The exponential development of biotechnology has lead to a quasi unlimited number of potential products going from biopolymers to vaccines. Cell culture has therefore evolved from the simple cell growth outside its natural environment to its use to produce molecules that they do not naturally produce. This rapid development could not be continued without new control and supervising tools as well as a good process understanding. This requirement involves however a large diversity and a better accessibility of process measurements. In this framework, software sensors show numerous potentialities. The objective of a software sensor is indeed to provide an estimation of the system state variables and particularly those which are not obtained through in situ hardware sensors or laborious and expensive analysis. In this context, This work attempts to join the knowledge of increasing bioprocess complexity and diversity and the time scale of process developments and favours systematic modelling methodology, its flexibility and the speed of development. In the field of state observation, an important modelling constraint is the one induced by the selection of the state to estimate and the available measurements. Another important constraint is the model quality. The central axe of this work is to provide solutions in order to reduce the weight of these constraints to software sensors development. On this purpose, we propose four solutions to four main questions that may arise. The first two ones concern modelling uncertainties. 1."How to develop a software sensor using measurements easily available on pilot scale bioreactor?" The proposed solution is a static software sensor using an artificial neural network. Following this modelling methodology we developed static software sensors for the biomass and ethanol concentrations in a pilot scale S. cerevisae cell culture using the measurement of titrating base quantity, agitation rate and CO₂ concentration in the exhaust gas. 2."How to obtain a reaction scheme and a kinetic model to develop a dynamic observation model?". The proposed solution is to combine three elements: a systematic methodology to generate, identify and select the possible reaction schemes, a general kinetic model and a systematic identification procedure where the last step is particularly dedicated to the identification of observation models. Combining these methodologies allowed us to develop a software sensor for the concentrations of an allergen produced by an animal cell culture using the discrete measurement of glucose, glutamine and ammonium concentrations (which are also estimated in continuous time by the software sensors). The two other questions are dealing with kinetic model uncertainty. 3 "How to correct kinetic model parameters while keeping the system observability?". We consider the possibility to correct some model parameters during the process observation. We propose indeed an adaptive observer based on the theory of the most likely initial conditions observer and exploiting the information from the asymptotic observer. This algorithm allows to jointly estimate the state and some kinetic model parameters. 4 "How to avoid any state observer selection that requires an a priori knowledge on the model quality?". Answering this question lead us to the development of hybrid state observers. The general principle of a hybrid observer is to automatically evaluate the model quality and to select the appropriate state observer. In this work we focus on kinetic model quality and propose hybrid observers that evolves between the state observation from an exponential observer (free convergence rate tuning but model error sensitivity) and the one provided by an asymptotic observer (no kinetic model requirement but a convergence rate depending on the dilution rate). Two strategies are investigated in order to evaluate the model quality and to induce the state observation evolution. Each of them have been validated on two simulated cultures (microbial and animal cells) and one real industrial one (B. subtilis). ∙ In a first strategy, the hybrid observer is based on the determination of a parameter that drives the state estimation from the one obtained with an exponential observer (exponential observation) when the model is of good quality to the one provided by an asymptotic observer (asymptotic observation) when a kinetic model error is detected. The evaluation of this driving parameter is made either with an a priori defined function or is coupled to the identification of the initial conditions in a most likely initial conditions observer. ∙ In another strategy, the hybrid observer is based on a statistical test that compares the state estimations provided by an exponential and an asymptotic observer and corrects the state estimation according to it./ Le rapide développement des biotechnologies permet actuellement d'envisager un nombre quasi illimité de produits potentiels allant du biopolymère au vaccin. La culture cellulaire a dès lors évolué de la simple croissance de cellules en dehors de leur environnement naturel à son exploitation pour la production de molécules qu'elles ne produisent pas naturellement. Un tel développement ne peut se poursuivre sans l'utilisation de nouvelles technologies de contrôle