🔋 AI Battery Management System (AI BMS)

---

🚀 Overview

The AI Battery Management System (AI BMS) is an advanced solution for real-time monitoring, state prediction, and optimization of battery performance. Using machine learning models, it estimates key parameters like SOC (State of Charge), SOH (State of Health), and temperature. Integrating with edge platforms for fast data processing and supporting Over-The-Air (OTA) updates for remote learning, the AI BMS is designed to manage battery packs efficiently, enhance safety, and extend lifecycle for applications such as electric vehicles and energy storage systems.

---

🖼️ Prototype Interface

Here is a screenshot of the AI BMS Prototype Interface showcasing various monitoring, control, and alert features:

---

🔄 Pipeline Description

1. 📊 Raw Data Collection:

   • Battery sensors gather real-time data, including voltage, current, temperature, and other parameters, forming the foundation for accurate model predictions.

2. ⚙️ Data Pre-processing:

   • Collected data is cleaned, resampled, and normalized to ensure accuracy in machine learning inputs, making it suitable for model training.

3. 🧠 Training ML Models:

   • Machine learning models such as LSTM and FNN are trained on historical data to estimate SOC, SOH, and temperature, enabling predictive insights and monitoring.

4. ✅ Validation Step:

   • Model validation fine-tunes hyperparameters to improve performance and ensure the model’s reliability.

5. 🌐 Deployment on Edge and Cloud:

   • Models are deployed on edge devices for real-time SOC predictions and cloud platforms for remote monitoring, balancing low latency with flexible remote processing.

6. 📡 OTA Updates:

   • OTA updates enable the model to adapt based on new data, ensuring the AI BMS remains accurate and up-to-date with continuous learning.

7. 📈 Prediction and Results:

   • The AI BMS delivers accurate SOC predictions under various temperature conditions, crucial for real-time battery health management.

8. 📊 Visualization and Analysis:

   • Visualization tools display SOC predictions, battery temperatures, and other metrics, providing insights for efficient monitoring and decision-making.

---

🌟 Features

• 🔋 High-Precision SOC Estimation: Predicts state of charge with high accuracy, reducing risks of unexpected power loss.
• 🩺 SOH, SOE, and SOI Estimation: Provides machine learning-based predictions for battery health, energy, and imbalance states.
• 🌡️ Temperature Prediction: Real-time temperature estimation prevents overheating.
• 💧 Immersion-Cooled Pack Management: Integrated with cooling systems to ensure safe, stable battery operation.
• 📡 Remote Learning & OTA Updates: ECM models enable continuous battery performance improvements with over-the-air parameter updates.
• ⚡ Charging Algorithm Optimization: Optimizes charging cycles to enhance battery longevity and performance.
• 🌐 Edge & Cloud Integration: Supports seamless integration with edge platforms for low-latency operations and cloud platforms for advanced simulations and monitoring.

📁 Project Structure

📂 Folders:

• data/:

  • raw/: Contains raw data fetched from AWS Timestream or other sources.
  • cleaned/: Stores cleaned and processed data ready for model training.

• scripts/:

  • aws_integration/:
    • AWS_timestream_to_python.py: Fetches data from AWS Timestream using boto3.
  • preprocessing/:
    • data_collection_preprocessing.py: Collects and preprocesses cloud data.
    • extract_combine_current_data.py: Extracts and combines current data for analysis.
    • resample_current_data.py: Resamples the current data.
    • rescale_current_data.py: Rescales the current data for consistency.
  • model_training/:
    • train_LSTM_Avg_temp_seasonal.py: Trains LSTM for average temperature with seasonal data.
    • train_FNN_winter_finetune.py: Fine-tunes FNN model with winter data.
    • train_LSTM_temp.py: Trains LSTM for temperature prediction.
    • train_LSTM_temperature_prediction.py: Another script for LSTM temperature prediction.
    • Re-train LSTM on simulated data.py: Retrains LSTM on simulated data.
  • simulations/:
    • PyBamm_simulation.py: Simulates battery behavior using the PyBaMM library.
    • octave_simulation_full.py: Full simulation using Octave.
  • evaluation/:
    • Test_model.py: Tests the saved models.
    • test_saved_model.py: Evaluates a specific saved model.
    • life_cycle_data_analysis.py: Plots dQ/dV life cycle curves for analysis.

• README.md: Project documentation.

• requirements.txt: Lists dependencies for running the project.

