法规解读| ECA关于灯检可见异物的FAQ

问答指南聚焦9大主题

What are the differences between the qualification test set, the particle test set (Knapp test), and the function test set?

确认测试集、粒子测试集（Knapp测试）和功能测试集之间有什么区别？

The three test sets serve different purposes:

三个测试集用于不同的目的：

• Function test set: Used for routine functional testing before batch inspection (and preferably after batch completion) to confirm correct system functionality. It typically contains easily detectable defects and is not intended to challenge the detection capability of the inspection system.

功能测试装置：用于批检验前（最好是批检验完成后）的常规功能测试，以确认系统功能正确。它通常包含易于检测的缺陷，并且不打算挑战检测系统的检测能力。

• Qualification test set (non-particulate defects): Used during performance qualification to demonstrate the detection of deterministic defects (e.g. missing stoppers or cosmetic defects). Qualification may be performed against predefined acceptance criteria and therefore does not necessarily require the inclusion of conforming units.

确认测试集（非颗粒性缺陷）：用于性能确认期间，以证明对确定性缺陷的检测（例如缺少瓶塞或外观缺陷）。确认可以根据预先确定的验收标准进行，因此不一定需要包括合格品。

• Particle test set (Knapp test): Used during performance qualification to demonstrate particle detection capability. Since particle detection is typically evaluated by man-machine comparison, the set contains both particulate defects and conforming units to allow comparison of the overall inspection performance.

Should qualification and function test sets be prepared artificially or should they represent real production defects?

是否应该人为地准备确认和功能测试集，或者它们应该代表真正的生产缺陷？

Both approaches are acceptable. Whenever possible, test sets should contain representative defects originating from the manufacturing process, as these best reflect real production conditions. Where this is not feasible, artificially prepared defects may be used, provided they are representative of the expected defect characteristics.

两种方法都是可以接受的。只要有可能，测试集应该包含来自制造过程的代表性缺陷，因为这些缺陷最好地反映了真实的生产条件。如果这是不可行的，则可以使用人工制备的缺陷，只要它们是预期缺陷特征的代表。

Test sets may be prepared either in-house or by specialized external providers. However, product- and process-specific defects should be representative of the manufacturing process and appropriately justified.

What aspects should be considered when assessing artificially created test kits, and how should their suitability be justified?

在评估人工创建的测试套件时，应该考虑哪些方面，以及如何证明它们的适用性？

Artificially created test kits are often used when representative production samples are unavailable, technically unsuitable, or cannot be obtained in sufficient quantity for qualification, training, or evaluation purposes.

当无法获得有代表性的生产样品，技术上不合适，或无法获得足够数量的确认、培训或评估时，通常使用人工制造的测试包。

Their suitability should be assessed based on the inspection technology used and the required level of representativeness. Artificially created defects and conforming units should exhibit optical, physical, and mechanical characteristics comparable to those of the real product under routine inspection conditions.

应根据所使用的检验技术和所要求的代表性水平对其适用性进行评估。在常规检查条件下，人为制造的缺陷和合格部件应表现出与真实产品相当的光学、物理和机械特性。

For AI-based inspection systems, higher requirements regarding representativeness may apply, as these systems often utilize very subtle features and characteristics to distinguish between acceptable and defective units. For example, a genuine lyophilized cake with crystalline reflective properties can generally be distinguished from a matte plaster-based imitation if these differences are visible in the image data.

对于基于人工智能的检测系统，可能会对代表性提出更高的要求，因为这些系统通常利用非常微妙的特征和特征来区分可接受和有缺陷的单元。例如，如果这些差异在图像数据中可见，通常可以将具有晶体反射特性的真正的冻干粉饼与基于哑光石膏的仿制品区分开来。

If artificially created defects differ in relevant characteristics—such as texture, reflectivity, contrast, size, movement behaviour, or other optical or physical properties—there is a risk that the neural network learns features introduced by the artificial manufacturing process rather than the actual defect itself. Likewise, particle detection may be influenced by factors such as viscosity, turbidity, fill height particle size, or particle contrast. These factors should therefore be considered when assessing the suitability of artificially created test kits and may also affect the performance of conventional rule-based inspection systems.

如果人工制造的缺陷在相关特征（如纹理、反射率、对比度、大小、运动行为或其他光学或物理特性）上有所不同，那么神经网络就有可能学习到人工制造过程引入的特征，而不是实际缺陷本身。同样，颗粒检测可能受到粘度、浊度、填充高度、粒径或颗粒对比度等因素的影响。因此，在评估人工制造的样品集的适用性时，应考虑这些因素，这些因素也可能影响传统的基于规则的检测系统的性能。

Conventional rule-based systems are generally less susceptible to such effects because they evaluate predefined image features rather than learning complex representations. Although these approaches are often less selective and may be less suitable for highly challenging inspection tasks, they typically demonstrate greater robustness against variations or realism drifts within artificially created test kits.

