1,720,967 research outputs found
Tecniche di Machine Learning per Macchine Smart e Impianti Smart: Applicazioni in Manutenzione Predittiva e Condition Monitoring Industriale
No full textIl crescente interesse per le tecniche di intelligenza artificiale è notevolmente accresciuto nell’ultimo decennio tra tutti i settori industriali. Tali metodologie supportano lo sviluppo di nuove funzionalità attraverso l’integrazione di sistemi cyber-fisici e unità di monitoraggio integrate in macchine e impianti, come ricorda il prefisso smart. Il lavoro di tesi si concentra sul Condition Monitoring e la manutenzione predittiva, i cui compiti più diffusi riguardano l’Anomaly Detection, la Fault Classification e la stima della Remaining Useful Life. Molte applicazioni “industry-class” si affidano a metodi di Supervised Machine Learning, in quanto possiedono metriche di valutazione più mature dei metodi unsupervised. Nonostante ciò, limiti addizionali aggiungono complessità allo sviluppo di una soluzione commerciabile. Il flusso di lavoro perciò necessita di gestire restrizioni provenienti dalla piattaforma di monitoraggio, come il carico CPU sopportabile, o diverse condizioni operative. Tali limitazioni spesso sfociano in trade-off di progetto, ma tale informazione nota a priori può portare effetti benefici al flusso stesso. Il lavoro di tesi analizza le limitazioni più diffuse e il loro rapporto con il flusso di lavoro, fornendo alcuni metodi per integrare in modo benefico tale informazione a priori a seconda del compito di monitoraggio e del passo di elaborazione dati. Per validare i metodi proposti sono analizzati due casi di studio riguardanti l’impiego di umidificatori industriali. In Feature Engineering la selezione delle feature è uno dei fattori maggiormente impattanti in quanto la maggioranza delle risorse computazionali e di memoria è impiegata per la loro estrazione. Una Feature Selection sostenibile deve perciò considerarne i costi di estrazione e l’utilità in campo. A tale scopo è proposto un algoritmo chiamato Feature Voting per operare una selezione multi-obiettivo considerando dataset appartenenti a condizioni operative differenti e attributi delle feature, come i costi di estrazione. È inoltre introdotta un’ottimizzazione del Feature Voting basata sul Design Of Experiments. Il Feature Voting incrementa la portabilità di un’applicazione sviluppata in molti contesti operativi diversi, come auspicato dalla pratica industriale. È inoltre proposta una ridefinizione della funzione target, basata su classi di manutenzione condivise, allo scopo di efficientare la Fault Classification riducendo la complessità di classificazione. L’approccio è esteso considerando i costi di manutenzione e sviluppando adattamenti cost-sensitive dei classificatori attraverso l’Ensemble Learning e il paradigma client-server. La procedura riduce i costi di errata classificazione senza riallenare i classificatori sviluppati. Un successivo compito di post-elaborazione riguarda l’utilizzo di predizioni instabili fornite da un classificatore black-box sull’azione di manutenzione da eseguire. È perciò proposta una tecnica di stabilizzazione delle predizioni basata sulla teoria dei Fuzzy Set, incorporando inoltre un meccanismo di isteresi dinamica. Il sistema proposto ha stabilizzato le predizioni e anticipato la richiesta di manutenzione, anche in condizioni di elevata incertezza. Un secondo tema è inerente all’utilizzo della stima della Remaining Useful Life e dell’incertezza associata per attività di post-prognostica. Il metodo proposto utilizza il Functional Profile Modelling per modellare lo stress del sistema, è poi applicato un opportuno adattamento di scala per la correzione a posteriori della stima. Il metodo proposto rappresenta un approccio alternativo guidato dalle richieste industriali legate all’apprendimento sul campo e all'integrazione di informazioni a posteriori. L’intera dissertazione offre inoltre una metodologia di analisi per correlare un flusso di progetto basato sul Machine Learning con i temi di interesse industriale nel campo della manutenzione smart.The increasing interest in Artificial Intelligence techniques has risen in the last decade among all the industrial knowledge domains. Such methods support new advanced functionalities assisted by the integration of cyber-physical systems and monitoring units inside machines and plants, as recalled with the smart appellative. The thesis focuses on Condition Monitoring and Predictive Maintenance in which Anomaly Detection, Fault Classification, and Remaining Useful Life estimation are the most common tasks to solve. Most industry-class applications rely on Supervised Machine Learning techniques, as the evaluation metrics are more consolidated than unsupervised methods. However, additional restrictions add complexity to the development of a marketable solution. During the development of such algorithms, a regular Machine Learning workflow needs to manage additional constraints coming from the monitoring platform, like the available CPU, memory, or different field conditions. Such restrictions often lead to design trade-offs, but such information is prior known, and the design process can benefit from it. This thesis aims to analyze some of the most diffused constraints, along with the Machine Learning workflow, to provide some techniques that enable beneficially embedding such prior information according to the monitoring task and the processing step. To validate the methods, two industrial case studies are analyzed regarding the employment of industrial humidifiers. In Feature Engineering, the selection of features is one of the most impacting factors because many computational and memory resources, when using a deployed model, are due to the extraction of features. A sustainable Feature Selection thus needs to consider their extraction costs and their validity for a given field condition. Feature Selection is addressed by proposing an algorithm called Feature Voting to perform a multi-objective selection that considers datasets belonging to different field conditions and feature attributes, like the computational