The use of AI methods enables computer systems to simulate human thinking behavior to be able to make decisions independently. There are numerous use cases in process and quality control in the industrial environment. Time series, images, and other data collected through measurement and inspection methods often form the basis for determining deviations from the norm. The evaluation of the data by humans is often incomplete due to the large amounts of data for cost reasons, and there is potential for improvement in the accuracy of the assessments.
The benefit of AI methods is high, especially for assignment and classification tasks and recognizing patterns and anomalies. In addition to the tried and tested analytical ‘tools,’ they are playing an increasingly important role in production control. The computing power required for this is ‘affordable’ today, and proven mathematics and powerful open-source software libraries are available. In this context, challenges such as data quality, changes in mindset, and costs are often discussed. I would like to consciously go into further challenges in introducing AI methods in production control and the strategies to be derived from them.
Table of Contents
A first step in introducing AI methods recognizes that a fundamental change in mindset is necessary: Away from the security and detachment of traditional rule-based analytical methods towards an open, learning, and experience-based system. The central framework that can be used company-wide or integrated systems?
This can be a first step if OEMs offer affordable, mature, and integrated AI applications sufficiently tailored to the specific use case. However, the following criteria should not be ignored:
- How well can the AI recipes be set up, versioned, and monitored?
- How well can the forecast quality be kept at a sufficient level, and how flexibly can models be retrained?
- How flexibly can they be extended to similar use cases without incurring additional costs?
- How well are the results reproducible/repeatable and explainable?
- How well does the system meet the requirements for traceability, further development, and upgrade options?
- How can data from other (non-plant) sources be integrated into the AI recipes if required?
- Which interfaces can result in data being made available to other systems?
- How well do the integrated AI applications meet the requirements of company-specific business processes such as release guidelines, data protection, data security, monitoring, and user management?
- How heterogeneous is the company’s plant park, and how many AI use cases should be implemented in the future?
If these questions are to be pursued as uniformly as possible in the company, a central, standardized AI framework that can be used company-wide for various applications and works independently of the data-collecting systems is recommended. As a result, the user can apply his expert knowledge in a uniform procedure, application-specific instead of plant-specific. Following aspects:
From a technical point of view, a modular, easily integrated approach should be used for a step-by-step and flexible connection to the existing data-generating systems. This also offers the advantage that a gradual implementation is possible.
From a process point of view, a consistent workflow should be available to implement the AI applications, which considers the criteria mentioned above. A standardized interface should enable model developers to provide application-specific model templates, which technical experts can then use and manage. The processes required for this are, for example, teaching, parameterization, verification, release, monitoring, and management of the AI recipes. All these tasks must be carried out without an AI/data science background to ensure sufficient flexibility in production.
Such a workflow can also be used for conventional analytical methods and form a basis for many use cases beyond AI. The user specifies which data is to be evaluated and how a data scientist develops the methods, and a consistent workflow enables fast and reliable roll-out in production. All upcoming data analysis tasks in production control can be implemented in a central system.