传统的基于规则的系统通常不太容易受到这种影响，因为它们评估的是预定义的图像特征，而不是学习复杂的表示。尽管这些方法通常较少选择性，并且可能不太适合具有高度挑战性的检查任务，但它们通常在人工创建的测试套件中表现出更强的耐用性，以对抗变化或实际的漂移。

Consequently, the suitability of artificially created test kits should always be justified and documented with regard to the inspection technology and the intended application. The decisive criterion is not whether defects are artificially created or taken from routine production, but whether the test kit reliably reproduces the behaviour of the final drug product under the selected inspection strategy.

因此，人工制造的测试集的适用性应始终根据检测技术和预期应用进行论证和记录。决定性的标准不是缺陷是人为造成的还是从日常生产中产生的，而是在选定的检验策略下，测试集是否可靠地再现了最终药品的行为。

How should test sets be maintained throughout their lifecycle?

在整个生命周期中应该如何维护测试集？

Test sets should be managed under a defined lifecycle programme to ensure that they remain representative and fit for their intended use. Individual samples may be replaced as required (e.g. due to damage, ageing, or expiry). Some defect types may remain suitable for many years (e.g. bad crimping), whereas others (e.g. particles or contaminated samples) may change over time and therefore require more frequent replacement.

测试集应该在定义的生命周期计划下进行管理，以确保它们保持代表性并适合其预期用途。个别样品可按要求更换（例如，由于损坏、老化或过期）。一些缺陷类型可能会保持多年（例如，不良轧盖卷边），而其他缺陷类型（例如颗粒或受污染的样品）可能会随着时间的推移而改变，因此需要更频繁地更换。

Where appropriate, qualification and training kits should be updated to include newly identified productor process-specific defects to ensure continued representativeness.

在适当的情况下，确认和培训材料应更新，以包括新发现的生产工艺特定缺陷，以确保持续的代表性。

All test sets should be periodically reviewed (at least annually) and formally released following successful review. In addition, their integrity and suitability should be verified before use, where appropriate. The review, any replacements, and the release of the updated test set should be documented in accordance with the site's quality management system.

What should a compliant inspection test set contain?

一个合格的检验测试集应该包含哪些内容？

A compliant inspection test set should contain representative defect samples covering all relevant defect categories together with a sufficient number of representative non-defect units. The composition of the test set should be appropriate for its intended purpose (e.g. training, qualification, or routine functional testing) and allow a meaningful assessment of inspection performance.

一个符合要求的检验测试集应该包含有代表性的缺陷样本，包括所有相关的缺陷类别，以及足够数量的有代表性的非缺陷单元。测试集的组成应适合其预期目的（如培训、确认或常规功能测试），并允许对检验性能进行有意义的评估。

As a practical example, qualification test sets used for manual inspection or man–machine comparison should predominantly consist of non-defect units, with the proportion of defect units typically limited to approximately 10% of the total test set in order to minimize inspection bias. Typically, around 20–30 representative defect samples per defect category are included for qualification purposes, whereas routine function test sets may contain fewer samples. Each defect category should cover the expected variability of the defect. For example, a crack category should include defects of different lengths, orientations, and locations, while scratches or poor crimping should represent different severities and appearances rather than only a single characteristic example. The exact composition should always be justified according to the intended application, inspection methodology, and required statistical confidence.

作为一个实际的例子，用于人工检测或人机比较的确认测试集应该主要由非缺陷单元组成，缺陷单元的比例通常限制在总测试集的10%左右，以尽量减少检测偏差。典型地，每个缺陷类别大约包含20-30个有代表性的缺陷样本，用于确认目的，而常规功能测试集可能包含更少的样本。每个缺陷类别都应该涵盖缺陷的预期可变性。例如，裂纹类别应包括不同长度、方向和位置的缺陷，而划痕或不良轧盖卷边应代表不同的严重程度和外观，而不仅仅是单一的特征示例。确切的成分应始终根据预期的应用、检验方法和所需的统计置信度进行证明。

Test sets should be blinded and randomized, where appropriate, to prevent memorization and bias. They should remain representative of the current product and manufacturing process and be maintained in accordance with the defined lifecycle programme.

☎400-888-0570

📱181-1645-6496

🌏www.zholion.com

📩marketing@zholion.com