Visual inspection &
AUTONOMOUSLY CONFIGURING AND REAL-TIME LEARNING PRODUCT
Avis high speed, real-time, inline inspection system
SYSTEM: ADD-ON FOR SEMI-AUTOMATED INSPECTION SYSTEMS
CONTAINER TYPE: VIALS, AMPULES, CARTRIDGES, SYRINGES, AND BOTTLES
PRODUCTS: WATER-LIKE, FREEZE-DRIED, SUSPENSIONS AND OILY
TIME TO CONFIGURE: < 60 MINUTES
SPEED: 1,980 PIECES PER HOUR
Less Than 60 Minutes to Configure
Record: Run 700 pre-inspected defect-free units through AVIS and create a recording.
Train: Select recording for AVIS to train. During this step AVIS builds a machine learning model to represent normal variation of the product in <60 minutes. This inspection recipe can then be saved for future use.
Run: In run mode, live camera feed is combined with the previously created inspection recipe. If a video frame shows variation, the anomalous area is highlighted and displayed and the defective unit is rejected.
Detect all defect types
Glass, fiber, and metal particles are a few of the many kinds of particulate often found in pharmaceutical parenterals. These critical defects, though seldom occurring, arise in even the best of production environments. AVIS replicates the human ability to learn normal variation and detect defects without prior knowledge of what to look for.
Scratches, glass bubbles, outer markings and other cosmetic defects can occur at various stages of the production process and reduce total product quality. AVIS is able to learn normal variations in good product such as bubbles in solution, variance in container thickness and fill levels. Defects outside of these normal variations are flagged and rejected allowing product quality to remain high while minimizing false rejects.
Under and overfill, missing seal, cross-threaded cap and many other flaws in production can jeopardize the integrity of a containers closure. Closure integrity defects occur infrequently and have diverse appearance making them hard for traditional approaches to learn and identify. Like any other defect, flaws in closure integrity deviate from a product’s normal variation enabling AVIS to identify and reject them.
Any Product and Container
- Run 500 defect-free units through AVIS and make high-definition recording in which each frame is divided into thousands of subcells.
- Assign Each subcell assigned its grayscale histogram of magnitudes … (HOM).
- Each HOM is clustered based on similar histograms for a given vertical region of the frame.
- The size of each cluster is the number of histograms assigned to it during training. Most clusters are very large (millions of members) and their anomaly index is close to 0.
- Some clusters are very small indicating a histogram that is very uncommon, and their anomaly index is large (close to 1000).
- By setting an anomaly index threshold we determine the sensitivity of the defect detection, balancing false accepts against false rejects. We detect defects as any unit having a subcell with an anomaly index greater than the anomaly index threshold