METHOD, SYSTEM AND COMPUTER PROGRAM PRODUCT FOR MONITORING A SHAPING PROCESS

Description

[0001] The present invention relates to methods for monitoring a forming process carried out in cycles according to the features of the preamble of claim 1 and to a computer program product for monitoring a forming process carried out in cycles according to the features of the preamble of claim 19.

[0002] Methods are known whereby at least the following data is provided in a data collection: for past cycles, cycle data of the forming process carried out on a forming machine and, for at least some of the past cycles, quality data of molded parts produced with the forming machine.

[0003] During further operation of a forming machine to carry out the forming process, at least one further cycle of the forming process is carried out, whereby cycle data is again generated, which are referred to in this document as current cycle data.

[0004] It is known to use this data to predict the quality of the molded parts produced in the current cycle. In principle, this can be used to monitor the molding process, for example, to alert users when production is imminent or already underway.

[0005] Various approaches are known for calculating or determining this prediction. DE 112011103778 T5 and its US family member US 2012 / 0123583 A1 disclose the use of a multivariate model for controlling an injection molding process. DE 102017104884 A1 discloses the monitoring of a process based on defect patterns. DE 102018102299 A1 discloses an optical inspection system for detecting features of injection-molded parts.

[0006] US 2006 / 0224540 A1 discloses a system for monitoring an injection molding machine using multiple neural networks. US 2021 / 0008774 A1 discloses anomaly detection including automated root cause analysis based on a qualitative model. DE 19514535 A1 discloses the collection of process data to perform monitoring in conjunction with a predictive model.

[0007] The disadvantage of the systems known in the state of the art is that either great effort is required to provide testing systems or that effective monitoring is not ensured because the predictions of the quality of the molded part are not good enough.

[0008] Especially in injection molding processes, predicting the quality of the molded part is anything but trivial, not only because injection molding processes are inherently complex and nonlinear, but also because injection molding processes are subject to numerous changes in reality. For example, even a change in the outside temperature or slight adjustments in the molding machine settings lead to changes that are clearly reflected in the measured values ​​obtained from the machine. It is noteworthy, however, that such changes often do not have a negative impact on the manufactured molded part. In fact, a modified process can often simply continue because good parts continue to be produced.

[0009] The challenge with such molding processes is therefore to determine whether changes in the process have a negative impact on the quality of the molded part, which is difficult in practice. Simulation models are often difficult to compare with the minute changes that occur on the real machine. Most statistical models require large amounts of data to make reliable predictions, which are often unavailable, especially at the start of production.

[0010] The invention therefore has for its object to provide a method and a computer program product that allow improved monitoring of a forming process.

[0011] With regard to the method, the problem is solved by the features of claim 1, namely by the following steps: Determining at least two nearest neighbors in the form of the cycle data from at least two of the past cycles, such that the cycle data of the at least two nearest neighbors are closer to the current cycle data than those cycle data that do not belong to the at least two nearest neighbors, wherein only those past cycles for which quality data is contained in the data collection are used to determine the at least two nearest neighbors, Checking a predictability criterion, wherein it is checked whether a quality scatter of the quality data of the cycles of the at least two nearest neighbors is smaller than a maximum scatter and / or larger than a minimum scatter, and if the predictability criterion is not met, outputting a first message that a quality and / or a quality date of the molded part produced with the at least one further cycle cannot be reliably predicted.

[0012] Preferably, it may be provided that, as part of an anomaly check, it is checked whether the current cycle data is an outlier and, if this is the case, at least one of the following is carried out for at least one subsequent cycle - preferably several subsequent cycles: Activating a device for collecting quality data and / or requesting an operator to collect the quality data and / or controlling a reject switch and / or depositing at least one molded part produced in at least one further cycle in a test deposit and / or marking at least one molded part produced in at least one further cycle.

[0013] With regard to the computer program product, the problem is solved by the features of claim 19, namely by the computer program product containing instructions for performing the following steps: Determining at least two nearest neighbors in the form of the cycle data from at least two of the past cycles, such that the cycle data of the at least two nearest neighbors are closer to the current cycle data than those cycle data that do not belong to the at least two nearest neighbors, wherein only those past cycles for which quality data is contained in the data collection are used to determine the at least two nearest neighbors, Checking a predictability criterion, wherein it is checked whether a quality scatter of the quality data of the cycles of the at least two nearest neighbors is smaller than a maximum scatter and / or larger than a minimum scatter, and if the predictability criterion is not met, outputting a first message that a quality and / or a quality date of the molded part produced with the at least one further cycle cannot be reliably predicted.

