Sunday, 12 July 2026
On a foundry line, two operators look at the same metal part. Both identify a defect. Yet their conclusions differ. For one, the part can continue on its way. For the other, it must be rejected. This situation is far from exceptional in industry.
On a foundry line, two operators look at the same metal part. Both identify a defect. Yet their conclusions differ. For one, the part can continue on its way. For the other, it must be rejected. This situation is far from exceptional in industry.
Artificial intelligence applied to quality control refers to machine learning methods that detect, classify and trigger the ejection of non-compliant parts on an industrial production line, without human intervention and without interrupting line speed.
It relies mainly on convolutional neural networks (CNNs) trained to recognize defects from images captured by industrial cameras. The machine vision system analyzes each part in a few tens of milliseconds and sends an ejection signal to the PLC as soon as the probability of a defect exceeds the set threshold.
Unlike human visual inspection, limited by fatigue and line speed, an AI system maintains consistent judgment 24/7, whether detecting a 0.3 mm crack on a lidded tray, a surface defect on a nuclear part or an incorrect assembly on an automotive line.
What is AI applied to industrial quality control?
AI applied to quality control is a machine learning system that analyzes images in real time to identify any deviation from a compliant reference, then mechanically triggers the ejection of the defective part.
Two approaches coexist on today's production lines. Classical vision (hand-coded rules, fixed light thresholds) has shown its limits with variable defects: it over-rejects when the lighting changes, and lets through atypical defects that were not anticipated during configuration.
Supervised AI changes this model. A neural network is trained on thousands of annotated images, both compliant and non-compliant, and learns to generalize. It recognizes a defect even if it does not look exactly like those seen during training. That is why Psycle can deploy the same vision engine on a food line at Babynov and on a nuclear process at Orano Melox, with different models trained on data specific to each context.
Detectable defects mainly fall into four families: surface defects (scratches, stains, bubbles), shape defects (deformation, missing material), presence/absence defects (missing component, missing cap), and assembly defects (poorly sealed lid, incorrect positioning).
How does AI detect industrial defects?
AI detects defects in four sequential steps: image acquisition, preprocessing, CNN model inference, sending the ejection signal. The whole process runs in less than 20 milliseconds, compatible with production rates of several hundred parts per minute.
It all starts with optics. An industrial camera (area scan for a global image, line scan for continuous scanning) captures each part at the precise moment it passes under the lens, synchronized with the lighting trigger. Image quality at this stage determines everything that follows.
Preprocessing normalizes lighting variations and crops the region of interest. The image then passes through the CNN running on a GPU processor at the edge of the line. The model calculates a defect probability for each known class: scratch, bubble, deformation, missing lid.
If this probability exceeds the decision threshold, the system sends a digital signal to the PLC within the following milliseconds. The ejection arm or pneumatic reject acts. The part leaves the line without stopping production.
Step | Action | Component | Typical duration |
|---|---|---|---|
1. Acquisition | Image capture synchronized with line speed | Industrial camera + dedicated lighting | Variable, depends heavily on the sensor used |
2. Preprocessing | Luminance normalization, ROI cropping | Embedded CPU or GPU | 1 to 5 ms |
3. CNN inference | Calculation of defect probability per class | Embedded GPU | 5 to 15 ms |
4. Decision & ejection | PLC signal if score > threshold | PLC / robot controller | < 2 ms |
Which metrics evaluate the performance of an AI quality system?
Three metrics quantify the performance of an AI quality system: precision measures the share of defects among ejected parts, recall measures the share of real defects actually detected, and the F1 score combines the two. It is the confusion matrix that reveals their imbalances.
Choosing the wrong decision threshold has a direct cost. A threshold that is too low generates false positives: compliant parts wrongly ejected. The line loses material yield. A threshold that is too high produces false negatives: defects that get through and reach the customer. Both situations have a measurable cost.
The F1 score forces you to find the balance. It cannot be maximized by ignoring one of the two terms. In practice, lines with high regulatory stakes (food, nuclear, pharmaceutical) require very high recall, since it is better to over-reject than to let defects through, with a target recall > 0.98 and an F1 > 0.95.
Psycle integrates a real-time monitoring dashboard that displays these metrics by shift, by product and by defect. This setup makes it possible to adjust the decision threshold without retraining the model.
