Incorporating A-priori Knowledge into Convolutional Neural Networks for Impact Echo Frequency Estimation

Abstract Manual evaluation and interpretation of Impact echo (IE) data is often labor-intensive and time-consuming, motivating the growing interest in applying machine learning (ML) techniques to this non-destructive testing (NDT) method. However, the scarcity of labeled datasets limits the generalizability of ML models to new, unseen data. This study investigates strategies to integrate a-priori knowledge into Convolutional Neural Networks (CNNs) for improved prediction of the S1 Lamb wave frequency from IE signals. To this end, time–frequency representations of IE signals, derived using the Short-Time Fourier Transform (STFT), are used as model inputs. A-priori knowledge is introduced in the form of initial frequency estimates obtained through manual evaluation. Additionally, transfer learning is employed to enrich the limited measurement dataset with data from 2D numerical simulations. The results demonstrate that, although training loss curves remain similar across models, incorporating additional information significantly enhances performance on unseen datasets. Furthermore, pre-training with simulation data accelerates convergence during early fine-tuning stages. The highest predictive accuracy was achieved when the initial guess was directly embedded into the loss function.