The use of artificial intelligence offers possibilities where classical machine vision reaches its limits.
VITRONIC took a closer look at the topic of artificial intelligence (AI) in machine vision in the blog article "Machine vision gets intelligent." In this article, I discuss the integration of AI in machine vision systems and look at it using a concrete example from weld seam inspection — the detection and classification of weld spatter.
Weld spatters reduce the quality of visible surfaces or surfaces relevant for assembly. Therefore, it is essential to have an extremely reliable detection.
Classical machine vision (MV) has its limitations when dealing with very heterogeneous backgrounds and a high variation of a defect feature. Therefore, the result of classical edge detection is not always conclusive in the case of spatters. AI eliminates this disadvantage because AI autonomously detects patterns that humans themselves do not perceive or cannot quantify. In this way, AI makes MV more flexible, and the system learns from new conditions.
But how can the assessment of "non-acceptable" weld spatter be expressed in terms of the tolerance of an inspection system? To do this, experts evaluate which irregularities the AI should classify as weld spatter. Convolutional neural networks (CNNs) are trained with this pre-assessed data. Such a trained network can classify other spatters better after only a very short training phase compared to the classical approach.
When the customer buys our VIRO WSI system, it already knows 50,000 variations of weld spatter. So, the customer doesn't have to train it first, VITRONIC has already done the work for him.
It is important to have a large number of different data sets for the defect type so that later a high bandwidth of different splashes can be recognized reliably. The more input data is available, the better the network can be trained.
The training data set, or more precisely, each abnormality is assigned a label. These labels form the target values for the training.
It should be considered that neural networks are not deterministic. There is a high probability in error detection, but no uniqueness. For these probabilities, you can then set a threshold value at which classification as a spatter takes place. The lower the threshold value is set, the greater the uncertainty (100% minus threshold value). This in turn means that there can be pseudo errors in the classification even with neural networks.
Inline inspection requires compliance with the given cycle times of the production line. Therefore, in practice, the time for required for inspection is also essential.
For evaluation purposes, we have compared total inspection times of classical and AI-supported MV. Certain AI algorithms (such as SSD, Single Shot Detection) achieved almost the same speeds as classical machine vision.
Taking into account the higher accuracy of the SSD method, it is thus the better alternative for inspecting weld seams.
We are constantly working on optimizing our solutions, including the SSD for weld spatter. My focus is on further reducing pseudo defects while maintaining at least the same processing speed.
Of course, AI can be employed for more than just classifying weld spatter. In the future, VITRONIC will also use AI for detecting other defects such as pores and indentations. There is great potential to train AI on other criteria.
The use of artificial intelligence offers possibilities where classical machine vision reaches its limits.