and memory extraction costs. Feature Voting tuning is also performed based on the Design Of Experiments. Feature Voting boosts the usage performance of a deployed machine in plenty of working environments, as found in the industrial practice. A maintenance-based target redefinition has been proposed to efficiently improve Fault Classification, i.e. grouping fault types according to shared maintenance interventions to lower the classification complexity. The approach is extended by considering the maintenance costs, leading to cost-sensitive adaptations of classifiers exploiting Ensemble Learning and the client-server paradigm. The procedure lowers the misclassification cost without retraining the deployed classifiers. In postprocessing, an additional task focuses on the practical use of the predictions given by a black-box unstable classifier on whether maintenance action to perform on the system. A prediction stabilization technique is proposed by exploiting the Fuzzy Set theory. A dynamical hysteresis mechanism is also introduced to increase the scheduling margin. The proposed system provided more stable predictions over time and anticipated the intervention alarm, even in highly uncertain conditions. A second postprocessing task deals with the post-prognostic usage of the Remaining Useful Life estimate and the management of its uncertainty. A method for the posterior correction of such an estimate is proposed. The technique exploits Functional Profile Modeling to describe the stress experienced by the monitored system on the field; then, a proper scaling is applied. The method represents an alternative approach, driven by the industrial requirement of learning from the field and integrating the extracted posterior information. All the thesis corpus offers an analysis methodology that relates a regular Machine Learning workflow with open issues of industrial interest in the smart maintenance field
Impact of Noise on Machine Learning-based Condition Monitoring Applications: a Case Study
No full textIn the paper, application of Machine Learning (ML) techniques for the continuous monitoring of the health status of mild mission-critical industrial equipment is considered. A meaningful real-life case-study is presented in order to show how acquisition conditions may severely impact on the performance of the system. In particular, it is shown that a wrong estimate of noise effects in the deployed system may induce a wrong choice of the best features feeding the ML monitoring algorithm, hence affecting accuracy of the target devices. The discussed results may provide an useful guidance to the practitioner in the field during the design phase of ML-based devices depending of the equipment specifications and environmental conditions
Type-2 Fuzzy Set Aggregation and Health Status Mining from Condition-Monitoring Applications
No full textFeature Ranking under Industrial Constraints in Continuous Monitoring Applications based on Machine Learning Techniques
No full textThe design work-flow of machine learning techniques for continuous monitoring or predictive maintenance in an industrial context is usually a two step procedure: the selection of features to be computed from the observed signals and training of a suitable algorithm with real-life meaningful data, that will be next deployed in the second step. Feature selection is a relevant task since it provides a powerful optimisation of the deployed algorithm performance, for the given training data-set. The paper provides a method for feature ranking and selection that embeds constraints coming from real-life applications, including sensing device specifications, environmental noise, available processing resources, being all these latter aspects not considered in the currently available literature methods for feature selection. A practical case-study in the field on anomaly detection of machines is reported and discussed, in order to show the good properties of the provided method
Impact of Usage Profiles on Remaining Useful Life and Post-Prognostic Maintenance Decisions
No full textCost-Sensitive Fault Identification in Predictive Maintenance Applications: a Case Study
No full text
Going Beyond Counting First Authors in Author Co-citation Analysis
Get PDFThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
Variations on the Author
Get PDF“Variations on the Author” discusses two of Eduardo Coutinho’s recent films (Um Dia na Vida, from 2010, and Últimas Conversas, posthumously released in 2015) and their contribution to the general question of documentary authorship. The director’s filmography is characterized by a consistent yet self-effacing form of authorial self-inscription: Coutinho often features as an interviewer that rather than express opinions propels discourses; an interviewer that is good at listening. This mode of self-inscription characterizes him as an author who is not expressive but who is nonetheless markedly present on the screen. In Um Dia na Vida, however, Coutinho is completely absent form the image, while Últimas Conversas, on the contrary, includes a confessional prologue that moves the director from the margins to the center of his films. This article examines the ways in which these works stand out in the filmography of a director who offers new insights into the notion of cinematic authorship
Appropriate Similarity Measures for Author Cocitation Analysis
Get PDFWe provide a number of new insights into the methodological discussion about author cocitation analysis. We first argue that the use of the Pearson correlation for measuring the similarity between authors’ cocitation profiles is not very satisfactory. We then discuss what kind of similarity measures may be used as an alternative to the Pearson correlation. We consider three similarity measures in particular. One is the well-known cosine. The other two similarity measures have not been used before in the bibliometric literature. Finally, we show by means of an example that our findings have a high practical relevance.information science;Pearson correlation;cosine;similarity measure;author cocitation analysis
- …
