In the framework of industrial quality management, traditional Statistical Process Control (SPC) procedures depend on quality characteristics measured on the product of manufacturing processes. They also assume that a number of parts may be collected during In-Control (IC) operations to estimate the process parameters and to design the control charts. Nevertheless, the evolving market demands and the development of novel technologies have been leading to productive scenarios where traditional SPC methods are no more appropriate or even not applicable. In different discrete-part manufacturing applications (e.g., in the aerospace sector), the production of high-value-added products implies extended machining times (e.g., several hours for a single part, possibly longer than the tool life). It also involves expensive tools and materials, together with high precision requirements. The use of traditional SPC procedures, based on post-process measurements, implies a delay between the possible occurrence of a fault and the detection of its effects on the product. This yields unacceptable costs for wasting expensive materials and time-consuming re-manufacturing operations. In addition, high customization requirements in various sectors impose small lot productions or even one-of-a-kind productions (i.e., the production of lots that consist of a single item). In that case, there is no possibility to perform a control chart design phase based on repeated processes, and hence novel quality control procedures must be considered. A viable solution consists of sensorizing the machine tools and the production systems in order to collect data about the quality and stability of the process during the process itself. This is possible thanks to the continuous technological developments that are leading to data-rich industrial environments, where several sources of potentially useful information are easily available. The result is a paradigm-shift from product-based SPC to in-process SPC. A quality monitoring based on in-process data may provide a faster reaction to out-of-control shifts, thanks to condition-based control strategies aimed at quickly mitigating (or even suppressing) the effects of faults, with a consequent reduction of wasted parts. Furthermore, in-process SPC provides a potential 100% production coverage as it allows collecting data during each process run, instead of evaluating the quality characteristics on sampled products at the end of the process. However, such a paradigm shift implies a number of novel challenges and critical issues with respect to the traditional SPC practice, because of the high sampling frequencies of sensor signals, the computational constraints, the complexity of the signal patterns, the time-varying properties of industrial processes, the streams of data from different sources, and the multiplicity of operating conditions. Those challenges and critical issues push the need to develop novel SPC approaches and to improve the existing ones. This thesis is aimed at studying and developing novel signal-based SPC methods in order to deal with those challenges. A particular family of industrial processes in considered, i.e., the family of discrete-part manufacturing operations that exhibit a cyclical behaviour. In that case, the IC state of the process can be described by cyclically repeating patterns of acquired signals, known as “profiles”. Therefore, this study deals with the use of profile monitoring methods for in-process sensor signals. Different inter-related research problems are discussed and faced. In the first part of the study, the focus is on signals from a single sensor that exhibit complicated patterns or undesired misalignments. In this frame, the two following problems were faced: (i) the integration of registration information in a profile monitoring framework, to guarantee a proper management of different signal variability sources; (ii) the enhancement of profile monitoring performances in the presence of complicated signal patterns characterized by information contributions on different time-frequency scales. The analysis is then extended to signals coming from multiple sensors, which must be properly integrated and fused together in order to achieve a better and more synthetic representation of the on-going process. In the last part of the study, the analysis is focused on the development of profile monitoring methods for processes that exhibit multiple IC states, which represents a challenging violation of traditional SPC assumptions. This analysis if motivated by the fact that, in different manufacturing applications, parts of the same quality can be produced by processes can not be characterized by a unique IC state. The existence of multiple IC states is due to operating modes that vary in time (e.g., different cutting parameters, different tools, different ambient conditions, etc.), and hence novel profile monitoring methods are required. All the proposed approaches are driven by actual industrial challenges and designed for in-process utilization. They were tested by means of Monte Carlo simulations and compared with benchmark techniques. Real data from different industrial case studies, including waterjet cutting, grinding of cylindrical rolls and end-milling processes, were used to demonstrate the performances of the proposed methods in actual industrial scenarios.
