Recognition of low frequency disturbances by wavelet transform and neural network

U ovome radu promatrani su propadi napona, prenaponi, oscilatorne prijelazne pojave, harmonička izobličenja i naponski urezi. Za svaku kategoriju umjetno je generirano 200 signala. Nad ovim signalima provela se diskretna valična transformacija koja omogućava da se pri različitim frekvencijama koristi različita vremensko-frekvencijska rezolucija. Pri dekompoziciji signala koristili su se db1, db4 i db6 valići. Dekompozicija signala napravljena je u 5 nivoa čime se dobiva jedan koeficijent aproksimacije i 5 koeficijenta detalja. Za svaki koeficijent računaju se tri statistička parametra: standardna devijacija, entropija i asimetrija signala čime se dobiva ukupno 18 varijabli koje predstavljaju jedan signal. Za prepoznavanje navedenih smetnji koristila se neuronska mreža s 18 ulaznih neurona, 3 skrivena i 6 izlaznih neurona. Klasifikacijski modeli s različitim valićima analizirani su na temelju točnosti, matrica konfuzije i drugih parametara. Analizom je ustanovljeno da model s db6 valićem pruža najbolje prepoznavanje navedenih smetnji.In this thesis voltage sag, swell, oscillatory transient, harmonic distortion and voltage notches were examined. For each category 200 signals were artificially generated. On those signals discrete wavelet transform was carried out which for different frequencies enables different time-frequency resolution. For signal decomposition db1, db4 and db6 wavelets were used. Signal decomposition was made in 5 levels which provides one approximation coefficient and 5 detail coefficients. For each coefficient three statistical parameters were calculated: standard deviation, entropy and skewness, thus obtaining 18 variables which represents one signal. For disturbance recognition neural network with 18 input neurons, 3 hidden and 6 output neurons was used. Classification models with different wavelets were analyzed based on accuracy, confusion matrix and other parameters. Analysis founded that model with db6 wavelet provides best recognition of mentioned disturbances