Metric | What it measures | Simplified formula | Indicative industrial threshold |
|---|---|---|---|
Precision | Rate of real defects among ejections | TP / (TP + FP) | > 0.95 |
Recall | Rate of real defects correctly detected | TP / (TP + FN) | > 0.98 (critical lines) |
F1 score | Precision / recall balance | 2 × (P × R) / (P + R) | > 0.95 |
Overall ejection rate | Process stability indicator | Ejections / Total parts | Variable, to be calibrated per product |
In which industrial sectors does AI improve quality control?
AI quality inspection applies wherever a visual or dimensional defect has a measurable cost: food, nuclear, automotive, electronics, pharmaceuticals and metallurgy. Each sector imposes its own constraints in terms of line speed, certification and the definition of an acceptable defect.
Sector | Main application | Psycle example or specific constraint |
|---|---|---|
Food | Inspection of lids, trays, cans | Babynov (ovality and impact inspection of cans), natural variability of organic products |
Nuclear | Process monitoring, surface inspection | Orano Melox, zero tolerance for false negatives, mandatory image archiving for regulatory audits |
Automotive & mechanical engineering | Inspection of machined parts, painted surfaces, assemblies | Speeds > 100 parts/min, inference necessarily at the edge of the line (edge computing) |
Electronics | Inspection of PCBs, solder joints, components | High-resolution line scan camera, continuous scanning of each board |
Pharmaceuticals / cosmetics | Fill level, cap and label inspection | Strict regulatory compliance, mandatory batch traceability |
Metallurgy & plastics | Sheet metal surface defects, porosity in injection-molded parts | Coaxial lighting to reveal defects invisible under diffuse lighting |
What conditions ensure the success of an AI quality project?
An AI quality project succeeds if six conditions are met before deployment: representative training data, rigorous annotation, a compatible cycle time, tested PLC integration, a threshold that can be adjusted without retraining, and a model update plan.
Most failures do not come from the model. They come from data preparation and line integration. Here are the six conditions Psycle systematically checks before each deployment.
Condition | What it requires | What happens if it is missing |
|---|---|---|
1. Representative data | Images covering the real variants of defects: variable lighting, different batches, tool wear | The model drifts as soon as conditions change (e.g. winter vs. summer light) |
2. Rigorous annotation | An operator able to distinguish a real defect from an acceptable normal variation | A roughly annotated model produces rough results, whatever the amount of data |
3. Cycle time compatibility | Inference + ejection signal within the time window available between two parts (e.g. 300 ms at 200 parts/min) | The system lags one part behind, and the ejection hits the wrong part |
4. Tested PLC integration | OPC-UA protocol or digital I/O qualified before go-live | Non-reproducible reject signal, false positives on the real line |
5. Threshold adjustable without retraining | An interface exposing the decision threshold to the quality manager | The slightest drift = integrator intervention = unplanned downtime |
6. Retraining plan | Documented annotation + retraining + validation procedure | Without an update procedure, a new defect or a change in raw material makes the model obsolete |
Frequently asked questions about AI for quality control
What is quality control using artificial intelligence?
Quality control using artificial intelligence consists of using machine learning algorithms, notably convolutional neural networks (CNNs), to analyze images or sensor data in real time and automatically detect any non-conformity on a production line. AI replaces or assists the human inspector with greater repeatability and speed.
Which metrics can be used to evaluate an AI system for quality control?
Three metrics are essential: precision (the ratio of correctly detected defects to total detections), recall (the ratio of real defects correctly identified), and the F1 score (the harmonic mean of the two). The confusion matrix summarizes these results. An F1 score above 0.95 is generally required for deployment in industrial production.
How much training data is needed for a reliable AI quality system?
The minimum volume depends on the complexity of the defect and its variability. In practice, 500 to 5,000 annotated images per defect class make up an initial training set. Data augmentation techniques (rotation, noise, lighting changes) can enrich smaller datasets. The rarer the defects in production, the more critical targeted annotation becomes.
Can AI completely replace human quality control?
No, in the majority of current industrial applications. AI excels at repetitive, visual and measurable defects at high speed. It performs less well on defects that are subjective, new (not seen during training) or that require contextual reasoning. The hybrid AI + human supervision model remains the recommended standard for lines with high regulatory or safety stakes.