6 cl lightly aged Cuban rum, approx. 2.5 cl freshly squeezed lime juice and approx. 2 teaspoons of white cane sugar – that’s all you need for a classic Daiquiri. There are lots of different recipes, but how you mix it is important – everything depends on consistency. But what does a Cuban cocktail have to do with a research project on predictive maintenance? You can read about it in the following blog post.
Daiquiri or DAIKIRI - the data cocktail for industry 4.0
Tasty chaos? To create a new taste experience, bartenders have to try everything, break down the individual ingredients, and combine them with something new. Not much different in the (almost) identically named research project DAIKIRI, which is concerned with filtering relevant data from a huge cocktail of machine data, data that can then be used to diagnose problems such as a machine failure. In both cases, it is a matter of distilling the essence and keeping consistent.
But enough comparisons. Let's look at the research project DAIKIRI - Explainable Diagnostic AI for Industrial Data. The aim of the project, which is funded by the German Federal Ministry of Education and Research, is to use artificial intelligence (AI) to automatically generate diagnostic reports for mechanical and plant engineering in the future. These diagnostic reports will also provide comprehensible explanations for why a mechanical component is in danger of failing, for instance. As a result, machine operators will be able to understand and evaluate the significance of the diagnostics and make decisions and take actions accordingly.. The project coordinator is USU Software AG. Other partners include the Data Science working group at the University of Paderborn, AI4BD Deutschland GmbH and pmOne AG. The 24-month research project was launched at the beginning of 2020.
Predictive maintenance is the key to productivity
With sales of over 225 billion euros and more than 1 million employees, mechanical and plant engineering is one of the most important segments of the German economy. Due to the increasing automation of industrial processes and the rising productivity in this sector, gigabytes to petabytes of event data accumulate daily. Machine learning (ML) is already making important contributions to the optimization of production processes because of its ability to intelligently process huge amounts of data. This helps to reduce production costs. Unplanned interruptions in production remain a worst-case scenario. The consequences for quality, costs, and brand can be significant. According to Aberdeen Group calculations, the average cost of downtime is as high as $260,000 per hour. With predictive maintenance, machine downtime and damage can be avoided. The potential is huge. According to McKinsey, it can reduce production plant downtime by up to 50 percent and maintenance costs by 20 to 40 percent.
However, this requires integrated, digitized and networked machine monitoring - and the processing of maintenance logs, configuration data, sensor and telemetry data. These must be combined for predictive maintenance.
Large companies in the industry in particular make intensive use of AI technologies for this purpose. This puts medium-sized companies under massive pressure and is forcing them to react. Often, however, there is a lack of know-how to use the valuable industrial mass data through intelligent data processing.
Making industrial data comprehensible - ways out of the black box
Today, large volumes of industrial data for mechanical and plant engineering are processed using modern machine learning platforms, e.g. Azure ML-Studio or Katana Flow. However, these platforms have one major disadvantage: currently, the results of the algorithms can only be understood and interpreted by machine learning experts. Machine Learning (ML) algorithms therefore only provide clues as to "what" should be exchanged, but not the "why". However, the service technician responsible for replacing the machine part must also know why the algorithm suggests replacing a ventilation system, for example. The reasoning provided, e.g. “airflow not sufficient” or “service life of the component exceeded”, helps the service technician to react adequately.
The consequences of this "black box behavior" are obvious:
- Errors in ML models can be detected only with difficulty or not at all. Therefore, these approaches cannot be used for system-critical diagnoses.
- Acceptance of these black-box models is often low even for non-critical diagnoses, because domain experts are unable to comprehend the results of these systems and are therefore skeptical about them.
- There’s no guarantee that the results will be validated because it is unclear how data is combined into models.
These data-driven predictions must be transparent in order to make reliable decisions and avoid costly mistakes. This is the aim of the DAIKIRI research project.
DAIKIRI: Self-explanatory condition monitoring of machines
For the first time, DAIKIRI will develop AI procedures that are self-explanatory and automatically "verbalize" the results of AI, thus making them transparent. The project aims to significantly improve methods for the "semantification" of data. In this way, intrinsically explainable procedures for so-called Statistical Machine Learning (SML) will be made available for practical applications. To do this, data will be converted semi-automatically into description logic. The biggest hurdle in this conversion is the automatic creation of ontologies. For learning ontologies, DAIKIRI uses novel embedding and clustering approaches. The "training" is done with real data from real use cases, so that DAIKIRI learns these ontologies directly from machine data. In the end, the system will automatically predict the "what" and "why" in a way that allows the service technician to understand the result. This creates a reliable and trustworthy foundation for making decisions. In the future, these methods will allow new business fields to open up in other industries, such as medical technology, in addition to mechanical and plant engineering.
Henrik Oppermann is Business Unit Manager at USU Software AG and heads the USU Group's research department. There, he is responsible for the development of Industrial Big Data solutions and coordinates the work of the Smart Data projects.