Three-Stream Heterogeneous Fusion with Graph Representation for Acoustic Recognition of Concrete Components in Complex Working Conditions

Civil engineering is essential to the transportation industry, and the maintenance of concrete components is a critical aspect. Recently, the IE (Impact-Echo) method has become a significant technique for void detection. However, conventional IE approaches have limitations and are susceptible to subjective factors. Consequently, AI (artificial intelligence) methods have attracted increasing attention. Although progress has been made, many issues persist in real-world situations. To address these shortcomings, such as weak transient characteristic comprehension, insufficient analysis of periodic features, and inadequate handling of impact-induced resonance and boundary reflection, we propose a novel three-stream input heterogeneous fusion framework. Our contributions are as follows: (1) GAF (Gramian Angular Field) representations that refine the analysis of energy distribution and asymmetric acoustic fingerprint variation rate; (2) MTF (Markov Transition Field) representations that enhance analysis of state transitions; (3) a BiLSTM (Bidirectional Long Short-Term Memory) module with dynamic constraints and the Flood Algorithm to improve time-frequency domain information comprehension. Experimental results demonstrate an average recognition accuracy of up to 96% across four tasks in complex working conditions, approaching competitive performance and indicating practical application value.