Nell'ambito del controllo industriale della qualità, le tradizionali tecniche di Statistical Process Control (SPC) dipendono da caratteristiche di qualità misurate direttamente sul prodotto del processo manifatturiero. Tali tecniche assumono che sia possibile raccogliere un numero sufficiente di parti prodotte in condizioni In-Controllo (IC) per stimare i parametri di processo e progettare le carte di controllo. Tuttavia, esigenze di mercato in continua evoluzione e lo sviluppo di nuove tecnologie portano a scenari produttivi in cui i tradizionali metodi SPC sono non sufficientemente efficaci o addirittura inapplicabili. In diverse applicazioni di discrete-part manufacturing (e.g., nel settore aerospaziale), la produzione di prodotti ad alto valore aggiunto comporta tempi di lavorazione molto lunghi (e.g., diverse ore per una singola parte, anche più della durata di un singolo utensile). A ciò si accompagna l'uso di materiali e utensili costosi e requisiti stringenti sulla qualità della lavorazione. L'uso di procedure tradizionali di SPC, basate su misure post-process, comporta un ritardo tra il possibile verificarsi di un guasto e la sua segnalazione basata sull'osservazione degli effetti lasciati sul prodotto. Ciò comporta costi inaccettabili per lo scarto di materiali costosi e lunghi tempi per la produzione di nuove parti. Inoltre, le esigenze di elevata customizzazione in diversi settori impongono la produzione di lotti ridotti o addirittura di lotti costituita da una singola parte (one-of-a-kind production). In tali casi, non è possibile effettuare una fase di progettazione delle carte di controllo basata su ripetizioni dello stesso processo, e quindi nasce la necessità di studiare e sviluppare metodi nuovi. Una possibile soluzione consiste nella sensorizzazione delle macchine utensili, con l'obiettivo di raccogliere dati sulla qualità e stabilità del processo, durante il processo stesso. Ciò è reso possibile dai continui sviluppi tecnologici che stanno portando ad ambienti industriali "data-rich", in cui sono potenzialmente disponibili molte sorgenti di informazioni utili. Ne consegue un cambio di paradigma, da SPC basato sul prodotto (post-process) ad un nuovo approccio di SPC in-process. Un sistema di monitoraggio della qualità basato su dati in-process può fornire una reazione più rapida a condizioni fuori-controllo, grazie a strategie di controllo " su condizione", con un beneficio notevole in termini di riduzione dei non conformi. Inoltre, il monitoraggio in-process permette una copertura fino al 100% della produzione, siccome si basa su dati raccolti in ogni fase del processo, invece che su caratteristiche di qualità misurate a valle del processo stesso. Nonostante questo, però, tale cambio di paradigma comporta nuove sfide e vari problemi rispetto all'SPC tradizionale, a causa delle elevate frequenza di acquisizione dati, dei vincoli computazionali, della complessità dei segnali, delle proprietà tempo-varianti dei processi, del flusso di dati provenienti da più sensori, e della molteplicità di condizioni operative. Tali sfide e tali criticità spingono il bisogno di studiare approcci nuovi e/o di migliorare e adattare quelli esistenti. L'obiettivo di questa tesi consiste nello studio e sviluppo di nuovi metodi SPC basati su segnali, considerando una particolare famiglia di processi industriali, i.e., quei processi caratterizzati da comportamenti ciclici. In questo scenario, lo stato IC del processo può essere descritto da pattern di segnali che si ripetono ciclicamente nel tempo, e che vengono chiamati "profili". Quindi, questo studio si focalizza su metodi di "profile monitoring" per segnali acquisiti in-process. Vengono discussi e affrontati diversi problemi di ricerca, tra loro inter-correlati. Nella prima parte dello studio, l'attenzione è rivolta a segnali mono-sensore che mostrano pattern complessi e/o disallineamenti indesiderati. In questo contesto, vengono affrontati due problemi: (i) l'integrazione delle informazioni relative all'operazione di registrazione in un framework di profile monitoring, per garantire una corretta caratterizzazione delle varie sorgenti di variabilità; (ii) il miglioramento delle prestazioni di monitoraggio in presenza di pattern complessi in cui i vari contributi informativi sono rappresentabili su scale in tempo-frequenza distinte. L'analisi è poi estesa a segnali provenienti da sensori multipli, che devono essere propriamente integrati e "fusi" insieme per ottenere una migliore e più sintetica comprensione del processo. Nell'ultima parte dello studio, l'attenzione si concentra sullo sviluppo do metodi di profile monitoring per processi caratterizzati da molteplici stati IC, situazione che rappresenta una violazione critica delle tradizionali ipotesi alla base dell'SPC. Questa analisi è motivata dal fatto che, in molte applicazioni manifatturiere, parti di eguale qualità possono essere prodotte da processi che non comportano un unico stato IC. L'esistenza di molteplici stati IC è dovuta a condizioni operative che variano nel tempo (e.g., diversi parametri di lavoro, diversi utensili, diverse condizioni ambientali, etc.). Tutti i metodi proposti sono guidati da reali sfide industriali e progettati per un impiego in-process. Essi sono stati verificati attraverso l'uso di simulazioni Monte Carlo e confrontati con tecniche benchmark. Sono stati inoltre utilizzati dati reali acquisiti in vari casi di studio industriali, che includono processi waterjet, di fresatura e di rettifica, per dimostrare le prestazioni dei metodi proposti in reali contesti industriali.
