PSI - Issue 64

Philipp Kähler et al. / Procedia Structural Integrity 64 (2024) 1248–1255 Kähler / Petryna / Structural Integrity Procedia 00 (2019) 000 – 000

1250

3

suitable for many areas of SHM. They can be applied from simple classification tasks to complex deep learning damage detection and localization schemes, see Avci et al. (2021). Here, the goal of the present CNN is not just the classification of the load into predefined load groups, but the extraction of precise load characteristics. Therefore, the CNN must solve a regression task, which is trained through supervised learning. In order to process multiple time series from different sensors simultaneously, every time series for each sensor is reshaped into a two-dimensional matrix, which is further extended into the third dimension by the sensor number. Thus, the input is a ( × × ) matrix with and describing the window size of the reshaped sensor data and being the number of sensors. This data is then processed by multiple two-dimensional convolutional layers with different filter sizes to extract the desired features. An overview of the CNN’s structure is provided in Fig . 1.

filter size x

filter size x

filter size x

filter size x

filter size x

Convolutional a er atch Normalization a er

filters with different sizes

input size x x

x

x

Full Connected a er

egression a er

x

x

x

x

x

x

x

x

ea e a er ax ooling a er

ea e a er

Fig. 1. Overview of CNN’s structure for feature extraction of the load from acceleration sensor data .

As described previously, different types of information need to be extracted from the characteristics of the signals. Although it is possible to determine all desired properties using a single CNN, accuracy generally increases if each output of interest is processed by a separate CNN. Therefore, several CNNs are arranged in a cluster structure, where each CNN performs a different task. The cluster of the load identification process is shown in Fig. 2.

num erof vehicles

owman vehicles are on the ridge

input data

first second first second

hich traffic lane is loaded

CNNs

load magnitudes load velocities time dela order of vehicles

feature extraction

CNNs

Fig. 2. Cluster structure of the load identification process.

After establishing the architecture of a CNN, the dimension of its output vector remains fixed. Nevertheless, this dimensionality is directly linked to the number of vehicles present on the bridge at the same time. Consequently, various CNNs with distinct output vector sizes must be trained to accommodate diverse scenarios. To address this, a two-step approach was developed: initially, to classify the acceleration data within the decision-making networks, and subsequently, to extract the load properties utilizing the suitable feature-extracting networks. For each CNN, an individual hyperparameter optimization was conducted regarding the influence of the reshaping of the input size, filter size, number of filters, number of layers, batch size, activation functions and epochs to ensure the best possible convergence.