A method for detecting leaks in a container, particularly for aseptic products

Abstract

Described is a method (1) of detecting leaks in a container (C), particularly for aseptic products, which comprises an inference step of: - collecting data items in real time by means of at least one pressure transducer (T) connected to a data collection device (DCU); the pressure transducer (T) having a predetermined frequency "F", - primarily processing the collected data items (DR) by means of algorithms, - providing the clean data items (DP) as input data items (Xi) for the first neural network (PRN); said first neural network being able to reproduce the characteristic curve of a container (C) conforming to the one being examined, - reconstructing the input data items (Xi) by means of the first neural network, obtaining output data items (Xu), - carrying out a linear combination on the input (Xi) and output (Xu) data items so as to create a tensor that is a function of said input (Xi) and output (Xu) data items, - entering the tensor into a second neural network (SRN), so as to obtain a classification (Y).

Description

[0001] A method for detecting leaks in a container, particularly for aseptic products

[0002] DESCRIPTION

[0003] Technical field

[0004] The present invention relates to a method of detecting leaks in a container, particularly for aseptic products.

[0005] Background art

[0006] The Container Closure Integrity Testing (CCIT) is a fundamental process in the pharmaceutical world, because it serves to ensure that the containers used to store drugs are perfectly sealed and do not leak.

[0007] In the same way, this method is also intended for other sectors and is applied in the field of metallic materials (cans, aerosols, etc.) as well as in the so-called "food and beverage" sector.

[0008] If, in fact, even a minimum quantity of air or contaminants were to enter bottles, vials or syringes which must contain sterile solutions, the product could lose effectiveness or, worse, become hazardous for the patient. For this reason, the CCIT ensures that the content remains intact, safe and protected until it is used.

[0009] There are several ways to check if a container is intact.

[0010] One od the known test uses helium gas: the container is filled with this very light gas and a check is performed to see whether it leaks out, which would reveal even the smallest cracks or fissures.

[0011] There are also more sophisticated techniques, such as infrared spectroscopy, which checks the concentration of a gaseous species or the pressure (partial or total) of gas in the headspace of a container, or the electrical conductivity test, which is often used for ampoules and strips produced in an aseptic environment with “blow-fill-seal” technology (so-called “BFS strip”), which measures whether there are voltage variations due to micro-holes.

[0012] Another very precise method, of a deterministic type, which is based on the measurement of the pressure inside a watertight chamber in which theproduct is placed, is the pressure or vacuum test: the container is put under pressure or in a depressurized environment and it is observed whether there are abnormal variations, which could indicate a leak.

[0013] An example is the Vacuum Decay Method (VDM), described in ASTM F2338-09, together with its variant, the Pressure Decay Method (PDM). These methods, together with other deterministic methods recognized by regulatory bodies such as USP and ASTM, are widely established and reliable, but present some critical issues.

[0014] Firstly, they do not adequately analyse the transitional phases, such as the suction phase at the beginning of the test cycle. These phases are evaluated only at predefined and fixed times, reducing the accuracy in the dynamic evaluation of the system.

[0015] Further, the compliance assessment is based on an arithmetic difference between two pressure values, ignoring non-linear behaviours potentially present between the two points. This approach may fail to capture more complex variations that distinguish a compliant product from a defective one.

[0016] Thirdly, the acceptance thresholds are determined only statistically, by analysing averages and standard deviations of the collected data items. However, this does not necessarily reflect the actual differences between a compliant and a defective product, causing possible false rejects, that is, compliant products which are rejected in order to avoid the acceptance of defective products, with a negative economic impact on production.

[0017] Another problem is due to the that, in the tests based on negative pressures, the presence of liquids, in particular water-based liquids, can affect the results. During the testing, the liquid may evaporate, increasing the pressure in the chamber and leading to false rejects.

