Lost data of non-sparse construction-induced ground vibrations using bidirectional gated recurrent unit method

Construction-induced ground vibrations (CIGV) present serious challenges in urban development, with potential impacts ranging from structural damage to environmental disturbance. Continuous monitoring is essential for effective assessment and mitigation, and wireless sensor systems have become a common solution due to their flexibility and scalability. However, such systems are vulnerable to data loss caused by communication interruptions, hardware failures, and power issues. This study addresses the problem of recovering both random and continuous data losses in non-sparse CIGV time series using real vibration records from multiple construction projects in Toronto. Due to the non-sparse and highly nonlinear nature of CIGV signals, traditional compressive sensing (CS) methods demonstrate limited performance. Adaptive wavelet thresholding (AWT) provides improved reconstruction but fails to consistently capture peak particle velocity (PPV) values, which are critical for evaluating vibration impact. A bidirectional gated recurrent unit (BGRU) model is implemented to overcome these limitations. By learning complex temporal patterns, the BGRU approach demonstrates superior accuracy, particularly in handling extended periods of data loss—an issue frequently encountered in CIGV monitoring. The findings highlight the effectiveness of deep learning-based recovery methods for ensuring data continuity and reliability in CIGV monitoring applications.