Autonomous structural health monitoring (SHM) platforms for bridge networks produce continuous streams of damage indicators but lack principled procedures for setting the statistical thresholds that govern when an alert is raised. This dissertation addresses three interconnected gaps in the state of the art, using the P3P platform deployed by ANAS across the Italian highway network as the operational context. The first gap is the absence of a generalizable upper control limit (UCL) procedure. P3P sets the UCL of its Hotelling T ² control chart at the empirical 90 th percentile of the training distribution, without theoretical justification. This dissertation characterizes the T ² distribution across 106 monitoring segments and evaluates four UCL methods: the P3P empirical percentile, the exact Beta distribution formula of Shabbak and Midi [1], the Generalized Extreme Value (GEV) Block Maxima method, and the Generalized Pareto Distribution (GPD) Peaks Over Threshold method. The Shapiro- Wilk test rejects normality in all 106 segments, and the Gamma distribution provides the best fit in 34% of cases under the Akaike Information Criterion, with Lognormal (27%), GEV (24%), and Weibull (14%) accounting for most of the remainder, and Beta adequate in only 3%; UCL methods assuming Gaussian residuals are therefore fundamentally misspecified for this context. Two segments with insufficient block maxima for stable GEV fitting are excluded from the method comparison, which is conducted across the remaining 104 segments. At a nominal false alarm rate of 10%, GEV Block Maxima is the only method satisfying the bound, with a mean monitoring false alarm rate of 5.2%, against 18.7%, 20.0%, and 34.9% for the GPD, P3P empirical, and Beta methods, respectively. A percentile sweep from the 90 th to the 99 th confirms that GEV meets the bound at every calibration percentile, supporting retention of the platform' s standard 90 th percentile. The advantage of GEV is operational rather than raw sensitivity: at a common false alarm rate, the viable methods detect the same minimum damage, but only GEV delivers the target rate at the standard percentile from a principled tail fit, whereas the empirical and GPD methods reach it only when recalibrated to a percentile knowable in hindsight, and the Beta method never reaches it. The second gap is the lack of a threshold transfer protocol for uninstrumented bridges. Starting from the Normal Condition Alignment framework of Giglioni et al. [2], the GEV UCL derived from the 18 P3P segments monitoring six frequencies is transferred to the Ponte di Pizzighettone, a five-span concrete girder bridge on the SP CR EX SS 234 in Lombardia. Because the Hotelling T² statistic is scale-invariant through its covariance matrix, the alignment reduces to an identity for equal-dimensional segments, and the transferred UCL of 30.26 is simply the network mean, requiring no assumed target training size; establishing this scale- invariance is itself the transfer result. The target bridge is represented by a three-dimensional grillage finite element model in OpenSeesPy, calibrated to the first experimental frequency of 3.919 Hz, with the effective elastic modulus of the composite section as the single calibration parameter.The third point is that there is no connection between statistical thresholds and damage which has a physical meaning. A parametric study involving 144 scenarios on the Pizzighettone model, with the severity and spatial extent of the local stiffness reduction varied, determines the smallest amount of damage that can be reliably detected under each UCL method.
Le piattaforme autonome di monitoraggio della salute strutturale (SHM) per reti di ponti producono flussi continui di indicatori di danno, ma mancano di procedure consolidate per definire le soglie statistiche che determinano quando generare un allarme. Questa tesi affronta tre lacune interconnesse nello stato dell’arte, utilizzando come contesto operativo la piattaforma P3P impiegata da ANAS sulla rete autostradale italiana. La prima lacuna riguarda l’assenza di una procedura generalizzabile per la definizione del limite di controllo superiore (UCL). P3P imposta l’UCL della propria carta di controllo T² di Hotelling al 90° percentile empirico della distribuzione di addestramento, senza una giustificazione teorica. Questa tesi caratterizza la distribuzione del T² su 106 segmenti di monitoraggio e valuta quattro metodi: il percentile empirico di P3P, la formula esatta della distribuzione Beta di Shabbak e Midi [1], il metodo Block Maxima della distribuzione Generalizzata dei Valori Estremi (GEV) e il metodo Peaks Over Threshold della distribuzione Generalizzata di Pareto (GPD). Il test di Shapiro-Wilk rifiuta la normalità in tutti i 106 segmenti e la distribuzione Gamma fornisce il miglior adattamento nel 34% dei casi secondo il criterio di Akaike, con Lognormale (27%), GEV (24%) e Weibull (14%) a coprire gran parte del resto e Beta adeguata solo nel 3%; i metodi UCL che assumono residui gaussiani sono quindi fondamentalmente mal specificati per questo contesto. Due segmenti con un numero insufficiente di massimi di blocco per un adattamento GEV stabile sono esclusi dal confronto tra metodi, condotto sui restanti 104 segmenti. A un tasso di falso allarme nominale del 10%, il metodo GEV Block Maxima è l’unico a rispettare il limite, con un tasso medio di falso allarme del 5,2%, contro 18,7%, 20,0% e 34,9% rispettivamente per GPD, percentile empirico P3P e formula Beta. Un’analisi di sensibilità dal 90° al 99° percentile conferma che GEV rispetta il limite a ogni percentile di calibrazione, il che giustifica il mantenimento del 90° percentile standard della piattaforma. Il vantaggio di GEV è operativo più che di pura sensibilità: a parità di tasso di