[0018] In addition to the above, these systems often require the use of algorithms that compensate for the behaviour of the testing heads. The algorithms perform calculations at each stage to follow the trend of compliant products in the system, trying to maintain the effectiveness of the process.Finally, the methods currently used are not able to accurately distinguish compliant products that, due to outside factors, have values similar to those of non-compliant products. This makes it more difficult to properly separate the two types of product, increasing the risk of false positives (compliant containers reported as defective) and false negatives (defective containers reported as compliant).

[0019] The importance of these tests is clear: keeping container products (drugs, cans, aerosols, etc.) safe, stable and free from contamination is a top priority. The regulators, such as the FDA in the United States and the EMA in the European Union, require companies to demonstrate that their containers are capable of protecting the product throughout its life cycle. The leak detection methods currently used require that the data items and information about the container to be analysed are collected by pressure transducers that record, store and use only a few points at fixed, finite and predetermined instants of the pressure curve of the test cycle.

[0020] This results in a large quantity of lost information that can lead to false negatives or, even worse, false positives.

[0021] Further, in the methods currently in use, the data items are processed extensively, carrying out many analytical compensations, often comparing not only the raw values of the process but also those data items that undergo various calculation steps before being usable by the system and thus able to guarantee good leak detection performance levels.

[0022] Finally, in these known methods, experimental thresholds are determined on a statistical basis to be subsequently compared with the above-mentioned point pressure values. Once again, this comparison, although generally effective, tends to greatly simplify the problem by reducing it to a few point values that, although representing the process, fail to fully capture the complexity of the entire testing cycle.

[0023] Summary of the invention

[0024] The aim of the invention is to eliminate the above-mentioned drawbacks in prior art leak detection methods for containers, particularly for asepticproducts, which allows the development of more advanced testing systems which are capable of reducing the number of false rejects and improving the overall production efficiency.

[0025] Within the scope of the above-mentioned aim, an aim of the invention is to produce a leak detection method for a container for aseptic products, which takes into account the transitional phases by improving the accuracy in the dynamic evaluation of the system and reducing the risk of false positives or negatives.

[0026] Another aim of the invention is to provide a complete and optimal leak detection result, giving the user certainty of the final outcome, avoiding false positives and false negatives.

[0027] Another aim of the detection method is to simplify distinguishing a compliant product from a defective one,

[0028] A further aim of the above-mentioned method is to take into account the influence of any liquid contamination.

[0029] Another aim of the invention is to ensure that the containers are correctly sealed and that the product remains safe from the time of production until its use by the patient, avoiding the risk of contamination.

[0030] The purpose of the invention, including these and other aims, which are described in more detail below, is achieved by a method for detecting leaks in a container for aseptic products, according to the invention, comprising the technical features described in one or more of the appended claims. The dependent claims correspond to possible different embodiments of the invention.

[0031] In particular, according to a first aspect, the present invention relates to a method of detecting leaks in a container for aseptic products comprising the following steps, in particular the inference step intended to replace and / or complement the testing method.

[0032] Firstly, data items are collected in real time by means of one or more pressure transducers, connected to a data collection device. Each pressure transducer has a predetermined frequency. A primary processing of thedata items collected is carried out by means of algorithms.

[0033] The clean data items are provided in the correct format to a first neural network as input data. The first neural network is able to reproduce the characteristic curve of a container for aseptic products conforming to the one being examined.

[0034] Subsequently, the input data items are reconstructed by means of the first neural network, obtaining output data items.

[0035] A linear combination is then carried out on the input and output data items, so as to create a tensor that is a function of both said input and output data items.

[0036] The tensor is input into at least a second neural network, so as to obtain a certain class.

[0037] Obviously, there can also be more than one secondary neural network. Depending on whether the class thus determined falls within a predetermined range, a degree of confidence attributable to the second neutral network is expressed.

[0038] Therefore, detection of the anomaly is based on the use of two deep neural networks, the first of which is an autoencoder-type network, while the second is in fact a classifier / discriminator.

