Integrating Process Simulation Modeling and Predictive Analytics to Gain Deeper Insights into Machine Health and Performance

Abstract

The efficacy of machine learning (ML) algorithms for turbomachinery condition monitoring can be compromised by the lack of robust historical data for training. While unsupervised or deep learning (DL) algorithms may be used when sufficient volumes of ‘labeled’ data are unavailable, they offer limited insights into detected anomalies or outliers. Additionally, the inherent dependency on data volume and variety delays the deployment of these algorithms, making an ML-only approach unsuitable for situations such as a machine's first operating run. The paper discusses how a combination approach utilizing both first-principle based performance algorithms and ML algorithms can address several shortcomings of the ML-only approach. Examples are provided to demonstrate that one type of algorithm can outperform the other in the detection of specific anomalies. Therefore, when deployed in parallel, they provide the ability to predict / detect a larger universe of machine faults. The combination approach can also address the lack of interpretability inherent to ML algorithms in cases wherein both sets of algorithms show anomalous behavior. To further address the issue of data inadequacy and poor data quality, the concept of simulation-based transfer learning is introduced. A thermodynamic simulation model is used to generate performance data for a multistage, variable composition centrifugal pump. This data is then used to train two deep stateful LSTM neural network models to predict pump discharge pressure. In the first model only simulation data is used for training while for the second model both simulation and historical data are used. Prediction results from both models are compared with those from a performance algorithm and an LSTM model trained solely on historical data. Test results demonstrate that the LSTM model trained on both simulation and historical data outperforms the other algorithms. This methodology can be applied successfully to accelerate the deployment and enhance the value of deep learning algorithms for machine performance analysis. An additional benefit of training the model on simulated data derived from well proven thermodynamic/aerodynamic principles, is that the insightfulness of performance algorithms may be ‘inherited’ by the deep learning algorithms.