[0014] Preferably, the computer program product may include further instructions for performing the following steps: Carrying out an anomaly check, checking whether the current cycle data is an outlier and, if this is the case, for at least one subsequent cycle - preferably several subsequent cycles - causing at least one of the following: activating a device for collecting quality data and / or requesting an operator to collect the quality data and / or controlling a reject switch and / or depositing at least one molded part produced in the at least one further cycle in a test tray and / or marking at least one molded part produced in the at least one further cycle.

[0015] A basic aspect of the invention is that the available data are not only used to make a prediction of the quality of the currently produced molded part, but that - preferably first - it is checked whether an appropriate prediction is even possible with the available data, i.e. checking a predictability criterion.

[0016] In particularly preferred embodiments, for example, a review of the quality of the produced molded part can be requested if the predictability criterion is not met.

[0017] This provision or request of quality data can be done in various ways, namely by Activating a device for collecting quality data and / or requesting an operator to collect the quality data and / or controlling a reject switch and / or depositing at least one molded part produced in at least one further cycle in a test deposit and / or marking at least one molded part produced in at least one further cycle.

[0018] In other words, the quality data should be collected in whatever way possible, or the information should be disseminated that the quality data must be collected.

[0019] According to the invention, it is neither necessary to use complex systems for the continuous, direct monitoring of the quality of the currently produced molded part, nor is it necessary to completely forego effective monitoring of the molding process.

[0020] A further advantage of the invention is that it can be implemented in a relatively simple manner by simply checking the predictability using data from those previously performed cycles that are as close as possible to the current cycle, ie by using the cycle data of the cycles of at least two nearest neighbors.

[0021] The maximum and minimum scatter can each be considered limit values. If the quality scatter of the quality data from the cycles of at least two nearest neighbors is below the minimum scatter and / or above the maximum scatter, a reliable prediction of the quality of the molded part produced in at least one subsequent cycle cannot be expected. If the scatter is too small (quality scatter below the minimum scatter), this is due to the molded part quality being too dependent on the recorded cycle data. If the scatter is too large (quality scatter above the maximum scatter), a reliable prediction is obviously also not possible.

[0022] The fact that a quality or quality date of the molded part produced with at least one additional cycle cannot be reliably predicted can be understood to mean that the accuracy of the prediction with the available data and, if applicable, the prediction model used is likely too low to be useful. The fact that there could in principle exist (e.g., more elaborate or complex) models or data with which a reliable prediction is possible, or that the prediction is accurate for one cycle even though the predictability criterion is not met, is irrelevant.

[0023] Preferably, it can be provided that the cycle data and / or the quality data are each uniquely assigned to a certain past cycle. However, investigations by the applicant have shown that the method according to the invention can verify predictability with high accuracy even if the cycle data and the quality data are assigned only to certain groups of cycles (i.e., the information regarding the exact cycle in which a certain cycle date or a certain quality date occurred is lost and replaced by an approximate assignment to a group of cycles).

[0024] Past cycles may be cycles that were performed immediately before the next cycle and / or that were performed some time before (for example, during a previous production batch).

[0025] The determination of at least two nearest neighbors is based on a metric that defines the distance between the elements of the cycle data. Elements of the cycle data are defined as those data associated with a cycle (where, as explained above, it is not necessary to know which past cycle this actually was).

[0026] For example, if three measurements are taken in each cycle, each element of the cycle data consists of three measurements. The distances between the elements must therefore be determined in three-dimensional space in this example. In real forming processes, far more than just three measurements are taken per cycle, which is why the cycle data space usually has many more dimensions than three.

[0027] The distances in this space can be determined, for example, using the well-known p-norm. The Euclidean norm (p=2, "normal" Euclidean geometry) or the sum norm (p=1, Manhattan or taxi geometry) can be preferred.

[0028] The quality data may contain at least one (continuous or discrete, in particular binary) quality parameter.

[0029] The at least one quality parameter can serve as a criterion for the quality of the molded parts produced, in particular for whether it is a good part or a reject part.

[0030] For this purpose, it can be provided, for example, that operators specify a permissible tolerance range, whereby the molded part is classified as a good part if the at least one quality parameter lies within this range, and is classified as a reject part if the at least one quality parameter lies outside this range.

