The current QASS measuring instruments are all equipped with pattern recognition methods.
The software searches for known patterns in the data streams delivered by structure-borne sound and other sensors.
This allows different signals to be recognized by their signatures.
With many machine emissions it can of course be difficult and tedious to obtain the correct patterns.
Especially in bending straightening, one often has to wait a long time for a sufficient amount of crack signals, which then appear distributed over thousands of components.
So it is obvious to use methods from machine learning.
In order to obtain new patterns as effectively as possible, we use a specially developed signal clustering method.
First of all, we determine all noises that can be meaningfully distinguished from the background noise of the process.
The QASS Optimizer4D remembers the respective noise section of the emission.
Then we generate a sample of each unknown noise and compare their similarity to each other.
This gives us a map of all their similarity distances, which we then use to group them by degree of similarity, cluster them.
We can either specify a similarity level or a number of clusters.
The signals, each of which is grouped into one cluster, ideally represent the typical manifestations of a particular process.
We then combine the different manifestations of the cluster signals into a representative pattern, we calculate a mean pattern that fits all signals of the cluster as well as possible.
Our golden sample.
All these calculations are carried out by special operator networks, graphically programmable analysis modules, within the automatic measurement sequence on the QASS Optimizer4D hardware.
When new representative patterns have been calculated, the user must classify them as machine noise or crack patterns.
Often this is done in close cooperation between QASS application engineers and the QA departments of the customer. If the signal situation is uncertain, a laboratory analysis of the components concerned is also gladly carried out. Here it is sufficient to test signals that the system or the user believes to be machine noises.
Crack signals should in any case lead to the rejection of the component.
We have had good experience with our new optimization concept.
The Optimizer4D equipped with pattern recognition software collects data in the straightening process for about 2 weeks (while the recognition works with the previous sensitivity). Afterwards we subject the found “crack signals” to the clustering process and get the different noise types of the straightening process.
In the course of optimization, noise types are then stored in the pattern recognition system, which can then immediately recognize them when they occur and no longer misinterpret them as crack signals.