Smart refinery is built on structured data
Tomi Lahti, PRODUCT MANAGER / 20 May 2021
Last time I wrote how AI will shape refinery operations and support humans in decision making and safety improvement. Now we take a glance to basic assumptions behind futureproof ML / AI solutions for the industrial production environments. Because there is a difference between “typical” AI implementation and Industrial AI meant for OT (operational technology) environments.
All is based on good quality data
Smart Refinery or any intelligent production facility of the future will in its operations rely on processed, real-time information and related future predictions. The basic requirement is good quality data which can be streamed. Data quality is the most common stumbling block why ML fails to deliver its promises. If the quality of the data is not good enough to make a working application, the only option is to fix the data. This should not be forgotten, because everything you build relies on good quality data.
ML models create estimations and predictions from streaming data (continuously generated by different sources, and incrementally processed at each moment of time). The same data can also be applied for simulation models. So the basic requirement is to have access to online data, which is good enough. This can be implemented in many ways. When operating in the process industry, the best way is to use OPC UA and it’s Information models for structuring data and including metadata to each tag. With OPC UA, data is its own documentation and is easy for data scientists and maintenance functions to operate; changes in data structure will be automatically propagated all the way to ML tools.
Streaming data is consumed by prediction models, running on an efficient machine learning pipeline, which can be optimized for process industry on-premise model orchestration, lifespan management and production deployment. ML models can also be run on the cloud, but the existing cloud ML tools miss the benefits of OPC UA structured data. This may make model development burdensome and system maintenance devilish, if the model count is big. Meaning that a true Smart Refinery insists on structured data.
Production AI requires structured data with OPC UA Information Models
For true production digital transformation, structured data is essential. Currently most of the ML projects use flat data, which needs to be manually maintained e.g. when there are changes in process or the process equipment. When approaching a true, holistic Smart Refinery, the maintenance burden caused by the number of models and changes in production becomes cumbersome.
Considering the emerging race for carbon neutrality in terms of changing raw materials and production units, the importance of structured data for ML / AI and competitiveness needs to be even further emphasized. To be really data-driven, companies must make data access and management as easy as possible for the data end users (data scientists and so on). For process industry production environments, this is where OPC UA Information models are meant and ideally suited. Especially if there exists a Companion specification (mutually agreed, industry-specific information models), those should be used, as they define common elements for industry and enable true interoperability between different vendors. No more siloed data or vendor lock-ins!
This is how NAPCON has built its Industrial ML Pipeline; leading the process industry towards Industry 4.0 and Smart Refinery. Using OPC UA and Information Models, NAPCON ML Pipeline is fully compatible with Industry 4.0, acting as an engine for full scale Smart Refinery production estimation and prediction services. With Information models, model development is rapid and library maintenance is easy.
Many current exercises for company digitalization may seem nice and make boards happy for now, but are not really delivering the step changes reachable by Industry 4.0. Problem is that they are not compatible with future floor level M2M communication and automation and thus miss the future benefits of Industry 4.0. Also, e.g. ML models predicting asset conditions double their value when they communicate also with Supply Chain Management, production and blending and are easily maintainable and scalable.
For these reasons, successful production digital transformation should use standard OPC-UA for data accessibility extensively, and plant management should pay attention to create structure for data with Information Models or Companion Specifications.
Next time I will tell you more about our Industrial AI solution, Advisor, how it responds to wicked problems existing in the process industry and how it can be implemented for your refinery or plant. Stay tuned!