Global warming is an increasingly destructive phenomenon, whose effects are now clearly evident and prompt worldwide institutions to seek urgent solutions. In this context, the automotive industry is undergoing a profound energy transition, moving away from internal combustion engines toward electric propulsion systems to reduce environmental impact and greenhouse gas emissions. This transition is expected to accelerate further in the coming decades, driven by increasingly stringent regulatory frameworks and consumer demand for sustainable mobility. This doctoral thesis focuses on the optimization and design of electric motors, supported by artificial intelligence techniques. Particular emphasis is placed on addressing NVH (Noise, Vibration, and Harshness) issues, which emerge as critical challenges in electric vehicles, often leading to discomfort for passengers due to high-frequency tonal noise emissions. Electric motor design requires the concurrent management of numerous design variables, engineering constraints, and objective functions, such as maximizing efficiency, torque, and power while minimizing NVH emissions. Given the inherent trade-offs among these competing objectives, it is generally impossible to identify a unique optimal solution; instead, designers must seek an appropriate balance that best fits the specific application requirements. In this context, multi-objective optimization provides a powerful framework to manage all relevant factors concurrently, generating a Pareto front of optimal trade-off solutions from which designers can select the most appropriate configuration. Moreover, electric motor design is inherently multiphysical, involving the coupled interaction of electromagnetic, thermal, and structural domains. To enable computationally efficient optimization, surrogate models are leveraged to approximate the behavior of high-fidelity simulations. Their accuracy, however, is fundamentally determined by the distribution, density, and informativeness of the training dataset, underscoring the pivotal role of low-discrepancy sampling strategies for effective design space exploration. Low-discrepancy sequences, such as the Sobol sequence, are employed to maximize information content with a minimal number of simulation runs, enabling efficient surrogate model training. To address these challenges, this thesis proposes the development of a novel optimization strategy based on the Adaptive Pareto Algorithm (APA), capable of producing well-distributed Pareto fronts with significantly reduced computational cost through the dynamic integration and iterative refinement of artificial neural network-based surrogate models. Unlike traditional evolutionary algorithms, the APA dynamically adapts the surrogate model during the optimization process, concentrating computational efforts where the Pareto front evolves more rapidly. This adaptive process demonstrates that the final Pareto-optimal solutions are predominantly generated by the neural network itself, underlining the critical role of AI in accelerating the electric motor design process. Since electric motors are subject to various uncertainties, such as dimensional tolerances, operational conditions, and material property variations, this thesis further incorporates Reliability-based Optimization (RBO) strategies into the design framework. Two complementary robust methods are investigated: the beta-efficient approach and the direct minimization of the objective functions standard deviation. These methods prove capable of achieving designs that are not only optimal under nominal conditions but also highly reliable under real-world variability, thus ensuring consistent performance across a wide range of operating scenarios. Different confidence levels, specifically 95% and 99%, are explored in the beta-efficient approach to evaluate the trade-off between performance and robustness. A substantial part of the research effort is devoted to improving the prediction of NVH performance. The analysis of internal cooling systems is identified as a crucial factor, and a simplified prototype is developed to investigate the fluid-structure interaction. The results highlight that neglecting the dynamic effect of the coolant leads to significant underestimation of vibrational modes and radiated noise, emphasizing the necessity of including fluid--structure coupling in numerical simulations. Following this, a complete case study is proposed, where a predictive NVH simulation framework is developed. Particular focus is placed on fully predictive structural modeling, eliminating the need for experimental modal testing. The modeling chain incorporates the flow of electromagnetic forces, structural dynamics, and acoustic radiation. Since electromagnetic excitations serve as the initial input, accurate force modeling proves critical, as small variations can substantially affect the final NVH outputs. Excellent correlation between numerical predictions and experimental results is achieved, validating the predictive capability of the developed framework. Furthermore, the acoustic analysis is conducted using a decoupled simulation strategy, where structural vibrations are first computed independently and subsequently serve as boundary conditions for the acoustic solver. This decoupled approach significantly reduces computational costs compared to fully coupled structural-acoustic simulations while maintaining high accuracy in the prediction of radiated noise. The contributions of this work are validated through multiple case studies involving traction electric motors for automotive applications, demonstrating tangible improvements in performance, robustness, and acoustic comfort. The methods developed herein are readily applicable to industrial electric motor design processes, supporting the accelerated development of high-performance, reliable, and acoustically comfortable electric drive systems. Overall, this thesis advances simulation-driven, multi-objective, and reliability aware design methods for the desing of electric motors, providing a strong foundation for the development of next-generation sustainable and acoustically optimized electric powertrains. The results and innovations presented in this research have been disseminated through a series of peer-reviewed scientific publications, which collectively form the structure of this doctoral thesis.
