Business Requirements for AI Systems
management expectations
There are specific use cases for AI methods. After all, it’s not a one-size-fits-all solution, it won’t solve all problems. Some applications are better suited for traditional computer vision, some may require both, and some may only require artificial intelligence. AI systems are expensive—both in terms of cost and resources required upfront. Open source tools require a lot of development time, while external tools tend to be expensive. Also, a GPU is usually required to achieve sufficient performance on the system. Many manufacturers tend not to have GPUs or equivalent processing power. Therefore, it is important to determine which applications are well suited for AI with strong business needs.
The importance of vision system settings
Before diving into AI, it is recommended to have a solid foundation in vision system setup. That’s less important for AI, though, because it can often handle worse conditions than traditional systems. All the usual machine vision system rules apply here – good lighting, camera resolution, focal length, etc. If any of these factors fall short of the mark, it’s worth going back and addressing them before diving into AI. Make sure to have a strong vision system setup for the best results.
Reference to human performance
AI systems are most successful where human performance is strong. Once the system is set up, operators can easily identify/classify images by eye so they can determine if they are suitable for AI. However, if human performance is insufficient, then the AI model is likely to underperform. Using human performance as a reference point for what an AI model can achieve, if an operator is only 70% correct in recognizing an image, it is unlikely that the AI will perform any better than that. Therefore, if human performance is not good enough for an application, that performance issue should be addressed first and brought to an acceptable level. Once the operator has achieved the desired performance, AI can be considered.
time and resources
Collecting images and training the model requires considerable effort. Often, collecting high-quality images is the hardest part, as many manufacturers have very low defect counts. Without data, it can be difficult to train a model for defective parts. Training tools are helpful, providing pretrained models that require fewer samples to train. Training is an iterative process, spanning multiple steps, to find the ideal parameters for the model to run. Optimizing a model usually takes time and experimentation. Also, if new data comes up in the field, the model will need to be trained and deployed again.
Examples of AI applications:
One example application of AI in machine vision is for final assembly inspection, another is for printed circuit board or PCB inspection.
❶ Assembly inspection:
background
Final inspection of a part/product or assembly is usually done by an operator, or a traditional machine vision system, or both. Teledyne cameras will be highlighted here as example products. Final inspections may check for bent pins, scratches on surfaces, correct placement of connectors, alignment of stickers, correct printing of text, distance between mechanisms, and more. Basically, any exceptions that occurred during the build need to be found. But then, the list of criteria that needs to be looked up quickly can become very long. Traditional rule-based systems struggled to handle all corner cases, and it was difficult to train new operators.
Why AI?
There are often too many rules to determine what is “passed”. This makes it difficult for traditional machine vision systems to achieve good performance. The alternative is that manual inspections are time-consuming for many companies and difficult to make ambiguous judgments for new operators. Traditional rule-based systems often do not have sufficient performance, and manufacturers rely on operator judgment to help. There may be different lighting conditions, as well as a high degree of variation in defect location, shape and texture. Often, only a simple “good/bad” qualitative output is required. However, this can also be combined with traditional rule-based algorithms if desired.
benefit
With AI, setup is much easier. After collecting a large number of images to train a model, getting a system to run is usually much less development work than a rule-based system, especially with AI tools. With a proper system, usually a GPU, the check is much faster, on the order of milliseconds. AI systems should also perform more reliably than humans if presented with good data, and are a good way to standardize inspection procedures. The algorithm is typically trained on data provided by multiple operators, which reduces human error. This helps reduce human bias or fatigue that a single operator can experience. In this example, AI can help manufacturers reduce out-of-the-box failures and improve inspection quality and throughput.
❷ PCB inspection:
background
PCB manufacturers need to inspect their boards for any defects. It could be a bad solder joint, a short circuit, or some other abnormality. Usually an AOI (Automated Optical Inspection) machine is used. However, it is difficult to handle all edge cases due to so many variations of defects. And the performance of rule-based systems is not accurate enough, and manufacturers will have operators perform manual inspections, which is time-consuming and expensive.
Why AI?
Traditional AOI systems have difficulty identifying defects. It either overshoots or falls short of performance, causing a defective PCB to pass or a good PCB to fail. Similar to other situations, there are too many rules to determine a “good board”. Depending on the application, AI can be used here to classify defects that vary widely in size and shape, like shorts, opens, wrong components, solder defects, and more.
benefit
With artificial intelligence, manufacturers can improve the accuracy and quality of inspections. This helps reduce the number of defective PCBs that pass inspection. It also saves time and labor costs for any human-assisted inspections, and increases throughput by automating tasks that operators take longer to complete.
GIPHY App Key not set. Please check settings