Development of validation standards for deep learning models in the context of industrial modelling and optimisation of production lines for defect detection in industrial facilities

High-speed visual inspection in modern manufacturing suffers when laboratory metrics fail to predict f ield outcomes, causing waste, rework, and safety risks under domain shift and tight cycle-time budgets. The aim of this study was to establish an industry-oriented validation standard for deep-learning defect detection that is explicitly tied to production risk and deployability, with emphasis on operation at very low false-positive rates, high-percentile latency limits, and reproducible procedures for robustness, calibration, and explainability. The methodological design combined analysis of prevailing practices with a digital-twin evaluation of high-speed inspection streams, linking laboratory trials to an anonymised production line. Results dominated the findings: recall measured at very low false-positive rates showed the strongest absolute correlations with field-proximal outcomes (≈ 0.85 -0.88), partial area over the 0-5% false-positive band ranked second, and overall accuracy was weak (≈ 0.38-0.42). Latency acted as an acceptance gate: a configuration near 75 ms (p95) and 100 ms (p99) achieved either maximum realised throughput (≈ 810 parts per minute) or minimum errors, whereas a slower configuration at 141 ms and 171 ms coincided with ≈ 540 parts per minute and higher error counts. Structured validation reduced false positives from 84 to 66 per ten thousand units (-21.4%) and false negatives from 61 to 49 (-19.7%). Under controlled operating shifts, recall fell by 12.9-16.2% without adaptation and by 4.1-6% with threshold-only online calibration; the harmonic mean of precision and recall fell by 8.7-11.5% without adaptation and stabilised to 3.1-4.4% with calibration. Reproducibility and integration were evidenced by coefficients of variation of 2.1 2.4%, inter-operator threshold variance of 0.07-0.09, explainability compliance of 92.8-94.1% of batches at an overlap threshold of 0.5, service uptime of 99.6-99.8%, and deterministic rollback. The resulting standard yields practical value by specifying acceptance targets (prioritise recall at very low false positive rate; enforce p95 ≈ 75 ms/p99 ≈ 100 ms to sustain ≈ 810 ppm), prescribing lightweight online threshold calibration to cap shift-induced losses, and supplying auditable key performance indicators (variation bounds, explainability compliance, uptime, rollback) for commissioning checklists and supplier contracts to reduce deployment risk and life-cycle cost