Profile monitoring of multi-stream sensor data
Abstract
In the framework of industrial quality management, traditional Statistical Process Control (SPC) procedures depend on quality characteristics measured on the product of manufacturing processes. They also assume that a number of parts may be collected during In-Control (IC) operations to estimate the process parameters and to design the control charts. Nevertheless, the evolving market demands and the development of novel technologies have been leading to productive scenarios where traditional SPC methods are no more appropriate or even not applicable. In different discrete-part manufacturing applications (e.g., in the aerospace sector), the production of high-value-added products implies extended machining times (e.g., several hours for a single part, possibly longer than the tool life). It also involves expensive tools and materials, together with high precision requirements. The use of traditional SPC procedures, based on post-process measurements, implies a delay between the possible occurrence of a fault and the detection of its effects on the product. This yields unacceptable costs for wasting expensive materials and time-consuming re-manufacturing operations. In addition, high customization requirements in various sectors impose small lot productions or even one-of-a-kind productions (i.e., the production of lots that consist of a single item). In that case, there is no possibility to perform a control chart design phase based on repeated processes, and hence novel quality control procedures must be considered. A viable solution consists of sensorizing the machine tools and the production systems in order to collect data about the quality and stability of the process during the process itself. This is possible thanks to the continuous technological developments that are leading to data-rich industrial environments, where several sources of potentially useful information are easily available. The result is a paradigm-shift from product-based SPC to in-process SPC. A quality monitoring based on in-process data may provide a faster reaction to out-of-control shifts, thanks to condition-based control strategies aimed at quickly mitigating (or even suppressing) the effects of faults, with a consequent reduction of wasted parts. Furthermore, in-process SPC provides a potential 100% production coverage as it allows collecting data during each process run, instead of evaluating the quality characteristics on sampled products at the end of the process. However, such a paradigm shift implies a number of novel challenges and critical issues with respect to the traditional SPC practice, because of the high sampling frequencies of sensor signals, the computational constraints, the complexity of the signal patterns, the time-varying properties of industrial processes, the streams of data from different sources, and the multiplicity of operating conditions. Those challenges and critical issues push the need to develop novel SPC approaches and to improve the existing ones. This thesis is aimed at studying and developing novel signal-based SPC methods in order to deal with those challenges. A particular family of industrial processes in considered, i.e., the family of discrete-part manufacturing operations that exhibit a cyclical behaviour. In that case, the IC state of the process can be described by cyclically repeating patterns of acquired signals, known as “profiles”. Therefore, this study deals with the use of profile monitoring methods for in-process sensor signals. Different inter-related research problems are discussed and faced. In the first part of the study, the focus is on signals from a single sensor that exhibit complicated patterns or undesired misalignments. In this frame, the two following problems were faced: (i) the integration of registration information in a profile monitoring framework, to guarantee a proper management of different signal variability sources; (ii) the enhancement of profile monitoring performances in the presence of complicated signal patterns characterized by information contributions on different time-frequency scales. The analysis is then extended to signals coming from multiple sensors, which must be properly integrated and fused together in order to achieve a better and more synthetic representation of the on-going process. In the last part of the study, the analysis is focused on the development of profile monitoring methods for processes that exhibit multiple IC states, which represents a challenging violation of traditional SPC assumptions. This analysis if motivated by the fact that, in different manufacturing applications, parts of the same quality can be produced by processes can not be characterized by a unique IC state. The existence of multiple IC states is due to operating modes that vary in time (e.g., different cutting parameters, different tools, different ambient conditions, etc.), and hence novel profile monitoring methods are required. All the proposed approaches are driven by actual industrial challenges and designed for in-process utilization. They were tested by means of Monte Carlo simulations and compared with benchmark techniques. Real data from different industrial case studies, including waterjet cutting, grinding of cylindrical rolls and end-milling processes, were used to demonstrate the performances of the proposed methods in actual industrial scenarios.File | Dimensione | Formato | |
---|---|---|---|
2015_01_PHD_Grasso.pdf
Open Access dal 16/12/2015
Descrizione: Tesi completa
Dimensione
6.96 MB
Formato
Adobe PDF
|
6.96 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/10589/99785