et de supervision ainsi q'une bonne compréhension et maîtrise du biprocédé. Cette exigence nécessite cependant une meilleure accessibilité et une plus grande variabilité des mesures des différentes variables de ce procédé. Dans ce contexte, les capteurs logiciels présentent de nombreuses potentialités. L'objectif d'un capteur logiciel est en effet de fournir une estimation des états d'un système et particulièrement de ceux qui ne sont pas mesurés par des capteurs physiquement installés sur le système ou par de longues et coûteuses analyses. Cet objectif peut être obtenu en combinant un modèle du système avec certaines mesures physiques au sein d'un algorithme d'observation d'état. Dans ce domaine de l'observation des bioprocédés, ce travail tente de considérer, à la fois, l'augmentation de la complexité et de la diversité des bioprocédés et l'exigence d'un développement rapide en favorisant le caractère systématique, flexible et rapide des méthodes proposées. Dans le cadre de l'observation des bioprocédés, une importante contrainte de modélisation est induite par la sélection des états à estimer et des mesures disponibles pour cette estimation. Une seconde contrainte est la qualité du modèle. L'axe central de ce travail est de fournir certaines solutions afin de réduire le poids de ces contraintes dans le développement de capteurs logiciels. Pour ce faire, nous proposons quatre réponses à quatre questions qui peuvent survenir lors de ce développement. Les deux premières questions concernent l'incertitude de modélisation. Quant aux deux questions suivantes, elles concernent l'incertitude du modèle lui-même. 1."Comment développer un capteur logiciel exploitant des mesures facilement disponibles sur un bioréacteur pilote?". La réponse que nous apportons à cette question est le développement d'un capteur logiciel statique basé sur un réseau de neurones artificiels. Cette structure nous a permis de développer des capteurs logiciels de concentrations en biomasse et éthanol au sein d'une culture de S. cerevisae utilisant les mesures en ligne de quantité de base titrante, de vitesse d'agitation et de concentration en CO₂ dans le gaz sortant du réacteur. 2."Comment obtenir un schéma réactionnel et un modèle cinétique pour l'identification d'un modèle dynamique d'observation". Afin de répondre à cette question, nous proposons de combiner trois éléments: une méthode de génération systématique de schémas réactionnels, une structure générale de modèle cinétique et une méthode d'identification dont la dernière étape est particulièrement dédiée à l'identification de modèles d'observation. La combinaison de ces éléments nous a permis de développer un capteur logiciel permettant l'estimation continue de la concentration en un allergène produit par une culture de cellules animales en utilisant des mesures échantillonnées de glucose, glutamine et ammonium (qui sont elles aussi estimées en continu par le capteur logiciel). 3."Comment corriger certains paramètres cinétiques tout en maintenant l'observabilité du système?". Nous considérons ici la possibilité de corriger certains paramètres du modèle cinétique durant le procédé de culture. Nous proposons, en effet, un observateur d'état adaptatif exploitant la théorie de l'observateur par identification des conditions initiales les plus vraisemblables et l'information fournie par un observateur asymptotique. L'algorithme proposé permet ainsi de fournir une estimation conjointe de l'état et de certains paramètres cinétiques. 4."Comment éviter la sélection d'un observateur d'état nécessitant une connaissance, a priori, de la qualité du modèle?". La dernière contribution de ce travail concerne le développement d'observateurs d'état hybrides. Le principe général d'un observateur hybride est d'évaluer automatiquement la qualité du modèle et de sélectionner l'observateur d'état approprié. Au sein de ce travail nous considérons la qualité du modèle cinétique et proposons des observateurs d'état hybrides évoluant entre un observateur dit exponentiel (libre ajustement de la vitesse de convergence mais forte sensibilité aux erreurs de mesures) et un observateur asymptotique (ne nécessite aucun modèle cinétique mais présente une vitesse de convergence dépendante du taux de dilution). Afin de réaliser cette évaluation et d'induire l'évolution de l'observation d'état entre ces deux extrémités, deux stratégies sont proposées. Chacune d'elle est illustrée sur deux cultures simulées (une croissance bactérienne et une culture de cellules animales) et une culture réelle de B. subtilis. ∙ Une première stratégie est basée sur la détermination d'un paramètre de pondération entre l'observation fournie par un observateur exponentiel et un observateur asymptotique. L'évaluation de ce paramètre peut être obtenue soit au moyen d'une fonction définie a priori soit par une identification conjointe aux conditions initiales d'un observateur par identification des conditions initiales les plus vraisemblables. ∙ Une seconde stratégie est basée sur une comparaison statistique entre les observations fournies par les deux types d'observateurs. Le résultat de cette comparaison, lorsqu'il indique une incohérence entre les deux observateurs d'état, est alors utilisé pour corriger l'estimation fournie par l'observateur exponentiel.

