Traditional fault detection/tolerance approaches introduce relevant costs to achieve unconditional correctness during data processing. For example, the classical Duplication With Comparison (DWC) approach achieves fault detection by duplicating the system and discarding data as soon as any error occurs, while the Triple Modular Redundancy (TMR) approach triplicates the system and achieves fault tolerance by means of a voting scheme. However, many application environments are inherently tolerant to a certain degree of inexactness or inaccuracy. In this thesis, we focus on the practical scenario of image processing in space, a domain where faults are a threat, while such applications are not safety-/mission-critical and are inherently tolerant to a certain degree of errors. We first propose a shift from the concepts of corrupted/uncorrupted the concept of usability of the processed image to relax the traditional requirement of unconditional correctness, and to limit the computational overheads related to reliability. We then introduce our new flexible and lightweight fault management scheme designed for application environments where inexactness is tolerated up to a given limit. The new scheme is based on the DWC approach but employs a flexible checker module that is able to deal with inexactness generated by faults and noise. A key novelty of our scheme is the utilization of CNNs to discriminate between usable and unusable images and reduce the costs associated with the detection and the management of faults. Experiments on two aerospace image processing case studies show overall time savings of 14.90% and 34.73% for the two applications, respectively, as compared with the baseline classical DWC scheme. From the dependability point of view, the two hardened applications are able to correctly produce a sable output 99.7% and 98.4% of the times a fault occurs, which is well within acceptable margins in non safety-/mission-critical scenarios.
I tradizionali approcci impiegati per il rilevamento e la tolleranza ai guasti introducono costi elevati per ottenere la correttezza assoluta durante l'elaborazione dei dati. Per esempio, il classico approccio Duplication With Comparison (DWC) permette di rilevare i guasti duplicando il sistema e scartando i dati non appena avviene un errore, mentre il Triple Modular Redundancy (TMR) prevede la triplicazione del sistema e permette di tollerare un guasto grazie ad uno schema di voto. Tuttavia, molte applicazioni hanno una intrinseca tolleranza ad un determinato grado di inesattezza. In questa tesi, ci concentiramo sullo scenario pratico dell'elaborazione di immagini nello spazio, un ambiente in cui i guasti sono un pericolo, mentre applicazioni di questo tipo non sono critiche per la sicurezza o per la missione e possono tollerare alcuni tipi di errori. Il primo passo è proporre un cambiamento dai concetti di corrotto/non corrotto al concetto di usabilità dell'immagine in elaborazione per poter rilassare i tradizionali requisiti di assoluta correttezza e limitare i costi computazionali relativi all'affidabilità. Successivamente, proponiamo un innovativo e flessibile schema di gestione dei guasti progettato per applicazioni in cui le inesattezze sono tollerate entro determinati limiti. Il nuovo schema si basa sull'approccio DWC ma utilizza un nuovo modulo controllore in grado di gestire le inesattezze generate dai guasti e dal rumore. Una importante novità del nostro schema è l'utilizzo delle Reti Neurali Convoluzionali per discriminare tra immagini utilizzabili e inutilizzabili e ridurre i costi associati al rilevamento ed alla gestione dei guasti. Gli esperimenti condotti su due casi di studio evidenziano un risparmio di tempo ottenuto del 14.90% e 34.73% per le due applicazioni, rispettivamente, in comparazione con il classico schema DWC. Dal punto di vista dell'affidabilità, le due applicazioni irrobustite con il nostro schema producono correttamente un output utilizzabile nel 99.7% e 98.4% delle volte in cui un avviene un guasto, risultando in margini accettabili per scenari non critici per la sicurezza o la missione.
A neural network based fault detection/management scheme for reliable image processing applications
BIASIELLI, MATTEO
2018/2019
Abstract
Traditional fault detection/tolerance approaches introduce relevant costs to achieve unconditional correctness during data processing. For example, the classical Duplication With Comparison (DWC) approach achieves fault detection by duplicating the system and discarding data as soon as any error occurs, while the Triple Modular Redundancy (TMR) approach triplicates the system and achieves fault tolerance by means of a voting scheme. However, many application environments are inherently tolerant to a certain degree of inexactness or inaccuracy. In this thesis, we focus on the practical scenario of image processing in space, a domain where faults are a threat, while such applications are not safety-/mission-critical and are inherently tolerant to a certain degree of errors. We first propose a shift from the concepts of corrupted/uncorrupted the concept of usability of the processed image to relax the traditional requirement of unconditional correctness, and to limit the computational overheads related to reliability. We then introduce our new flexible and lightweight fault management scheme designed for application environments where inexactness is tolerated up to a given limit. The new scheme is based on the DWC approach but employs a flexible checker module that is able to deal with inexactness generated by faults and noise. A key novelty of our scheme is the utilization of CNNs to discriminate between usable and unusable images and reduce the costs associated with the detection and the management of faults. Experiments on two aerospace image processing case studies show overall time savings of 14.90% and 34.73% for the two applications, respectively, as compared with the baseline classical DWC scheme. From the dependability point of view, the two hardened applications are able to correctly produce a sable output 99.7% and 98.4% of the times a fault occurs, which is well within acceptable margins in non safety-/mission-critical scenarios.| File | Dimensione | Formato | |
|---|---|---|---|
|
MatteoBiasielliThesis.pdf
accessibile in internet per tutti
Descrizione: Thesis PDF
Dimensione
12.1 MB
Formato
Adobe PDF
|
12.1 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/10589/149902