Improving Battery Thermal Management Using a Model-Based Virtual Temperature Sensor

Overview

This repository contains the implementation of a PCNN-LSTM model for virtual temperature sensing of lithium-ion batteries. The model is trained using the NASA battery dataset and deployed on an ESP32-S3 microcontroller after post-training quantization (PTQ). The primary goal of this work is to optimize battery thermal management by predicting battery surface temperature using voltage and current readings.

Features

• Hybrid PCNN-LSTM Model: Combines CNN for feature extraction and LSTM for sequential modeling.
• Dataset Handling: Processes NASA battery dataset for training and testing.
• Quantization for Edge Deployment: Applies full integer (INT8) quantization for microcontroller execution.
• Deployment on ESP32-S3:
  • Memory-efficient model conversion using TensorFlow Lite.
  • Integration with microcontroller hardware for real-time inference.
• Comparison Metrics:
  • Mean Absolute Error (MAE)
  • Root Mean Squared Error (RMSE)
  • Evaluation of quantization impact on accuracy.

---

Implementation Details

Dataset

• Source: NASA Battery Aging Dataset
• Selected Batteries:
  • Training: Batteries 30, 31, 32
  • Testing: Battery 29
• Features Used:
  • Measured voltage
  • Measured current
  • Measured temperature
  • Computed mean voltage and current
  • Integrated current over time

Data Preprocessing

1. MATLAB to CSV Conversion:
   • Extracts raw MATLAB (.mat) files.
   • Converts necessary features into CSV format.
2. Data Cleaning & Normalization:
   • Standardized using Z-score normalization: [ X_{\text{norm}} = \frac{X - \mu}{\sigma} ]
   • Missing values handled and sequences generated for time-series modeling.
3. Dataset Preparation:
   • Concatenation of relevant charge/discharge cycles.
   • Removal of impedance measurement cycles.
   • Saving processed data in .npy format for efficient model training.

---

Model Architecture

The PCNN-LSTM model follows this structure:

1. 1D Convolutional Layer (CNN):
   • Extracts local features from the time-series input.
2. LSTM Layers:
   • Captures temporal dependencies in battery signals.
3. Fully Connected Layers:
   • Outputs predicted battery surface temperature.

---

Training Process

• Framework: TensorFlow 2.18.0
• Batch Size: 32
• Epochs Tested: 10, 20, 30, 40
• Optimizer: Adam
• Loss Function: Mean Squared Error (MSE)
• Early Stopping: Applied for optimal convergence

---

Quantization & Edge Deployment

1. TensorFlow Lite Conversion

• Model converted from .keras format to TensorFlow Lite (.tflite).
• Applied full integer quantization (INT8) for edge efficiency.

2. Model Deployment on ESP32-S3

• Board Used: ESP32-S3-N16R8
• Specifications:
  • Flash Memory: 16 MB
  • RAM: 512 KB (SRAM) + 16 MB (PSRAM)
  • Clock Speed: 240 MHz
• Conversion to C Format:
  • xxd tool used to convert .tflite to C array (model.h).
  • Model stored in Flash Memory (PROGMEM) to optimize RAM usage.
• Integration with Arduino IDE:
  • Used EdgeNeuron library for inference on ESP32.

---

Performance Evaluation

Error Metrics (Before & After Quantization)

Epochs	Model	MAE (°C)	RMSE (°C)
10	TensorFlow	0.64	1.61
	TFLite on ESP	0.78	1.08
20	TensorFlow	1.60	2.61
	TFLite on ESP	0.63	0.94
30	TensorFlow	1.43	2.91
	TFLite on ESP	2.73	3.43
40	TensorFlow	1.42	2.49
	TFLite on ESP	0.79	1.41

• Key Findings:
  • Quantization improves accuracy in some cases due to implicit regularization.
  • Reduced RMSE in epochs 10, 20, and 40, showing better generalization.
  • Some accuracy loss at 30 epochs, indicating dependency on training dynamics.

---