[0039] Advantageously, if the result of the classification complies with the pre-established range, then the container for aseptic products is safe and has no leaks.

[0040] Otherwise, if the result of this classification does not comply with the predetermined range, the container for aseptic products has a leak and therefore is not safe.

[0041] Advantageously, the device for collecting data items is independent within the measurement system.

[0042] This means that the device can also be added to existing systems which use the traditional, prior art vacuum decay method.

[0043] Not to mention that a stand-alone device can have its own power supply and data storage system, making the loss of essential data items less likelyand ensuring continuous data collection even in the event of faults or interruptions of the main system.

[0044] Preferably, the data collection device is integrated into the measurement system, improving coordination and synchronisation between data collection and analysis, with a lower likelihood of time delays or discrepancies.

[0045] Further, the integration reduces the complexity of the system, simplifying the installation, the configuration and the use since all the components are designed to work together.

[0046] Obviously, with the device directly integrated, there are fewer connection points and vulnerabilities, reducing the possibility of data transmission errors or malfunctions caused by outside components.

[0047] Advantageously, the pressure transducer collects the data on the entire curve of the test cycle, thus ensuring a complete and accurate detection of anomalies: by acquiring data on the entire curve, the system can identify not only the obvious losses but also small pressure variations throughout the cycle. This helps detect minimal losses that could be overlooked if data items were collected only at specific points in the cycle.

[0048] This also increases diagnostic reliability, since the continuous collection allows the entire pressure profile to be analysed, making it possible to identify gradual changes or unusual pressure spikes that may indicate progressive leaks or structural defects of the container.

[0049] With data items along the entire curve, it is possible to construct a detailed model of the response of the container to pressure over time. This model allows the neural network to compare the real curve with the theoretical one, identifying deviations that signal leakage or safety problems.

[0050] Consequently, by collecting data items throughout the cycle, the system has a complete picture of container behaviour, reducing the likelihood of evaluation errors and increasing confidence in the results.

[0051] Finally, the availability of continuous data along the entire test curve allows the measurement system to be better calibrated, optimizing the detectionfor each type of container and improving the sensitivity of the test.

[0052] Advantageously, the first neural network is of the autoencoder type and is suitable for receiving the input data items relating to the pressure curve of the test cycle and some process parameters, returning the output data items relating to the reconstructed curve of the associated compliant product. Using a first neural network of the autoencoder type to receive the input data (pressure curve and process parameters) and return the output data (reconstructed curve of the compliant product) offers several key advantages in the detection of the leaks and in the assessment of the compliance of the containers.

[0053] An autoencoder is designed to learn the optimal representation of the data items of a compliant product. When applied to the pressure curve and to the process parameters, it accurately reconstructs only the data items that correspond to a compliant container.

[0054] The anomalies are then detected by the second neural network and by the supporting analytical parameters: if the curve obtained differs greatly from that of the “theoretical compliance”, there will be reconstruction anomalies. In addition to this, the autoencoder can learn to ignore irrelevant variations and noise in the data items, focusing on the most relevant characteristics of the pressure curve of a compliant container. This reduces the risk of false positives caused by noise or fluctuations unrelated to actual losses.

[0055] During the encoding phase, the autoencoder reduces the size of the input data, extracting only the essential information. This process helps to isolate the distinctive features of the curve of a compliant container, reducing the volume of data and facilitating the processing by the second neural network. The autoencoder can also be trained on different pressure curves of compliant products to adapt to various types of containers and process conditions. This makes the system versatile, capable of detecting leaks on a wider range of products and situations.

[0056] The output data items generated by the autoencoder represent, in practice, a “filtered” version of the compliant pressure curve. This provides thesecond neural network with cleaner and more representative input, improving its ability to correctly classify containers as compliant or non-compliant.

[0057] Advantageously, the second neural network is a classifier or discriminator and it is suitable for receiving as input the tensor representing the real product and identifying the degree of similarity of this tensor with respect to that of a compliant product, providing as output the class and the degree of confidence.

