TEP Fault Diagnosis: Comparative Experiments of MLP, CNN, and Graph Neural Networks

This article introduces an open-source project that compares the fault diagnosis performance of Multilayer Perceptron (MLP), 1D Convolutional Neural Network (CNN), and Graph Neural Network (GNN) on the Tennessee Eastman Process (TEP) dataset. The project covers data preprocessing, model architecture design, and experimental result analysis. The key finding is that under the current implementation, GNN's performance is weaker than MLP and CNN, emphasizing the importance of a fair comparison framework. The project is maintained by faiazu, with source code available on GitHub (link: https://github.com/faiazu/TEP-fault-diagnosis), and was released on June 6, 2026.

The Tennessee Eastman Process (TEP) is a classic benchmark dataset in the field of process control, containing 21 fault types and 52 process variables, used to test monitoring and diagnostic algorithms. Fault diagnosis is crucial for industries such as chemical engineering, as it can avoid equipment damage and safety accidents, reduce downtime and maintenance costs. Deep learning applications in this field are increasing, but there is a lack of fair comparisons between different models. This project aims to fill this gap.

Sliding window is used to process time series: window size of 60 steps, step size of 10, and 52 variables (41 measured variables xmeas and 11 manipulated variables xmv). Two types of datasets are constructed: flattened windows (shape N×3120) for MLP (losing local temporal structure); 2D windows (shape N×60×52) for CNN and GNN (preserving temporal correlation). Data partitioning is based on simulation runs (instead of random), avoiding data leakage caused by adjacent windows from the same run appearing in both training/validation sets, ensuring reliable evaluation.

1. MLP Baseline Model: Receives flattened 3120-dimensional feature vectors, performs classification via fully connected layers, and establishes a performance baseline;
2. 1D CNN: Takes 2D windows as input, performs 1D convolution along the time dimension, and automatically extracts local temporal features;
3. GNN: Constructs a spatiotemporal graph (nodes are sensor-time pairs, edges include spatial dependencies and temporal connections). It has strong theoretical expressive power, but its performance is the weakest in the current implementation.

Complete scripts are provided: dataset construction (build_small_dataset.py for MLP, build_small_dataset_2d.py for CNN/GNN), model training (train_baseline.py for MLP, train_cnn1d.py for CNN, train_gnn.py for GNN), and evaluation (evaluate_model.py supports all three models). Evaluation metrics include Top1-3 accuracy, per-fault accuracy, and confusion matrix. Complete results can be viewed on the online page (https://faiazu.github.io/projects/tep-results/).

The core contribution of the project is the fair comparison framework. Possible reasons for GNN's poor performance: the fixed EDGES_75 edge list does not accurately capture variable dependencies, the message-passing mechanism needs optimization, GNN training is difficult (many hyperparameters), and the data scale is insufficient to leverage its advantages.

Industrial value: Reuse the code framework to establish a TEP fault diagnosis benchmark environment, test new models, verify GNN applicability, and learn time series processing to avoid leakage. Future directions: Explore graph structure learning, spatiotemporal feature fusion, transfer learning, etc., to improve TEP fault diagnosis performance.