Page 95 - Read Online

Wang et al. Intell Robot 2022;2(4):391-406 https://dx.doi.org/10.20517/ir.2022.25 Page 401

experimental results showed that the accuracy achieved 91.96% with a training time of 123 minutes with the
MS-Celeb-1 M face database and that the RMSE was reduced to 121.9 on air quality dataset prediction. The
BLS performs abstract representation and incremental learning of dynamic high-dimensional data and
achieves a high diagnostic accuracy, short time and strong real-time performance. However, the BLS is
inferior to deep networks in terms of the deep data mining performance of feature extraction and has a poor
effect on detection problems for time-dependent data.

The deep neural network model uses not only a deep neural network to represent the multilayer features of
the identified object but also the extracted high-level features to reflect the intrinsic nature of the data,
which has better robustness and diagnostic ability than the shallow network. Classic deep intelligent
diagnosis methods include CNNs, deep belief networks and recursive neural networks. Li et al. employed
the sliding window algorithm to divide not only the bridge crack image into slices with a size of 16 pixels ×
16 pixels but also the slices into the bridge crack surface element and bridge background surface element .
[69]
The authors then proposed a deep bridge crack classification model based on CNN for the identification of
bridge background surface elements and bridge crack surface elements. The proposed algorithm achieved
an average accuracy of 94.5% with 2000 bridge crack images. Islam et al. established a deep CNN using an
encoder and decoder framework for semantic segmentation to realize pixel-level automatic detection of
bridge cracks, obtaining scores of approximately 92% for both the recall and F1 value . Liang et al. utilized
[70]
a double CNN model to identify cracks in actual concrete bridges, which highly improved the reliability and
accuracy of identification with accuracies of 98.6% and 99.5% by the CNN and FCN models, respectively .
[71]
Li et al. proposed a new type of fully connected state, long-term and short-term memory neural network,
which was employed to discriminate sensor faults and structural damage without knowing the fault details,
[72]
and obtained excellent performance with an RMSE of 0.03 . The deep network model has made progress
in big data processing. Although the training takes a long time due to the numerous network layers,
complex structure, and many super parameters, it can process the data with complex internal information
characteristics and is more suitable for the engineering needs of timely warning for dangerous situations in
bridge health monitoring.

For intelligent crack detection model construction, a shallow network model can mine complex data for
bridge crack detection, but its feature extraction ability is weak, and it cannot carry out multilayer feature
representation for recognized objects. Thus the recognition rate is low and the generalization performance
is mediocre. The BLS can mine and integrate multilevel and multimodal, complex state information by
widening the network and incremental learning. The BLS achieves high diagnostic accuracy, fast speed and
strong real-time performance for the research target but has a poor effect on the time-dependent data
detection problem. The deep network model can represent the multilayer features of the recognized object
and use the extracted high-level features to reflect the intrinsic nature of the data. The model has the ability
of deep information mining and achieves better robustness, generalization and recognition performance.

The structure of the deep neural network mapping model is shown in Figure 3, and the methods utilized for
intelligent detection in the above review articles are summarized in Table 3.

3. MAIN CHALLENGES
Despite great developments in the research on concrete bridge crack identification based on data, certain
challenges exist.

As a cavity, the crack has three-dimensional information with length, width and depth. However, traditional
crack identification is usually based on the recognition with the two-dimensional surface information