Abstract: The pipeline materials of thermal power plants are prone to nonlinear behaviors such as creep and phase transformation in high-temperature environments, leading to degradation of their mechanical properties over time. There is a logical gap between the continuous state of crack propagation and discrete decision-making (i.e., a fault layer between continuous state estimation and discrete decision-making), resulting in high rates of micro crack missed detection and high risks of misjudgment. Therefore, a method for detecting microcracks in high-temperature pipeline welding of thermal power plants using multimodal sensing fusion technology is proposed. Pipeline welding data is collected through the collaboration of ultrasound, infrared, and machine vision sensors. After A/D conversion to digital signals, the position state, propagation velocity state equation, and observation model of microcracks are constructed using Kalman filtering. Variance calculation and linear weighted average are combined to fuse multimodal information, and the fusion result is optimized through error covariance matrix. Design a collaborative framework of Kalman filtering and CART classification regression tree to fill the logical chain from continuous state to engineering decision-making. Use classification regression tree to construct a micro crack category classifier, and input the fused state estimation values and multi-source features into the classifier to achieve accurate detection of micro crack location and type. The experimental results show that multimodal sensing fusion technology can accurately detect the location of welding microcracks in high-temperature pipelines, and the multi class logarithmic loss value is small.