Taking AI into production. Total’s AI-based ESP predictive maintenance.

Octo on the ‘curse’ of moving artificial intelligence into production. ‘Hype is in the air!’ Avoiding the data science/IT/business ‘War of Thrones’. AI prolongs Total’s subsea pump run-life.

French IT consultancy Octo Technology (an Accenture partner company) held a one-day event in Paris recently that set out to ‘break the spell’ of translating AI initiatives into production. Octo’s Karim Sayadi recommends not building a data lake, data science does not need one to start and projects ‘will fail many times before succeeding’. ‘Exploration, uncertainty and failure is the norm’. AI has given rise to several myths. It is not a solution to everything and one size does not fit all. Such myths are a ‘bad pattern’, making it harder to promote AI and leading to projects that end at proof-of-concept stage or perhaps as the ‘never ending PoC’.

Actually, the solution is simple. Build, observe and evaluate (with the business) and then start over. Start with a data lab and a couple of terabytes of information. A proper data lake comes later and is a choice not to be taken lightly. Use simple ‘explainable’ algorithms and evaluate with good metrics including false positives and false negatives. Finally, do not force AI where more simple approach may be all that is needed.

Yacine Benabderrahmane added that data science may initially represent a simple workflow but that the passage into production is much harder. Enter the Octo Academy, a methodology for producing maintainable and reusable data science solutions and (perhaps) a ‘data-driven’ enterprise. However, ‘hype is in the air’ and much ML is rolled-out on a shaky foundation. Octo’s approach is to deploy minimal ML (just what is needed) on a solid data foundation. In the end, it is ‘craftmanship’ that is required. This implies standard coding techniques, adapted for AI and avoiding building a software ecosystem ‘that is so complicated that you may as well just shoot yourself now!’ Other insights included the need for good quality data (AI is not a data quality tool). Using big IT systems is not an objective. Real time should not be a priority because it is very hard to implement.

There is often a ‘War of Thrones’ between IT, data science and the business, with each blind to the others’ capabilities and needs. This ‘leads to failure and recriminations’. Often data science and the business get along, but IT is more problematic, as it can be distant from the business and peripheral to data science. Octo’s claim is to bring all three together. All need to be involved collectively. One final warning was ‘beware of the super-polyvalent data scientist’ although this was not expanded upon. Octo’s approach is to build collaborative ‘feature teams’ to drive harmonious development and deployment.

Total’s Arnaud de Almeida and Sophânara de Lopez presented on predictive maintenance of industrial equipment. Total is beginning to integrate digital and AI in its production operations. The initial use case revolved around predicting failure on subsea electrical submersible pumps, widely deployed on Total’s Angolan developments in up to 2000 meter water depths. Early detection of abnormal pump behavior (drift) and knowledge of failure modes can maximize pump life and optimize logistics.

Total started with a PoC using historical data on pump pressure, motor speed and logistics, operational and failure post mortem reports. An in-house PoC using data from 10 wells proved successful. This has now been scaled-up to run in a hosted data lake. Today, some 100 wells are monitored with between 30 and 100 sensors per pump. All sensors are modeled and compared with reality. Even weak signals may indicate an issue. Random forest supervised learning was the main AI deployed. The Apache AirFlow workflow manager is key, running in Docker containers for scalability. ESP models are provisioned in a continuous delivery process, with weekly training and daily predictions. Data is time stamped with Python Arrow and saved in the Azure data lake. Grafana, PyCharm and Docker/Jenkins and Octo’s Data Driver also ran. Total recommends that data scientists brush-up on software craftmanship by adopting devops. The solution is now deployed in Total’s local well operations centers and smart rooms for remote monitoring by experts. More from the Octo blog (in French).

This article originally appeared in Oil IT Journal 2018 Issue # 6.

For more information or to comment on this topic email here.