Method for performing an integrity test on a test consumable of a

Abstract

The invention relates to a method for performing an integrity test on a test consumable (1) of a bioprocess plant using a test arrangement (2), in which a predetermined test program is applied to the test consumable (1) via the test arrangement (2) during a test routine (26) performed by a test control unit (3) of the test arrangement (2), and the integrity state is derived from the process data acquired during the test routine (26). It is proposed that, during the test routine (26), early process data (27) generated by a test sensor arrangement (4) of the test arrangement (2) in an early stage of the test program is used to determine the integrity state of the test consumable (1) by associating early process data (27) with a system model (29) of a reference module (30) in an association step (28), the reference module (30) emulates a test consumable (1), in particular a complete test consumable (1), taking into account only predetermined process data.

Description

[0001] The present invention relates to a method for performing integrity testing on test consumables of a bioprocessing facility according to the general part of claim 1, to a method for training a machine learning model according to the general part of claim 10, to a test control unit suitable for the first proposed method according to claim 18, to a training arrangement for performing the second proposed method according to claim 19, and to a data storage device with a trained machine learning model according to claim 20.

[0002] The term "bioprocess" currently refers to any biotechnological process, particularly a biopharmaceutical process. An example of such a bioprocess is the cultivation of microbial or mammalian cells under defined conditions using a bioreactor, where the cell culture fluid is transferred from the bioreactor to a downstream process. Such a downstream process might involve the separation of cells and supernatant via a filter. Before or after use, such filters must typically be integrity tested to ensure their integrity. In particular, most regulatory bodies mandate both pre-use and post-use integrity testing.

[0003] The method for performing integrity testing on test consumables of bioprocess facilities discussed can be applied to various fields of biotechnology. The increasing demand for biological products (such as biopharmaceutical drugs) has driven the need for high efficiency in this field. Efficiency in this sense is not only about the cost effectiveness of the components to be used, but also about the controllability of the processes associated therewith. The integrity testing of consumables (such as filters) is an important, usually mandatory but time-consuming part in many bioprocesses.

[0004] Today, integrity tests are typically performed as follows: First, the integrity testing device pressurizes an external volume to a predetermined test pressure. The integrity testing device then maintains the test pressure for a predetermined stabilization time. Subsequently, after measurements lasting the predetermined test time, the integrity test is assessed as passed if the measured quantity remains below or above a predetermined threshold, or within a predetermined range. Depending on the test procedure performed, the measured quantity may be at least one of the following: the amount of air diffused through a wet filter membrane, the amount of water flowing through a hydrophobic filter membrane, and / or the pressure at which the amount of air diffused through a wet filter membrane exhibits a disproportionate increase (thus identifying the so-called "bubble point" of the filter).

[0005] The known prior art (WO2016 / 030013A1) relates to a method according to the general part of claim 1. During a test routine performed by a test control unit of the test arrangement, a predetermined test program is applied to a test consumable via the test arrangement. Before the invention in the mentioned prior art, this test program had to be fully run in order to derive the integrity status from the process data collected during the test routine. Today, in many biotechnology production and / or quality control lines, this method for integrity testing of test consumables still constitutes a bottleneck in terms of process efficiency. The resulting combination has only limited efficiency with regard to process time and accuracy as well as the controllability of the entire process. EP2425886A1 further shows an integrity testing device that can be used with the proposed method.

[0006] The mentioned prior art provides a solution for predicting the results of integrity tests at an early stage of the test procedure by analyzing early process data based on known early process data of the complete consumable. However, deriving the data necessary for this prediction is itself time-consuming.

[0007] An object of the present invention is to provide a method for performing integrity testing on test consumables in a bioprocess facility, which method increases the accuracy and timeliness of the integrity testing with low workload.

[0008] The invention is based on the problem of improving the known methods in such a way that the accuracy and timeliness of the integrity test are increased with low effort.

[0009] With respect to a method for performing an integrity test on a test consumable having the features of the general part of claim 1 , the above-mentioned problem is solved by the features of the characterizing part of claim 1 .

[0010] A fundamental implementation of the present invention is that integrity testing need not be performed in full. It has been discovered that the integrity status of the tested consumable can be determined based on earlier process data by comparing it with a system model describing earlier process data for known intact consumables. To enable this comparison, a large amount of data on intact consumables is necessary. For example, if the consumables are filters, this means measuring a large number of filters. This typically requires wetting a large number of filters, running lengthy test procedures on the filters, drying them, and so on.

[0011] Machine learning is particularly desirable for the task of identifying intact consumables (especially filters) from early process data, because the process of identifying intact consumables has well-defined criteria, but even for different consumables considered intact, early process data can deviate from each other. To train a machine learning model, the number of reference measurements needs to be extremely large, making the actual collection of this data very tedious and time-consuming.

[0012] The key aspect of the present invention is that such a system model can be derived from a reference module that is not a "real" test consumable. For at least some process parameters, the reference module (e.g., a needle valve) can simulate the test consumable, while for other process parameters, the needle valve will not be able to simulate the test consumable. Here, and preferably, the needle valve is an automated, preferably electric, needle valve.

[0013] A reference program (particularly a test program) is applied to a reference module instead of a test consumable, thereby generating process data for the process parameters. The process parameters (for which the reference module simulates the test consumable) are measured and then form generic early process data that can be used to derive a system model (e.g., of a complete filter).

