The quality of infrastructure elements such as bridges and viaducts is crucial to ensure public safety and stable economic development. Most infrastructure structures were built long time ago and now operate under conditions different from those expected during their design phase. Visual and partial inspection procedures can be costly, invasive, and ineffective when it comes to detecting early signs of failures. Early identification of problems in infrastructural elements is still an unresolved issue, especially in the context of noisy environment and irregular inspections.
The current paper introduces the development and evaluation of an AIoT-based solution for identifying infrasonic vibration patterns in order to detect potential problems with the structure in the initial stages. The presented model involves using ML and DL solutions in a single pipeline that includes the entire process of collecting and processing data until making decisions.
The two main approaches used in building the models were tested. The first one uses the combination of an LSTM-based autoencoder and a random forest classifier for anomaly and category prediction. The second approach utilizes the fully connected neural network with dual output for classification and risk prediction tasks. Results show that both solutions have high recall (>90%) with relatively low levels of false positives (~0.5%).
The research shows that using ML and DL models for detecting the initial signs of problems with infrastructure structures based on infrasonic vibrations is possible and efficient. Moreover, the use of severity policies allows translating outputs of ML models into practical decisions.