[0058] The use of a second neural network of the classifying or discriminating type for receiving the tensor generated and determining the degree of similarity with respect to that of a compliant product firstly offers a high level of precision in the recognition of anomalies. The discriminating network is trained to accurately distinguish the characteristics of a compliant product from a non-compliant one and, by analysing the input tensor, is able to identify even small differences with respect to the expected behaviour, improving the detection of leaks or defects.

[0059] The discriminating network can only classify the product as compliant or non-compliant, but also provide a degree of confidence for each classification. This helps determine the safety of the assessment, making the system more reliable and allowing informed decisions to be made on which containers to pass or exclude.

[0060] In addition, the interpretability of the data by a human operator is improved. By receiving a tensor, the discriminating network is able to evaluate both the original input data and the data reconstructed by the autoencoder, exploiting the complex non-linear relationships between this information. This approach allows for a more robust and accurate classification, as it takes into account all the relevant characteristics of the test cycle.

[0061] Due to its discriminating nature, the network is particularly effective in avoiding classification errors, reducing the risk of false positives and false negatives.

[0062] Moreover, since the network is specifically trained to identify the “similarity”with the compliant product, it is also able to recognize subtle and gradual deviations in the structure of the tensor, which could indicate latent problems or initial damage which is not immediately visible.

[0063] The classification with a degree of confidence allows a more sophisticated decision-making process to be implemented. For example, different confidence thresholds can be established for various degrees of intervention: directly rejecting containers with a low confidence, subjecting those with a medium confidence to further checks, or accepting only those with a high confidence.

[0064] Over time, the discriminator can be continuously trained with new data items to refine its ability to recognise even imperceptible changes compared with the compliant product, keeping the system aligned with any production updates.

[0065] This training is preferably carried out “off-line”, that is, not on the container products, so as not to compromise the validation status of the machine and / or modify the specifications.

[0066] Description of the drawings

[0067] Further features and advantages of the invention are more apparent in the detailed description below, with reference to a preferred, non-limiting embodiment of the method for detecting leaks in a container for aseptic products, illustrated by way of example and without limiting the scope of the invention, with the aid of the accompanying drawings, in which:

[0068] Figure 1 shows schematically a diagram of the testing system, applicable to inline, semi-automatic and manual solutions. In the semi-automatic or manual systems, the compliant / non-compliant selection can be handled by the operator or also manually. The diagram applies to all types of container C, the STC testing chamber, and, in general, to testing cycles based on a pressure differential;

[0069] Figure 2 shows schematically the first neural network, which receives as input the testing cycle pressure curve and some process parameters, returning the reconstructed curve Xu of the associated compliant product;Figure 3 shows schematically the second neural network, which takes as input a tensor and determines the degree of similarity of this tensor to that of a compliant product, providing as output a classifier Y and a confidence level P.

[0070] Detailed description

[0071] The above-mentioned drawings show a preferred embodiment of a method for detecting leaks in a container C for aseptic products, according to the invention, which is identified in its entirety by reference numeral 1 and comprises an inference step comprising the following steps.

[0072] With reference to Figure 1 , the container C is located in a hermetically sealed chamber STC suitable for use in testing.

[0073] The chamber STC is connected to a vacuum suction pump V.

[0074] Data items are collected in real time and over the entire curve of the test cycle, by means of a pressure transducer T mounted on the chamber STC, which has a predetermined frequency F, connected to a data collection device DCU, which is independent and integrated into the measurement system. The data items collected DR are primarily processed by means of algorithms, providing clean data items DP.

[0075] For the detection of anomalies, two deep neural networks are provided: a first autoencoder type neural network PRN and a second neural network SRN (a classifier), which operate via a Neural Network Computer Unit (NNCC).

[0076] These clean data items DP are used as input data Xi for the first neural network PRN, which can reproduce the characteristic curve of a container C for aseptic products compliant with the one under examination.

