(37a) Digital Twins for Predicting Early Onset of Failures Flow Valves

Unplanned shutdowns in petrochemical refineries due to equipment failures cause significant financial losses. Across our operations, Shell has been collecting real-time data for decades. At present, more than 10 million operational variables per minute are collected, streamed, archived, integrated with operational control systems. There is enormous potential to exploit this data further. With predictive analytics and machine learning algorithms, unexpected failures and unnecessary maintenance can be avoided, saving assets millions of dollars per annum in optimized maintenance and deferment avoidance.

This paper presents the results of two different proof-of-concept study carried out in Shell Pernis and Shell Martinez manufacturing sites, to use unlabelled historical process control data to develop a digital twin algorithm which predicts failures. Experiments for the use case were performed on multiple control valves. The target was to verify, whether the machine learning methods are capable of distinguishing between normally behaving valves from abnormally behaving valves. The difference between the prediction and the measured values, i.e. the error rate, should be as low as possible for the normal valves while being as large as possible for the abnormal valves.

Multiple solutions based on artificial neural networks and statistical approaches were implemented to model the normal behavior of the monitored systems in Pernis. Mismatches with predictions of the modelled systems were then used to predict failures in advance, where the prediction horizon reached more than a month for some use cases. 4-layer GRUs (Gated Recurrent Unit) with tanh activation functions and an input sequence length of 4 samples provided the best results. GRUs were 7% faster to train than LSTMs (Long Short-Term Memory) while reducing the prediction error by 15%. Furthermore, the approach allows for predicting failure with high accuracy, characterising failing signals with deviations that are 5x larger than the deviations during normal operation. This indicates that machine learning models can predict failures in petrochemical refineries for the studied use case, without the need for industry-specific knowledge, if the model is trained with data representing fault- free operation.

In case of Martinez, first, a traditional mass balance with meters upstream and downstream of the target flow element was utilized. A secondary accuracy check occurred by developing a novel, machine learning based predictive deterioration model that relied on first principles and statistical multivariate regression to augment and validate traditional mass balancing. This model successfully verified that the mass balance approach provided acceptable meter accuracies in this instance and allowed engineers to track predicted flow meter performance vs. actual metering. Going forward, these models will enable equipment deterioration analysis and be the catalytic step change towards predictive maintenance.