Data Science - Man and machine united by quality control
For businesses in Central Europe, product quality is a key factor in ensuring international competitiveness. However, the price for this advantage is an often laborious – and expensive – quality control. Automation can help to some extent, but the trend towards batch sizes of 1 significantly increases complexity. Now a new approach promises help – and opens up new ways of human-machine collaboration in quality assurance.
Insight in brief
- product quality is a key factor in ensuring international competitiveness
- costs for quality control explode
- new ways of human-machine collaboration in quality assurance promise help
Faced with increasing global competition, industrial manufacturers are not only subject to constantly growing price pressure, they must also meet ever higher quality standards. As a result the efficacy and, above all, the efficiency of the quality assurance process are becoming key success factors. Automated systems are already used extensively to minimize time-consuming manual quality control steps. These systems are taught to distinguish between good and insufficient quality by providing training data. One downside is that they are often laser-based and therefore expensive. Consequently, cheaper camera systems in combination with AI are increasingly being used. For the training of this AI, “supervised learning” is commonly used.
The problem with this approach is that the training data for supervised learning needs to consist of labelled sample data which includes items of both good and insufficient quality. In this context, “labelled” means that in addition to the data itself, information is also available as to whether the product is intact or damaged. Each individual component or variant needs its own training data and training runs. For Industry 4.0 this presents a major problem, since increased customisation and advancing specialisation mean greater product variety and much more frequent product changes. Faced with this increased variability, the shortcomings of the current generation of intelligent quality control assistant systems quickly become apparent. Especially with small production runs, the effort involved can quickly explode.
Supervised learning needs lots of image data
Imagine that, for quality assurance purposes, we want to distinguish between visually intact and visibly faulty car body components moving along a conveyor belt. For a machine to be able to perform this task automatically following supervised learning, we need to provide the machine with lots of images of intact car body components as well as lots of images of potentially faulty components. For a simple component with a small number of characteristic fault types, this might be commercially feasible. But how to deal with product innovations from R&D, customisations or previously unknown fault types?
In cases like this, the quality assurance department will need to determine whether there is a need for further training and new training data must be produced where necessary. This requires close coordination between the R&D and production departments. In addition to providing the training data, this entails substantial extra work. If the product comes in lots of variants or changes are made too frequently, investing the time and capacity involved in performing this work may simply not be worthwhile.
This is where ‘unsupervised learning’ comes in. Unsupervised learning means that training data no longer need to be labelled by a human. For quality assurance departments, the potential benefits are huge, since unsupervised learning significantly reduces both the labelling workload and the level of coordination required.
Unsupervised learning – the next stage in quality assurance evolution
Unsupervised learning has been around for a while already and, after supervised learning, represents nothing less than the next evolutionary step towards autonomous machine learning. The use of the term unsupervised learning is not limited to quality assurance – it is used to refer to any kind of machine learning approach which doesn’t need labelled data. Our previous example of a visual inspection of a car body component helps to illustrate this concept: The training data consists solely of images of fault-free body components. This means that no manual labelling is required, as all images are automatically labelled as acceptable.
So why aren’t we making more use of this concept? Autonomous learning requires specific algorithms. In the field of machine learning, these algorithms are often not derived strictly mathematically but are arrived at by trial and error. There are lots of supervised learning quality control algorithms around which have already proven their utility in real world use. For unsupervised learning algorithms that’s not yet the case.
Unsupervised learning – a hot field for research
With demand for unsupervised learning algorithms increasing, the last few years have seen a lot of research in this field, resulting in a number of potentially useful algorithms. We have tested a few using sample data and verified their utility in a manufacturing context, with some very positive results. We were, for example, able to identify anomalies on steel surfaces with no manual image labelling involved.
This work used a very clean data set intended solely for research purposes. So we decided to go one step further. Using low-cost webcams, we created a more realistic dataset of simple wood panels in-house. Even in these low quality images, the machine was able to autonomously identify anomalies with no human intervention.
This offers the prospect of new, low-cost options for human-machine collaboration in quality assurance. One possibility is that machines could significantly reduce the number of cases requiring human evaluation. Taking our car body component example, this might mean that the machine marks components suspected of being faulty for subsequent human inspection. All other parts can be verified with no human intervention or using samples selected using a statistical sampling method. This would reduce the human workload and allow staff to focus on the small number of potentially faulty units.
Unsupervised learning – quality assurance and beyond
One thing is clear – demand for new quality control solutions is growing steadily. Unsupervised learning is not just the latest buzzword – it promises to enable the next level of automation. Companies have an opportunity to become pioneers and gain a valuable competitive advantage, with potential multiplier effects. That’s because unsupervised learning could also offer utility in many other areas, such as automatic detection of anomalies in time series data (e.g. IoT sensor data).
It’s time to unleash the power of algorithms!
You are interested in Data Science and unsupervised learning. Please contact us!
Kevin Henrichs is Lead Data Scientist and an expert in Machine Learning. He holds a PhD in experimental atomic physics and has worked for many years analysing huge amounts of data. He looks back on several years of hands-on Machine Learning experience, among others in the fields of mechanical engineering and HR, specialising in computer vision and natural language processing.