(487a) A Novel Loss Function for Key Process Variables Prediction of Chemical Process Fault Prognosis

With the transformation of industrial production digitization and automation, process fault detection and diagnosis (FDD) has been an indispensable technical method to realize the safe and efficient production of chemical process [1]. FDD may have a long detection delay for some faults, and fault prognosis could detect the occurrence of faults in advance, which would reduce the impact [2]. Accurate prognosis of key process variables in chemical process can indicate the possible change to reduce the probability of abnormal conditions. Current state-of-the-art deep learning forecasting methods, often trained with variants of the Mean Squared Error (MSE), might lack the ability to provide sharp predictions in deterministic data context [3]. The smooth and delayed prediction results could be obtained by the classical models trained by MSE loss function, which could bring great challenges to the predictions of key variables and the fault prognosis of chemical processes.

To handle these challenges, Cuturi et al. [4] proposed a learning loss building upon the dynamic time warping (DTW) discrepancy, which computed the soft-minimum and took advantage of a smoothed formulation of DTW. On this basis, Le Guen et al. [3] introduced a distortional loss including shape and time, which tried to predict a sudden change using features of shape and time of process data. The loss can improve the performance and predictions in synth and traffic datasets. Inspired by the shape criteria and soft-DTW, we proposed a novel loss function for key process variables prediction of chemical process fault prognosis. Specifically, we introduced a loss function including trend changes and shape features (LITAS) to deep learning methods. In practical industrial applications, chemical plants are in a smooth operating state in the vast majority of cases, with only rare instances of erratic fluctuations. Firstly, we increased the proportion of data containing abnormal fluctuations by dividing the datasets. Then we applied the proposed LITAS to deep learning models, such as LSTM, BiLSTM and Transformer, and trained the models based on the divided datasets. Finally in the online part, the chemical process fault prognosis was achieved by predicting the changes of process variables.

This work used a real fluid catalytic cracking (FCC) dataset of a petrochemical company for application. The process data was collected in a reactor and a regenerator located in the FCC process, which assume an important role in the generation of products from the fluidization reaction and regeneration of the catalyst. The temperature of the reactor was chosen as the key process variable by the chemical plant. The experiments carried out on the real FCC dataset showed good behavior of LITAS compared to models trained with the L1, L2 and Smooth L1 loss function. LITAS could get similar results to Euclidean loss when MSE is used as the judgment criterion in the test dataset, but LITAS could obtain closer prediction results to the ground truth when the process had abnormal fluctuations. And LITAS was proved to be robust to the choice of model, which showed the ability to improve the prediction performance for fully connected networks as well as recurrent networks.

In conclusion, we proposed a novel loss function LITAS for key process variables prediction of chemical process fault prognosis. And we applied the loss function to a real FCC process dataset to improve the fault prognosis capability for the key variables.

References

[1] Bi, X. et al. (2022). One step forward for smart chemical process fault detection and diagnosis. Computers & Chemical Engineering. DOI: 10.1016/j.compchemeng.2022.107884.

[2] Bai, Y. et al. (2022). Chapter Three - Data-driven approaches: Use of digitized operational data in process safety, in Methods in Chemical Process Safety, vol. 6, Elsevier, pp. 61–99.

[3] Le Guen, V. and Thome, N. (2023). Deep Time Series Forecasting With Shape and Temporal Criteria. IEEE Transactions on Pattern Analysis and Machine Intelligence. DOI: 10.1109/TPAMI.2022.3152862.

[4] Cuturi, M. and Blondel, M. (2017). Soft-DTW: a Differentiable Loss Function for Time-Series. Proceedings of the 34th International Conference on Machine Learning.