[0014] This is facilitated by the use of a reference module that can simulate a complete test consumable (such as a filter setting) taking into account predetermined process data (particularly a fluid flow rate corresponding to a leak caused by the test consumable). The container volume used in the reference module can be varied to simulate various filter settings. By using the reference module, the time required to train a machine learning model can be significantly reduced, as training does not require a "real" test consumable. In this way, the generation of a system model can be performed more quickly.

[0015] During a test routine for integrity testing of the test consumable, the measured early process data is correlated with the simulated predetermined early process data to determine the corresponding integrity status. Thus, process time can also be minimized during the test routine.

[0016] Ultimately, by reducing the time required to generate system models and reducing process time during testing routines, the overall time required to perform integrity testing can be minimized.

[0017] It is proposed in detail that during a test routine, early process data generated by a test sensor arrangement of a test arrangement at an early stage of the test procedure are used to determine the integrity state of the test consumable (by associating the early process data with a system model of a reference module in an association step), which reference module simulates the test consumable, in particular the complete test consumable, only taking into account predetermined process data.

[0018] Claim 2 describes a preferred component of a test arrangement. This test arrangement allows integrity testing of test consumables with low effort and optimized manufacturing footprint. In particular, a reusable, preferably single-device integrity test device can be used to easily handle integrity testing.

[0019] The embodiment according to claim 3 involves steps for optimizing the test routine. In a preferred embodiment, the typical test procedure is shortened, and the integrity status derived by the proposed method can be used as the final integrity status, thereby deriving the integrity status with minimal effort. In particular, if a test consumable is determined to be incomplete, waiting for the remainder of the integrity test may be considered a waste of time. If a failure in the integrity test is not detected early on, it may be necessary to wait for the results of the test procedure.

[0020] In an embodiment according to claim 4, the sensor arrangement generates process data during some or all steps of the test procedure. For example, the process data can be used to determine the success or failure of the test procedure based on subsequent process data, in particular process data from the test step. Furthermore, claim 4 provides a preferred definition of the early process data.

[0021] Claim 5 relates to a preferred manner in which the reference module can be used to derive the system model during the reference routine and thus directly describes the preferred system model used in the proposed method.

[0022] A preferred embodiment according to claim 6 involves deriving the system model from generic process data in a reference routine. In this preferred case, at least one process parameter simulating earlier process data of the complete test consumable is measured in the same way by the test sensor arrangement and the reference sensor arrangement.

[0023] Claim 7 provides a preferred embodiment of the reference module. In particular, it has been found that needle valves and calibrated leaks can be used effectively as substitutes for real test consumables, in particular filters. Both devices generate process data that simulates a complete test consumable in many relevant aspects.

[0024] A preferred embodiment according to claim 8 relates to a system model which is a trained machine learning model. This is particularly advantageous because the accuracy of the system model can be increased and the complexity of deriving the system model can be reduced.

[0025] A low-complexity solution for the correlation results between early process data and system models is defined in claim 9. In particular, the trained machine learning model can be used to classify test consumables into a complete or incomplete category.

[0026] The preferred embodiment according to claim 10 relates to a preferred option for determining the integrity status based on measured, predetermined physical properties of the test consumable 1. For example, a machine learning model can be trained to recognize different filter sizes. If the machine learning model now identifies a filter as having a higher flow rate than it actually has, it can be determined that the filter is not intact.

[0027] The embodiment according to claim 11 relates to preferred options regarding different possible test procedures, which are compatible with the proposed solution.

[0028] Another equally important teaching according to claim 12 relates to a method for training a machine learning model.

[0029] All explanations given with respect to the proposed method for performing integrity testing are fully applicable. In particular, all features of the method for training a machine learning model can be features of the trained machine learning model and thus combined with the teachings of the first method. The two methods can be combined into a joint method, and / or features of the first method can affect the purpose of the second method and therefore also be described with respect to parts of the second method.

[0030] All explanations given about the training of machine learning models can, where applicable, also generally be used to generate models of systems with or without machine learning.

[0031] According to claim 13, the training arrangement can be run in its main aspects simultaneously with the test arrangement. In this way, the generic early process data can be closer to the data of the actual test consumables, thereby enabling better training of the machine learning model.

[0032] Claim 14 relates to the possibility of repeating a reference procedure to collect multiple measurement results. The measurement result boundaries can be changed between reference procedures, for example for each reference procedure or for each number of reference procedures. These changes can affect the reference volume and / or the fluid pressure and / or the fluid temperature and / or the parameterization of the reference module, and / or can be changes to the reference module itself.

[0033] Particularly preferred embodiments with regard to the reference module are defined in claim 15. These measures are particularly advantageous because they ensure comparability between the behavior of the test consumable and the reference module and thus ensure the accuracy of the system model.

[0034] Claim 16 relates to preferred parameters for generating generic process data. These aspects increase accuracy and reliability.

[0035] Claim 17 defines a particularly preferred embodiment for training a machine learning model. Using data from actual measurements of test consumables to train the machine learning model helps increase the accuracy of the machine learning model in reliably deriving the integrity status. This can be achieved by either blending data from test consumables or by fine-tuning the machine learning model using this data. Both approaches can be used with small amounts of test consumable data, thereby increasing accuracy with less time, effort, and cost.

[0036] A preferred embodiment according to claim 18 associates a method for training a machine learning model with a method for performing integrity testing on a test consumable.

[0037] An equally important further teaching according to claim 19 relates to a test control unit suitable for the first proposed method.

[0038] All explanations given regarding the proposed method for performing integrity testing and the proposed method for training machine learning models are fully applicable.

[0039] An equally important further teaching according to claim 20 relates to a training arrangement for carrying out the proposed second method.

