Benchmarking Machine Learning Fault Detection Methods on the Tennessee Eastman Process Dataset

The Tennessee Eastman Process (TEP) dataset is a standard benchmark for process control and fault diagnosis due to its complex, multivariate dynamics. We evaluate a broad range of machine learning and deep learning methods for multivariate time-series anomaly detection on the dataset, covering 17 of its 20 fault types. After preprocessing the dataset, we benchmark classical models such as XGBoost and sequence models including LSTM classifiers, LSTM-FCN, a convolutional transformer, and a proposed Transformer-Kalman (TransKal) classifier. We also assess hybrid LSTM and convolutional autoencoder approaches. Supervised models were trained on combined normal and faulty data, while semi-supervised models used only normal data. Finally, we evaluate the models using an independent dataset to assess generalizability. Several models achieved 99%+ multiclass accuracy, but showed fragility on the independent dataset.