Engineering Challenges in Industrial AI

This talk summarizes important experiences we cultivated in several projects where we developed AI methods for industrial customers like chemical production plants or gas fired power plants.

One might think that applying leading edge AI methods to large quantities of industrial data will automatically yield valuable results. In our experience there are many "traditional" obstacles that need to be removed, before the magic can happen [2].

For many use cases, data from various sources need to be integrated. For example, time series data stored in industrial Data Historians need to be labeled according to quality data stored in Laboratory Information Management Systems (LIMS). The process data like setpoints and raw material characteristics need to be contextualized using the production data, e.g. production orders, raw material lots, from ERP systems. Although many companies are currently building up enterprise data lakes, the reality in many plants still is that data is locked in separate silos and a real-time data integration requires significant efforts.

Building, automating and operating a large industrial plant requires major (traditional) engineering efforts and expertise. Data scientists usually have different backgrounds and often lack knowledge, for example in control systems engineering and process engineering. In addition, many important engineering artifacts like Piping and Instrumentation Diagrams (P&ID) and process design document are either not up to date or even unavailable at all. However, in order to define relevant use cases and design powerful AI solutions, quite a high level of both traditional engineering and data science know-how is required. This means that traditional plant experts need to closely cooperate with data science experts. In the long run, software tools like ABB Ability^TM BatchInsight will be able to offer easy to use solutions for a broad class of use-cases, so that plant experts can use AI solutions without the need for additional consulting [3].

Many industrial AI methods work with time series data. A larger industrial plant like a refinery can have many tens of thousands different time series. In the best case a consistent hierarchical naming schema is available and for each time series a good description exists. In reality, it can be cumbersome to identify the best set of time series related to the use case at hand [1]. In addition, data quality needs to be ensured. For example, if sensors are not properly calibrated, using this data can destroy AI models. In one of our projects, a sensor was being replaced at a certain point in time and the measured value changed from power to current. For convenience and because of the typical time pressure during daily operation, the new value was stored in the old variable. Our AI methods identified this abrupt change in system behavior, however this finding had no value for the customer.

Industrial plants and their subsystems have different operating states that often need to be clearly separated. For example, if a time interval where the plant was in fault condition is accidentally included to train a reference model, the quality of this model will degrade.

In summary, various challenges exist in the application of industrial AI solutions. Our talk will provide our learnings in addressing these challenges and will help to ask the following five questions before starting an industrial AI project:

• Can we define a use case with sufficient business value?

• Are people both with domain expertise and with expertise in AI committed to the project?

• Will the communication between domain experts and AI experts be good enough?

• Is all required data available for the project in sufficient quantities?

• Is the quality of the data sufficient or are additional preparation and cleaning efforts required?