Industrial inspection is on the cusp of a revolution. In conventional “rule-based” inspection, an expert decides which features — for example, in the case of images, common features include edges, corners, contours, patches, etc. — are relevant for inspection, and then creates a rule-based set of inspection criteria.
With advances in machine learning, especially deep learning networks, this landscape is about to change fundamentally. Instead of relying on a human expert to design features and create rules, the user instead collects a variety of samples, and trains a deep learning network. The network can be easily and quickly deployed for immediate use.
As with most claims and promises about AI, the reality is more complex. Real industrial inspection applications pose many difficult challenges, even for AI. As one example, training data is rarely plentiful, and when available is highly imbalanced (usually not enough defective samples). Reducing both false negatives and false positives is of utmost importance, as is reducing training times. Achieving real-time speeds typically requires extremely light and sparse networks, which limit their expressiveness. Finally, the human labor involved in training and setting up a network should be minimal.
What is often left unsaid are the many important caveats for the above paradigm to work, and the many difficult challenges posed by real industrial inspection applications.
Accuracy
The rate of both false negatives and false positives must be kept to less than 1~3% for typical industrial inspection tasks.
Speed
Inspection of a part or component must be completed in 10~20 msec or less.
Flexibility
New defect classes, and additional training data, must be incorporated into the network quickly and efficiently, without retraining from scratch.
Scalability
The inspection system should be modular and extensible for easy integration and expansion
Our philosophy at Saige is to focus on the data. The data — whether it be RGB images, x-ray, CT, MRI or other imaging data, or whether it be time series data like electrical or acoustic measurement signals or vibration patterns — should dictate everything. Real-life data is insufficient, noisy, incomplete, error-filled and only partially labelled (if at all). The advanced proprietary tools developed by Saige for cleaning, augmenting, and labelling the data provide a powerful edge over the competition; they are the key to making AI inspection work even in the presence of noisy and limited data.
A second key element of our approach is optimization. Our model reduction and network compression algorithms are key to meeting the stringent speed requirements of real-life inspection tasks. Our optimal transfer and directed attention learning technology can speed up training by orders of magnitude compared to conventional methods. Our team is completely invested in keeping up with the latest advances in machine learning, and how to properly leverage these advances to the application domain.
Finally, the cost of human labor needed to train, deploy, and maintain a deep learning solution should be minimal, and that the solution should be scalable and easily integrated into any existing legacy system or infrastructure. The ROI for deploying an AI solution is of critical importance.
With the lessons learned over years of experience working closely with equipment manufacturers, suppliers, and especially our customers, Saige understands what works and what doesn’t, how to honestly evaluate a customer’s return on investment in AI, and how to provide the right solution at the lowest cost.