Automotive

AI Machine Vision for Hard-to-Detect Defects in Injection Molding

Deep learning detects the scratches, gas marks and stains that rule-based vision cameras missed, while production and equipment data predict quality and optimize process conditions. The process defect rate fell from 8% to 4%.

Industry
Automotive
Solutions applied
v2a2
  1. Problem

    The rule-based vision cameras on the injection line could not detect scratches, gas marks or stains, so a separate visual inspector was stationed to re-check every part. Inspection became the bottleneck, and with no visibility of optimal production conditions there was little scope to predict defects in advance or reduce cost.

  2. Data collected

    Manufacturing and production data from the injection equipment were collected alongside image data from the vision inspection system. A data assessment was carried out first so that quality inspection results could be read against process variables.

  3. Modules applied

    V² (AI machine vision) and A² (quality prediction and process optimization). XGBoost was used for quality prediction and a genetic algorithm for process optimization. PaDiM was applied for vision anomaly detection, with Efficient-AD added to address PaDiM over-detection.

  4. Result

    Defects the rule-based cameras had been missing are now detected automatically, easing the burden of duplicate visual inspection. Correlation analysis identifies the factors affecting inspection results automatically, so the conditions that drive defects can be seen in the data and adjusted.

Why this process

The hardest defects on an injection line are the ones that are difficult to define as defects at all — a fine scratch that appears and disappears with the lighting angle, a stain sitting on the specification boundary.

That is where rule-based vision runs into a dilemma. Tighten the threshold and good parts get rejected; loosen it and defects escape. In the end a person had to look again.

Detection and prevention together

However good the inspection, the defect rate itself does not fall — you are only filtering out what was already made. So both sides were addressed.

Detection — An anomaly detection model trained on images of good parts (PaDiM) found the defects rule-based inspection missed. PaDiM’s weakness was over-detection, which Efficient-AD was added to correct.

Prevention — Production and equipment data were combined into a quality prediction model (XGBoost), and a genetic algorithm derived optimal process conditions. Once you can see which combinations of conditions raise defects, you can change the conditions themselves.