Characterization of Mechanical Discontinuities Using Machine Learning and Knowledge-Driven Causal Model

Abstract

Mechanical wave transmission through a material interacts with the propagation of mechanical discontinuity in the material. Machine learning can be used to monitor the propagation of embedded discontinuities by analyzing the resultant waveforms recorded by a multipoint sensor system placed on the surface of the material. Our study accomplishes a first-of-its-kind monitoring of mechanical discontinuity propagation by using data-driven model to process the multipoint waveform measurements resulting from a single impulse source. While conventional data-driven methods, especially supervised learning, rely primarily on statistical correlation/association and lack domain knowledge and causality. The primary objective of this work is to discover new geophysical causal signatures relevant to the multipoint waveform measures caused by mechanical discontinuity propagation inside a solid material. The use of causal signatures also led to the development of a novel knowledge-driven model that creates a versatile and resilient data system for both linear and physical crack propagation samples. The newly discovered causal signatures confirm that the statistical correlations/associations and conventional feature rankings are not statistically significant indicators of causality. The new developments presented in this work, especially the causal-based knowledge-driven model, have both theoretical and practical implications that can improve fracture monitoring, prediction, and early warning. In near future, similar knowledge-driven approaches will gain traction in mainstream applications of data analytics to overcome the drawbacks of current machine learning approaches such as lack of generalizability and explainability.