---

🛠️ Setup Instructions

🔧 Requirements

To run this project, you’ll need the following dependencies:

• Python 3.7+
• TensorFlow
• PyTorch (optional for model experiments)
• PyBaMM (for battery simulations)
• Pandas, NumPy, Matplotlib (for data processing and visualizations)
• boto3 (for AWS integration)

Install dependencies with:

pip install -r requirements.txt

🏃 Running the Models

🔄 Training

To train the SOC, SOH, and temperature prediction models, use the following command:

python scripts/Re-train\ LSTM\ on\ simulated\ data.py --model soc --input_path data/input_data.csv

📈 Visualizations

Notebooks for visualizing predictions and model performance are not explicitly provided in this repository, but you can create custom visualizations based on the provided scripts using libraries such as matplotlib and seaborn.

☁️ AWS Integration

The system is integrated with AWS services like Timestream and SageMaker for model deployment and data storage. The boto3 library is used for querying Timestream data.

Make sure your AWS credentials are configured using:

• aws configure

🚀 Usage Examples

🔋 SOC Prediction

To predict the State of Charge (SOC) of a battery pack using the pre-trained SOC model:

python scripts/inference.py --model soc --input_path data/test_data.csv

🌡️ Temperature Estimation

To estimate the temperature of battery cells based on real-time voltage, current, and environmental temperature data:

-python scripts/inference.py --model temp --input_path data/temperature_data.csv

🌐 Edge and Cloud Deployment

The AI BMS supports deployment on edge platforms like NXP MCUs and cloud platforms like AWS SageMaker. The OTA update feature allows for remote learning and parameter adjustments to continuously optimize performance.

📊 Results

Model performance metrics, graphs, and comparisons between different configurations are saved in the /results folder. Here you can find:

• SOC vs. Actual Graphs
• Temperature Prediction Accuracy
• Model Loss Curves (Training vs Validation)

---

🔍 SOC Prediction Analysis with Charge-Discharge Cycles at Different Temperatures

The plots below demonstrate State of Charge (SOC) predictions under various temperature conditions, showcasing how SOC behavior changes during charge-discharge cycles. The actual SOC values are represented by blue lines, while predicted SOC values are shown in red dashed lines.

🔑 Key Insights

1. 🌡️ Test Data at 0°C:

   • SOC cycles exhibit a regular pattern, but the charge-discharge rate is slightly slower compared to higher temperatures.
   • Predicted SOC closely follows the actual SOC, with only minor deviations observed, indicating strong model performance under cold conditions.

   

   • Analysis: The prediction model shows stable performance, tracking the actual SOC curve accurately even at low temperatures. This suggests that the model effectively captures battery behavior under cold conditions, crucial for applications where battery performance may be impacted by external cold environments.

2. 🌡️ Test Data at 10°C:

   • The SOC curve appears more dynamic, with slightly higher amplitude oscillations than at 0°C, reflecting improved battery performance and efficiency at this moderate temperature.
   • Predictions are highly accurate, aligning consistently with actual SOC values throughout the charge-discharge cycles.

   

   • Analysis: At 10°C, the model's predictive capability continues to excel, indicating enhanced battery response. The close alignment between predicted and actual SOC values shows the model's robustness and accuracy under moderate temperatures, illustrating its adaptability to temperature variations.

3. 🌡️ Test Data at 25°C:

   • SOC cycles are smoother with well-defined peaks and troughs, as the battery operates near its optimal temperature.
   • The predicted SOC closely matches the actual SOC, highlighting the model’s accuracy in ideal temperature conditions.

   

   • Analysis: The model demonstrates peak accuracy at 25°C, operating optimally within the ideal temperature range for battery performance. The close match between actual and predicted SOC values reinforces the model’s reliability under favorable temperature conditions.

---

📊 Temperature Prediction Results

The model's predictions for immersion-cooled battery pack temperature reveal several important trends and observations. Below is a detailed analysis of short-term, cyclic, and long-term predictions, along with insights into deviations.

---

🌡️ 1. Short-Term Predictions

• Observation: The plot reveals three significant points (e.g., July 26, July 30) where the predicted temperature drops noticeably compared to the actual values.
• Cause:
  • These drops are associated with sudden operational changes, likely due to abrupt variations in Pump Duty Cycle or Liquid Level, which may not have been adequately represented in the training data.
  • Interaction features such as MxTmp_PumpDutyCycle_Interaction may need further refinement to handle these edge cases more effectively.
• Result: While the model performs well overall, these isolated deviations suggest areas for improvement in feature engineering and data coverage.

---

🌡️ 2. Cyclic Behavior Across Operational Periods

• Observation: The model effectively captures the cyclic patterns observed during charge-discharge cycles. Predicted peaks and troughs align closely with actual values.
• Analysis: Features like SOC and SOH contribute significantly to the model's ability to predict temperature dynamics during cyclic battery operations.

