Merge pull request #969 from spandana2004/main

Denoising sine wave signal

Denoising Sine wave signal/Dataset/README.md

@@ -0,0 +1,25 @@
+# Dataset for Signal Denoising Autoencoder
+
+This dataset contains synthetic or real-world signals with added noise, used for training and testing a denoising autoencoder. It includes columns representing the original amplitude of the signal and the noisy version of the signal.
+
+## Dataset Structure
+
+The dataset file (`ex1.xlsx`) includes the following columns:
+- **Amplitude**: The original, clean signal amplitude.
+- **Amplitude_noise**: The amplitude of the signal after noise has been added.
+
+The data is sequential and each entry represents one time step.
+
+### Loading the Dataset
+To load the dataset in Python, use the following code:
+
+```bash
+import pandas as pd
+
+data = pd.read_excel('ex1.xlsx')
+original_amplitude = data['Amplitude'].values
+noised_amplitude = data['Amplitude_noise'].values
+```
+
+### Usage
+This dataset is used to train and evaluate denoising models, which will learn to filter noise from signals, providing a cleaned signal that approximates the original.

Denoising Sine wave signal/Model/README.md

@@ -0,0 +1,119 @@
+# Sine Wave Signal Denoising
+
+## Project Overview
+
+This project demonstrates how to denoise a noisy signal using deep learning models, primarily through a Convolutional Autoencoder, LSTM Autoencoder, and an enhanced Conv1D Autoencoder. Each model is designed to take a noisy signal as input and reconstruct the original (denoised) signal. After denoising, the Savitzky-Golay filter is applied to further smooth the output.
+
+### Methodology
+
+1. **Preprocessing**: 
+   - The signal is normalized using `MinMaxScaler`.
+   - Both noisy and original signals are reshaped into sliding windows for training.
+
+2. **Model Architectures**:
+   - **Conv1D Autoencoder**: A Convolutional Autoencoder built to map noisy signals to clean ones, using convolutional and pooling layers for feature extraction and dimensionality reduction.
+   - **LSTM Autoencoder**: A recurrent model utilizing LSTM layers to capture temporal dependencies, effective for time-series data.
+   - **Enhanced Conv1D Autoencoder**: An improved version of the Conv1D Autoencoder with deeper layers for better feature extraction and reconstruction capabilities.
+
+3. **Postprocessing**:
+   - The Savitzky-Golay filter is applied to the output to further reduce residual noise.
+
+4. **Visualization**:
+   - The original, noisy, and denoised signals are plotted for a comprehensive visual comparison.
+
+### Dependencies
+
+- Python 3.x
+- `pandas`
+- `numpy`
+- `scikit-learn`
+- `tensorflow`
+- `scipy`
+- `matplotlib`
+
+### Installation
+
+1. Clone the repository:
+   ```bash
+   git clone https://github.com/yourusername/signal denoising autoencoder.git
+   ```
+
+2. Install dependencies:
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+### Usage
+
+1. Prepare your dataset in an Excel file (`ex1.xlsx`) with two columns: `Amplitude` and `Amplitude_noise`.
+
+2. Run the denoising script:
+   ```bash
+   python denoise_signal.py
+   ```
+
+3. The script will train the autoencoders on the noisy signal, output the denoised signals, and plot them alongside the original and noisy signals.
+
+### Model Details
+
+#### Conv1D Autoencoder
+
+The model architecture:
+```
+Layer (type)                    Output Shape         Param #
+=================================================================
+input_signal (InputLayer)        [(None, 4500, 1)]    0
+_________________________________________________________________
+conv1d (Conv1D)                  (None, 4500, 16)     64
+_________________________________________________________________
+max_pooling1d (MaxPooling1D)     (None, 2250, 16)     0
+_________________________________________________________________
+conv1d_1 (Conv1D)                (None, 2250, 8)      392
+_________________________________________________________________
+max_pooling1d_1 (MaxPooling1D)   (None, 1125, 8)      0
+_________________________________________________________________
+conv1d_2 (Conv1D)                (None, 1125, 8)      200
+_________________________________________________________________
+up_sampling1d (UpSampling1D)     (None, 2250, 8)      0
+_________________________________________________________________
+conv1d_3 (Conv1D)                (None, 2250, 16)     400
+_________________________________________________________________
+up_sampling1d_1 (UpSampling1D)   (None, 4500, 16)     0
+_________________________________________________________________
+conv1d_4 (Conv1D)                (None, 4500, 1)      49
+=================================================================
+Total params: 1,105
+Trainable params: 1,105
+Non-trainable params: 0
+```
+
+#### LSTM Autoencoder
+
+The LSTM Autoencoder captures sequential dependencies, which is particularly effective for time-series data. This model consists of:
+- A stack of LSTM layers for encoding the sequence.
+- An intermediate dense layer for feature extraction.
+- LSTM layers for decoding the signal back to its denoised form.
+
+The model is effective for denoising tasks with complex temporal dependencies and uses `Mean Squared Error (MSE)` as the loss function.
+
+#### Enhanced Conv1D Autoencoder
+
+An advanced convolutional model that expands upon the basic Conv1D Autoencoder:
+- Deeper architecture with additional convolutional layers for enhanced feature extraction.
+- Additional UpSampling and convolutional layers for improved reconstruction capabilities.
+- MaxPooling and UpSampling for dimensionality adjustments and noise removal, offering better denoising for high-frequency noise.
+
+### Results
+
+After training each model, the denoised signals are plotted and visually compared with the original and noisy signals. Each model significantly reduces noise, preserving the main structure of the original signal.
+
+### Comparison of Model Performance
+
+- **Conv1D Autoencoder**: Ideal for signals with moderate noise levels; balances simplicity and accuracy.
+- **LSTM Autoencoder**: Better suited for sequential data, capturing dependencies over longer periods.
+- **Enhanced Conv1D Autoencoder**: Offers the best denoising performance among the models due to its deeper architecture and robust feature extraction capabilities.
+
+### Conclusion
+
+This project demonstrates how deep learning models, particularly convolutional and recurrent autoencoders, can be used for signal denoising. Further improvements can be made by experimenting with model architectures, window sizes, and smoothing techniques to enhance the denoising process.
+