FAQs

How can I use AI in manufacturing?

Every aspect of manufacturing stands to gain from AI, specifically the use of AI-based algorithms for quality control, predictive maintenance and management of smart devices on the shop floor.

How does AI detect defects?

Unlike traditional quality control where feature- engineering was used to define defective units, AI algorithms learn to recognize them by being shown examples. Instead of identifying detailed and specific measurements and programming your system to make the pass/fail decision based on these measurements, the algorithm learns from examples, e.g. pictures or other physical measurements, and then determines whether a new product it is shown is good or defective.

Why is defect detection with AI superior to feature-engineering?

Feature-engineering is more labor- and time-intensive than training an algorithm. This is particularly relevant in cases where the specs change frequently or even occasionally. While retraining the algorithm can be done quickly, even small changes to the product require a whole new set of features to be programmed.

What type of defects does AI detect?

Algorithms can be trained to detect any defect that can be recognized visually or other physical measurements. This ranges from large and obvious defects, e.g. a wine bottle without a cork to the small and subtle, e.g. a hairline crack in the glass of the wine bottle.

Can AI classify different defect types or classes?

Yes, when trained appropriately, AI algorithms can differentiate between different types of defects and identify defective units accordingly, e.g. wine bottles a) without corks, b) with the label upside down or otherwise misaligned, c) with cracks in the glass or d) insufficient fill volume. To correctly classify a defect, the algorithm needs an adequate “understanding” of how it looks and with that a larger training set.

What is a training set?

To learn and AI-based algorithm needs to “see” examples of both defective and good units. The set of pictures or other sensor data that is used to teach the algorithm the difference between good and defective products is referred to as training set. A training set can also be refered to as training library.

How is a training set generated?

Existing sensor data are used or collected, e.g. pictures of wine bottles before they are boxed up. These pictures are tagged good or defective or good and type of defect by an experienced team member. These tagged data are then used to train the algorithm.

What is a validation set?

After the algorithm has been trained using the training set, it is shown a set of pictures or sensor data it has never seen before. This set is used to gauge the performance of the algorithm. If the desired accuracy has not been achieved yet, additional training is needed.

How large do training sets need to be?

This depends very much on the rarity and type of defect(s) and the required accuracy. The rarer a defect, the more training data are needed so the algorithm “sees” enough examples to learn. A large, obvious defect, e.g. a wine bottle without a cork, requires only a few hundred sets of input data such as pictures, for more subtle defects, e.g. tiny scratches on a metal surface thousands of pictures might be needed. If defects have to be classified into different categories more input data are required to teach the algorithm the differences than are needed for a simple pass/fail decision. Accuracy also impacts training set size, the more accurate the answer has to be, the more training is required.

What are the quality requirements for a training set, e.g. of pictures?

A good training set contains a sufficient number of high-quality, relevant pictures. Just like a human, what the algorithms doesn’t see, it doesn’t recognize. In the case of pictures good quality speaks to resolution, light, sharpness and relevancy to what is shown. If critical features are outside the plane of focus or insufficiently lit, the algorithm will not be able to learn their relevancy.