Il riscaldamento globale è un fenomeno sempre più distruttivo, i cui effetti sono ormai chiaramente visibili e spingono le istituzioni a livello mondiale a ricercare soluzioni urgenti. In questo contesto, l'industria automobilistica sta affrontando una profonda transizione energetica, abbandonando i motori a combustione interna a favore dei sistemi di propulsione elettrica, con l'obiettivo di ridurre l'impatto ambientale e le emissioni di gas serra. Si prevede che questa transizione acceleri ulteriormente nei prossimi decenni, spinta da normative sempre più stringenti e dalla crescente domanda dei consumatori per una mobilità sostenibile. Questa tesi di dottorato si concentra sull'ottimizzazione e la progettazione di motori elettrici, supportate da tecniche di intelligenza artificiale. Un'attenzione particolare è rivolta alle problematiche NVH (Noise, Vibration, and Harshness), che rappresentano una delle principali sfide nei veicoli elettrici, causando spesso disagio ai passeggeri a causa delle emissioni acustiche tonali ad alta frequenza. La progettazione di un motore elettrico richiede la gestione simultanea di numerose variabili progettuali, vincoli ingegneristici e funzioni obiettivo, come il massimo rendimento, la coppia e la potenza, insieme alla minimizzazione delle emissioni NVH. Data la natura conflittuale di questi obiettivi, non è possibile identificare una soluzione ottimale univoca; è invece necessario ricercare un equilibrio adeguato che soddisfi al meglio i requisiti specifici dell'applicazione. In questo contesto, l'ottimizzazione multi-obiettivo fornisce un efficace quadro metodologico per gestire simultaneamente tutti i fattori rilevanti, generando un fronte di Pareto di soluzioni ottimali in termini di compromesso, da cui il progettista può selezionare la configurazione più adatta. Inoltre, la progettazione di motori elettrici presenta una natura intrinsecamente multifisica, che coinvolge l'interazione accoppiata tra domini elettromagnetico, termico e strutturale. Per rendere l'ottimizzazione computazionalmente efficiente, vengono utilizzati modelli surrogati per approssimare il comportamento di simulazioni ad alta fedeltà. Tuttavia, l'accuratezza di questi modelli dipende in modo fondamentale dalla distribuzione, densità e qualità informativa del dataset di addestramento, evidenziando il ruolo cruciale di strategie di campionamento a bassa discrepanza per una efficace esplorazione dello spazio di progetto. Sequenze a bassa discrepanza, come la sequenza di Sobol, vengono impiegate per massimizzare il contenuto informativo con un numero minimo di simulazioni, facilitando l'addestramento efficiente dei modelli surrogati. Per affrontare queste sfide, la tesi propone lo sviluppo di una nuova strategia di ottimizzazione basata sull'Adaptive Pareto Algorithm (APA), capace di generare fronti di Pareto ben distribuiti con un costo computazionale significativamente ridotto, grazie all'integrazione dinamica e al raffinamento iterativo di modelli surrogati basati su reti neurali artificiali. A differenza degli algoritmi evolutivi tradizionali, l'APA adatta dinamicamente il modello surrogato durante il processo di ottimizzazione, concentrando gli sforzi computazionali nelle regioni in cui il fronte di Pareto evolve più rapidamente. Questo processo adattivo dimostra che la maggior parte delle soluzioni finali ottimali viene generata direttamente dalla rete neurale, sottolineando il ruolo centrale dell'intelligenza artificiale nell'accelerare la progettazione di motori elettrici. Poiché