[0040] All explanations given regarding the proposed method for performing integrity testing and the proposed method for training a machine learning model and testing a control unit are fully applicable.

[0041] Another equally important teaching according to claim 21 relates to a data storage device having a trained machine learning model obtained by the proposed second method.

[0042] All explanations given regarding the proposed method for performing integrity testing and the proposed method for training machine learning models, the test control unit and the training arrangement are fully applicable.

[0043] In the following, embodiments of the present invention are explained with reference to the accompanying drawings. The accompanying drawings show:

[0044] Figure 1 ,Test arrangement for the proposed method,

[0045] Figure 2 , as two examples of typical pressure versus time curves that can be expected from a complete test consumable (solid line), and curves for a non-complete test consumable (dashed line),

[0046] Figure 3 ,Flowchart of the proposed method,

[0047] Figure 4 , the proposed training arrangement, and

[0048] Figure 5 , the proposed training routine for training machine learning models.

[0049] like Figure 1 As shown, the proposed method for performing an integrity test on a test consumable 1 of a bioprocess facility uses a test arrangement 2. The test arrangement 2 includes a test control unit 3 and a test sensor arrangement 4. For the integrity test, a predetermined test program is applied to the test consumable 1 via the test arrangement 2. The test program can be applied automatically and / or manually.

[0050] Figure 1 A preferred embodiment of a test arrangement 2 is shown, on which a preferred embodiment of the proposed method can be explained. The test arrangement 2 can include a test gas inlet 5 for supplying a test gas, such as air, nitrogen or helium, to the test arrangement 2. During the integrity test, the test control unit 3 opens a first valve 6, which is preferably designed as a proportional valve. The first valve 6 is used to control the flow rate or mass flow of the test gas through a gas fluid line 7 of the test arrangement 2. It is particularly preferred that the gas fluid line 7 includes a first pressure sensor 8 for determining the pressure of the test gas, a temperature sensor 9 for determining the temperature of the test gas and / or a flow sensor 10 for determining the flow rate of the test gas. Preferably, the gas fluid line 7 includes all of the aforementioned sensors. The sensors of the test sensor arrangement 4 generate process data which are provided to the test control unit 3.

[0051] The gas fluid line 7 preferably connects the test gas inlet 5 of the test arrangement 2 with a test gas outlet 11, which leads to the test consumable 1, such as a filter to be tested. Exemplarily, to completely shut off the test gas supply to the test gas outlet 11 to perform an integrity test (in one example, a pressure decay test), the fluid line further includes a second valve 12 (preferably a shut-off valve). The gas fluid line 7 further includes a second pressure sensor 13 downstream of the second valve 12 to determine changes in the test gas pressure during the integrity test.

[0052] In order to perform an integrity test of the test consumable 1, and depending on the test to be performed, a liquid may be provided. The liquid may be provided in a liquid container 14, which may be connected to a liquid supply 15 of the test arrangement 2. In order to transport the liquid from the liquid container 14 to the liquid outlet 16, the test arrangement 2 may comprise a liquid pump 17, in particular a peristaltic pump, which may be arranged downstream of the liquid supply 15. Alternatively, to a pump, the test arrangement 2 may comprise a second test gas inlet 18, which is connected to a second fluid line 19 having a third valve 20, which is preferably designed as a proportional valve. The second fluid line 19 terminates in the liquid container 14. The third valve 20 is used to control the flow rate or mass flow of the test gas through said fluid line of the test arrangement 2 in order to convey the liquid. Figure 1 Only two alternatives are shown schematically in one figure.

[0053] Thus, liquid is conveyed from the liquid supply 15 through the third fluid line 21 to the liquid outlet 16 via the fourth valve 22, which leads to the test consumable 1, such as a filter to be tested. The fourth valve 22 may be an on / off valve, a proportional valve or a proportional control valve.

[0054] The described components can be collected inside a corresponding integrity testing device 23, which preferably has a single housing containing the described components (when they are present). However, all explanations given with respect to a single housing or a single integrity testing device 23 can also generally apply to the integrity testing arrangement 2. Although the use of the above-described compact integrity testing device is preferred in terms of ease of handling, the proposed method steps can alternatively be performed using one or more discrete components (such as a pressure generator or the like).

[0055] The test consumable 1 is placed inside the outer volume 24 (also in Figure 1 Downstream of the external volume 24 , a fifth valve 25 is shown.

[0056] The term "external volume" means any housing, box or the like comprising a predetermined volume in which the test consumable 1 is placed. Thus, the external volume 24 may refer to the volume in which the test consumable 1 ( Figure 1 ), or the outer volume 24 may refer to a housing in which the test consumable 1 is placed by the manufacturer. According to the preferred embodiment, the outer volume 24 is already part of the test consumable 1 (not shown).

[0057] Ultimately, the integrity status is derived from the process data which are acquired by the test sensor arrangement 4 during the test routine 26 .

[0058] Figure 2Exemplary pressure versus time curves are shown for a complete test consumable 1 (solid line) and for an incomplete test consumable 1 (dashed line).

[0059] What is proposed in detail is a method for performing integrity testing of test consumables 1 of a bioprocess facility using a test arrangement 2, wherein during a test routine 26 executed by a test control unit 3 of the test arrangement 2, a predetermined test procedure is applied to the test consumable 1 via the test arrangement 2, and the integrity status is derived from process data acquired during the test routine 26.

[0060] The proposed method for performing integrity testing on test consumables 1 of a bioprocess facility is preferably assigned to a quality control phase of a test consumable 1 production line (not shown) and / or to pre- and / or post-testing of a bioprocess using the test consumable 1 .

