Page 86 - Read Online

Page 392 Wang et al. Intell Robot 2022;2(4):391-406 https://dx.doi.org/10.20517/ir.2022.25

are affected by overload, temperature change, reinforcement corrosion, construction defects and other
[1]
factors . If cracks cannot be detected and maintained in a timely manner, traffic safety will gradually be
affected, with the potential for bridge collapse and related accidents. Once bridge collapse occurs, it will
reduce its structural bearing capacity, which will affect the reliability and safety of the structure, causing
immense social repercussions .
[2-4]
In engineering practice, the health monitoring of bridge cracks has attracted the attention of relevant
national departments and research institutions , corresponding norms and standards have been issued,
[5,6]
and the monitoring methods and monitoring parameters for bridge deformation and cracks have been
clearly specified.

In recent years, there have been many significant developments in research on the detection of concrete
bridge cracks [7-15] . It has been shown that the applications of machine learning in bridge cracks have
achieved good results [16-18] . Some studies utilize convolutional neural networks (CNNs) to detect and
segment cracks in civil infrastructure with multiple objects [19-23] . It is significant to carry out data-driven
research on feature extraction, data fusion and intelligent detection of concrete bridge cracks, which could
provide not only a scientific basis for intelligent maintenance of bridges but also data support and a
theoretical basis for bridge defect detection, which could have an important role in improving social and
economic benefits.

After more than 20 years of scientific research and practice, health monitoring systems have been installed
on at least 300 bridges in China. The health monitoring system of a long-span bridge is composed of at least
dozens or even hundreds of sensor measuring points. Therefore, numerous monitoring data have been
accumulated.

As crack evolution is a gradual and multiscale dynamic process, research is challenging to some extent.
There are some common technical difficulties, which will be described from the following three levels:

(1) Feature extraction of multimodal parameters
The characterization information related to bridge cracks includes bridge crack shape information,
structural mechanics index information, and crack environment information. Bridge crack shape
information comprises the multimodal parameters of bridge cracks, such as the length, width and depth of
cracks. Structural mechanics index information consists of the dynamic and static elastic modulus,
compressive strength, and stress distribution. Crack environment information includes the load,
temperature, humidity, and foundation settlement.

From the level of feature extraction for multimodal parameters of bridge cracks, the characterization
information of bridge cracks has multiscale and diversity, which makes it difficult to extract multimodal
parameter characteristics. The bridge service environment is complex and changeable, and the crack
formation mechanism of concrete bridges varies, which generates multiscale and diverse characterization
information. The large amount of information hinders feature extraction.

Due to the influence of background interference information, equipment accuracy, data acquisition mode,
signal propagation path and propagation medium, there is considerable redundancy in the data, resulting in
a weak effective signal in the monitored multimodal data and difficulty in extracting its sensitive features.