Internship-IIT-Roorkee

Real Time Anomaly Detection in DC-DC Boost Converter Signal

Project Overview

This project implements real-time anomaly detection for signals from a DC-DC Boost Converter (and DC Motor) using deep learning models. The workflow includes data collection from STM32, model training on Kaggle, deployment to Raspberry Pi, and real-time inference with data logging.

---

1. Data Collection from STM32

• Hardware Used: STM32 Nucleo-F303ZE
• Signal Acquisition:
  • ADC Input: PF4 (ADC1_IN5) is used to read the analog signal.
  • UART Output: PC4 (USART2_TX) and PC5 (USART2_RX) are used for UART communication with the Raspberry Pi.
• Firmware: The STM32 firmware reads ADC values, applies a digital filter, and transmits filtered values over UART at 115200 baud.

Pin Connections:

• STM32 PF4: Connect to the analog signal source (e.g., output of Boost Converter).
• STM32 PC4 (TX): Connect to Raspberry Pi RX (GPIO15, physical pin 10).
• STM32 PC5 (RX): Connect to Raspberry Pi TX (GPIO14, physical pin 8).
• Common GND between STM32 and Raspberry Pi.

---

2. Uploading Data to Kaggle

1. Collect Data:

   • Use the Raspberry Pi to receive UART data and log it as CSV files (see Raspberry Pi/uart.py).
   • Example files: normal.csv, random.csv, load.csv in Dataset/DC Motor/ or Dataset/Boost Converter/.

2. Upload to Kaggle:

   • Go to Kaggle Datasets.
   • Click "New Dataset" and upload your CSV files.
   • Fill in the dataset details and make it public or private as needed.

---

3. Model Training on Kaggle

1. Open the Notebook:

   • Use Kaggle/model-training.ipynb as your starting point.

2. Load the Dataset:

   • Use Kaggle's data path, e.g.:

     import pandas as pd
     normal = pd.read_csv('/kaggle/input/your-dataset/normal.csv')
     load = pd.read_csv('/kaggle/input/your-dataset/load.csv')
     random = pd.read_csv('/kaggle/input/your-dataset/random.csv')

3. Preprocess and Generate Training Data:

   • The notebook provides functions to segment, label, and augment the data.

4. Activate GPU:

   • In the Kaggle notebook, go to Settings (right sidebar) and set "Accelerator" to "GPU".

5. Train the Model:

   • The notebook includes code for training several models (ResNet, CNN, VGG, U-Net, MLP).
   • Training uses PyTorch and scikit-learn.

6. Save the Model and Scaler:

   • After training, save the model and scaler:

     import torch
     import joblib
     traced_model = torch.jit.trace(model, dummy_input)
     traced_model.save('model.pt')
     joblib.dump(scaler, 'scaler.pkl')

   • Download these files from the Kaggle notebook output.

---

4. Deploying to Raspberry Pi

1. Setup Raspberry Pi:

   • Install dependencies:

     pip install -r Raspberry\ Pi/requirements.txt

   • Ensure pyserial, torch, joblib, matplotlib, tk, etc. are installed.

2. Copy Model Files:

   • Create a new folder in Raspberry Pi/ (e.g., model_1/).
   • Add your model.pt and scaler.pkl to this folder.

3. Update Model Paths:

   • Edit Raspberry Pi/inference.py and update MODEL_SCALER_MAP with your new model folder and file names.

---

5. Running Real-Time Inference

• Run the main application:

  python Raspberry\ Pi/app.py

• The app:
  • Reads UART data from STM32.
  • Applies the trained model for anomaly detection.
  • Displays real-time plots and predictions.
  • Records all incoming data and predictions to CSV logs.
• Retraining: If new data is collected, you can upload it to Kaggle and repeat the training process.

---

6. Notes on Data Logging and Retraining

• Every run of the inference app logs both the raw voltage and the model's predictions.
• These logs can be used to further improve or retrain your models.

---

7. Hardware Pin Summary

Function	STM32 Pin	STM32 Peripheral	Raspberry Pi Pin
ADC Input	PF4	ADC1_IN5	Analog Signal
UART TX (to Pi)	PC4	USART2_TX	Pin:8 (RX)
UART RX (from Pi)	PC5	USART2_RX	Pin:10 (TX)
GND	GND	-	GND

---

8. Troubleshooting

• Serial Port: Ensure /dev/serial0 is enabled on Raspberry Pi (raspi-config > Interface Options > Serial).
• Baud Rate: Both STM32 and Pi must use 115200 baud.
• Permissions: You may need to add your user to the dialout group on the Pi for serial access.

---

9. References

• STM32 Nucleo-F303ZE Datasheet
• Kaggle Documentation
• PyTorch Documentation