Physics-Informed Neural Network for damage localization using Lamb waves - Thesis for Master’s Degree in Aeronautical Engineering

Aug 4, 2026

Upload by Eden

Lamb waves enable structural damage assessment through their sensitivity to defects, but conventional pipelines often require extensive signal processing to extract single-valued damage indices from the measurements. Traditional probabilistic tomographic approaches can yield damage probability maps while suffering from artifacts, subjective parameter choices, and limited quantification capability. Machine-learning methods have improved guided-wave diagnostics, yet many still rely on damage-index extraction, risking information loss. This work proposes a physics-informed neural network framework for damage localization with Lamb waves, avoiding damage-indices extraction and validating performance through multiple case studies.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages30

Reading Time1 h 16 min

AccessFree to download

Introduction

Preventive vs condition-based maintenance and SHM

Lamb wave-based damage diagnosis workflow

Limits of tomographic algorithms and damage-index extraction

Machine learning approaches for guided-wave damage diagnosis

Motivation for physics-informed neural networks

Physics-Informed Neural Network for damage localization using Lamb waves Tesi di Laurea Magistrale in Aeronautical Engineering-Ingegneria Aeronautica Jacopo Ghellero, 996039 Advisor: Prof. Francesco Cadini Co-advisors: Dr. Luca Lomazzi Eng. Lucio Pinello Academic year: 2022-2023 Abstract: Lamb waves have been widely utilized for assessing structural damage due to their sensitivity to defects. Despite their ease of excitation and acquisition, significant processing is often necessary to derive single-valued indicators, known as damage indices, from the acquired signals. Traditionally, damage indices have been developed using tomographic algorithms to create damage probability maps, though this approach is subject to limitations. Recently, machine learning has been employed to enhance the accuracy of guided wave frameworks for damage diagnosis. However, many existing methods still involve extracting damage indices from the acquired signals, potentially leading to the loss of diagnostic information and decreased accuracy. Recently, a new approach within the machine learning field has risen in popularity for its flexibility and explainability: physics-informed neural networks. These networks allow embedding some known physical laws in the algorithm to make sure the predictions adhere to the physics of the problem. However, little to no applications can be found in the field of damage diagnosis. In this context, the present work aims to present a physics-informed framework to perform damage diagnosis using Lamb waves avoiding damage indices extraction. Various case studies were considered for evaluating the performance of the proposed framework. Key-words: PINN, Machine Learning, Neural Network, SHM 1. Introduction Every component during its lifespan is subjected to different types of loads that inevitably lead to damagesand safety concerns. In order to operate engineering structures in safe conditions, different strategies have been developed and tested during years of technological advancement. The first one employed is preventive maintenance which is a time-based approach where equipment or systems are regularly serviced or replaced regardless of their current condition. The primary goal is to prevent breakdowns and extend the lifespan of assets. This strategy involves scheduled inspections, routine replacements of components, and general maintenance activities. Following this approach, unnecessary shutdown and replacement are often performed in order to avoid risks that are not quantifiable due to the limited information on component status available. In recent years thanks to technological development a new approach has raised in popularity that is condition-based maintenance, particularly in the context of Structural Health Monitoring (SHM), which relies on real-time data and sensor information to assess the current health of equipment or structures. Instead of performing maintenance at predetermined intervals, actions are taken based on the actual condition of the asset. This new way of approaching maintenance allows both cost reduction and increased safety. SHM is becoming very common in safe critical applications, e.g., for dealing with oil and gas pipelines, wind turbines, aeroplanes and many more, where the correct operation of a component is crucial to avoid accidents. Also, SHM has been adopted for cost-critical applications where the cost related to replacing or repairing the component is large enough that minimizing downtimes is crucial to save money. In this context, among the many SHM methods proposed in the literature, Lamb wave-based algorithms for damage diagnosis have shown satisfactory performance when thin-walled structures are considered [1–6] . Typically, these waves are generated and detected by placing a network of piezoelectric (PZT) devices on the structure [7–13] . The process of diagnosing damage generally involves two stages. Initially, signals are recorded when the stru

FAQs

Why are Lamb waves useful for damage localization?

Lamb waves are sensitive to structural defects, making them suitable for diagnosing damage using guided-wave sensing and acquired signals.

What limitation exists in many traditional and ML-based damage diagnosis methods?

Many methods require extracting damage indices from the measured signals, which can lose diagnostic information and reduce accuracy.

What is the key goal of the proposed framework?

The framework aims to perform damage diagnosis using Lamb waves while avoiding damage-index extraction by embedding physical laws into the learning process via physics-informed neural networks.

Aug 4, 2026

Upload by Eden