Efficient and Reliable Predictive Maintenance in Trains based on BiLSTM Model

Overview

This project implements a predictive maintenance system to estimate the Remaining Useful Life (RUL) of train components, specifically air compressors, by combining the MetroPT dataset from Zenodo and the MetroPT-3 dataset from Kaggle. It compares traditional machine learning models (Linear Regression, Random Forest, SVR, KNN) with a Bidirectional Long Short-Term Memory (BiLSTM) neural network. SHAP (SHapley Additive exPlanations) is used to interpret model predictions, focusing on the BiLSTM model for enhanced explainability.

Project Structure

Efficient-and-Reliable-Predictive-Maintenance-in-Trains-based-on-BiLSTM-Model/
├── config/
│   └── config.yaml                # Configuration file for paths and parameters
│
├── data/
│   ├── raw/                       # Raw datasets (MetroPT.csv, MetroPT3(AirCompressor).csv)
│   ├── predictions/               # Stored model predictions
│   └── processed/                 # Processed data (X_train, X_test, y_train, y_test)
│
├── models/                        # Store trained models (pickle, HDF5)
│
├── notebooks/
│   ├── 01_data_exploration.ipynb        # Exploratory Data Analysis (EDA)
│   ├── 02_prepare_data.ipynb            # Data preprocessing and dataset combination
│   ├── 03_train_evaluate_models.ipynb   # BiLSTM and baseline models training and evaluation
│   ├── 04_shap_analysis.ipynb           # SHAP analysis (for BiLSTM only)
│
├── scripts/
│   ├── data_preprocessing.py      # Preprocessing functions
│   ├── evaluate_models.py         # Evaluation functions
│   ├── utils.py                   # Utility functions for plotting
│   └── model_training.py          # Functions for training different models
│
├── requirements.txt               # Project dependencies
└── README.md                      # Project documentation

Datasets

This project combines two datasets for robust predictive maintenance:

1. MetroPT

• Source: Download from Zenodo and save as data/raw/MetroPT.csv.
• Features: Analog sensors (pressure, temperature, current), digital signals, GPS (latitude, longitude, speed).
• Citation:

  Veloso, B., Ribeiro, R.P., Gama, J. et al. The MetroPT dataset for predictive maintenance. Sci Data 9, 764 (2022). https://doi.org/10.1038/s41597-022-01877-3

2. MetroPT-3

• Source: Download from Kaggle and save as data/raw/MetroPT3(AirCompressor).csv.
• Features: Analog sensors (pressure, temperature, motor current), digital signals (air intake valves).
• Citation:

  N. Davari, B. Veloso, R. P. Ribeiro, P. M. Pereira and J. Gama, "Predictive maintenance based on anomaly detection using deep learning for air production unit in the railway industry," 2021 IEEE 8th International Conference on Data Science and Advanced Analytics (DSAA), Porto, Portugal, 2021, pp. 1-10, doi: 10.1109/DSAA53316.2021.9564181.

Requirements

• Python 3.7 (required for tensorflow==1.14.0)
• Anaconda (optional, recommended for Conda environment management)
• Git (required to clone the repository)

Dependencies listed in requirements.txt:

tensorflow==1.14.0
shap==0.42.1
pandas>=1.5.0
numpy>=1.23.0
scikit-learn>=1.2.0
matplotlib>=3.5.0
pyyaml>=6.0
joblib>=1.2.0

Installation

1. Clone the repository:

   git clone https://github.com/YahiaouiLydia/Efficient-and-Reliable-Predictive-Maintenance-in-Trains-based-on-BiLSTM-Model.git
   cd Efficient-and-Reliable-Predictive-Maintenance-in-Trains-based-on-BiLSTM-Model

2. (Optional) Create a virtual environment:

   • Using Conda (recommended for managing Python 3.7):

     conda create -n pyota_env python=3.7
     conda activate pyota_env

   • Using Python venv:

     python -m venv pyota_env
     source pyota_env/bin/activate  # On Windows: pyota_env\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

Usage

This project is structured into four steps, each represented by a notebook. Execute all cells in each notebook sequentially to avoid runtime errors. Notebooks can be run via terminal or manually through Jupyter Notebook, JupyterLab, or an IDE like VS Code.

⚠️ Important: Ensure datasets are downloaded and config.yaml is configured before running notebooks.

