PSI - Issue 44

Valentina Giglioni et al. / Procedia Structural Integrity 44 (2023) 1948–1955 Valentina Giglioni et al. / Structural Integrity Procedia 00 (2022) 000–000

1953

6

at the top of one pier, respectively, S3-L is aimed at collecting data in the middle of the main span along the longitudinal direction, while S4-L acquires the longitudinal accelerations at the top of the second pier. It should be mentioned that more than 50 time-histories analysis are carried out to provide pseudo-accelerations time-histories, which are standardized, rearranged into short sequences and then normalized between -1 and 1. The whole dataset, organized in 6240 sequences, is split into data for training (64%), validation (7%), testing-healthy (12%) and testing-damage (17%). 4.4. Application of the autoencoder-based technique The MLP autoencoder, composed of five hidden layers, is trained for the -th sensor to reconstruct 5-seconds-long acceleration sequences. The hyperbolic tangent function (tanh) is used as activation function and the Adam optimizer is employed in the updating process to minimize the mean squared error, selected as loss function. The number of training epochs is set to 15 and the batch size equal to 128. Each sequence’s reconstruction error is evaluated by means of the ORSR index, whose distribution in the training period allows to define the threshold # ( ) corresponding to the 95-th percentile.

Fig. 2. General view of the Z24 FEM model (a), description of sensors position (b) and damaged elements with a reduced elastic modulus (c). Then, according to the proposed damage detection approach, 24 sequences are grouped into a unique macro sequence of 2 minutes length and a second threshold $ ( ) is adopted to quantify the percentage of the inner damaged short sequences needed to classify the macro-sequence as damaged. Fig. 3 illustrates the four control charts of the previously selected sensors, by setting the classification threshold $ ( ) to the 95-th percentile. Results highlight the effectiveness of the proposed methodology in clearly identifying the occurrence of seismic events-induced damages by notably avoiding the presence of false negatives errors, which may be extremely dangerous in civil engineering applications. It is also interesting to point out that this approach is quite promising for damage localization, by deducing a near correlation between the single sensor and the type of damage. The earthquake induced scenario simulated on the piers is expected to modify the transversal and longitudinal dynamic response. Indeed, in this application, sensors located in both these directions are deemed effective to recognize the investigated damage. The procedure is also quite fast, taking less than one minute to train and test the ML algorithm with the described dataset. The goal of future works would be to apply the developed technique with a larger dataset, different simulated damage scenarios and different sensors location in order to build a robust damage detection technique, traying also to further corrupt the generated signals according to the observed environmental and operational effects. 5. Conclusions This paper illustrates an unsupervised ML-based approach for the post-earthquake diagnosis of roadway bridges. The proposed procedure, focused on a Multi-Layer Perceptron autoencoder as feature extractor and damage detector,