In order to differentiate from the competition on national and international markets, high product quality is indispensable. Quality control in production is usually sample-based in the form of statistical process control. However, statistical process control can only inspect a small proportion of the components produced, and thus holistic and complete proof of inspection is not provided. Increasing the sample size, however, is associated with high inspection costs. But what about the use of automated inspection processes? Is the use of artificial intelligence in quality control beneficial?
Possibilities of quality control in the production process
During the production process, there are different points in time at which a quality inspection can be performed (see Fig. 1):
- On the one hand, a so-called pre-product inspection can take place so that no defective products are loaded into the process.
- On the other hand, the inspection of product characteristics or machine parameters can be conducted during as well as after each production step.
- The last possibility to check a product for quality is after the production process in the so-called quality assurance. It should be noted that merely testing after the production process only allows defects to be detected, not prevented. Data collected through the inspections are returned to production to counteract any quality deviations.
Fig. 1: Inspection positions in the production process
Quality control with the help of AI
The usage of AI for quality control implies certain requirements for the data infrastructure. As already described in our last blog post, data is the foundation for a successful AI model. All relevant process data must be collected, stored and processed – then optimal results can be achieved with AI used for quality control.
AI is already being used occasionally at the above-mentioned inspection positions in the production process for quality inspection:
- Pre-product inspection (1) and quality assurance (3)
Image-processing AI systems are mostly used here. Reasons for this are, on the one hand, that image processing is easy to implement in contrast to the use of deep learning models. On the other hand, complex data analysis can often be dispensed with in image recognition. This is because all that is needed to train the system is precise training data, which the AI can use to learn to detect errors itself.
- During / after (2) the production steps
Until now, data-processing AI systems during the production process often failed due to a lack of sensors or equipment connectivity. This is because a large number of built-in sensors are important for data collection, constantly providing data on machine status and manufactured objects. This data can then be checked by AI systems for certain properties that deviate significantly from the expected values.
Goals of using AI
One goal when using AI during the production process is inspection optimization. Thereby, the inspection process as well as the inspection quality are to be improved.
In addition, the AI system can be used for quality predictions. In other words, it aims to make predictions about the future quality of a product. Machine data from the production process can be used for this purpose, for example. By monitoring the performance, quality and condition of the machines in real time, it is also possible to make detailed predictions about machine behavior. To make concrete forecasts, data from many machines, the expertise of the machine manufacturer, and knowledge about the usual behavior of the machine are needed. All this data forms a basis on which AI systems can make forecasts.
Insight into one of our AI projects
To illustrate the advantages of using AI for quality inspection, let’s take a closer look at a use case (see Fig. 2).
A medium-sized company in the metalworking industry manufactures complex metal parts in high volumes. Defects of just a few micrometers in size are often critical – so each processed part must be inspected individually. The company uses an optical 3D inline measuring system for quality inspection. However, despite the optical measurement, 10% pseudo rejection still remains. In order to reduce the 10% of false rejects remaining in the pseudo rejection rate, a time-consuming manual reinspection of the supposedly bad rated metal parts is necessary.
The existing measuring system is now to be expanded to include AI-based component classification. In this way, pseudo rejections will be recognized as such directly in the production process with the help of AI. However, the AI must be permanently trained due to batch variations and the resulting possible changes in the pseudo rejection features. Since this is not possible by means of manual labeling of image data, a strategy must also be developed that enables the AI to learn autonomously and during production to recognize pseudo rejection features.
In order to train the AI with the parts that are not OK, the individual parts have to be identified and checked individually by the 3D measuring system. The parts that are OK can be delivered directly. The AI now comes into play for the defective parts. In the image data of the supposedly bad parts, the AI can identify whether any of the previously trained measurement artifacts are present. If this is the case, it is highly likely that the part is a good one. To rule out the possibility that a real defect is present despite the measurement artifact, the components that have been checked by the AI run through the 3D measurement again. Thanks to a previously created digital fingerprint, the 3D measurement system knows whether the component has already been inspected. If this component now has a defect in the second measurement, it can be declared directly as a wastage. If it does not have a defect, it is in fact a pseudo-reject and the component can therefore be delivered. Afterwards, the 3D image data of the first measurement of the pseudo rejection is transmitted to the AI as training data. This way, the AI is constantly trained and adapted to the changing defect characteristics.
Thanks to the AI, pseudo rejections can be recognized as such directly during the production process. This enables the company to deliver nearly 1 million more parts annually that would otherwise be mistakenly discarded. Manual rework can be eliminated, significantly increasing productivity. CO2 emissions and resource consumption are also significantly reduced.
Fig. 2: Procedure of the inspection process
Conclusion
The ongoing digitalization and the constantly increasing cost and competitive pressure in production make automated quality inspection almost inevitable. The use of an AI for quality inspection makes sense both in the production of large quantities and in small and medium quantities. It is crucial that the characteristics to be inspected are similar or the same. Especially when many features of a component need to be inspected, quality control with AI is very beneficial. AI systems can be used at all inspection positions in the production process for quality inspection. Goals such as process optimization or quality prediction are of particular importance. This is because productivity can be increased, resources and costly rework can be reduced, and defects can be detected in advance.
