Neural Network Technology for Monitoring the Damaged State of an Extended Structure Based on Analysis of Non-stationary Waves

Abstract

Introduction .  Destruction of the surface layer of concrete structures caused by an aggressive environment is one of the most common types of damage that require remote and automatic monitoring. Acoustic non-destructive testing methods can address this problem. However, the existing approaches, including the contour line intersection method we previously developed, demonstrate instability in the low-porosity region (less than 15%) because of the weak dependence of signal amplitude on time. So, we identified a problem that existing methods cannot reliably identify damage parameters across the whole range of values. The objective of this work is to develop a neural network technology for the robust identification of the depth and porosity of the damaged layer of an extended structure based on the analysis of non-stationary waves, which determines the uniqueness and scientific novelty of the study. Materials and Methods .  The study is based on numerical simulation. The mechanical model is a two-layer elastic anisotropic strip of thickness H = 1 cm with an upper damaged layer modeled as a porous composite whose effective properties calculated in the ACELAN-COMPOS package. The damage parameters are layer thickness h (0.05–0.3 cm) and porosity p (10–80%). A piezoelectric actuator and sensor made of PZT-4 ceramics with vertical polarization are placed on the surface of the strip. Non-stationary waves are excited by a potential difference on the actuator and recorded by the sensor. To obtain training data, a series of direct problems for 48 parameter combinations were solved using the finite element method in the ACELAN package with quadratic triangular elements (mesh: 5,616 elements; 12,155 nodes). Based on cubic interpolation, training (10,000 points) and test (40,000 points) datasets were generated. A feedforward neural network, a multilayer perceptron with four hidden layers (128-64-32-16 neurons), ReLU activation function, dropout layers (20%), and Adam optimizer was constructed and a weighted loss function was used to reduce positive asymmetry in predictions. Results .  On the independent test set, the trained network achieves a mean absolute error of 2.433% for porosity and 0.121 mm for depth ( R 2 = 0.977 and 0.910, respectively). In the low-porosity region ( p < 15%), the error is 0.405% for porosity and 0.173 mm for depth with a bias of only +0.188%, confirming the reliability of the approach in the region. All 40,000 test points yield unique two-dimensional signatures, empirically confirming the uniqueness of the inverse problem solution. Discussion.  The obtained data confirm that the proposed neural network technology overcomes the fundamental limitation of the contour line intersection method, providing stable predictions across the entire porosity range. The use of a weighted loss function effectively reduces systematic overestimation of porosity predictions. A limitation of the approach is the assumption of constant porosity along the depth and length of the layer. Accounting for inhomogeneities would require solving an inverse problem of significantly higher complexity. In future work, we plan to move beyond just two extracted features and instead use the full time-domain signal, possibly with convolutional architectures. Conclusion .  The neural network technology developed in this research provides an effective method of identification of damage parameters across the entire range of values, overcoming the limitations of existing methods. The obtained results confirm the applicability of the proposed approach for acoustic monitoring systems of extended concrete structures.