95% reduction in rejects and €500,000 in savings? Increased product quality through automated quality control
It is actually a very simple formula: The poor quality of a product leads to lasting damage to the relationship with the customer. If the quality of a product is high, a lasting customer relationship is usually established. However, keeping the quality of a product permanently high depends on many variables. Ideally, it is based on automated quality control.
Product quality is primarily dependent on excellent process quality. Process quality means reducing rejects. Predictive quality, i.e. the data-supported optimization of process and product quality, is therefore an indispensable tool for increasing product quality. The American company Hill's Pet Nutrition, for example, has long recognized this.
AI-supported predictive quality system using the example of pet food products
Hill's Pet Nutrition manufactures pet food products. The business strategy dictates frequent new launches and product changes. With each launch and changeover, the product development team faces the challenge of producing the new product to strict quality standards for moisture, fat, protein content and density. In the past, manual tests were required for each changeover, each of which took 40 minutes. Only after this analysis could adjustments be made to the system settings.
The AI-supported predictive quality system provides a prediction for the result of the manual test. Based on this, suggestions are provided to the production team in real time on how the system settings should be adjusted to ensure the required product quality, even before final laboratory data is available. Once the suggested settings have been made, only final and less time-consuming tests are required. The use of predictive quality technologies and automated quality control led to a significant reduction in rejects and increased the CpK value (the most important parameter for describing the capability of processes) by half. This means that rejects were reduced by around 95%.
What is predictive quality?
Predictive quality, also known as predictive quality assurance, is used in the context of data science and predictive analytics. The aim is to continuously optimize processes on the basis of automated predictions. Various parameters are used to determine the quality of a product. From the specified tolerance limit of individual part production to environmental conditions such as relative humidity, individual parameters can lead to problems with the product.
Predictive quality enables automated quality control and therefore monitoring of all production parameters. A prediction of any quality defects in the product is made on the basis of the interactions. It is therefore possible to make corrections and change settings before a product is finished. This leads to massive time and cost savings.
Key objectives of predictive quality
The general aim of predictive quality is always to optimize quality by using predictions as a basis for decisions on further measures. However, there are two main objectives in particular: automated quality control and process parameter optimization.
Automated quality control enables an assessment to be made in 50 to 600 milliseconds, depending on the respective requirements, and thus of course absolutely trumps human judgment. It is optimized in terms of costs and benefits, with the AI model also incorporating general quality requirements.
In process parameter optimization, a model learns the optimum model parameters and applies them directly. Depending on the material or component, for example, individually optimized parameters are used. Process parameter optimization enables automatic parameter setting for minimized rejects.
Deep learning AI model using the example of technical plastic film production
Toray Plastics is another impressive example of the benefits of predictive quality. Toray Plastics manufactures a wide range of technical plastic film products for industrial applications. Each application, each customer has its own specifications and requires the corresponding ideal operating conditions of the system. Each film break costs tens of thousands of dollars.
An IoT approach was used to record 400 key figures in real time. The relevant key figures for each system were identified and evaluated using a deep learning AI model. As a result, the production teams receive continuous status reports on possible sources of error and a risk analysis for film breakages. Adjustments can be made accordingly quickly. Automated quality control has improved plant efficiency from around 85% to over 90% and saves over 500,000 euros in material and operating costs every year.
{{whitepaper}}
Standardized approach for AI-supported predictive quality solutions
Production managers, site managers and predictive quality experts agree that the medium-term goal must be better product quality through better process quality. Every company sets highly individual parameters for the optimization of its products. However, a basic prerequisite for increasing product quality is a standardized approach to data collection and processing. Such a standard enables the quick and cost-effective implementation of a proof of concept, which can be used to check the effect of the planned measures in advance.
However, data management is not originally the responsibility of production and site managers. The collection, processing, analysis and use of data is an independent area of responsibility that can also be taken on by external specialists if internal resources are lacking. A trained data team, consisting of data consultants and data engineers, collects data and tests AI solutions during ongoing operations and achieves rapid and in-depth results during this test phase.
Predictive quality processes require cloud-based background architecture
Another important building block for establishing predictive quality processes is an individualized background architecture. Systems are constantly changing, for example when new machines are added, new production lines are introduced or suppliers are added to the portfolio. If companies only build for today, they will be left behind tomorrow. Self-learning systems, on the other hand, adapt continuously.
The background architecture should be cloud-based. This allows it to be organized flexibly and data processing costs are only incurred for actions that are really necessary. The costs for the entire system are allocated to individual production lines according to actual expenditure and thus remain visible. This ensures maximum transparency.
Challenges for initial projects with predictive quality
Data collection, data preparation and data combination are fundamental topics that need to be organized in advance of any predictive quality project. This is because data forms the basis for every learning algorithm. The article "Predictive maintenance as a project" describes in detail how the associated challenges are mastered.
Further components for successful predictive quality projects
Predictive quality is based on company-specific requirements and therefore requires a detailed view of what is feasible and sensible. The following components of a corresponding project depend on numerous factors and must always be reassessed.
Automated quality control
Automated quality control is not always relevant and applicable. Its use is particularly interesting when quality control is visual and not yet automated.
Machine learning
Machine learning is suitable for identifying the correlation between machine parameters, settings and external parameters. This allows complex patterns to be discovered. An Excel analysis, for example, cannot discover regularities. AI can do this very well. This results in a prediction of product quality.
Reversal
A new AI can learn suitable settings for industrial systems in order to deliver the required key figures. Both quality indicators and general process indicators can be set as targets. In this way, a system can not only deliver the required quality, but also keep maintenance costs to a minimum.
And now? Proof of concept as an optimal project start
If the decision for predictive quality is clear, although there is still uncertainty about the existing or required framework conditions, a test balloon is worthwhile. As already mentioned, a proof of concept is prepared and implemented relatively quickly, integrates the topic of data organization and rolls it out again if necessary, quickly shows initial results within an average of four months and is within a calculable time and cost framework.
