| Abstract | U ovom radu opisan je značaj točne estimacije temperature rotora PMSM-a za automobilsku primjenu zbog utjecaja na učinkovitost, performanse i trajnost stroja. Odrađen je pregled literature koja je estimirala temperaturu PMSM te je predloženo korištenje umjetne inteligencije, odnosno naprednih algoritama strojnog učenja. U radu je korišten programski jezik Python te biblioteke Numpy, Pandas, Matplotlib, Scikit-Learn. Analiziran je javno dostupan set podataka koji sadrži mjerene varijable PMSM te je napravljena predobrada, korelacijska analiza, skaliranje seta podataka i uklanjanje netipičnih vrijednosti.
Početna razmatranja odrađena su korištenjem MLP, a zatim su testirani razni algoritmi strojnog učenja sa zadanim postavkama. Testirani su na obrađenom setu podataka koristeći podjelu seta podataka na set za treniranje i set za testiranje uz evaluacijske metrike R2, MAE i RMSE te su MLP, DT i KNN postigli dobre rezultate. Implementirana je peterostruka unakrsna validacija i nasumično pretraživanje hiperametara kako bi se poboljšala izvedba izdvojenih algoritama.
Zbog dodatnog poboljšanja implementirane su ansambl tehnike poput ET, Bagging i voting ansambla, te su postignuti odlični rezultati. Verzije voting ansambla koristile su četiri do osam modela te je model evaluiran srednjom vrijednosti metrika prilikom unakrsne validacije na setu za treniranje i testu na setu za test te su postignuti izvrsni rezultati. |
| Abstract (english) | In this paper, the significance of accurate rotor temperature estimation for PMSM in automotive applications is described, due to its impact on machine efficiency, performance, and durability. A literature review of PMSM temperature estimation was conducted, proposing the use of artificial intelligence, specifically advanced machine learning algorithms. The Python programming language and libraries such as Numpy, Pandas, Matplotlib, and Scikit-Learn are utilized in the research. A publicly available dataset containing measured PMSM variables is analysed, involving data preprocessing, correlation analysis, dataset scaling, and the handling of outliers.
Initial investigations were performed using MLP. Subsequently, various machine learning algorithms with default settings were tested on the pre-processed dataset, using a split into training and testing sets. The evaluation metrics R2, MAE, and RMSE were employed, where MLP, DT, and KNN yield favourable results. To enhance the performance of selected algorithms, five-fold cross-validation and random hyperparameter tuning are implemented.
Additional improvements are achieved by employing ensemble techniques such as ET, Bagging, and a Voting ensemble. Excellent results are obtained, with the Voting ensemble employing four to eight models and being evaluated based on the mean values of metrics during cross-validation and testing. This approach results in outstanding performance. |
