Injection Molding

Injection moulding defect detection is no longer constrained by sensor technology. Cavity-pressure sensors, machine PLCs, and downstream vision systems each generate the data needed to predict and detect every common defect type. The constraint is data plumbing — the absence of a part-level record that links every vision frame to its cavity-pressure trace and machine setpoints at the moment of production.

The Six-Phase Cycle and Its Control Variables

Resin drying at 60–120 °C for 2–6 hours removes moisture that would otherwise cause splay and degrade polymer molecular weight. Plasticisation at 200–320 °C barrel temperature with 50–200 rpm screw speed melts and homogenises the charge. Injection at 50–200 MPa fills 95–99% of the cavity volumetrically in 0.5–5 seconds. Hold pressure at 60–80% of peak injection pressure packs out volumetric shrinkage. Cooling at 30–120 °C mould temperature solidifies the part to ejection temperature.

The four critical control variables are melt temperature (±2 °C tolerance), cavity pressure (±5% at peak), fill time (±0.05 s), and hold pressure (±2%). Of these, cavity pressure is the most information-dense — it encodes the entire fill and pack history of that specific shot, and its deviation from the validated baseline is the earliest leading indicator of an impending defect.

What Sensors Detect and How

CNN-based vision models provide high-accuracy detection across the common defect types: short shots, flash, sink marks, burn marks, warpage, weld lines, splay, and contamination. These are geometric defects that appear in the finished part and are detectable by optical inspection — but only after the part is produced.

ANN models operating on in-mould temperature and cavity pressure data achieve high defect prediction accuracy — predicting the defect before the part is ejected. A multivariate cavity sensor combining piezo pressure, IR thermopile temperature, and derived melt viscosity can predict part mass, wall thickness, and tensile strength from the in-cycle sensor trace. The prediction happens while the part is still in the mould. The vision system confirms the prediction after ejection.

The two systems — cavity-pressure prediction and vision confirmation — are complementary, not redundant. Cavity pressure catches process-origin defects before they become parts. Vision catches any defects the process model missed and provides the geometric ground truth needed to validate and retrain the process model.

The Cost-at-Point Escalation Problem

Defect cost at the press is approximately 1×: one part, one cavity, one shot worth of cycle time. At assembly, where a defective component causes a downstream assembly to fail or requires disassembly and rework, the cost is 5–10×. At the customer, where a warranty return or field failure occurs, the cost is 50–100×. The arithmetic makes a compelling case for catching defects at the press rather than at final inspection or in the field.

The operational reality is that most injection moulding facilities catch the majority of defects at final inspection or at the customer — not at the press. The gap between press-side detection and final-inspection detection is the 5–10× cost multiplier that AI-linked inline inspection is designed to close.

Where Injection Moulding Shops Get This Wrong

The most common failure mode is running vision and cavity sensors as independent systems. The vision system generates a defect alert. The cavity-pressure system generates a process alert. Neither system knows what the other saw. The engineer who wants to understand whether the sink-mark increase on Cavity 3 was caused by a hold-pressure drift, a late velocity-to-pressure transfer, or check-ring wear has to manually correlate timestamps across two systems. That investigation takes hours to days, and produces a hypothesis rather than a verified cause. In the meantime, the rejected lot has already shipped or is sitting in quarantine accumulating cost.

The second failure mode is the ~5% sampling rate problem. Human visual inspection of injection-moulded parts covers approximately 5% of production volume on most lines. A defect that affects 3% of shots — easily escapable through 5% sampling — produces thousands of defective parts per shift before it is detected. Inline vision running at high coverage rates detects the defect rate change at the first statistical excursion, not after a statistically significant defect cluster reaches inspection.

What AI-Linked Inspection Changes

The part-level data record is the enabling infrastructure. Every vision frame linked to the cavity-pressure trace and machine setpoints from that specific shot, indexed by part ID, cavity number, and production timestamp. When a defect appears, the RCA query — what changed in the cavity-pressure trace in the 20 shots before the defect appeared? — is answered in seconds rather than requiring a manual cross-system investigation.

New defect types are catalogued automatically. When a defect morphology appears that doesn't match any existing category, it enters the defect library with its visual signature, the cavity-pressure trace at the time of detection, and the machine state — so the institutional memory of every defect the facility has encountered accumulates in a searchable record rather than in the working memory of whoever was on shift when it happened.

For the ten most common injection moulding defects with their physical mechanisms and structured diagnostic sequences, see Ten Injection Molding Defects and Their Root Causes. For the parameter interactions that seed these defects during process drift, see Why Single-Variable Thinking Fails in Injection Molding Process Optimization.

References

1. 1. Chen, J., Guo, Z., & Wang, Y. (2020). Machine learning prediction of injection molding defects from process parameters. International Journal of Advanced Manufacturing Technology, 110(7–8), 2023–2033. https://doi.org/10.1007/s00170-020-06011-4
2. 2. Gordon, G., et al. (2015). Multivariate cavity sensor for injection molding quality prediction. International Journal of Advanced Manufacturing Technology, 78(9–12), 1381–1391. https://doi.org/10.1007/s00170-014-6706-6
3. 3. Liu, X., et al. (2024). Deep learning-based vision inspection for injection molded parts. Scientific Reports, 14, 30751. https://doi.org/10.1038/s41598-024-81430-x
4. 4. White, R.T., et al. (2024). Bridging the gap: in-line quality control for battery manufacturing. Frontiers in Manufacturing Technology. https://doi.org/10.3389/fmtec.2024.1392038
5. 5. Lee, H., & Ryu, K. (2020). Real-time injection molding process monitoring using artificial neural networks. Applied Sciences, 10(22), 8171. https://doi.org/10.3390/app10228171