AI-based visual quality inspection – like the solution from 36ZERO Vision – uses a self-learning deep learning algorithm that learns and classifies real defect patterns with pixel-level accuracy. Rule-based systems work with pre-defined thresholds and statistics, where any deviation from the ‚golden sample‘ is classified as an error – which can lead to increased false positive rates. 36ZERO Vision reliably distinguishes good parts from genuine defects and achieves up to 31 times higher detection accuracy for known error types in industry. The decisive operational advantage: The deep learning platform adapts to new product variants or defect images with typically 5–20 training images, whereas rule-based systems require complex reprogramming by machine vision specialists when changes occur.

36ZERO Vision's no-code platform combines two complementary detection approaches: supervised deep learning for precise classification of known defect types – with up to 31 times higher detection accuracy – and unsupervised anomaly detection for unknown defects that have never been seen during training. This hybrid solution is particularly valuable in automated optical inspection (AOI) when new, previously unknown defect patterns emerge during ongoing series production, for example, due to tool wear or changes in material batches. The human-in-the-loop feedback system uses corrections from quality personnel for continuous model improvement, ensuring the platform remains flexible in operation. Thus, the platform covers both stable series inspection of known defects and early detection of new quality deviations.

For a production-ready AI inspection model with 36ZERO Vision, 5-20 labeled training images per error type are typically sufficient—up to 20 times fewer than comparable computer vision solutions and traditional AOI systems. Optionally, CAD design data can be used as a training basis or synthetic defects can be generated, further reducing the need for image acquisition in production—especially valuable for new products without existing defect examples. Initial model training takes about 2 hours; a first productive inspection model is ready for use after approximately 45 minutes of onboarding. Training itself takes place on the cloud platform, and the finished deep learning algorithm is then deployed as an edge model locally into the production line—without a permanent internet connection during operation.

No. 36ZERO Vision is designed as a no-code self-service platform that can be fully operated without programming or AI knowledge. Quality assurance staff can independently train, validate, and adapt AI inspection models to new product lines or error types by uploading and annotating sample images in the web-based interface. This eliminates the dependence on external machine vision experts for every product or variant change. 36ZERO Vision makes powerful deep learning-based visual quality inspection accessible to manufacturing companies of all sizes – without proprietary development environments or specialized knowledge.

The standard is a hybrid cloud-edge architecture: AI model training and management take place on the 36ZERO Vision Cloud platform (hosted by Microsoft Azure in Frankfurt), while the actual real-time inspection runs entirely locally on the 36ZERO Vision Edge client on an industrial PC – without a permanent internet connection. Image processing is thus performed autonomously in under 20 milliseconds, fast enough for high-speed manufacturing processes. This hybrid architecture combines the simplicity of cloud-based model management with the data security and latency advantages of an on-edge AI solution. For industries requiring fully on-premise model training, 36ZERO Vision also offers a specialized solution, for example in cooperation with partner DELL.