How AI Identifies Previously Unknown Defects
In industrial environments, anomaly detection (also referred to as anomaly identification or anomaly recognition) enables automated surface inspection by identifying significant deviations from the expected normal condition.
Unlike other AI approaches that require predefined defect classes, anomaly detection operates without prior knowledge of specific error types. This simplifies implementation, especially in dynamic or highly variable production settings.
Machine Learning
Supervised Learning
Goal: Classification or regression based on labeled data
Common Use Cases in Industrial Quality Control:
Unsupervised Learning
Goal: Discovering patterns or structure in unlabeled data
Common Use Cases in Industrial Quality Control:
The model is trained exclusively on defect-free examples and, in doing so, develops an internal representation of the standard product’s appearance. In practical terms, the model learns what is considered “normal”. Any deviations from this learned baseline are flagged during operation as potential anomalies.
Anomaly detection offers a quick entry point into automated quality inspection, as it does not require extensive defect annotations. However, in industrial environments, it quickly reaches its limits – especially when high precision or detailed error classification is required.
Easy to Implement
Because no defect annotations are needed, the manual effort is significantly reduced. In addition, defect-free parts occur far more frequently than defective ones in most production environments.
Detects Unknown Defects
An anomaly detection model can reliably identify previously unknown defects, as it detects unusual patterns without relying on predefined defect classes.
Lower Precision
In complex image data, segmentation models are often more efficient than anomaly detection. Anomaly-based systems tend to flag even minor, non-critical deviations as defects, which can result in unnecessary false rejects in practice.
Less Informative
Compared to anomaly detection, segmentation provides more detailed insight into production quality by enabling analysis of specific defect types and locations.
Anomaly detection and segmentation each have specific strengths and are most effective when combined in a unified quality control strategy.
While anomaly detection operates without the need for annotated defect data and is valuable for identifying new or unexpected deviations, segmentation, classification, and object detection models provide more precise informationabout the type and location of known defects.
At Maddox AI, we strategically combine both approaches:
This ensures that no relevant deviation goes unnoticed – even as products, processes, or defect types evolve over time.
“Anomaly detection can be used as a standalone solution or as a complement to existing inspection processes. It helps our customers identify unknown defect patterns at an early stage. This allows datasets to be continuously expanded and system quality to be improved in the long term.”
ML Engineer
Anomaly detection identifies deviations from the “normal state” without requiring prior knowledge of specific defects. It is particularly well-suited for detecting previously unknown defect patterns.
Typically, defect-free product data without annotations is sufficient for training (unsupervised learning), which keeps labeling effort low. However, a small number of labeled defect images is helpful for calibrating the model’s sensitivity and validating its performance.
In most industrial applications, anomaly detection is used as a complementary technology within a broader inspection system. It provides valuable insights into unknown deviations, but it generally does not replace models trained on manually labeled defect data that are optimized for precise classification of known error types.
We are ISO 27001 certified.