[0061] Typically, integrity testing is currently used during the production of test consumables 1 and / or during the production of pharmaceutical or biopharmaceutical drugs that need to use test consumables 1. As mentioned above, most regulatory agencies mandate that test consumables 1 be integrity tested before and after use. This test consumable 1 is preferably a filter, bag, etc., further preferably designed for disposable use. The integrity test of filters etc. is a basic requirement for key filtration applications in biotechnology and biopharmaceutical industries. FDA regulations require that all types of filters for the production of sterile solutions (such as injections) be integrity tested. In addition, official regulations require that corresponding documents be added to each production protocol of the filter and / or biological product (such as biopharmaceutical drugs) produced using any batch of this filter production.

[0062] The term "during the test routine" means that the test program is, if necessary, at least partially executed simultaneously with the test routine 26 in order to collect process data. However, this does not necessarily mean that the test program runs during the entire test routine 26, nor does it mean that the test program is executed by the test control unit 3. In particular, the test program can be executed completely or partially by an operator, even if it is executed automatically in the described embodiment. Data processing can be performed after the test program in the test routine. The mentioned test control unit 3 can be a local control unit or can include a cloud service provided by a cloud server instead of or in addition to local hardware. Therefore, the term "control unit" should be understood in a broad sense.

[0063] In this case and preferably, the integrity status is determined by the test control unit 3 during the test routine 26. Alternatively, it can also be determined by a different computer unit or the like in a subsequent and / or external data analysis step.

[0064] Importantly, during the test routine 26, early process data 27 generated by the test sensor arrangement 4 of the test arrangement 2 at an early stage of the test procedure are used to determine the integrity state of the test consumable 1 (by associating the early process data 27 with a system model 29 of a reference module 30 in an association step 28), which reference module 30 simulates the test consumable 1, in particular the complete test consumable 1, only taking into account predetermined process data.

[0065] Early process data is any data generated before the regular test program has advanced far enough to draw conclusions about the integrity of the test. For multi-step test programs, early process data can be defined as data collected during predefined steps of the test program, where these steps are distinct from one another, each serving a different physical purpose.

[0066] The term "correlation" does not imply correlation as a mathematical concept, but rather relates to any concept whereby earlier process data 27 is input into a system model 29, the system model 29 is applied to the input data, or is otherwise used to determine integrity status.

[0067] Reference module 30 is a device or device arrangement that can be used to simulate test consumable 1 with respect to certain process parameters. It can be used efficiently to collect data about the reaction of an intact test consumable 1 to certain test procedures without having to perform laborious experiments using real test consumables 1. Although this example is about simulating an intact test consumable 1, all explanations given can alternatively or additionally apply to incomplete test consumables 1. Therefore, system model 29 can be a system model of intact and / or incomplete test consumables 1.

[0068] Test routine 26 is used as an example in Figure 3 . After the early process data 27 are transmitted from the test sensor arrangement 4 to the test control unit 3, in an association step 28, the test control unit 3 associates the early process data 27 with the system model 29. In a classification step 31, an integrity class for the integrity status is derived (preferably by a machine learning model serving as the system model 29). Depending on the derived integrity status (which is verified in a verification step 32), the test consumable 1 is deemed to have passed or failed the integrity test. The association step 28 and the classification step 31 can be repeated continuously or can be initiated after a certain amount of time has passed since the early process data 27 has been completely collected.

[0069] The machine learning model has been trained as part of the first proposed method or is being trained as part of the second proposed method. Figure 5. Accordingly, the training is preferably based on an unsupervised single-class classification method. Currently, two classes should be detected: "complete" and "incomplete". However, the proposed training data are mainly or only training data related to complete test consumables 1. In order to avoid having to collect data about incomplete test consumables 1 in some way, a single-class classification model is used. The model is trained to learn to determine a single class and to derive a second class by detecting significant deviations from the learned class. For example, a state vector machine can be used. Unsupervised methods are particularly useful because it may already be known that all data relate to complete test consumables 1. On the other hand, this knowledge can be used to simply label all data as "complete" and use supervised learning methods. Of course, data on incomplete test consumables 1 can also be used.

[0070] According to one embodiment, it is proposed that the test arrangement 2 comprises a test apparatus for performing Figure 1 The integrity testing device 23 of the test procedure shown. The integrity testing device 23 preferably comprises a test control unit 3. As already mentioned, the test arrangement 2 may further comprise an external volume 24 positioned downstream of the integrity testing device 23, in which the test consumable 1 may be placed. The integrity testing device 23 and the external volume 24 are fluidically connected.

[0071] Here, and preferably, the integrity testing device 23 performs the test procedure by generating at least one predetermined process parameter in the test arrangement 2, in particular by generating a predetermined fluid pressure in the external volume 24. The external volume 24 can be designed as a multi-purpose or single-purpose housing. Here, the integrity testing device 23 is directly connected to the external volume 24 via a fluid line.

[0072] like Figure 2 As shown, preferably, in a pressurization step 33 of the test procedure, the integrity testing device 23 pressurizes the external volume 24 with a fluid to a predetermined test pressure 34 within a pressurization time 35, and in a stabilization step 36 of the test procedure, the integrity testing device 23 maintains the predetermined test pressure 34 for a stabilization time 37. As mentioned above, instead of the integrity testing device 23, a pressure regulator or the like can also be used to pressurize the external volume 24 to the predetermined test pressure 34 within the pressurization time 35 and perform the subsequent stabilization step 36.

