PSI - Issue 64

Robby Weiser et al. / Procedia Structural Integrity 64 (2024) 492–499 Robby Weiser / Structural Integrity Procedia 00 (2019) 000 – 000

493

2

1. Introduction The bridge infrastructure is increasingly in need of retrofitting and repair due to the occurrence of significant damage during its service life. Especially bridges from the 1960s, were dimensioned for far lower traffic loads than they have to cope with today. The main reason is the increase in goods and freight traffic. On steel bridges heavy goods traffic with its high axle loads lead to fatigue cracks in welded details. Other causes include the use of low grade steel with limited weldability and insufficient execution quality of welds. This quality issues results from the large manufacturing tolerances in sheet metal cutting at the time as described by Marzahn and Hamme (2014). In order to assess and repair fatigue cracks, it is necessary to know the stress situation on the component or structure. It depends mainly on the composition of the traffic, which leads to movements at the location of the crack. Structural Health Monitoring (SHM) is used to identify the gap opening of the monitored crack. In addition to the SHM installation at the bridge, a suitable approach for data analysis of the traffic compositions is required. There are various approaches to identify traffic loads and its patterns. For example, the approach of Steffens and Geißler (2021) is based on the filtering of individual vehicle signals from representative gap opening data. Earlier approaches were based on the establishment of feature vectors of the desired events. Subsequently, the characteristics of the respective features were determined and a realization-value was calculated by combining the different characteristics, cf. Mehdianpour (2002). New developments in in the field of data analysis, show the potential of Neural Networks (NN). These NNs have been used recently in the area of speech recognition. Audio signals and the used gap opening data are comparable due to its signal structure. In speech recognition, Graves et al. (2013) showed that bidirectional Long Short-Term Memory Networks (BiLSTM) were able to provide good results to identify specific features and are a more performant alternative to combined approaches of Recurrent Neural Networks (RNN) with Hidden Markov Models. The aim of this paper is to identify specific traffic compositions using a BiLSTM from representative measurement data of a crack. The results of these investigations provide a valuable input for the maintenance and repair of bridges. Currently, a bridge must be closed for traffic for the repair of fatigue cracks. Repairs under traffic loads would be advantageous and are sometimes even unavoidable, due to mobility restrictions and economic damage. In addition, investigations by Begemann et al. (2024) have shown, that welding under traffic loads is possible with certain restrictions of gap opening. The identification and reproduction of realistic traffic load conditions is necessary in order to determine the influences on the welding seam under cyclic traffic loads and to find representative time periods for crack repair. 2. Method and Material 2.1. Method A BiLSTM is used to classify the measurement data in order to recognize and classify different vehicle types in the measured dataset. This Deep Learning algorithm was introduced by Graves et al. (2005). It is based on the further development of RNNs into bidirectional RNNs based on Schuster and Paliwal (1997). For a meaningful classification, suitable vehicle classes must be defined first. For this purpose the Technische Lieferbedingungen für Streckenstationen (technical delivery conditions for route stations, TLS) by the Bundesministerium für Verkehr, Bau und Stadtentwicklung (2012) were used as reference. The TLS provides an overview of the various vehicle types and axle load combinations in road transport and currently describes around 75 different vehicle types. However, it is not practical to assign each vehicle type to its own class, as there are overlaps in the different axle configurations. To train a NN to classify all individual vehicle types would result in a significantly larger volume of training data. Therefore, similar vehicle types are grouped into a common class. Based on the German Nachrechnungsrichtlinie für Brücken im Bestand (recalculation guideline for existing structures) of the Bundesministerium für Verkehr, Bau und Stadtentwicklung (2011), a categorization was made into eight different vehicle classes. The recalculation guideline divides damage-relevant traffic into two groups based on the number of vehicle axles and the total weight. Group 1 includes vehicles with a high proportion of local traffic, such as buses or trucks with up to three axles. Group 2 includes vehicles with three axles or more that have a high proportion of long-