Weld Seam Anomaly Detection by Neural Networks – What’s the Story?
The breaking news from Mapvision last week was the official launch of the latest application for the Mapvision Multi-Camera Technology: Weld Seam Inspection.
In case you missed the news, you can read more about Mapvision Weld Seam Detection in our previous article and watch a recording of the Weld Seam Inspection launch webinar.
The launch was a success and has resulted in a flow of questions, ranging from solutions availability to many technical details and practical questions from the field. It seems that there are plenty of issues with the existing old-school weld inspection products and a lot of demand for something new!
In this article, we give more information relating to the most typical questions we have been getting (and repeat some of the basics). We are hoping this will clarify how Mapvision Weld Seam Inspection works. After all, we are talking about a completely new way of inspecting weld seams so it is easy to understand that you may need to unlearn some of the principles of how weld seam analysis has traditionally been performed.
We inspect from images.
As with all Mapvision solutions, we do not inspect the physical part. Instead, we grab a lot of images of all parts and then create a model – a visual digital twin – from which we measure from image details that are in the specification. In the case of other applications, such as dimensional measurements and presence checks, these can be distances, angles, and nuts and bolts. In the case of weld seam inspection, it naturally is weld seams we check from the images.
In the Mapvision Weld Seam Inspection solution, there are no scanning sensors, robots, or any other moving parts for the inspection (the cameras are stable in a rigid frame). Questions about characteristics such as sensor travel speed are therefore not relevant at all. We use some time for image grabbing. This is done only once, measuring all seams at the same time. After that, everything else is analytics and software magic.
No measurements – really!
The major difference between our previous applications and this new one for weld seams is the use of artificial intelligence (more specifically deep learning and neural networks). The system learns from data, by looking at a large number of weld seams, and detects anomalies. We do not do or get numerical measurements from the weld seams, so essentially we do not classify different defects. We do provide a notification for seams that are out of the ordinary range of variation, based on the visual appearance of the seam.
So, if you ask about the accuracy of the system or how small differences the system can detect (in millimeters), we really can’t tell. The sensitivity of the system can be adjusted to avoid false positives, but you will not get numbers. We appreciate that is a bit difficult to grasp in the engineering world, but it can be discovered empirically by giving it different inputs and by using artificial intelligence. On the other hand, you don’t get precision of micrometers from human inspectors of weld seam either. WSI in this regard is more human-like in behaviour.
That said, we are investigating the possibilities of combining dimensional measurements and anomaly detection so that we could in future releases also measure for seam lengths.
Defect categorization when we have enough data.
Mapvision Weld Seam Inspection supports externally visible defects such as seam presence, skip/gap, burn-through, craters, and porosity. The 1.0 version of the solutions do not, however, categorise the defects yet. The main reason for this is that we simply have not gathered enough data about all the typical defect types yet.
Defect categorization is on our development roadmap, but we need to get more data from real weld seams into our cloud weld database. Once we have managed to collect enough samples, defect categories will be introduced. This will also make the training of new systems quicker because we can then rely on universal data when building the current system models.
Higher-resolution cameras are required.
The new solutions are available now for all new Mapvision systems and as a software upgrade to systems with the so-called WSI hardware readiness pack already installed. Unfortunately, the reverse side of this is that it will not be available for old Mapvision systems. The reason for this is that weld seam inspection requires higher resolution cameras than the ones in our standard systems.
Can we upgrade the cameras then, you may ask? In principle yes, but it would be quite a costly operation that would require a lot of work on location. For now, we have decided that this option will not be offered. We will concentrate on new systems instead. If you think this would be an absolute must for your older Mapvision quality inspection system, let your account manager know.
Want to know more, please get in touch!