Software sensors are a potent tool to improve biotechnological real time process monitoring and control. In the current project, algorithms for six partly novel, software sensors were established and tested in a microbial reactor system. Eight batch and two fed-batch runs were carried out with a recombinant Escherichia coli to investigate the suitability of the different software sensor models in diverse cultivation stages. Special respect was given to effects on the sensors after recombinant protein expression was initiated by addition of an inducer molecule. It was an objective to figure out influences of excessive recombinant protein expression on the software sensor signals. Two of the developed algorithms calculated the biomass on-line and estimated furthermore, the specific growth rate by integration of the biomass changes with the time. The principle of the first was the application of a near infrared probe to obtain on-line readings of the optical density. The other algorithm was founded on the titration of ammonia as only available nitrogen source. The other two sensors analyzed for the specific consumption of glucose and the specific production of acetate and are predicted on an in-line HPLC system. The results showed that all software sensors worked as expected and are rather powerful to estimate important state parameters in real time. In some stages, restrictions may occur due to different limitation affects in the models or the physiology of the culture. However, the results were very convincing and suggested the development of further and more advanced software sensor models in the future.

<p>Software sensors are a potent tool to improve biotechnological real time process monitoring and control. In the current project, algorithms for six partly novel, software sensors were established and tested in a microbial reactor system. Eight batch and two fed-batch runs were carried out with a recombinant <em>Escherichia coli</em> to investigate the suitability of the different software sensor models in diverse cultivation stages. Special respect was given to effects on the sensors after recombinant protein expression was initiated by addition of an inducer molecule. It was an objective to figure out influences of excessive recombinant protein expression on the software sensor signals.</p><p>Two of the developed algorithms calculated the biomass on-line and estimated furthermore, the specific growth rate by integration of the biomass changes with the time. The principle of the first was the application of a near infrared probe to obtain on-line readings of the optical density. The other algorithm was founded on the titration of ammonia as only available nitrogen source. The other two sensors analyzed for the specific consumption of glucose and the specific production of acetate and are predicted on an in-line HPLC system.</p><p>The results showed that all software sensors worked as expected and are rather powerful to estimate important state parameters in real time. In some stages, restrictions may occur due to different limitation affects in the models or the physiology of the culture. However, the results were very convincing and suggested the development of further and more advanced software sensor models in the future.</p>

Esta dissertação propõe o desenvolvimento de uma rede neural artificial (RNA) direcionada a ambientes foundation fieldbus para realização do cálculo de vazão em dutos fechados. Para tanto, a metodologia proposta utiliza-se de medidas de pressão, temperatura e pressão diferencial, as quais normalmente estão disponíveis em plantas industriais. A principal motivação do emprego das redes neurais reside no seu baixo custo e simplicidade de implementação, o que possibilita o emprego de apenas blocos fieldbus padrões tornando a metodologia independente do fabricante. Foi utilizada uma rede perceptron multicamadas com algoritmo de treinamento backpropagation de Levenberg-Marquardt. O treinamento foi realizado numa programação elaborada para o software Matlab TM. A arquitetura da rede neural foi determinada por métodos empíricos variando-se o número de neurônios e de camadas neurais até se atingir um erro aceitável na prática. Após esses treinamentos foi desenvolvida uma programação para realizar os cálculos de vazão em um ambiente foundation fieldbus utilizando-se para tanto o software DeltaV TM do fabricante Emerson Process Management. Foram obtidos resultados com erro relativo médio de valor de vazão em torno de 1.43% para um primeiro cenário utilizando uma placa de orifício e ar como fluido, e de 0,073% para um segundo cenário utilizando uma placa de orifício e gás natural como fluido, com relação aos valores obtidos através do instrumento multivariável 3095MV TM do fabricante Rosemount. Os valores de erro encontrados validam o método desenvolvido nessa dissertação. / This dissertation proposes the development of an artificial neural network (ANN) directed to foundation fieldbus environment for calculation of flow in closed ducts. The proposed methodology uses measurements of pressure, temperature and differential pressure, which are usually available in industrial plants. The main motivation of the use of neural networks lies in their low cost and simplicity of implementation, which allows the use of standard fieldbus blocks by just making the method independent of the manufacturer. It was used a multilayer perceptron network with backpropagation training and algorithm from Levenberg-Marquardt. The training was programmed in the software Matlab TM. The architecture of the ANN was determined by empirical methods by varying the number of neurons and neural layers until it reaches an acceptable error. After such trainings, it was developed a program to perform the flow calculations in an foundation fieldbus environment using Emerson Process Management\'s DeltaV TM software. The results were obtained with an average relative error of flow rate of 1.43% for the first scenario using an orifice plate and air as a process fluid, and 0.073% for a second scenario using an orifice plate and natural gas as the fluid related to the values obtained from Rosemount 3095MV TM multivariable instrument. The values of error found validate the method developed in this dissertation.