[0077] The input data items Xi are then reconstructed by means of the first neural network, obtaining output data items Xu,

[0078] In the preferred solution, this first neural network PRN is of the autoencoder type and is suitable for receiving the input data items Xi relating to the pressure curve of the test cycle and some process parameters, returning the output data items Xu relating to the reconstructed curve of theassociated compliant product (Figure 2).

[0079] The autoencoders are a type of neural network capable of compressing input data Xi into a compressed representation (generally called latent space) Z. The encoder performs this “compression”. Starting from this compressed representation, which in fact tries to capture the main characteristics of the input data Xi, the network then tries to reconstruct the original data, giving as output Xu: this ’’reconstruction” is carried out by the decoder.

[0080] As mentioned above, the autoencoder has an excellent ability to learn the characteristics of the input data Xi (for example, the shape of the curve, any variations in slope), then reproducing the learned characteristics when presented with a new input. These capabilities make it superior to a simple classification system, because the network of the autoencoder does not classify the compliant items into one category and the rejects into another, but learns what the pressure curve of a compliant product looks like. Thanks to this feature, it is possible to compare the “theoretical good curve” generated by the autoencoder with the curve actually read by the machine: if they are sufficiently similar, the product is compliant; otherwise, it is a reject.

[0081] Examples of autoencoders tested are as follows:

[0082] VAE (Variational AutoEncoder)

[0083] CAE (Conditional AutoEncoder)

[0084] PIAE (Physics Informed AutoEncoder)

[0085] CoAE (Convolutional AutoEncoder)

[0086] Combinations of various Autoencoder architectures

[0087] In practice, a linear combination is carried out on the input data items Xi and the output data items Xu so as to create a tensor as a function of the above-mentioned data items:

[0088] T=f(Xi,Xu)

[0089] This tensor is entered into a second neural network SRN, so as to obtain a classification Y.The above-mentioned second neural network SRN is a classifier or discriminator and it is suitable for receiving the tensor as input and identifying the degree of similarity of that tensor with respect to that of a compliant product, providing as input classification Y, as well as the degree of confidence P, by means of an appropriate automation (figure 3).

[0090] Although, as mentioned, it is a classifier, the second neural network SRN does not compare a compliant product with a reject (anomaly), but compares the curve reconstructed by the autoencoder with the input curve, learning which differences are characteristic of a compliant product and which are characteristic of a reject.

[0091] In theory, it would be possible to have a similar result even by setting fixed thresholds calculated dynamically during the training of the first network; in practice the separation between compliant and non-compliant products is not always strictly linear, which makes fixed thresholds less suitable for the aim.

[0092] As mentioned, the use of the second network, instead of “fixed” thresholds (whether calculated by the system itself or by developer of the system) enables the capacity of the neural network to be exploited to perform nonlinear separations, providing a much more flexible approach and with a greater ability to identify anomalies, especially in borderline cases where traditional techniques cannot distinguish compliant from non-compliant. The discriminator returns a classification Y (compliant or anomalous) and a P value which express the degree of confidence in the Y classification Y. This helps to identify with precision the containers which have leaks, with a reliable indication of the level of safety of the test.

[0093] Like all the neural networks, also in this case a large quantity of data items is processed efficiently.

[0094] This characteristic does not stem solely from the architectures chosen, but above all from the fact that neural network systems are mainly based on linear algebra operations such as multiplications between matrices.

[0095] The choice of using tensors is therefore advantageous, since they areexcellent tools in this context: in fact, the same tensor can represent matrices of various ranges based on its size. It is therefore necessary to convert the input data into tensors of appropriate order / size; linear combinations are then formed to create tensors representing the characteristics sought.

[0096] From the above, it can be seen how the invention achieves the proposed purpose and aims and in particular it should be noted that a method is implemented for the detection of leaks in a container, particularly for aseptic products, which, by collecting data across the entire curve of the testing cycle, enables a more robust and detailed evaluation, improving the accuracy and the sensitivity of the leak detection and the overall safety of the testing process, providing an improved analysis capability.

