X-Ray and AI Judge a 16,000-Ton Rear Floor in 90 s
Integrated die casting merges several parts into one, and the rear floor is formed on a 16,000-ton class machine, large and complex. The process leaves porosity and shrinkage inside the material. Appearance inspection cannot see them, and if they reach a later process or a complete vehicle they become a strength risk that only appears in service.
Internal porosity needs a different physics, not better lighting
Surface defects are imaged with visible light; internal defects require penetration. Porosity and shrinkage have lower density, so more radiation reaches the flat panel detector and the defect appears as a gray anomaly. This is not a visible-light system with adjusted settings; the method moves to another physical dimension, and the chain from source and detector to shielding and interpretation is rebuilt.
One exposure recipe cannot serve a wall that changes thickness
The rear floor varies widely in wall thickness. With one recipe, thick zones may be under-penetrated and defects drown in noise, while thin zones overexpose and flatten the gray difference. The cell sets exposure parameters by thickness zone so thick areas penetrate fully and thin areas do not saturate. The price is a more complicated imaging sequence and more time reserved inside the takt.
Low-contrast defects are bought with annotated samples
Porosity and shrinkage have low contrast in a radiograph, and the background is a complex internal structure. AI capability comes from training on a large set of annotated samples, covering defects of different sizes and positions as well as the many forms of normal structure; otherwise the model learns to call normal variation a defect. Annotation of radiographs needs specialists, and that investment is often underestimated.
Full inspection forces automatic interpretation
People read images slowly and fatigue quickly, which cannot sustain 100 percent coverage inside the takt. AI locates and classifies internal defects, outputs type, position and size, and decides pass or fail. Consistency is also a precondition: every part is judged by the same standard, so results are comparable. Loading, exposure, interpretation and output must all fit within 90 seconds, and pipelining exposure with interpretation is what keeps a single part inside the cycle.
The shielding door is part of the takt
X-ray inspection must meet radiation protection requirements, and shielding structures conflict with inline flow: doors open and close, parts enter and leave, and every movement takes time. This cell optimises the shielding structure within safety standards to cut auxiliary time, which is one of the foundations of inline operation. A cycle estimate that counts exposure and interpretation but ignores shielding motion will be optimistic.
Data has to return to the process to close the loop
Inspection data links defect information to casting batch and mold number, so process engineers can locate specific parameters. Defect distribution trends indicate the direction of process drift; more porosity in one region may point to temperature or venting there. Only at that level is inspection more than screening bad parts and actually part of process improvement. The 100 percent internal check runs inside the line cycle, with specific figures subject to site acceptance results.
Selection starts with part thickness and the material's attenuation to X-rays, which set source and detector configuration. Next comes whether AI interpretation can complete the full read within the takt. Finally, count the time taken by shielding and logistics. The route suits metal castings, forgings and weldments with high unit value, complex structure and strict internal quality requirements; where sampling is still acceptable, an inline X-ray system may not pay back.
Three questions X-ray cannot avoid
Can X-ray find every internal defect?
No. It excels at density-driven defects such as porosity, shrinkage and inclusions. Crack-like defects depend on orientation relative to the beam and may be missed when unfavourably aligned.
Is AI interpretation more accurate than a human reader?
Each has limits. AI wins on speed and consistency and suits full inspection; people win on rare morphologies, so the sound approach is AI plus human review.
Why is data return mandatory?
Because casting parameter changes need defect evidence. Judging pass or fail alone leaves inspection at the gate; returning defect type, position and count to process engineering closes the loop from inspection to process action.