Esta dissertação propõe o desenvolvimento de uma rede neural artificial (RNA) direcionada a ambientes foundation fieldbus para realização do cálculo de vazão em dutos fechados. Para tanto, a metodologia proposta utiliza-se de medidas de pressão, temperatura e pressão diferencial, as quais normalmente estão disponíveis em plantas industriais. A principal motivação do emprego das redes neurais reside no seu baixo custo e simplicidade de implementação, o que possibilita o emprego de apenas blocos fieldbus padrões tornando a metodologia independente do fabricante. Foi utilizada uma rede perceptron multicamadas com algoritmo de treinamento backpropagation de Levenberg-Marquardt. O treinamento foi realizado numa programação elaborada para o software Matlab TM. A arquitetura da rede neural foi determinada por métodos empíricos variando-se o número de neurônios e de camadas neurais até se atingir um erro aceitável na prática. Após esses treinamentos foi desenvolvida uma programação para realizar os cálculos de vazão em um ambiente foundation fieldbus utilizando-se para tanto o software DeltaV TM do fabricante Emerson Process Management. Foram obtidos resultados com erro relativo médio de valor de vazão em torno de 1.43% para um primeiro cenário utilizando uma placa de orifício e ar como fluido, e de 0,073% para um segundo cenário utilizando uma placa de orifício e gás natural como fluido, com relação aos valores obtidos através do instrumento multivariável 3095MV TM do fabricante Rosemount. Os valores de erro encontrados validam o método desenvolvido nessa dissertação. / This dissertation proposes the development of an artificial neural network (ANN) directed to foundation fieldbus environment for calculation of flow in closed ducts. The proposed methodology uses measurements of pressure, temperature and differential pressure, which are usually available in industrial plants. The main motivation of the use of neural networks lies in their low cost and simplicity of implementation, which allows the use of standard fieldbus blocks by just making the method independent of the manufacturer. It was used a multilayer perceptron network with backpropagation training and algorithm from Levenberg-Marquardt. The training was programmed in the software Matlab TM. The architecture of the ANN was determined by empirical methods by varying the number of neurons and neural layers until it reaches an acceptable error. After such trainings, it was developed a program to perform the flow calculations in an foundation fieldbus environment using Emerson Process Management\'s DeltaV TM software. The results were obtained with an average relative error of flow rate of 1.43% for the first scenario using an orifice plate and air as a process fluid, and 0.073% for a second scenario using an orifice plate and natural gas as the fluid related to the values obtained from Rosemount 3095MV TM multivariable instrument. The values of error found validate the method developed in this dissertation.

The exponential development of biotechnology has lead to a quasi unlimited number of potential products going from biopolymers to vaccines. Cell culture has therefore evolved from the simple cell growth outside its natural environment to its use to produce molecules that they do not naturally produce. This rapid development could not be continued without new control and supervising tools as well as a good process understanding. This requirement involves however a large diversity and a better accessibility of process measurements. In this framework, software sensors show numerous potentialities. The objective of a software sensor is indeed to provide an estimation of the system state variables and particularly those which are not obtained through in situ hardware sensors or laborious and expensive analysis. In this context, This work attempts to join the knowledge of increasing bioprocess complexity and diversity and the time scale of process developments and favours systematic modelling methodology, its flexibility and the speed of development. In the field of state observation, an important modelling constraint is the one induced by the selection of the state to estimate and the available measurements. Another important constraint is the model quality. The central axe of this work is to provide solutions in order to reduce the weight of these constraints to software sensors development. On this purpose, we propose four solutions to four main questions that may arise. The first two ones concern modelling uncertainties.<p><p>1."How to develop a software sensor using measurements easily available on pilot scale bioreactor?" The proposed solution is a static software sensor using an artificial neural network. Following this modelling methodology we developed static software sensors for the biomass and ethanol concentrations in a pilot scale S. cerevisae cell culture using the measurement of titrating base quantity, agitation rate and CO& / Doctorat en sciences agronomiques et ingénierie biologique / info:eu-repo/semantics/nonPublished

