This study investigates novel Physics-Informed Neural Network (PINN) approaches for solving the Near-field Acoustic Holography (NAH) problem. Serving as a near-field sound reconstruction task, NAH is a non-contact method used to analyze sound sources by utilizing sound pressure measurements in the near field. This approach avoids potential damage caused by accelerometers and prevents the added mass from influencing the vibrating surface. This technique enables the identification of vibrating regions, providing detailed insights into the behavior of the sound sources without physical interference. Three methodologies upon on the same framework are proposed in this study. First, Complex-Valued Neural Networks (CVNNs) are employed to the previous model, referred to as the Complex-Valued Kirchhoff–Helmholtz-based Convolutional Neural Network (CV-KHCNN). The result validates the effectiveness of CVNNs in solving the NAH problem, while also highlighting the suitability of Cardioid as the activation function for CVNNs. Furthermore, t-SNE analysis is conducted to visualize the features of the embedding layer. The results suggest that CV-KHCNN possesses the ability to differentiate between BCs and mode shapes, even without prior knowledge. Then we revisited the conventional Compressive-ESM (C-ESM) for NAH and proposed PINN-CESM and Virtual Planes aided C-ESM (PINN-VP-CESM), which employs the KH integral as the prior knowledge to solve NAH with C-ESM. Indeed, unlike CV-KHCNN, where NNs are used mostly as learning agents, this approach leverages NNs as physics-informed agents, integrating physical laws directly into the learning process. This self-supervised method is a one-shot procedure, eliminating the need for a large training dataset. The results show that adding virtual planes can improve the reconstruction of fine details especially in highly complex vibrational patterns. Moreover, PINN offer more consistent performance across different scenarios, making it a reliable choice for source field reconstruction tasks. Particularly, unlike the conventional C-ESM, PINN-VP-CESM is less sensitive with regularization parameter. Inspired by this one-shot routine, we propose a Transfer Learning enhanced KHCNN (TL-KHCNN) framework, which includes two stages: pre-training CV-CNN and Physics-Informed one-shot fine-tuning. This combination can generalize the model that trained on a limit dataset for another task. We pre-train the network on the rectangular plate dataset and then fine-tune it on the violin top plate dataset. The results further indicates that extending the model’s application to more complex practical scenarios is very promising.
Questo studio indaga nuovi approcci con Reti Neurali Informate dalla Fisica (PINN) per risolvere il problema dell'Olografia Acustica in Campo Vicino (NAH). Servendo come un compito di ricostruzione sonora in campo vicino, la NAH è un metodo senza contatto utilizzato per analizzare le sorgenti sonore mediante misurazioni della pressione sonora nel campo vicino. Questo approccio evita i potenziali danni causati dagli accelerometri e impedisce che la massa aggiunta influenzi la superficie vibrante. Questa tecnica consente di identificare le regioni vibranti, fornendo dettagliate informazioni sul comportamento delle sorgenti sonore senza interferenze fisiche. In questo studio vengono proposte tre metodologie basate sullo stesso framework. In primo luogo, vengono impiegate Reti Neurali a Valori Complessi (CVNN) per il modello precedente, denominato Rete Neurale Convoluzionale basata sul Kirchhoff-Helmholtz a Valori Complessi (CV-KHCNN). I risultati confermano l'efficacia delle CVNN nella risoluzione del problema NAH, evidenziando anche l'idoneità della funzione di attivazione Cardioid per le CVNN. Inoltre, viene condotta un'analisi t-SNE per visualizzare le caratteristiche del livello di embedding. I risultati suggeriscono che CV-KHCNN possiede la capacità di differenziare tra BC e forme modali, anche senza conoscenze preliminari. Successivamente, abbiamo rivisitato il convenzionale Compressive-ESM (C-ESM) per NAH e proposto PINN-CESM e C-ESM assistito da Piani Virtuali (PINN-VP-CESM), che utilizza l'integrale KH come conoscenza preliminare per risolvere il NAH con il C-ESM. Infatti, a differenza di CV-KHCNN, dove le reti neurali sono utilizzate principalmente come agenti di apprendimento, questo approccio sfrutta le reti neurali come agenti informati dalla fisica, integrando