PSI - Issue 64

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

496

5

2.3. Model construction and data analysis procedure The data set for the training of the NN contain the data that they should later be able to process independently, as well as the output or solution that is expected for the respective data. The weightings of the neurons or hidden units (HU) of the NN are initially initiated randomly. By familiarizing the NN with the training data in so-called training epochs, these weightings are adjusted with training, so that the NN will be able to predict data outside the trainings data set correctly after training. Based on the vehicle classes presented, the BiLSTM was trained with measurement data that corresponds to the quantity equals to the measurement over a period of 5 hours. A further increase in the training data did not lead to any recognizable improvement in accuracy. The selection of the training data was random. The classes in the training data were labelled manually beforehand in order to specify the corresponding class and to adapt the network structure using appropriate training algorithms. Areas that were not labelled were assigned to a ninth class, the "n/a" class. This class comprises the measurement data areas in which none of the eight vehicle classes occur. Time series have the disadvantage of requiring a large amount of training data. Therefore, data reproduction and augmentation of the training data was performed. For this, the original training data set is artificially multiplied by modifications so that more training data is available for the NN. The SpecAugment method according to Park et al. (2019) was used for the data augmentation. In this method, the augmentation does not take place on the raw data set, but on the associated spectrogram. The spectrogram was created using the Fourier-based synchrosqueezing transformation according to Oberlin et al. (2014). This was followed by frequency masking using the SpecAugment method. The used network architecture was first determined in a parameter study in which a suitable number of hidden layers (HL), HU per HL and the drop-out were reviewed. The parameter study was carried out sequentially by implementing the next step based on the previously determined parameters. The individual iteration steps for the final definition of the network architecture are summarized in Table 1. As evaluation criterion, the accuracies achieved by the BiLSTM with the respective parameter set were compared with each other. After completing the training process, accuracy was assessed for each parameter combination using an additional validation dataset distinct from the training data. A one-hour period from the measurement data was considered as the validation data set, which, similar to the training data, had been manually labeled beforehand. Accuracy was evaluated by comparing the manually labelled sections with the classification outcomes generated by the BiLSTM. Starting with a drop-out rate of 5 %, iterations were conducted to determine the optimal number of HL and HU per HL for achieving the highest classification accuracy. Subsequently, the optimal drop-out rate was determined using a fixed selection of HL and HU. Finally, a network architecture comprising two HL, each containing 50 HU and a drop out rate of 20 % was established.

Table 1. Iteration steps for the final network architecture. Iteration

Hidden Layer

Hidden Units

Drop-Out [%]

1

1, 2, 3

5

5, 10, 20, 30, 40, 50, 75, 100

2

2 2

50 50

5, 10:10:90

Final

20

Figure 5. depicts a schematic representation of the network architecture and the described classification process for the measurement data. The raw data undergoes processing through the Fourier-based synchronsqueezing transformation before being classified by the finalized BiLSTM. In the following, the fully trained BiLSTM utilized for vehicle identification is referred to as TU-BS-1 BiLSTM.