Robust Fault Detection in Refinery Pumps Using a Deep Learning Technique

Detecting faults in refinery pumps is particularly challenging due to the noisy operational conditions they endure. This study introduces a novel, noise-robust fault-detection model for pumps in oil and gas refineries that utilizes vibration signals. Given the temporal and spatial dependencies in pump vibration data, ConvLSTM, which combines the sequence-processing capabilities of Long Short-Term Memory (LSTM) with the pattern-recognition power of convolutional operations, is used. To further pinpoint the most important patterns for fault detection, an attention layer is integrated. A key innovation in this work is the development of a robust training and hyperparameter tuning strategy. This approach specifically addresses noise by incorporating realizations of noisy data into the loss function and selection criteria, thereby enhancing the model’s resilience. A dataset of vibration signals from industrial refinery pumps, operating under real-life conditions, was used to validate the proposed model. The model’s performance was then benchmarked against conventional vanilla LSTM and ConvLSTM models (without attention), as well as other state-of-the-art models. The results showed a significant performance loss in conventionally developed models when applied to noisy data. For example, the F۱-score of the vanilla LSTM model dropped from ۰.۸۳۷ to ۰.۴۴۴. However, the integration of ConvLSTM and attention layers substantially improved the performance and narrowed the mentioned gap, boosting the F۱-score on noisy test data from ۰.۴۴۴ to ۰.۷۶۲. The proposed robust strategy yielded models that are highly resilient to noise. The developed models demonstrated strong and consistent performance on both noisy and original data. Specifically, the ConvLSTM+Attention model achieved an F۱-score of ۰.۹۰۹ with the robust hyperparameter tuning strategy alone, and this further increased to ۰.۹۳۳ on noisy data when the robust training strategy was also incorporated. The primary novelty of this study lies in the robust tuning and training strategy, which equips predictive models with significant noise robustness. This robustness makes the models a reliable tool for fault detection in real-life environments.

کلیدواژه ها:

ConvLSTM ، LSTM ، fault detection ، refinery ، Robust loss