,mechanism embedding and closed-loop updating,were also discussed.Future work should strengthen unified dataset construction,physical-prior embedding and online updating mechanisms.Taking the"four layer embedding path of mechanism constraints"(feature layer embedding objective function layer embedding structural layer embedding system layer embedding)as the main line,the embedding form of rock mechanics mechanism knowledge in data-driven models,the special requirements and representative methods of deep environments were explained layer by layer.In typical tasks such as rock burst classification,damage stage identification,strength prediction and parameter inversion,drilling perception,and digital twin closed-loop,the form and link of mechanism constraints are clarified,and specific problems that are difficult to solve with pure data-driven methods were solved.