PSI - Issue 78

Akshay Rai et al. / Procedia Structural Integrity 78 (2026) 891–898

895

Fig. 1: Continuous monitoring configuration: accelerometers A1-A12; S-series linear variable transducers C1-C4; temperature sensors T1-T6 [4].

Fig. 2: Flowchart to compute Temperature-Compensated Spectral Metrics for SHM

4. Convolutional Autoencoder

The construction and tuning of the convolutional autoencoder (CAE) were conducted to establish an archi tecture and optimise training parameters suitable for processing structural health monitoring data. This CAE is designed to reconstruct dominant SV spectra and ambient temperatures from multivariate inputs, with each sam ple represented as a vector of shape (257, 3) that combines 257 dominant SV bins with internal ( T 0 ) and external ( T 1 ) temperature measurements. The architecture features an encoder–decoder configuration, as shown in Figure 3. The encoder employs four 1D convolutional (Conv1D) layers to extract hierarchical representations while compressing the input into a latent code. It uses LeakyReLU activations and 1D spatial dropout to promote stable gradients and reduce overfitting. The decoder mirrors this structure with 1D transposed convolutions (Conv1DTranspose), progressively upsampling to reconstruct the original (257, 3) dimensionality. Its output layer implements a split activation scheme using