This thesis demonstrates the successful development of a software sensor for Siemens Energy's SGT-700 gas turbines using machine learning algorithms. Our goal was to enhance the robustness of measurements and redundancies, enabling early detection of sensor or turbine malfunctions and contributing to predictive maintenance methodologies. The research is based on a real-world case study, implementing the Cross Industry Standard Process for Data Mining (CRISP DM) methodology in an industrial setting. The thesis details the process from dataset preparation and data exploration to algorithm development and evaluation, providing a comprehensive view of the development process. This work is a step towards integrating machine learning into gas turbine systems. The data preparation process highlights the challenges that arise in the industrial application of data-driven methodologies due to inevitable data quality issues. It provides insight into potential future improvements, such as the constraint programming approach used for dataset construction in this thesis, which remains a valuable tool for future research. The range of algorithms proposed for the software sensor's development spans from basic to more complex methods, including shallow networks, ensemble methods and recurrent neural networks. Our findings explore the limitations and potential of the proposed algorithms, providing valuable insights into the practical application of machine learning in gas turbines. This includes assessing the reliability of these solutions, their role in monitoring machine health over time, and the importance of clean, usable data in driving accurate and satisfactory estimates of different variables in gas turbines. The research underscores that, while replacing a physical sensor with a software sensor is not yet feasible, integrating these solutions into gas turbine systems for health monitoring is indeed possible. This work lays the groundwork for future advancements and discoveries in the field. / Denna avhandling dokumenterar den framgångsrika utvecklingen av en mjukvarusensor för Siemens Energy's SGT-700 gasturbiner med hjälp av maskininlärningsalgoritmer. Vårt mål var att öka mätkvaliten samt införa redundans, vilket möjliggör tidig upptäckt av sensor- eller turbinfel och bidrar till utvecklingen av prediktiv underhållsmetodik. Forskningen baseras på en verklig fallstudie, implementerad enligt Cross Industry Standard Process for Data Mining-metodiken i en industriell miljö. Avhandligen beskriver processen från datamängdsförberedelse och datautforskning till utveckling och utvärdering av algoritmer, vilket ger en heltäckande bild av utvecklingsprocessen. Detta arbete är ett steg mot att integrera maskininlärning i gasturbinssystem. Dataförberedelsesprocessen belyser de utmaningar som uppstår vid industriell tillämpning av datadrivna metoder på grund av oundvikliga datakvalitetsproblem. Det ger insikt i potentiella framtida förbättringar, såsom den begränsningsprogrammeringsansats som används för datamängdskonstruktion i denna avhandling, vilket förblir ett värdefullt verktyg för framtida forskning. Utvecklingen av mjukvarusensorn sträcker sig från grundläggande till mer komplexa metoder, inklusive ytliga nätverk, ensemblemetoder och återkommande neurala nätverk. Våra resultat utforskar begränsningarna och potentialen hos de föreslagna algoritmerna och ger värdefulla insikter i den praktiska tillämpningen av maskininlärning i gasturbiner. Detta inkluderar att bedöma tillförlitligheten hos dessa lösningar, deras roll i övervakning av maskinhälsa över tid och vikten av ren, användbar data för att generera korrekta och tillfredsställande uppskattningar av olika variabler i gasturbiner. Forskningen understryker att, medan det ännu inte är genomförbart att ersätta en fysisk sensor med en mjukvarusensor, är det verkligen möjligt att integrera dessa lösningar i gasturbinssystem för tillståndsövervakning. Detta arbete lägger grunden för vidare studier och upptäckter inom området. / Esta tesis demuestra el exitoso desarrollo de un sensor basado en software para las turbinas de gas SGT-700 de Siemens Energy utilizando algoritmos de aprendizaje automático. Esto con el objetivo de contribuir a las metodologías de mantenimiento predictivo. La investigación se basa en un estudio industrial que implementa la metodología de Proceso Estándar de la Industria para la Minería de Datos, cuyo acrónimo en inglés CRISP-DM. La tesis detalla el proceso desde la preparación del 'dataset', la exploración de datos hasta el desarrollo y evaluación de algoritmos, proporcionando una visión holistica del proceso de desarrollo. Este trabajo representa un paso hacia la integración del aprendizaje automático en turbinas de gas. Nuestros hallazgos exploran las limitaciones y el potencial de los algoritmos propuestos, proporcionando un analisis sobre la aplicación práctica del aprendizaje automático en turbinas de gas. Esto incluye evaluar la confiabilidad de estas soluciones, su papel en la monitorización de la salud de la máquina a lo largo del tiempo, y la importancia de los datos limpios y utilizables para impulsar estimaciones precisas y satisfactorias de diferentes variables en las turbinas de gas. La investigación sugiere que, aunque reemplazar un sensor físico con un sensor basado en aprendizaje automatico aún no es factible, sí es posible integrar estas soluciones en los sistemas de turbinas de gas para monitorear del estado de la maquina.