[0073] These steps may be steps of a conventional test procedure. Conventional test procedure steps may be further followed, or the test procedure may be stopped here to save time. In one embodiment, in a test step 38 of the test procedure, the integrity test device 23 monitors changes in the maintained test pressure 34 for a test time 39, and in a vent step 40 of the test procedure, the integrity test device 23 vents the test arrangement 2 for a vent time 41.

[0074] In another embodiment, the test procedure ends after the pressurization step 33 or the stabilization step 36. Again, both alternatives are available, depending on the outcome of the verification step 32. Preferably, the test procedure ends if the test consumable 1 is determined to be incomplete, and / or the test procedure does not end if the test consumable 1 is not determined to be incomplete.

[0075] According to one embodiment, it is proposed that during one or more of these steps, preferably during all of these steps, the test sensor arrangement 4 generates process data, preferably the test sensor arrangement 4 generates early process data 27 during the pressurizing step 33 and / or the stabilizing step 36 .

[0076] The data generated in this way are sent from the test sensor arrangement 4 to the test control unit 3. The test sensor arrangement 4 and / or the test control unit 3 may be part of an integrity testing device 23.

[0077] Now focusing on the generation of the system model 29, it is possible that a predetermined reference procedure has been applied to the reference module 30 via the training arrangement 43 during a reference routine that has been executed by the reference control unit 42 of the training arrangement 43. These steps describe the properties of the system model 29. Alternatively, these steps can be part of the proposed method. Thus, the reference routine can be executed and the reference procedure can be applied.

[0078] The training arrangement 43 comprises (meaning it currently may comprise or has comprised) a reference sensor arrangement 44. During a reference routine, the reference sensor arrangement 44 of the training arrangement 43 generates or has generated generic early process data 45, and the training control unit 46 has derived or has derived the system model 29 from the generic early process data 45 in the reference routine. Figure 4 A training arrangement 43 is shown.

[0079] The training control unit 46 may be the reference control unit 42 or a cloud service or the like.

[0080] Preferably, the test sensor arrangement 4 and the reference sensor arrangement 44 are identical in construction and / or measure the same process parameters (when these parameters are used for the association step 28). Furthermore, the test arrangement 2 and the training arrangement 43 are preferably identical in construction. In a first step, this enables the generation of common process data by using the training arrangement 43 with the reference sensor arrangement 44, which allows the derivation of the system model 29. In a second step, this enables the generation of process data by using the test arrangement 2 with the test sensor arrangement 4. These similar measurement results form the basis for the association step 28, in which the process data are associated with the system model 29. Due to the structural similarities between the test arrangement 2 and the training arrangement 43, the same reference numerals have been used for components that are not described again.

[0081] Thus, the generic early process data 45 may be similar to the early process data 27 in relevant respects.

[0082] According to one embodiment, it is proposed that the early process data 27 represent at least one parameter reflecting the integrity state of the test consumable 1, the universal early process data 45 represent at least the same parameter, and for this at least one parameter, the reference module 30 simulates the test consumable 1, in particular the complete test consumable 1.

[0083] The at least one parameter measured in the test arrangement 2 and the training arrangement 43 comprises one or more parameters from the group consisting of: fluid pressure, amount of fluid substance, temperature of the fluid and / or flow rate of the fluid (preferably the volumetric flow rate of the fluid). Figure 1 As shown, the fluid here is preferably a test gas.

[0084] As already mentioned, the reference module 30 may have a different structure than the test consumable 1. In detail, preferably, the reference module 30 comprises a valve, preferably a needle valve 47, and / or the reference module 30 comprises a calibrated leak (not shown).

[0085] like Figure 4 As shown, the training arrangement 43 can include the external volume 24, even if the needle valve 47 is positioned downstream of the external volume 24. However, this further facilitates the test arrangement 2 and the training arrangement 43 having similar physical properties. The presence of the external volume 24 also allows for simulating different test consumables 1 by adjusting the volume of the external volume 24. The volume of the external volume 24 can be adjusted, for example, by using external volumes 24 with different geometries. In addition to or in lieu of the temperature sensor 9 of the parameter adjustment device 48, the training arrangement 43 can include a temperature sensor 9 in or at the external volume 24.

[0086] The system model 29 describes the reference module 30 in such a way as to allow the integrity status of the test consumable 1 to be determined by correlating its earlier process data with the system model 29. The system model 29 may be, for example, a trained machine learning model, a statistical model or an analytical model.

[0087] According to a preferred embodiment, it is proposed that the system model 29 is the above-mentioned trained machine learning model and that the machine learning model is applied by the test control unit 3 in order to determine the integrity status of the test consumable 1 based on the early process data 27 .

[0088] Alternatively, as also described above, the system model 29 is a statistical model or an analytical model.

[0089] The term "analytical model" refers to a quantitative model designed to answer a specific question. Its primary goal is to provide a closed-form formula for a specific property. Therefore, an analytical model is a mathematical model with a closed-form solution—that is, the solutions to the equations describing the system's behavior can be expressed as mathematical functions.

[0090] The term "statistical model" means a mathematical model in which some or all of the input data has some randomness (e.g., represented by a probability distribution), so that for a given set of input data, the output is not reproducible, but is instead described by a probability distribution. The output set is obtained by running the model multiple times, where new input values are sampled from the probability distribution each time the model is run.

[0091] Here and preferably, the integrity state is represented by an integrity class from a predetermined set of integrity classes. In the test routine 26, the integrity state is assigned one of the integrity classes. Preferably, this predetermined set of integrity classes only includes the integrity classes "Integrity" and "Non-Integrity".

