Automation of Quality Control

As consumers, we all know the disappointment of having purchased a product, only to subsequently discover that quality was not what we had hoped for.

In many ways, however, we are fortunate here in the West, most of the goods we can buy are made to high quality standards.

In fact, ever since W. Edwards Deming in the 1980s got the idea that focusing on quality over time will lead to lower costs, but that focus on costs will lead to lower quality - most manufacturing companies have accepted that persistent focus on quality gives the best results.

For many production companies, quality control is a noticeable item in the budget, and therefore there is also a built-in risk that a company will be tempted to lower it, perhaps on the grounds that it is nevertheless rare that there are actually errors in the production. Unfortunately, quality control is one of those things that you really only appreciate when you experience how bad the product can become without it.

One way to lower costs while maintaining a high standard, or perhaps even improve quality - is by automating the process.

Automation of quality control can be done by many methods, it does not have to be complicated. However, there are some types of production where it is really difficult to automate using traditional methods. This is where machine learning techniques can shine.

A focus area for machine learning research in recent years has been to recognize and understand what images contain. This research has led to it today being one of the most proven and robust methods in machine learning.

An example of using these techniques in practice is quality control of oranges. A number of researchers from the University of Valencia have demonstrated how image-based techniques can help detect inappropriate spots on oranges. This knowledge can be used directly in the production to sort the ugliest oranges (which will surely be used for juice instead).

In the pictures above, you first see the camera image which is input to the system, then an image where you the spots are manually marked to have a comparison basis for assessing the system, and finally you can see the areas the system has detected as spots. You can see that it is not flawless, there are 2 false spots and 2 small spots that are not detected - overall the precision is good and will certainly be able to improve the average quality of the oranges that are not sorted.

Another example is inspection of woven fabrics. Textile weaving typically takes place in low-wage countries, but even though wages are low, it is still difficult for humans to not make mistakes if you have to look at woven fabric all day. On the other hand, automated quality control can take place 24 hours a day and with proper implementation is probably more accurate and robust than a human being.

Below are detected errors from a system developed by people from the Haute Alsace University of France. The red markings show different kinds of errors that the system detects on black woven fabric. In their experiments, even their system proved very effective, 100% of defects above 1.2mm were detected. Given that it was back in 2012 and that the development is going extremely fast, defective detection in woven fabrics must now be considered a solved problem.

It is not only in relation to algorithms that the progress is amazin, there are also many new sensors and special kinds of cameras that come to market.

One of the most important innovations in quality control is the 3D camera. You may know it from Microsoft's Kinect extension to the Xbox, and thanks to this kind of mass-produced device, 3D cameras have become so cheap that at least the price doesn't hold one back.

In addition to recording the color of each pixel in the image, a 3D camera can also capture the depth. It is done with a special infrared pattern that you cannot see with the naked eye, but which the camera captures and which can be used to provide depth in the image.

The image below shows an example of how it can be useful. It is from the production of apples, and the camera is used to detect if one or more apples are missing in a box.

I have here chosen some examples where cameras are used, partly because it is more interesting to look at - but you can of course use machine learning for many other types of sensors.

Some other interesting sensors for use in quality control are:

• Vibration sensors to detect unwanted frequencies in gearboxes, motors and more.

• Ultrasonic sensors for detecting cracks in metal and ceramics.

• X-ray photography to detect foreign matter in the food industry.

• Laser measurements to measure shape and surface of products.

• NIR spectrometer for measuring food, e.g. maturation level for apples.

Using these intelligent techniques in production offers advantages in terms of workload and predictability in the process - but it also offers some interesting possibilities in that the processes by being digitized can also communicate the current status for use in collection of data. This means that the company's processes and monitoring can be done in real-time and information about defects can be communicated quickly to the right decisionmaker.

By integrating with the company's other systems and possibly also developing new applications where you can continuously get charts and an overview of the most important parameters, you can enable completely new and much more flexible processes.

Therefore, I also see quality control as one of the really exciting use scenarios for machine learning, it will potentially be a step on the way to brand new ways of running a production company.