direttamente le leggi fisiche nel processo di apprendimento. Questo metodo auto-supervisionato è una procedura one-shot, eliminando la necessità di un ampio set di dati di addestramento. I risultati mostrano che l'aggiunta di piani virtuali può migliorare la ricostruzione dei dettagli fini, soprattutto nei modelli vibratori altamente complessi. Inoltre, PINN offre prestazioni più coerenti in diversi scenari, rendendolo una scelta affidabile per i compiti di ricostruzione del campo sorgente. In particolare, a differenza del C-ESM convenzionale, PINN-VP-CESM è meno sensibile al parametro di regolarizzazione. Ispirati da questa procedura one-shot, proponiamo un framework KHCNN migliorato dal Transfer Learning (TL-KHCNN), che include due fasi: pre-addestramento CV-CNN e fine-tuning one-shot informato dalla fisica. Questa combinazione può generalizzare il modello addestrato su un set di dati limitato per un altro compito. Pre-addestriamo la rete sul set di dati della piastra rettangolare e poi la perfezioniamo sul set di dati della piastra superiore del violino. I risultati indicano ulteriormente che estendere l'applicazione del modello a scenari pratici più complessi è molto promettente.
Physics-informed neural network approach for near-field acoustic holography
Luan, Xinmeng
2023/2024
Abstract
This study investigates novel Physics-Informed Neural Network (PINN) approaches for solving the Near-field Acoustic Holography (NAH) problem. Serving as a near-field sound reconstruction task, NAH is a non-contact method used to analyze sound sources by utilizing sound pressure measurements in the near field. This approach avoids potential damage caused by accelerometers and prevents the added mass from influencing the vibrating surface. This technique enables the identification of vibrating regions, providing detailed insights into the behavior of the sound sources without physical interference. Three methodologies upon on the same framework are proposed in this study. First, Complex-Valued Neural Networks (CVNNs) are employed to the previous model, referred to as the Complex-Valued Kirchhoff–Helmholtz-based Convolutional Neural Network (CV-KHCNN). The result validates the effectiveness of CVNNs in solving the NAH problem, while also highlighting the suitability of Cardioid as the activation function for CVNNs. Furthermore, t-SNE analysis is conducted to visualize the features of the embedding layer. The results suggest that CV-KHCNN possesses the ability to differentiate between BCs and mode shapes, even without prior knowledge. Then we revisited the conventional Compressive-ESM (C-ESM) for NAH and proposed PINN-CESM and Virtual Planes aided C-ESM (PINN-VP-CESM), which employs the KH integral as the prior knowledge to solve NAH with C-ESM. Indeed, unlike CV-KHCNN, where NNs are used mostly as learning agents, this approach leverages NNs as physics-informed agents, integrating physical laws directly into the learning process. This self-supervised method is a one-shot procedure, eliminating the need for a large training dataset. The results show that adding virtual planes can improve the reconstruction of fine details especially in highly complex vibrational patterns. Moreover, PINN offer more consistent performance across different scenarios, making it a reliable choice for source field reconstruction tasks. Particularly, unlike the conventional C-ESM, PINN-VP-CESM is less sensitive with regularization parameter. Inspired by this one-shot routine, we propose a Transfer Learning enhanced KHCNN (TL-KHCNN) framework, which includes two stages: pre-training CV-CNN and Physics-Informed one-shot fine-tuning. This combination can generalize the model that trained on a limit dataset for another task. We pre-train the network on the rectangular plate dataset and then fine-tune it on the violin top plate dataset. The results further indicates that extending the model’s application to more complex practical scenarios is very promising.File | Dimensione | Formato | |
---|---|---|---|
Master_thesis_XinmengLuan.pdf
accessibile in internet per tutti a partire dal 23/06/2027
Descrizione: main thesis
Dimensione
13.91 MB
Formato
Adobe PDF
|
13.91 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary_Master_thesis_XinmengLuan.pdf
accessibile in internet per tutti a partire dal 23/06/2027
Descrizione: executive summary
Dimensione
5.51 MB
Formato
Adobe PDF
|
5.51 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/223095