---

🌡️ 3. Long-Term Trends

• Observation: Over an extended period, the model successfully tracks the general trends in battery temperature (MxTmp) influenced by environmental temperature (Tx), including gradual increases and decreases. However:

  • A noticeable gap exists in predictions around January 2024, where predicted temperatures deviate significantly from actual values.
  • In most other segments, the model demonstrates strong alignment with actual values, showcasing its ability to effectively capture the relationship between MxTmp and Tx over time.

• Analysis of Environmental Temperature (Tx):

  • Key Role: As Tx represents the environmental conditions, it strongly impacts the thermal behavior of the battery system. The interaction between MxTmp and Tx is critical for accurate long-term predictions.
  • Observed Impact:
    • Higher Tx values correlate with elevated MxTmp, reflecting increased ambient influence on battery temperature.
    • Sudden drops in Tx appear to create delayed or less pronounced changes in MxTmp, possibly due to the thermal inertia of the battery system.

• Cause of January Gap:

  • The gap in predictions around January 2024 may stem from:
    • Underrepresented operational conditions (specific Tx ranges or seasonal patterns) in the training data.
    • Insufficient interaction features, such as MxTmp_Tx_Interaction, to fully capture the dynamic influence of Tx on MxTmp.
    • Unmodeled external factors, such as sudden environmental changes, impacting the system.

• Result:

  • While the model generally performs well, addressing inconsistencies like the January gap by:
    • Incorporating additional training data for seasonal patterns.
    • Refining features to include more detailed interactions between MxTmp and Tx.
    • Exploring external factors influencing temperature trends.
  • These improvements will enhance the model’s long-term prediction reliability and robustness to varying Tx conditions.

📈 Plot 1: Full Long-Term Trends

📉 Plot 2: Focused View on January 2024

---

✨ Summary

1. Short-Term Deviations: Prediction drops at specific points highlight the need for enhanced interaction features and better representation of edge cases in the training data.
2. Cyclic Patterns: The model excels in predicting cyclic temperature behavior during charge-discharge cycles.
3. Long-Term Accuracy: The model effectively captures overall temperature trends, demonstrating its reliability for real-world thermal management applications.
4. Future Work:
   • Refine interaction features to better handle sudden operational changes.
   • Incorporate additional training data to improve performance under edge conditions.

---

✨ Summary

1. Accuracy Across Time Scales: The model demonstrates high accuracy in predicting short-term fluctuations, cyclic patterns, and long-term temperature trends.
2. Role of Features: Key features and interaction terms significantly enhance the model’s ability to predict temperature, highlighting the importance of battery operational parameters and cooling system dynamics.
3. Real-World Application: These results validate the model’s reliability for real-time thermal management in immersion-cooled battery packs, ensuring safety and operational efficiency.

---

📊 Deployment Considerations

The SOC and temperature prediction models were benchmarked for deployment on both Edge (NXP NPU) and Cloud platforms. Below are the estimated results for performance and resource utilization:

---

1. Edge Deployment on NXP NPU

Hardware:

• Processor: NXP i.MX 8M Plus (Quad-core Cortex-A53 with Neural Processing Unit - NPU).
• Target Framework: TensorFlow Lite.

Results (LSTM and FNN Models):

Metric	SOC Model (LSTM)	SOC Model (FNN)	Temp Model (LSTM)	Temp Model (FNN)
Model Size	1.5 MB (Quantized)	1.1 MB (Quantized)	2.8 MB (Quantized)	2.3 MB (Quantized)
Inference Time	~15 ms	~8 ms	~20 ms	~10 ms
Latency	<25 ms	<15 ms	<30 ms	<20 ms
Memory Usage	~20 MB	~12 MB	~30 MB	~18 MB
Accuracy (Test Data)	~98%	~95%	~96%	~93%

Insights:

• LSTM models, with 100 sequences, provide higher accuracy for both SOC and temperature predictions but require more memory and slightly higher latency due to their recurrent nature.
• FNN models are faster and more resource-efficient, making them better suited for highly constrained edge devices.

---

Edge Deployment Setup

Below is the NXP i.MX 8M Plus Evaluation Kit used for deploying the SOC and temperature prediction models:

• NXP i.MX 8M Plus Processor: Executes real-time inference for SOC and temperature predictions.
• Peripheral Connections: Interfaces with sensors capturing voltage, current, and other parameters.
• Deployment: Models are optimized and deployed on this hardware for efficient edge computing.

🤝 Contributing

Contributions to this project are welcome! Feel free to open an issue or submit a pull request if you'd like to improve the project or add new features.

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

👤 Author

Usama Yasir Khan
AI Engineer
XING Mobility