Manufacturing companies often struggle to keep their customers 100% satisfied: Delayed deliveries or issues with quality are not always easy to put a stop to. At the same time, the increasing complexity in manufacturing, a wealth of variants and rising deadline pressure are raising the costs of manual quality analysis and rework.
However, this problem must not be ignored. After all, efficiency, planning stability and customer satisfaction are decisive competitive advantages for manufacturing companies.
Predictive quality – data-driven, automated quality analysis
There is a way, with the help of data science, to reduce or automate the manual time-consuming work caused by rework and waste in manufacturing: predictive quality. This method describes data-driven, automated defect detection and quality analysis, which can be used in a variety of processes.
Defect Prediction
First, there is defect prediction, a predictive method. It is divided into supervised and unsupervised defect prediction. Here, supervised means that historical data can be used to determine which material or part was defective in the past. In combination with the existing process and machine data, this information can then be used to teach the machine learning algorithm which influencing factors ultimately led to rejects. Unsupervised defect prediction lacks information about which part was rejected or hat to be reworked in the past. Anomaly detection is then used to determine which data pattern describes normal operation and when anomalies or outliers occur here. Both methods are closely related to the predictive maintenance use case but look at machine data from the standpoint of product quality rather than machine reliability.
Defect Detection
With the second method, defect detection, supervised and unsupervised defect detection are also discussed separately. In this method, rejects are not predicted, but instead detected in real time right after parts have been processed. In supervised defect detection, manufactured parts are usually measured using sensors or photographed so that the machine learning algorithm can check whether they are faulty. In unsupervised defect detection, on the other hand, the algorithm is no longer taught which part is intact or defective. It learns independently, based on all the parts produced, which part “looks normal” and which does not. However, this approach often requires a larger amount of data to achieve a adequate result.
Root cause analysis
The final method for preventing rejects and rework is root cause analysis. This methodology is used when nobody really knows where in the manufacturing process the quality issues arise. This procedure is rarely fully automated and in most cases is performed by data scientists in close cooperation with different departments close to production. It involves consolidating historical data and using causal discovery to look for commonalities in process instances that caused rejects or rework.
Advantages of predictive quality for companies
The choice of the exact procedure depends, among other things, on the data situation and the individual objectives of the respective company. Nevertheless, the advantages can be summarized in general terms:
- Reduced labor for rework
- Less waste of material
- Significantly increased planning reliability
- Lower lead times
- More reliable, efficient production
- Increased customer satisfaction
- Positive impact on sustainability
- Less operation disruptions caused by defective components
When companies discover quality defects only at the end of a manufacturing process, they usually have to go back or repeat multiple steps to fix it. This throws the entire production planning out of sync. Lead times for orders are extended. Deadlines promised to customers can no longer be met. Quality problems reduce the productivity of the entire operation.
This can be avoided with Machine Learning-aided solutions for predictive quality. The term describes several AI-based processes that either automatically sort out rejects right after they have been produced or detect whether defective parts are created during the manufacturing process. In defect detection, which is often used by large-scale manufacturers, AI uses image recognition processes, for example, to check whether individual manufactured parts have characteristic features of a defect and thus differ from the thousands of other screws or spark plugs that have been manufactured without defects.
It is very valuable when data on the manufacturing process can be clearly assigned to the components produced and whether they have quality issues. Companies achieve this, for example, by equipping assembly stations and machines at which they collect process data with RFID scanners that read the RFID tags attached to the parts produced there and the tools used for this purpose.
Closing Thoughts
Satisfied customers are usually the result of high-quality products and processes. To ensure this in the long term, modern technologies such as machine learning help to automatically reduce rejects and reworks – and therefore, costs. Which approach suits your company best depends on your current data maturity and your individual processes.
Would you like to get to know concrete use cases of the various processes? Learn more about predictive quality here.
