How to integrate full AI judgement automation on your SMT line

AOI inspection know-how are built on experience, attention to detail and skilled decision-making. With AI judgement automation, that expertise doesn’t disappear - it becomes part of a continuous support system that strengthens both quality and daily workflow.

Because AOI inspection verification is a critical part of any SMT line - and it relies on skilled people making the right decisions, every time. However, the work is repetitive and demanding, even for the most experienced inspectors. 

By training AI on real inspection data and decisions, you preserve the knowledge your team builds over time and make it accessible to everyone on the line. The result is a stronger, more consistent process where inspectors can focus on what they do best, while AI takes care of the repetitive evaluation. 

It’s about creating a smarter workflow, improving quality, and making everyday work more engaging for the people who keep your production moving.

One effective way to integrate Yamaha AI judgement automation is directly at the workstation level, where inspectors already review assemblies rejected by AOI. Here, the AI becomes part of the secondary judgement process - supporting, not replacing, the existing workflow (1st level).

When a PCB board is flagged as NG, experienced inspectors review the AOI images and data to confirm the judgement. If a false NG is identified, the board can be returned to production, while confirmed defects are routed for rework. Every decision made - both human and AI-supported - is captured and stored.

This is where the learning loop begins. The AI uses real inspection decisions to train its models continuously, improving its understanding of patterns, defects and issue cases over time. As confidence levels increase, the system gradually becomes capable of assisting in live inspection scenarios, offering a second opinion when needed.

Once the AI models reach a stable and high-confidence threshold, AI can begin supporting specific component types or defect categories with automated judgement. At this stage, the strength of the system lies in the combination: human experience guiding complex cases, and AI ensuring speed, consistency, and reduced operator fatigue in repetitive decisions.

For less experienced inspectors, this support is especially valuable. The AI Judgement Station can present a clear recommendation together with a confidence score, helping operators make more informed decisions while building their own expertise over time. The result is a stronger inspection process - where learning, quality and throughput improve together on the SMT line.

Software applications for AI-based inspection are now entering the SMT market and are already proving highly effective in identifying soldering defects such as insufficient solder or poor wetting from 2D AOI inspection data.

Solutions like Yamaha Robotics’ AI Judgement Station are designed to give manufacturers control over the entire AI learning process. Instead of relying on predefined models, users can build and train their own AI directly from real production data on the line.

The process starts by selecting representative OK and NG samples from existing AOI inspections. From here, the software automatically generates initial AI models, which are then trained using data collected continuously from daily production shifts. As more images are added, the model becomes increasingly refined, and users can monitor accuracy in real time to determine when performance stabilises.

• In practice, even a relatively small dataset - around 30 NG images combined with OK samples - can be enough to begin generating usable models. 
  
  
• With continued learning, performance improves steadily, and after processing larger datasets (for example 1000+ images), the model can reach a very high level of accuracy.

Once trained, AI judgement operates extremely fast, typically delivering inspection feedback in around one second per case. This makes it a practical support tool for live production environments, where speed and consistency are critical.

One area where AI already demonstrates reliable, human-like judgement - and directly reduces inspector workload - is OCR (optical character recognition).

In traditional rules-based machine vision systems, character reading can quickly become a challenge. Low contrast, surface scratches, contamination, or partial obstruction from foreign material often lead to misreads or require manual verification. These conditions are common in surface-mount assembly environments and can directly affect the readability of critical information such as component values, date codes, polarity markings, and board identification.

AI-OCR can read markings that otherwise would require visual inspector confirmation.

AI-based OCR approaches this problem differently. Instead of relying solely on fixed rules, it builds on learned patterns - making inferences in a similar way to human operators when text is partially damaged or difficult to read. This allows the system to interpret uncertain or degraded markings with a higher level of robustness.

The Yamaha AI Judgement Station package includes pre-trained AI-OCR models that can be used directly out of the box, without the need for additional training or fine-tuning. This enables manufacturers to introduce AI-supported character recognition quickly into existing inspection workflows, improving consistency while freeing up inspectors to focus on more complex judgement tasks.