i motori elettrici sono soggetti a molteplici incertezze, come tolleranze dimensionali, condizioni operative e variazioni nelle proprietà dei materiali, la tesi integra nel framework progettuale anche strategie di ottimizzazione affidabilistica. Vengono studiate due metodologie robuste e complementari: l'approccio beta-efficiente e la minimizzazione diretta della deviazione standard delle funzioni obiettivo. Queste metodologie permettono di ottenere progetti non solo ottimali in condizioni nominali, ma anche altamente affidabili in presenza di variabilità reale, garantendo prestazioni consistenti su un ampio spettro di scenari operativi. Nell'approccio beta-efficiente vengono esplorati diversi livelli di confidenza, in particolare 95% e 99%, per valutare il compromesso tra prestazioni e robustezza. Una parte sostanziale dell'attività di ricerca è dedicata al miglioramento della predizione delle prestazioni NVH. L'analisi dei sistemi di raffreddamento interno viene identificata come un fattore cruciale, e viene sviluppato un prototipo semplificato per investigare l'interazione fluido-struttura. I risultati dimostrano che trascurare l'effetto dinamico del liquido refrigerante porta a una sottostima significativa dei modi vibrazionali e del rumore irradiato, evidenziando la necessità di includere l'accoppiamento fluido-struttura nelle simulazioni numeriche. Segue quindi uno studio completo di caso in cui viene sviluppato un framework predittivo per la simulazione NVH. Particolare attenzione è dedicata alla modellazione strutturale completamente predittiva, eliminando la necessità di prove sperimentali modali. La catena di modellazione include la propagazione delle forze elettromagnetiche, la dinamica strutturale e la radiazione acustica. Poiché le eccitazioni elettromagnetiche costituiscono l'input iniziale, una loro corretta modellazione risulta fondamentale, in quanto anche piccole variazioni possono influenzare significativamente i risultati NVH. È stata ottenuta un'eccellente correlazione tra le predizioni numeriche e i risultati sperimentali, validando le capacità predittive del framework sviluppato. Inoltre, l'analisi acustica è condotta mediante una strategia di simulazione disaccoppiata, in cui le vibrazioni strutturali sono calcolate indipendentemente e successivamente utilizzate come condizioni al contorno per il solver acustico. Questo approccio consente una riduzione significativa del costo computazionale rispetto a una simulazione strutturale-acustica completamente accoppiata, mantenendo al contempo un'elevata accuratezza nella predizione del rumore irradiato. I contributi di questo lavoro sono validati tramite molteplici casi studio riguardanti motori elettrici per trazione automobilistica, dimostrando miglioramenti tangibili in termini di prestazioni, robustezza e comfort acustico. Le metodologie sviluppate risultano facilmente applicabili ai processi industriali di progettazione dei motori elettrici, supportando lo sviluppo accelerato di sistemi di trazione ad alte prestazioni, affidabili e acusticamente confortevoli. Nel complesso, questa tesi propone metodi di progettazione multi-obiettivo, affidabilistica e guidata dalla simulazione per i motori elettrici, fornendo una solida base per lo sviluppo di powertrain elettrici di nuova generazione, sostenibili e ottimizzati dal punto di vista acustico. I risultati e le innovazioni presentate sono stati divulgati attraverso una serie di pubblicazioni scientifiche, che costituiscono la struttura portante di questa tesi di dottorato.