falso allarme i metodi validi rilevano lo stesso danno minimo, ma solo GEV fornisce il tasso obiettivo al percentile standard a partire da un adattamento della coda teoricamente fondato, mentre i metodi empirico e GPD lo raggiungono solo se ricalibrati a un percentile noto solo a posteriori, e il metodo Beta non lo raggiunge mai. La seconda lacuna riguarda l’assenza di un protocollo di trasferimento della soglia verso ponti non strumentati. A partire dal framework di Normal Condition Alignment di Giglioni et al. [2], il limite GEV derivato dai 18 segmenti P3P che monitorano sei frequenze viene trasferito al Ponte di Pizzighettone, un ponte a cinque campate in cemento armato sulla SP CR EX SS 234 in Lombardia. Poiché la statistica T² di Hotelling è invariante di scala attraverso la propria matrice di covarianza, l’allineamento si riduce a un’identità per segmenti di uguale dimensione e l’UCL trasferito di 30,26 coincide semplicemente con la media di rete, senza richiedere l’assunzione di una dimensione del campione di addestramento del ponte bersaglio; stabilire questa invarianza di scala è esso stesso il risultato del trasferimento. Il ponte è rappresentato da un modello agli elementi finiti a graticcio tridimensionale in OpenSeesPy, calibrato sulla prima frequenza sperimentale di 3,919 Hz con il modulo elastico efficace della sezione composta come unico parametro di calibrazione.La terza lacuna riguarda l’assenza di un collegamento tra le soglie statistiche e un danno fisicamente significativo. Uno studio parametrico di 144 scenari sul modello di Pizzighettone , variando la severità e l’estensione spaziale della riduzione locale di rigidezza, individua il danno minimo affidabilmente rilevabile per ciascun metodo UCL.
Advanced Monitoring and Assessment of Existing Bridges: Statistical threshold generalisation, UCL transfer, and damage detection in an autonomous SHM platform for the Italian highway bridge network
JURADO DAVILA, DAMARIS EMILIA
2025/2026
Abstract
Autonomous structural health monitoring (SHM) platforms for bridge networks produce continuous streams of damage indicators but lack principled procedures for setting the statistical thresholds that govern when an alert is raised. This dissertation addresses three interconnected gaps in the state of the art, using the P3P platform deployed by ANAS across the Italian highway network as the operational context. The first gap is the absence of a generalizable upper control limit (UCL) procedure. P3P sets the UCL of its Hotelling T ² control chart at the empirical 90 th percentile of the training distribution, without theoretical justification. This dissertation characterizes the T ² distribution across 106 monitoring segments and evaluates four UCL methods: the P3P empirical percentile, the exact Beta distribution formula of Shabbak and Midi [1], the Generalized Extreme Value (GEV) Block Maxima method, and the Generalized Pareto Distribution (GPD) Peaks Over Threshold method. The Shapiro- Wilk test rejects normality in all 106 segments, and the Gamma distribution provides the best fit in 34% of cases under the Akaike Information Criterion, with Lognormal (27%), GEV (24%), and Weibull (14%) accounting for most of the remainder, and Beta adequate in only 3%; UCL methods assuming Gaussian residuals are therefore fundamentally misspecified for this context. Two segments with insufficient block maxima for stable GEV fitting are excluded from the method comparison, which is conducted across the remaining 104 segments. At a nominal false alarm rate of 10%, GEV Block Maxima is the only method satisfying the bound, with a mean monitoring false alarm rate of 5.2%, against 18.7%, 20.0%, and 34.9% for the GPD, P3P empirical, and Beta methods, respectively. A percentile sweep from the 90 th to the 99 th confirms that GEV meets the bound at every calibration percentile, supporting retention of the platform' s standard 90 th percentile. The advantage of GEV is operational rather than raw sensitivity: at a common false alarm rate, the viable methods detect the same minimum damage, but only GEV delivers the target rate at the standard percentile from a principled tail fit, whereas the empirical and GPD methods reach it only when recalibrated to a percentile knowable in hindsight, and the Beta method never reaches it. The second gap is the lack of a threshold transfer protocol for uninstrumented bridges. Starting from the Normal Condition Alignment framework of Giglioni et al. [2], the GEV UCL derived from the 18 P3P segments monitoring six frequencies is transferred to the Ponte di Pizzighettone, a five-span concrete girder bridge on the SP CR EX SS 234 in Lombardia. Because the Hotelling T² statistic is scale-invariant through its covariance matrix, the alignment reduces to an identity for equal-dimensional segments, and the transferred UCL of 30.26 is simply the network mean, requiring no assumed target training size; establishing this scale- invariance is itself the transfer result. The target bridge is represented by a three-dimensional grillage finite element model in OpenSeesPy, calibrated to the first experimental frequency of 3.919 Hz, with the effective elastic modulus of the composite section as the single calibration parameter.The third point is that there is no connection between statistical thresholds and damage which has a physical meaning. A parametric study involving 144 scenarios on the Pizzighettone model, with the severity and spatial extent of the local stiffness reduction varied, determines the smallest amount of damage that can be reliably detected under each UCL method.| File | Dimensione | Formato | |
|---|---|---|---|
|
Jurado_Damaris.pdf
accesso aperto
Dimensione
3.06 MB
Formato
Adobe PDF
|
3.06 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/114729