[Python] Problem with filtering frequencies and IFFT

I am new to signal procession, so I found myself stuck. I have a script that performs an FFT with zero-padding and attenuation Hann window on a waveform. Then, I need to filter out frequencies above a certain threshold and perform an IFFT to get a reconstructed waveform. I think I was able to reverse all my steps and get a reconstructed waveform identical to the original waveform. The problem started when I input a filtered frequency spectrum into my IFFT function. Here are some problems I encountered: In this, there were three frequencies present, none of which were intented to be filtered out, and the reconstructed waveform looked like it had some frequencies removed

In this image, I increased the sampling rate to what I will use later, and it resulted in a waveform that looked like this

Here's my code:

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# zero-padding function
def nextpow2(x):
return (x-1).bit_length()

# Hann window and FFT function
def preprocessing(waveform, duration):
NS = len(waveform) # length of waveform in samples
delF = 1/duration # stepsize in frequency domain
Fmax = NS * delF / 2 # maximum frequency possible in f-domain
NFFT = 2 ** (nextpow2(NS)+1) # Next power of 2 based on waveform length
attHann = sum(np.hanning(NS)) # Attenuation Hann window
y = (1/attHann) * np.fft.fft(np.hanning(NS) * waveform, NFFT) # spectrum of the input signal
mag0 = abs(y[0]) # magnitude of DC component
mag = abs(y[0:int(NFFT/2)])/mag0 # relative magnitude
f = Fmax * np.linspace(0, 1, int(NFFT/2)) # frequency spectrum vector
return f, mag, y, NFFT, Fmax

def inverse_processing(y, waveform):
NS = len(waveform) # length of waveform in samples
NFFT = 2 ** (nextpow2(NS)+1) # Next power of 2 based on waveform length
attHann = sum(np.hanning(NS)) # Attenuation Hanng window
y = attHann * y # reverse attenuation Hann window
waveform_windowed = np.fft.ifft(y, NFFT) # inverse FFT accounting for zero-padding
waveform_windowed = waveform_windowed[:NS] # remove zero-padding
reconstructed_waveform = waveform_windowed * (1 / np.hanning(NS)) # reverse Hann window
reconstructed_waveform = reconstructed_waveform.real
return reconstructed_waveform

# A simulated sin wave
# Parameters
fs = 186567 # Sampling frequency (Hz)
T = 1 # Duration in seconds
timevector = np.linspace(0, T, int(fs*T), endpoint=False)

# Frequencies to combine (Hz)
freqs = [50, 120, 300]

# Generate combined sine wave
waveform = np.zeros_like(timevector)
for f in freqs:
waveform += np.sin(2 * np.pi * f * timevector)

duration = timevector[-1]
Srate = len(waveform)/duration

# FFT, zero-padding, Hann window
f, mag, y, NFFT, Fmax = preprocessing(waveform, duration)

###############################################################################
# band-pass filter
y_filtered = y.copy()
y_filtered = y_filtered[:len(f)]

mask = (f >= 0) & (f <= 400)
y_filtered[~mask] = 0

###############################################################################

reconstructed_waveform = inverse_processing(y, waveform)
filtered_waveform = inverse_processing(y_filtered,waveform)
time_reconstructed = np.linspace(0, duration, len(reconstructed_waveform))


I tried doing my inverse process without the extra steps for zero-padding and Hann window.

Continue reading...