AspenTech: Risk of Using ‘Clean Data’

Mike Brooks

In recent years, data analysis has played a key role in many industries, including industrial production and engineering design. In combination with substantive knowledge, analytics can be indispensable in determining the causes of interruptions and loss of profit.However, the results are highly dependent on the data context, and the conclusions may turn out to be false. 

Need for guidance. The
technical director of a young machine learning company once said: “Give me the data and I will solve any problems.” I would like to repeat these words, but, unfortunately, this does not work like that. Data analysis technologies, including machine learning, are universal for all industries, but subject knowledge is not. Therefore, for successful work both terms are needed.

An analytical solution should distinguish causation from simple correlation and report only real problems. But data analysis, including machine learning, is not a lifesaver. In order to find the correct answers to questions with the help of analytics, guidance is needed.Otherwise, meaningless correlations arise, for example, the famous assertion that increased consumption of margarine leads to divorces in Maine. Guidance is subject knowledge that limits contextual data, defines reasonable expectations, and excludes meaningless relationships.

Machine learning helps detect all kinds of data correlations, many of which are completely useless. To establish a causal relationship, knowledge and experience are needed. What skills and experience will you need to create a solution, how long will it take and will the solution be scalable? In a sense, this is a natural limitation of machine learning. 

Using clustering when working without human intervention, a machine learning system can detect and remember patterns of behavior. In the design and production process, using clustering, you can determine the standard signals coming from the sensors in and around the plants. And then, based on deviations from the norm, called anomalies, it is possible to detect equipment malfunctions.

Another technology of machine learning – machine learning with a teacher – requires the participation of a person who reports on the event and gives the date and time when it happened. However, the machine learning system does not understand what happened.She knows only the date and time. To determine the meaning of an event, you need subject knowledge and understanding of the data context. Having learned about the event, the machine learning system remembers the signs of a certain behavior that preceded the event. For example, when used in heavy industries, the machine may fail due to damage to the bearing. Having remembered the exact behavior model during wear or failure, the AI ​​analyzes the new data to detect a repetition of this model in them before the failure occurs.Proactive notifications allow you to not expect complete wear and tear and repair before breakage.

The company’s specialists understand the relationship between the behavior patterns of machines and the mechanics of wear. Based on this knowledge, they guide the machine learning system, helping to identify the correct patterns of behavior in the event of a failure.In addition, using empirical and non-empirical models, we can predict an approximate range of results, and then determine machine learning guidelines that will help find accurate wear patterns. The data context is very important when marking events, selecting variables, and managing data cleansing. Effective solutions combine knowledge of the processes that serve as a data source and the experience of using analytical technologies.Therefore, directions should be firm and reliable. 

Practical application
How does it work in practice? Take a two-step approach. Let’s start with the design.Examine the data-generating process, correctly label important events, and calculate the most significant of them, for example, known physical limitations. Use this information as guidance to clean up data and appropriate behaviors based on equipment operating conditions. After completing the design process, switch to data analysis mode. 

At this point, you provide a data context: now the algorithms do not take into account specific problem areas. Now, data, algorithms and behavior models do not know their sources: data is just data. The scope, units of equipment and data sources are diverse and not important. In this context, we do not need rigorous technical models and complex differential equations.

Simply put, data entry guidelines play an important role. To arrive at the right conclusions, clearly defined data sets are needed. Subject knowledge defines the data context.Therefore, you need to study the intricacies of each production process, and then move from design to analytics using directions.