Digital Signal Processing and Deep Learning

---

This repository is a study of Digital Signal Processing techniques for Audio signals

1. DSP helps creation of features for ML

   • We look at creation of features for traditional ML - for example a SVM classifer
   • And, we look at creating features for Deep learning techniques such as MLP, CNN, RNN and LSTM

2. For traditional ML feature engineering, we look at Time domain, Frequency domain and time-frequency features

3. Time Domain features:

   • A-D-S-R model: Attack-Decay-Sustain-Release model for audio
   • Amplitude envelope (AE)
   • Root-mean-square Energy (RMS-E)
   • Zero-crossing rate (ZCR)

4. Frequency Domain features:

   • FFT - Fast Fourier Transform
   • STFT - Short Time Fourier Transform
   • MFCC - Mel Frequency Cepstral Coefficents
   • MFCC: MFCC feature extraction technique basically includes windowing the signal, applying the DFT, taking the log of the magnitude, and then warping the frequencies on a Mel scale, followed by applying the inverse DCT.

5. For deep learning, we use MFCC converted to images

6. We build a CNN using TensorFlow Keras

7. To-Do: Build RNN, LSTM for DSP

8. For additional technical notes please see: https://github.com/Rajesh-Siraskar/Digital_Signal_Processing_and_Deep_Learning/blob/main/README_DSP-Notes.ipynb

Project

Valve fault diagnotics based on sound files

• 20-May-2021
• Attempt to train to recognize faulty valves
• Unseen data will be from another folder