Al-based optimization and NVH-oriented design of traction electric motors
Barri, Dario
2024/2025
Abstract
Global warming is an increasingly destructive phenomenon, whose effects are now clearly evident and prompt worldwide institutions to seek urgent solutions. In this context, the automotive industry is undergoing a profound energy transition, moving away from internal combustion engines toward electric propulsion systems to reduce environmental impact and greenhouse gas emissions. This transition is expected to accelerate further in the coming decades, driven by increasingly stringent regulatory frameworks and consumer demand for sustainable mobility. This doctoral thesis focuses on the optimization and design of electric motors, supported by artificial intelligence techniques. Particular emphasis is placed on addressing NVH (Noise, Vibration, and Harshness) issues, which emerge as critical challenges in electric vehicles, often leading to discomfort for passengers due to high-frequency tonal noise emissions. Electric motor design requires the concurrent management of numerous design variables, engineering constraints, and objective functions, such as maximizing efficiency, torque, and power while minimizing NVH emissions. Given the inherent trade-offs among these competing objectives, it is generally impossible to identify a unique optimal solution; instead, designers must seek an appropriate balance that best fits the specific application requirements. In this context, multi-objective optimization provides a powerful framework to manage all relevant factors concurrently, generating a Pareto front of optimal trade-off solutions from which designers can select the most appropriate configuration. Moreover, electric motor design is inherently multiphysical, involving the coupled interaction of electromagnetic, thermal, and structural domains. To enable computationally efficient optimization, surrogate models are leveraged to approximate the behavior of high-fidelity simulations. Their accuracy, however, is fundamentally determined by the distribution, density, and informativeness of the training dataset, underscoring the pivotal role of low-discrepancy sampling strategies for effective design space exploration. Low-discrepancy sequences, such as the Sobol sequence, are employed to maximize information content with a minimal number of simulation runs, enabling efficient surrogate model training. To address these challenges, this thesis proposes the development of a novel optimization strategy based on the Adaptive Pareto Algorithm (APA), capable of producing well-distributed Pareto fronts with significantly reduced computational cost through the dynamic integration and iterative refinement of artificial neural network-based surrogate models. Unlike traditional evolutionary algorithms, the APA dynamically adapts the surrogate model during the optimization process, concentrating computational efforts where the Pareto front evolves more rapidly. This adaptive process demonstrates that the final Pareto-optimal solutions are predominantly generated by the neural network itself, underlining the critical role of AI in accelerating the electric motor design process. Since electric motors are subject to various uncertainties, such as dimensional tolerances, operational conditions, and material property variations, this thesis further incorporates Reliability-based Optimization (RBO) strategies into the design framework. Two complementary robust methods are investigated: the beta-efficient approach and the direct minimization of the objective functions standard deviation. These methods prove capable of achieving designs that are not only optimal under nominal conditions but also highly reliable under real-world variability, thus ensuring consistent performance across a wide range of operating scenarios. Different confidence levels, specifically 95% and 99%, are explored in the beta-efficient approach to evaluate the trade-off between performance and robustness. A substantial part of the research effort is devoted to improving the prediction of NVH performance. The analysis of internal cooling systems is identified as a crucial factor, and a simplified prototype is developed to investigate the fluid-structure interaction. The results highlight that neglecting the dynamic effect of the coolant leads to significant underestimation of vibrational modes and radiated noise, emphasizing the necessity of including fluid--structure coupling in numerical simulations. Following this, a complete case study is proposed, where a predictive NVH simulation framework is developed. Particular focus is placed on fully predictive structural modeling, eliminating the need for experimental modal testing. The modeling chain incorporates the flow of electromagnetic forces, structural dynamics, and acoustic radiation. Since electromagnetic excitations serve as the initial input, accurate force modeling proves critical, as small variations can substantially affect the final NVH outputs. Excellent correlation between numerical predictions and experimental results is achieved, validating the predictive capability of the developed framework. Furthermore, the acoustic analysis is conducted using a decoupled simulation strategy, where structural vibrations are first computed independently and subsequently serve as boundary conditions for the acoustic solver. This decoupled approach significantly reduces computational costs compared to fully coupled structural-acoustic simulations while maintaining high accuracy in the prediction of radiated noise. The contributions of this work are validated through multiple case studies involving traction electric motors for automotive applications, demonstrating tangible improvements in performance, robustness, and acoustic comfort. The methods developed herein are readily applicable to industrial electric motor design processes, supporting the accelerated development of high-performance, reliable, and acoustically comfortable electric drive systems. Overall, this thesis advances simulation-driven, multi-objective, and reliability aware design methods for the desing of electric motors, providing a strong foundation for the development of next-generation sustainable and acoustically optimized electric powertrains. The results and innovations presented in this research have been disseminated through a series of peer-reviewed scientific publications, which collectively form the structure of this doctoral thesis.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_Barri.pdf
non accessibile
Descrizione: Tesi
Dimensione
46.76 MB
Formato
Adobe PDF
|
46.76 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/10589/241217