[0092] according to Figure 5 As already mentioned, the machine learning model is preferably trained using "single-class classification" training. This means that the machine learning model is trained using a large amount of data on known intact test consumables 1, in this case simulated by reference module 30. As a result, the model is able to increasingly accurately identify whether a test consumable 1 is intact. This trained machine learning model preferably forms the basis for deriving system model 29, or is system model 29.

[0093] Alternatively, a machine learning model is applied by the test control unit 3 in order to determine predetermined physical properties of the test consumable 1 , in particular the filter size of the filter, and the integrity status is derived from the predetermined physical properties.

[0094] The machine learning model can, for example, be trained to detect different filter sizes in the same or different test settings and / or test environments (such as different temperatures). The integrity test can be performed as follows. A filter with an exemplary flow rate of 10 ml / min can be tested. Early process data 27 is measured and fed into the machine learning model. The trained machine learning model (exemplarily trained using data from reference modules 30 for 10 ml / min, 12.5 ml / min, and 15 ml / min) then determines which flow rate the filter has. If the flow rate is determined to be 10 ml / min, the filter is intact. If the flow rate is determined to be 12.5 ml / min or 15 ml / min, the filter is not intact. Of course, the machine learning model can be trained to detect a small or large number of different flow rates, such as at least three different flow rates, at least 10 different flow rates, or at least 50 different flow rates. In particular, the flow rates can be based on a training arrangement 43 that is essentially identical for different flow rates. By using the reference module 30, generating data for training the machine learning model for those different flow rates is greatly simplified.

[0095] Thus, the test consumable 1 may be determined to be intact if the determined physical properties match the nominal physical properties, and / or determined to be not intact if the determined physical properties do not match the nominal physical properties.

[0096] According to one embodiment, it is proposed that the test procedure is a test procedure from the group of diffusion tests, in particular a forward flow test and a multi-point diffusion test and a bubble point test. It should be noted here that different test procedures (e.g., a forward flow test and a bubble point test) can also be combined.

[0097] Alternatively, further test procedures are applicable, in particular water intrusion tests, water flow tests, etc. Preferably, all test procedures that are possible for the integrity testing device 23 used are applicable.

[0098] Another teaching of equal importance relates to a method for using a training arrangement 43 to train a machine learning model that can be used as a system model 29 for the proposed method for performing integrity testing.

[0099] During a reference routine executed by a reference control unit 42 of the training arrangement 43, a predetermined reference program is applied to the reference module 30 via the training arrangement 43. The training arrangement 43 comprises a reference sensor arrangement 44, wherein early process data 45 are generated by the reference sensor arrangement 44 during the reference routine.

[0100] During a training routine 49 performed by a training control unit 46 of the training arrangement 43 , the generic early process data 45 are used to train a machine learning model.

[0101] The training control unit 46 may be the reference control unit 42 or any different control unit. Here and preferably, the reference sensor arrangement 44 is identical to the test sensor arrangement 4 and the predetermined reference program may be a predetermined test program or may be or include the pressurization step 33 and / or the stabilization step 36.

[0102] All explanations given regarding the proposed method for performing the integrity test are fully applicable.

[0103] According to one embodiment, it is provided that the training arrangement 43 comprises a parameter adjustment device 48 for executing a reference program.

[0104] The parameter adjustment device 48 can be provided by the integrity test device 23 (also used in the first proposed method). Although the use of the integrity test device 23 (as described in relation to the first teaching) is preferred, the proposed method steps can alternatively be performed using one or more discrete components (such as a pressure generator or the like) that perform at least the functions of the parameter adjustment device 48. Preferably, the same type of device is used to perform the first and second methods. The parameter adjustment device 48 can include a reference control unit 42 and / or a reference sensor arrangement 44.

[0105] like Figure 4 As shown, the training arrangement 43 preferably further comprises the external volume 24 positioned downstream of the parameter adjustment device 48. The training arrangement 43 comprises a reference module 30 positioned downstream of the parameter adjustment device 48 and preferably downstream of the external volume 24.

[0106] In a preferred embodiment, the parameter adjustment device 48 executes a reference procedure by generating at least one predetermined process parameter in the training arrangement 43 , in particular generating a predetermined fluid pressure at the reference module 30 and / or in the external volume 24 .

[0107] The predetermined process parameters may be the same as the process parameters used in the test procedure for the integrity test.

[0108] In this case, no consumables are present in the outer volume 24 and the outer volume 24 exists only to better simulate the process parameters.

[0109] The parameter adjustment device 48 , the external volume 24 and the reference module 30 are fluidically connected.

[0110] It is contemplated that the reference program and reference routine are repeated multiple times, preferably at least 50 times, and more preferably at least 200 times. This repetition is advantageous for generating varying generic early process data 45. In a training routine 49, the varying generic early process data 45 is used to train the machine learning model. Preferably, the reference module 30 is parameterized differently during at least some of these repetitions to simulate physical changes in the complete test consumable 1.

[0111] As described, here and preferably, the reference module 30 , preferably a needle valve 47 or a calibrated leak, generates a fluid flow that simulates the fluid flow in the complete test consumable 1 at the training pressure value.

[0112] According to one embodiment, it is proposed that in the training routine 49 at least one parameter (preferably at least two parameters, further preferably all parameters, in particular the fluid pressure, the amount of fluid substance, the temperature and / or the flow rate of the fluid) measured by the test sensor arrangement 4 is used to generate generic early process data 45.