Step 1 – Exploratory Data Analysis (EDA)

• Purpose:
  • Understand the structure of MetroPT and MetroPT-3 datasets.
  • Visualize sensor data distributions.
  • Identify anomalies or missing values.
  • Explore correlations between variables.
• How to Run:

  jupyter notebook notebooks/01_data_exploration.ipynb

  Alternatively, open Jupyter Notebook, JupyterLab, or an IDE (e.g., VS Code) and navigate to notebooks/01_data_exploration.ipynb.
• Output: Visualizations and statistics (e.g., histograms, correlation matrices).

Step 2 – Data Preprocessing

• Purpose:
  • Load and combine MetroPT.csv and MetroPT3(AirCompressor).csv.
  • Standardize column names and units.
  • Merge datasets based on timestamps.
  • Compute RUL for each component.
  • Save preprocessed datasets for model training.
• How to Run:

  jupyter notebook notebooks/02_prepare_data.ipynb

  Alternatively, open Jupyter Notebook, JupyterLab, or an IDE and navigate to notebooks/02_prepare_data.ipynb.
• Output:
  • Baseline models: data/processed/x_train_baseline.csv, y_train_baseline.csv, x_test_baseline.csv, y_test_baseline.csv.
  • BiLSTM: data/processed/X_train_bilstm.joblib, y_train_bilstm.joblib, X_test_bilstm.joblib, y_test_bilstm.joblib, scaler_x_bilstm.joblib.

Step 3 – Model Training and Evaluation

• Purpose:
  • Train baseline models: Linear Regression, Random Forest, SVR, KNN, Gradient Boosting.
  • Train a BiLSTM model on time-series data.
  • Evaluate models using MAE, RMSE, and R².
  • Save trained models and results.
• How to Run:

  jupyter notebook notebooks/03_train_evaluate_models.ipynb

  Alternatively, open Jupyter Notebook, JupyterLab, or an IDE and navigate to notebooks/03_train_evaluate_models.ipynb.
• Output:
  • Trained models: models/bilstm_model.h5, models/random_forest_model.pkl, etc.
  • Predictions: data/predictions/ (model-specific files).
  • Plots (e.g., prediction comparisons).

Step 4 – SHAP Analysis (Explainability)

• Purpose:
  • Use SHAP to explain BiLSTM predictions.
  • Visualize global and local feature impacts.
  • Enhance model interpretability.
• How to Run:

  jupyter notebook notebooks/04_shap_analysis.ipynb

  Alternatively, open Jupyter Notebook, JupyterLab, or an IDE and navigate to notebooks/04_shap_analysis.ipynb.
• Output:
  • SHAP summary plots.
  • SHAP dependence plots.

Alternative: Run Scripts

For advanced users, scripts in scripts/ provide modular functions:

• data_preprocessing.py: Preprocessing functions.
• evaluate_models.py: Evaluation metrics.
• utils.py: Plotting utilities.
• model_training.py: Model training functions.

Configuration

The config.yaml file (config/config.yaml) defines paths and parameters. Example:

paths:
  raw:
    metropt: data/raw/MetroPT.csv
    metropt3: data/raw/MetroPT3(AirCompressor).csv
  processed:
    x_train_baseline: data/processed/x_train_baseline.csv
    y_train_baseline: data/processed/y_train_baseline.csv
    X_test_bilstm: data/processed/X_test_bilstm.joblib
    scaler_x_bilstm: data/processed/scaler_x_bilstm.joblib
  models:
    bilstm_model: models/bilstm_model.h5
bilstm:
  test_size: 0.3
  random_state: 42

Troubleshooting

1. FileNotFoundError: Missing Raw Data Files

   • Ensure data/raw/MetroPT.csv and data/raw/MetroPT3(AirCompressor).csv exist in data/raw/.
   • Download from Zenodo and Kaggle.

2. Dataset Mismatch: Feature Alignment

   • Check feature alignment in 02_prepare_data.ipynb (e.g., consistent sensor names, units).
   • Verify timestamps and sensor compatibility.

3. Dependency Issues: Python Version and Package Installation

   • Use Python 3.7 for compatibility with tensorflow==1.14.0.
   • Reinstall dependencies:

     pip install -r requirements.txt

   • Check TensorFlow version:

     pip show tensorflow

4. SHAP Errors: Explanation Model Issues

   • If shap.DeepExplainer fails, try shap.KernelExplainer.
   • Ensure tensorflow==1.14.0 is installed.

5. Output Verification: Check Generated Files

   • Verify:
     • data/processed/: Processed datasets.
     • models/: Trained models.
     • data/predictions/: Model predictions.

Contributing

This project is part of a doctoral research effort on predictive maintenance in trains. Contributions or suggestions are welcome via GitHub issues: Issues.