[0031] In particularly preferred embodiments, the method according to the invention is carried out during ongoing operation of the forming machine in each new cycle.

[0032] Particularly preferably, the method according to the invention can be implemented automatically and / or computer-implemented.

[0033] In addition to the information that a quality and / or quality date of the molded part produced with at least one additional cycle (at the current operating point) cannot be reliably predicted, the first message may contain, for example, possible reasons for the lack of predictability. Possible reasons for the lack of predictability could be, for example, insufficient quality recording or the fact that not all relevant process parameters are recorded as cycle data.

[0034] Of course, the first message does not have to literally state that a quality and / or a quality date of the molded part produced with the at least one additional cycle cannot be reliably predicted. What is meant is that the information is conveyed that either the quality of the molded part produced with the at least one additional cycle as a whole ("the quality") or one or more specific aspects of the quality (e.g., compliance with dimensional tolerances) cannot be reliably predicted. An aspect of the quality of a molded part can be quantified by a quality date, for example, as a quality parameter.

[0035] It may be provided that an evaluation of all those operating points (cycle data) of the forming process and / or the quality data is output for which the predictability criterion is not met, and / or whereby temporal changes in the relationships between the cycle data and the quality data are output.

[0036] Protection is also sought for a computer-readable storage medium on which a computer program product according to the invention is stored.

[0037] Furthermore, protection is sought for a shaping machine which is configured to carry out a shaping process to be carried out in cycles and which includes at least one memory and at least one processor unit, wherein a computer program product according to the invention is stored in the at least one memory and the at least one processor unit is configured to execute the computer program product according to the invention.

[0038] Forming machines can include injection molding machines, transfer presses, presses or the like.

[0039] The data collection is sometimes also referred to as a database.

[0040] It should be noted that in some machine configurations, multiple molded parts are produced in one cycle (e.g., in an injection mold with multiple mold cavities). Likewise, within the scope of the invention, more than one additional cycle can be performed, which are then tested as a group according to the invention. For the purposes of this document, for the sake of simplicity, reference is made to the molded part in the singular. However, particularly in the cases described, the possibility of multiple molded parts being produced in at least one additional cycle is also always to be understood.

[0041] Further advantageous embodiments of the invention are defined in the dependent claims. The optional measures described in connection with the method according to the invention can, of course, also be used in connection with or implemented in the computer program product according to the invention.

[0042] It may be provided that the quality data of the past cycles are collected by measuring the molded parts and / or by human inspection of the molded parts.

[0043] The quality dispersion can be calculated as the – preferably standardized – difference between a maximum quality value of the at least two nearest neighbors and a minimum quality value of the at least two nearest neighbors. The maximum quality value is understood to be the maximum value of the quality data of the cycles of the at least two nearest neighbors. Similarly, the minimum quality value is understood to be the minimum value of the quality data of the cycles of the at least two nearest neighbors.

[0044] The quality variance of the quality data can be used when checking the predictability criterion in a form where the quality variance is related to a cycle data variance of the cycle data of at least two nearest neighbors. For example, the quality variance can be divided by a cycle data variance before being compared with the maximum variance or the minimum variance to check the predictability criterion.

[0045] The cycle data scatter can be calculated as the - preferably normalized - maximum distance between at least two nearest neighbors.

[0046] Statistical values ​​of the cycle data and / or the quality data can be used for standardization.

[0047] As part of an anomaly check, it can be checked whether the current cycle data is an outlier. If so, a second message can be issued, for example, indicating that the current cycle data exhibits a significant deviation from the historical cycle data. The second message can, of course, also be issued within the scope of the variant of the invention.

[0048] The anomaly check can be performed using angle-based outlier detection (ABOD).

[0049] Alternatively or additionally, the anomaly check can determine the distances between the current cycle data and the cycle data (distances between elements, as above). For example, the current cycle data could be classified as an outlier if a minimum of these distances is greater than a specified threshold.

[0050] This threshold can be set by an operator and / or specified as a sensitivity parameter and / or the average intervals of the cycle data. This can be based on either all cycle data or only the cycle data during a defined training phase.

[0051] In this document, a sensitivity parameter is understood to be a value based on a maximum relative deviation, which can be calculated, for example, as an absolute deviation relative to a mean deviation.

[0052] Alternatively or additionally, operators may be provided with a sampling rate, for example a number s of samples to be taken per number Z of cycles, where Z could be fixed at 100, for example, so that operators only have to enter the number s of samples (per 100 cycles).