[0113] An interesting point is that the proposed synthetic data may not vary as much as the data of the test consumables. Therefore, training of the machine learning model may result in a model that incorrectly classifies some intact test consumables 1 as non-intact if they show variations that are not present in the generic early process data 45.

[0114] For this reason, it is possible that the trained machine learning model is used as a pre-trained machine learning model. Subsequently, during a fine-tuning routine performed by a training control unit 46 of the training arrangement 43, earlier process data of the test consumable 1, in particular a complete test consumable 1, is used to fine-tune the pre-trained machine learning model. The training control unit 46 can be the reference control unit 42 or a different control unit.

[0115] Additionally or alternatively, the generic early process data 45 and the early process data 27 of the test consumable 1 are blended to obtain blended early process data, and the blended early process data is used to train the machine learning model during a training routine 49 .

[0116] According to one embodiment, it is proposed to use a trained machine learning model in the proposed first method.

[0117] Another teaching of equal importance relates to a test control unit 3 suitable for the first proposed method.

[0118] All explanations given regarding the proposed method for performing integrity testing and the proposed method for training machine learning models are fully applicable.

[0119] Another teaching of equal importance relates to a training arrangement 43 for carrying out the proposed second method.

[0120] All explanations given regarding the proposed method for performing integrity testing and the proposed method for training a machine learning model and testing the control unit 3 are fully applicable.

[0121] Another teaching of equal importance relates to a data storage device having a trained machine learning model derived by the proposed second method.

[0122] The trained machine learning model may be suitable for use as the system model 29 in the proposed first approach.

[0123] All explanations given regarding the proposed method for performing integrity testing and the proposed method for training a machine learning model, the test control unit 3 and the training arrangement 43 are fully applicable.

[0124] Reference numerals

[0125] 1Test consumables

[0126] 2 Test Arrangement

[0127] 3 Test control unit

[0128] 4. Test sensor arrangement

[0129] 5Test gas inlet

[0130] 6First valve

[0131] 7Gas fluid pipelines

[0132] 8First pressure sensor

[0133] 9 Temperature sensor

[0134] 10 Flow sensor

[0135] 11 Test gas outlet

[0136] 12 Second valve

[0137] 13 Second pressure sensor

[0138] 14Liquid container

[0139] 15Liquid supply source

[0140] 16 liquid outlets

[0141] 17 Liquid Pump

[0142] 18 Second test gas inlet

[0143] 19 Second fluid pipeline

[0144] 20Third valve

[0145] 21 Third fluid pipeline

[0146] 22 Fourth valve

[0147] 23Integrity testing equipment

[0148] 24 External Volume

[0149] 25 Fifth Valve

[0150] 26 test routines

[0151] 27 Early Process Data

[0152] 28 association steps

[0153] 29 System Model

[0154] 30 reference modules

[0155] 31 Classification Steps

[0156] 32 Verification Steps

[0157] 33 Pressurization steps

[0158] 34Predetermined test pressure

[0159] 35 pressurization time

[0160] 36 stabilization steps

[0161] 37 stabilization time

[0162] 38 Test Steps

[0163] 39 test time

[0164] 40 Exhaust Steps

[0165] 41 Exhaust time

[0166] 42 Reference Control Unit

[0167] 43 Training Arrangement

[0168] 44 Reference sensor layout

[0169] 45 General Early Process Data

[0170] 46 Training Control Unit

[0171] 47 needle valve

[0172] 48 parameter adjustment equipment

[0173] 49 training routines

Claims

1. A method for performing an integrity test on a test consumable (1) of a bioprocess facility using a test arrangement (2), wherein during a test routine (26) executed by a test control unit (3) of the test arrangement (2), a predetermined test program is applied to the test consumable (1) via the test arrangement (2), and an integrity status is derived from process data acquired during the test routine (26), It is characterized by During the test routine (26), the early process data (27) generated by the test sensor arrangement (4) of the test arrangement (2) at an early stage of the test procedure are used to determine the integrity state of the test consumable (1) by associating the early process data (27) with the system model (29) of a reference module (30) in an association step (28), wherein the reference module (30) simulates the test consumable (1), in particular the complete test consumable (1), only taking into account predetermined process data.

2. The method according to claim 1, characterized in that The test arrangement (2) comprises an integrity test device (23) for performing the test procedure, wherein the integrity test device (23) preferably comprises the test control unit (3), wherein the test arrangement (2) further comprises an external volume (24) positioned downstream of the integrity test device (23), the test consumable (1) being placed in the external volume (24), the integrity test device (23) and the external volume (24) being fluidically connected, preferably, the integrity test device (23) performing the test procedure by generating at least one predetermined process parameter in the test arrangement (2), in particular a fluid pressure in the external volume (24).

3. The method according to claim 2, characterized in that In a pressurizing step (33) of the test procedure, the integrity testing device (23) pressurizes the external volume (24) to a predetermined test pressure (34) using a fluid, and in a stabilizing step (36) of the test procedure, the integrity testing device (23) maintains the test pressure for a stabilizing time (37), preferably, in a testing step (38) of the test procedure, the integrity testing device (23) monitors changes in the maintained test pressure within a test time (39), and in an exhausting step (40) of the test procedure, the integrity testing device (23) exhausts the test arrangement (2) within an exhausting time (41), and / or the test procedure ends after the pressurizing step (33) or the stabilizing step (36), preferably, the test procedure ends if the test consumable (1) is determined to be not intact, and / or the test procedure does not end if the test consumable (1) is not determined to be not intact.

