Fast and intelligent detection of concrete cracks based on sound signals and convolutional neural network

Introduction The traditional crack detection method usually requires a tedious process of sensor installation and removal, which seriously affects the efficiency of concrete structure management and maintenance. Methods For this reason, this paper develops a fast concrete crack detection method based on percussion with an improved convolutional neural network (CNN). By utilizing the percussion method, the sensors do not need to be coupled and installed on the concrete structure, which saves a lot of processes. The sound signals generated by percussion are collected by acoustic pressure sensors, while multiple data enhancement techniques are applied to enrich the data volume and diversity of the collected signals. The Mel-frequency cepstral coefficient (MFCC) of the sound signals are then extracted as inputs to the improved CNN model. The CNN used is mainly applied to initialize the weights by applying the transfer learning technique, and the Squeeze-and-Excitation Networks (SENet) attention mechanism is embedded to improve the model’s focus on important features. Finally, comparative experiments with different frame lengths, different models and different signal-to-noise ratios (SNR) are conducted using the improved CNN. Results The results show that the model validation process has the least loss and highest accuracy when the input frame length is 1024. The improved CNN has good feature learning ability for MFCC of percussion sound signals for effective recognition of concrete cracks. Compared with Resnet18, random forest and long short-term memory networks, the improved CNN has superior recognition accuracy and stability, and shows better noise robustness in high signal-to-noise ratio (SNR: −6 db∼6 db) environments. Discussion Therefore, the proposed method has a high potential for future crack detection in concrete structures.