Improving Refinery Safety and Efficiency with AI at the Edge

Installation of the data acquisition hardware and edge infrastructure was completed in spring of 2018. For the first time, the Texmark operations team was able to monitor live pump behaviour from wireless sensors and react to impending failures. SparkCognitionTM could not model the asset behaviour in its SparkPredict® solution immediately, as machine learning models depend on rich, historical data, but within two months of implementation, SparkPredict was able to create accurate asset models.

Once enough data was collected, SparkCognitionTM data scientists built a model of the pump operating behaviour based on features derived from accelerometer, pressure, and temperature data. They initially began with 125 features, and with input from Texmark’s operations and maintenance team, reduced them down to the 70 features most relevant to the model’s performance. With a still-limited data set and no labelled failure data, the data scientists used an unsupervised approach to train the model. Unsupervised learning finds patterns in unlabeled data so that a subject matter expert can interpret and label them appropriately. In total, the data cleansing, feature engineering, model building, and deployment took just over one month to complete. Another SparkCognition partner, leading pump manufacturer Flowserve, was consulted for validation of the approach.