4. The method according to claim 3, characterized in that During one or more of these steps, preferably during all of these steps, the test sensor arrangement (4) generates process data, preferably the test sensor arrangement (4) generates the early process data (27) during the pressurizing step (33) and / or the stabilizing step (36).

5. The method according to any one of the preceding claims, characterized in that During a reference routine performed by a reference control unit (42) of a training arrangement (43), a predetermined reference program is applied to the reference module (30) via the training arrangement (43), the training arrangement (43) comprising a reference sensor arrangement (44), during which the reference sensor arrangement (44) of the training arrangement (43) generates generic early process data (45), and in which the training control unit (46) derives the system model (29) from the generic early process data (45).

6. The method according to claim 5, characterized in that The early process data (27) represents at least one parameter reflecting the integrity state of the test consumable (1), the universal early process data (45) represents at least the same parameter, and for the at least one parameter, the reference module (30) simulates the test consumable (1), in particular the complete test consumable (1), preferably, the at least one parameter includes one or more parameters from the group of the following items: fluid pressure, amount of fluid substance, temperature of fluid and / or flow rate of fluid, the flow rate of fluid is preferably the volume flow rate of fluid.

7. The method according to any one of the preceding claims, characterized in that The reference module (30) has a different structure from the test consumable (1), preferably, the reference module (30) comprises a valve, preferably a needle valve (47), and / or, the reference module (30) comprises a calibrated leak.

8. The method according to any one of the preceding claims, characterized in that The system model (29) is a trained machine learning model, and the machine learning model is applied by the test control unit (3) to determine the integrity status of the test consumable (1) based on the early process data (27).

9. The method according to any one of the preceding claims, characterized in that The integrity state is represented by an integrity class from a predetermined set of integrity classes, and in the test routine (26) the integrity state is assigned one of the integrity classes, preferably the predetermined set of integrity classes comprising only the integrity classes "Integrity" and "Non-Integrity".

10. The method according to any one of claims 1 to 8, characterized in that The machine learning model is applied by the test control unit (3) to determine predetermined physical properties of the test consumable (1), in particular the filter size of the filter, and the integrity status is derived from the predetermined physical properties, preferably, if the determined physical properties match the nominal physical properties, the test consumable (1) is determined to be intact, and / or if the determined physical properties do not match the nominal physical properties, the test consumable (1) is determined to be not intact.

11. The method according to any one of the preceding claims, characterized in that The test procedure is a test procedure from the group of diffusion tests, in particular the forward flow test and the multi-point diffusion test, the bubble point test, the water intrusion test and the water flow test.

12. Method for training a machine learning model using a training arrangement (43), the machine learning model being usable as a system model (29) for the method for performing integrity testing according to any one of the preceding claims, wherein a predetermined reference program is applied to the reference module (30) via the training arrangement (43) during a reference routine executed by a reference control unit (42) of the training arrangement (43), wherein the training arrangement (43) comprises a reference sensor arrangement (44), wherein generic early process data (45) is generated by said reference sensor arrangement (44) during said reference routine, wherein the generic early process data (45) is used to train the machine learning model during a training routine (49) performed by a training control unit (46) of the training arrangement (43).

13. The method according to claim 12, characterized in that The training arrangement (43) comprises a parameter adjustment device (48) for executing the reference procedure, wherein the parameter adjustment device (48) preferably comprises the reference control unit (42), wherein the training arrangement (43) preferably further comprises an external volume (24) positioned downstream of the parameter adjustment device (48), wherein the training arrangement (43) comprises the reference module (30) positioned downstream of the parameter adjustment device (48) and preferably downstream of the external volume (24), preferably, the parameter adjustment device (48) executes the reference procedure by generating at least one predetermined process parameter in the training arrangement (43), in particular a fluid pressure at the reference module (30) and / or in the external volume (24).

14. The method according to claim 12 or 13, characterized in that The reference program and the reference routine are repeated a plurality of times, preferably at least 50 times, more preferably at least 200 times, to generate varying generic early process data (45), which are used to train the machine learning model in the training routine (49), preferably the reference module (30) is parameterized differently during at least some of these repetitions to simulate physical changes in the complete test consumable (1).

15. The method according to any one of claims 12 to 14, characterized in that The reference module (30), preferably a needle valve (47) or a calibrated leak, generates a fluid flow at a training pressure value, the fluid flow simulating the fluid flow in the complete test consumable (1).

16. The method according to any one of claims 12 to 15, characterized in that In the training routine (49), the following items measured by the test sensor arrangement (4) are used to generate the generic early process data (45): at least one parameter, preferably at least two parameters, further preferably all parameters, in particular pressure, amount of substance, temperature and / or flow rate.

17. Method according to any one of the preceding claims 12 to 16, characterized in that The trained machine learning model is used as a pre-trained machine learning model, and during a fine-tuning routine performed by a training control unit (46) of the training arrangement (43), early process data (27) of the test consumable (1), in particular the complete test consumable (1), is used to fine-tune the pre-trained machine learning model, and / or the generic early process data (45) and the early process data (27) of the test consumable (1) are mixed to obtain mixed early process data (27), and during the training routine (49), the mixed early process data (27) is used to train the machine learning model.

18. The method according to any one of claims 12 to 17, characterized in that The trained machine learning model is used in the method according to any one of claims 1 to 11.

19. Test control unit adapted for use in the method according to any one of claims 1 to 11.

20. A training arrangement for performing the method according to any one of claims 12 to 17.

21. A data storage device having a trained machine learning model derived by the method according to claims 12 to 17.

Images