[0097] This sequence of inference phases therefore allows the learning capacity of neural networks to be exploited, automating and perfecting the process for quality control of the container, with a high precision and optimised risk management.

[0098] On the contrary, according to the methods currently in use, the data items are processed extensively, carrying out various analytical compensations; the values compared are often not the raw ones from the process, but data items that undergo various calculation steps before being usable by the system and being able to guarantee good leak detection performance levels.

[0099] In particular, the use of an autoencoder as the first neural network provides an intelligent filter that isolates the essential characteristics of a compliant product, improving the precision and reliability of the system in detecting leaks or defects.

[0100] This is followed by the passage of the tensor in the second neural network SRN of a discriminating type that allows the system to be enriched with a highly specific and detailed classification, providing more reliable and interpretable results, which are crucial for a rigorous and effective quality controlUnlike the prior art leak detection methods, which involve defining experimental thresholds on a statistical basis to compare with the point pressure values, this approach gives a broader and more complete representation of the process, allowing the complexity of the entire test cycle to be fully captured.

[0101] The invention described can be modified and adapted in several ways without thereby departing from the scope of the inventive concept.

[0102] Moreover, all the details of the invention may be substituted by other technically equivalent elements.

[0103] In practice, the materials used, as well as the dimensions, may be of any type, depending on requirements, provided that they are consistent with their production purposes.

Claims

CLAIMS1) A method (1) of detecting leaks in a container (C), particularly for aseptic products, comprising an inference step of:- collecting data items in real time by means of at least one pressure transducer (T) connected to a data collection device (DCU); the pressure transducer (T) having a predetermined frequency “F”,- primarily processing the collected data items (DR) by means of algorithms, - providing the clean data items (DP) as input data items (Xi) for the first neural network (PRN); said first neural network being able to reproduce the characteristic curve of a container (C) conforming to the one being examined,- reconstructing the input data items (Xi) by means of the first neural network, obtaining output data items (Xu),- carrying out a linear combination on the input (Xi) and output (Xu) data items so as to create a tensor that is a function of said input (Xi) and output (Xu) data items,- entering the tensor into a second neural network (SRN), so as to obtain a classification (Y).2) The method (1) according to claim 1 , comprising the step of expressing a degree of confidence (P) that the second neural network has, as a function of whether the result of the classification (Y) conforms to or does not conform to said predetermined range.3) The method (1) according to claim 1, wherein if the result of the classification (Y) conforms to the predetermined range, then the container (C) is safe;if conversely said result of the classification (Y) does not conform to the predetermined range, the container (C) is not safe.4) The method (1 ) according to one or more of claims 1 -3, wherein the data collection device (DCU) is independent in the measurement system.5) The method (1 ) according to one or more of claims 1 -3, wherein the data collection device (DCU) is integrated into the measurement system.6) The method (1) according to one or more of claims 1-5, wherein the pressure transducer (T) collects the data (DR) over the entire curve of the test cycle.7) The method (1 ) according to claim 6, wherein said curve of the test cycle has, between the start “Ti” and the end“Tf” of the curve, an indefinite number “n” of points corresponding to the equation, depending on the predetermined frequency “F”:n=(Ti-Tf) / F8) The method (1) according to one or more of claims 1-7, wherein the first neural network (PRN) is of the autoencoder type and is suitable for receiving the output data items (Xi) relating to the pressure curve of the test cycle and some process parameters, returning the output data items (Xu) relating to the reconstructed curve of the associated compliant product.9) The method (1) according to one or more of claims 1-7, wherein the second neural network (SRN) is of classifying or discriminating type.10) The method (1 ) according to claim 9, wherein the second neural network (SRN) is suitable for receiving as input the tensor and identifying the degree of similarity of said tensor with respect to that of a compliant product, providing as output the classification (Y) and the degree of confidence (P).

Images