PSI - Issue 44

Sergio Ruggieri et al. / Procedia Structural Integrity 44 (2023) 1964–1971 Sergio Ruggieri et al./ Structural Integrity Procedia 00 (2022) 000–000

1966

3

et al., 2021b), empirical methods (Casolo et al., 2000, Casolo et al., 2019, Del Gaudio et al., 2020, Rosti et al., 2020), not forgetting hybrid and rapid visual screening methods (e.g., Ruggieri et al., 2020). 3. Transfer learning techniques for structural engineering problems The use of Machine Learning (ML) has recently taken hold in the field of civil and structural engineering, and is supporting several open issues, as reported in a recent state of the art proposed by Xie et al. (2020), such as earthquake engineering, structural health monitoring (Ierimonti et al., 2022) or structural damage detection (Gao and Mosalam, 2018; Cha et al., 2017, Zhu et al., 2020). With specific reference to the use of images for the detection of structural features and damages, ML methods are used according to the principles of deep learning, which is a class of ML algorithms that aims to extract information at a high level of abstraction from raw input using neural network architectures with multiple layers. Image processing is usually performed using multiple (hidden) layers between input and output, using convolutional neural networks, CNNs (Krizhevsky et al., 2012). A typical CNN layer consists of convolution filters, a nonlinear activation function, and a pooling step. The convolutions are the frame on which CNNs are built. Considering a bi-dimensional signal (e.g., an image), convolutions are expressed as a function of the number of picture elements, the width, the height, and the kernel square matrix. The nonlinear activation function (e.g., rectified linear units, or ReLU) is necessary to implement nonlinearities in CNNs, hence accounting for real-world effects. Finally, the pooling is a new layer added after the convolutional layer, which improves the output of the layer when the input is inaccurate through a sort of down sampling. Such networks can often achieve generalization on image classification. Therefore, their knowledge can be transferred to the problem under investigation via the transfer learning technique. Transfer learning exploits the internal structure of a pre-trained deep CNN since lower layers in a CNN extract generic features, such as edges and shapes, while higher layers in the network usually provide domain specific features. As a consequence, transfer learning aims to train only high-level layers specific to the problem under investigation: this may improve results even in the absence of a significant quantity of available data. Transfer learning is usually followed by a round of fine-tuning, where the whole network is retrained using a low learning rate. Following these concepts, Ruggieri et al. (2021) proposed VULMA , in which the module Bi VULMA allows the extraction of structural features from a building for which a photo is available. In particular, six existing CNNs models have been employed in the tool and trained in the dataset described in Cardellicchio et al. (2022), defined by the tool View VULMA . Thanks to the tool, it is possible to define several structural features, which will be defined in Section 4 and used in combination with the mechanical model generation. 4. Proposed methodology The methodology proposed in this paper aims to indicate seismic fragility for existing buildings for which a photo is available. The general framework is reported in Figure 1, where all steps are graphically defined. The first step consists in the application of Bi VULMA to the selected photo. The tool can extract some features with a specific assignment value, which can be used for the next steps. In particular, structural features for which the tool is trained are defined according to a type of assignment, such as an option in a list (identified through “List”), a specific number (identified through “Number”), or a Boolean value (identified through “Yes-No”), as specified in Ruggieri et al. (2021). The values provided by the above-mentioned tool are used as input data.The second step consists in the definition of the mechanical models that can represent the structural system of the pictured building with regard to the seismic behavior. It is worth noting that in this phase many geometrical and mechanical features are actually not available, as well as other features such as the year of construction, the area, the localization of the building. This information can be completed through the implementation of additional data sources, such as Census and Regional Technical maps. In the general application of the procedure, unknown parameters can be assumed with specific criteria and combined for a design simulation procedure. To define which unknown parameters should be combined, the possible values for the missing ones are fixed in predefined ranges and with a specific discretization. With reference to reinforced concrete (RC) structures, the unknown mechanical parameters are concrete compressive strength, steel yielding strength, and masonry infill parameters. Regarding geometrical parameters, the base area of the building and the aspect ratio are unknown and should be defined. The year of construction can also vary, and according to this parameter a different simulated design procedure can be followed. To perform the seismic analysis of simulated