Search

V sklopu pričujočega dela izdelamo simulacijski model generičnega električnega vozila, s katerim virtualno vozimo po naključno kreiranih cestah z različnimi na- klonskimi profili in hitrostmi vozila. Nato podatke o vožnji uporabimo za strojno učenje učnih modelov, kjer z računanjem optimiziramo izkoristek vožnje za po- ljubno pot. Delo je sestavljeno iz 7 sklopov. V prvem sklopu opišemo vozno- merilne cikle kot sta NEDC in WLTP ter kreiranje naključnih hitrostnih profilov, ki so hibrid med NEDC in WLTP. V drugem sklopu opišemo ravnovesje sil, ki vplivajo na osebno vozilo, tisto pogonsko ter ostale, ki delujejo v nasprotno smer. V tretjem sklopu opišemo model baterije ter njegovo implementacijo v Simulinku. Tukaj vključimo nekaj kronološko pomembnih podatkov ter uporabljene tehno- logije za izdelavo baterij. Najbolj obsežno poglavje oziroma sklop je četrti, kjer predstavimo sinhronski motor, ki je v večji meri predstavljen z vpoglednimi tabe- lami, ki so bile predhodno izračunane z metodo končnih elementov in definirajo lastnosti pogonskega motorja v posameznih obratovalnih točkah. Vključimo tudi ustrezne grafe simulacijskih rezultatov. V petem sklopu opišemo namen regulacije ali krmiljenja motorja ter PI regulator, ki je uporabljen v našem modelu vozila za regulacijo navora, kot vhod pa dobi dejansko hitrost in referenčno hitrost vozila. V šestem sklopu opišemo algoritme strojnega učenja in kreirane učne modele, kot so nevronska mreža, metoda podpornih vektorjev in k-najbližjih sosedov. V za- dnjem, sedmem sklopu analiziramo in ovrednotimo rezultate ter opišemo možno nadaljnje delo. As part of the present work, we create a simulation model of a generic electric vehicle with which we virtually drive on randomly created roads, with different inclination profiles and vehicle speeds. Later, we use the data of the ride itself to train learning models that optimize the calculation of route efficiency for any route. The work consists of 7 chapters. In the first part, we describe driving- measuring cycles such as NEDC and WLTP and how to create a random speed profile for the needs of the master’s thesis, which profile is a hybrid between NEDC and WLTP. In the second chapter, we describe the balance of forces on the road, ie the forces that affect the personal vehicle, the propulsion, and others that act in the opposite direction. In the third part, we describe the battery model and its implementation in Simulink. Here we include some chronologically important data and the technologies used to make the batteries. The most extensive chapter or section is the fourth, where we present a synchronous motor, which is largely represented by look-up tables. We also include the corresponding graphs of the simulation results. In the fifth chapter, we describe the purpose of control, and the PI regulator used in our vehicle model for torque control, where the input of the controller gets the actual speed and reference speed. In the sixth chapter, we describe machine learning algorithms and create learning models such as neural network, support vector machine, and k-nearest neighbors. In the last seventh chapter, we analyze the results and describe further work.

Srcna avskultacija je najstarejsa neinvazivna metoda za odkrivanje bolezni srcnih zaklopk. V diplomi smo se osredotocili na analizo fonokardiogramov s pomocjo metod digitalnega procesiranja signalov ter metod umetne inte- ligence za klasikacijo. Signal pridobljen iz elektronskega stetoskopa smo razdelili v segmente, kjer en segment ustreza enemu kardialnemu ciklu. Nad segmentom smo izracunali MFCC znacilke, katere smo uporabili kot vhod algoritmom strojnega ucenja v programskem sistemu Orange. Ciljni kla- sikacijski razred je stanje pacienta. Locevali smo med fonokardiogrami s prisotnim sumom in brez. Najboljso klasikacijsko tocnost smo dosegli z na- ivnim Bayesovim klasikatorjem. Dosegli smo 92,4 % tocnost pri pricakovani tocnosti vecinskega klasikatorja 75,2 %. Dosegli smo 76,9 % senzitivnost in 95,4 % specicnost, prav tako z naivnim Bayesovim klasikatorjem. Po nasem mnenju bi lahko bil tak sistem pomoc zdravnikom pri diagnosticiranju bolezni srcnih zaklopk. V teoreticnem delu diplome smo na kratko opisali algoritme, katere smo uporabljali ter kaksne so naravne omejitve pri delu s procesiranjem signalov. Heart auscultation is one of the oldest non-invasive method for detection valvular heart disease. In thesis we have focused on the analysis of phono- cardiograms with digital signal processing methods and methods of articial intelligence for classication. We have divided the signal obtained from elec- tronic stethoscope in to segments where one of the segment corresponds to one cardiac cycle. After that we calculate MFCC features on one of the segments. The features serve as an input to machine learning algorithms in system Orange. Target classication class is the condition of a patient. We have distinguish the phonocardiograms with a heart murmur and without it. The best classication accuracy that we achieved is with naive Bayes classicator of 92.4 %. The expected accuracy of majority class was 75.2 %. The best achieved sensitivity and specicity was 76.9 % and 95.4 % respec- tively, also with naive Bayes classicator. In our opinion such system could be used by physicians to help diagnose heart valve diseases. In theoretical part of the thesis we have described algorithms that we used and what are the limitations of processing a signal.

Heart auscultation is one of the oldest non-invasive method for detection valvular heart disease. In thesis we have focused on the analysis of phonocardiograms with digital signal processing methods and methods of artificial intelligence for classification. We have divided the signal obtained from electronic stethoscope in to segments where one of the segment corresponds to one cardiac cycle. After that we calculate MFCC features on one of the segments. The features serve as an input to machine learning algorithms in system Orange. Target classification class is the condition of a patient. We have distinguish the phonocardiograms with a heart murmur and without it. The best classification accuracy that we achieved is with naive Bayes classificator of 92.4 %. The expected accuracy of majority class was 75.2 %. The best achieved sensitivity and specificity was 76.9 % and 95.4 % respectively, also with naive Bayes classificator. In our opinion such system could be used by physicians to help diagnose heart valve diseases. In theoretical part of the thesis we have described algorithms that we used and what are the limitations of processing a signal.