[0053] It could then be provided that those of the Z cycles are selected for taking a sample (where the quality of the molded part is checked by an operator and / or a device for collecting quality data) whose cycle data have the greatest distance from the other cycle data and / or whose quality forecast is closest to a limit of the permissible tolerance range.

[0054] The second message may also include information about which process parameters contributed most or most to the current cycle data being classified as an outlier. In particular, those process parameters could be identified that contribute most or most to the difference between the current cycle data and the cycle data.

[0055] The anomaly check can preferably be performed before checking the predictability criterion. If the anomaly check detects an outlier, the predictability criterion no longer needs to be checked. In principle, however, the predictability criterion could still be checked.

[0056] The cycle data and / or the current cycle data may include setting data of the forming machine that are assigned to the past cycles or to the at least one further cycle.

[0057] If the cycle data and the current cycle data contain setting data of the forming machine, a setting check can be performed to determine whether a setting distance of the setting data assigned to at least one further cycle from the setting data of the cycle data is less than a predetermined maximum setting distance. If this is not the case, a third message can be output. Thus, it can be determined whether and / or when the setting of the forming machine has changed significantly.

[0058] This maximum adjustment interval can be set by an operator and / or as a sensitivity parameter and / or the average intervals of the adjustment data. This can be done using either all adjustment data or only the adjustment data during a defined training phase.

[0059] Alternatively or additionally, the setting check can also be carried out separately for each setting parameter and / or each other setting and it can be checked whether a setting has been changed beyond a certain extent.

[0060] The setting check can preferably be performed before the anomaly check and the predictability criterion check. The anomaly check and / or the predictability criterion check according to the invention can, for example, only be performed if the setting check has not detected any changes in the forming machine's settings (e.g., changes beyond the maximum setting range).

[0061] The cycle data and / or the quality data can be transformed and dimensionally reduced, and the transformed and dimensionally reduced cycle data and / or the transformed and dimensionally reduced quality data can be used for determining the at least two nearest neighbors and / or checking the predictability criterion and / or anomaly checking, whereby the necessary resources, for example for determining the nearest neighbors or calculating the quality variance, can be significantly reduced.

[0062] The transformed and dimensionally reduced cycle data and / or quality data may also be used for setting checks and / or any other calculations or data evaluations of the data in the data collection.

[0063] The cycle data and / or the quality data can preferably be transformed and dimensionally reduced by means of a multilinear regression model, in particular a partial least squares regression and / or a principal component regression.

[0064] The cycle data and / or the current cycle data may include process data collected during the past cycles and / or during the at least one further cycle by process measurement and / or reading from a machine control system, wherein the process measurement and / or the reading from the machine control system can be carried out on the forming machine itself or on a forming system that includes the forming machine.

[0065] Process measurement can be understood as any measurement of a physical or chemical quantity that occurs during the cycle.

[0066] The process data can include time series of measured or read data, which are generated during the cycle, for example, with a defined (or variable) sampling time. These time series can be directly incorporated into the cycle data and / or other cycle data. Alternatively or additionally, preferred parameters (e.g., maximum values, integrals, etc.) can be extracted from the time series and included in the data collection.

[0067] If the predictability criterion is met, a quality forecast based on the cycle data and the quality data can be output, for example in the form of at least one forecast value for the at least one quality parameter.

[0068] The quality forecast can be calculated using at least one of the following: Mean of the quality data of the cycles of the at least two nearest neighbors, preferably weighted according to the distances of the at least two nearest neighbors from the current cycle data Median of the quality data of the cycles of the at least two nearest neighbors multilinear regression model

[0069] The multilinear regression model can, for example, be a so-called "partial least squares" regression (PLSR) or a principal component regression (PCR), which are known in the state of the art.

[0070] Combinations of averaging and a multilinear regression model are also possible, for example if the quality data contain more than one quality parameter.

[0071] Together with the quality forecast, in particular the at least one forecast value for the at least one quality parameter, a confidence interval can be output within which the actual quality or the actual quality date lies with a given probability.

[0072] Alternatively or additionally, at least one of the following may be provided if the quality forecast, in particular the at least one forecast value, does not fall within the specified tolerance range: Outputting a fourth message and / or activating a device for collecting quality data and / or requesting an operator to collect the quality data and / or controlling a reject switch and / or depositing at least one molded part produced in at least one further cycle in a test deposit and / or marking at least one molded part produced in at least one further cycle.

[0073] This can be done for at least one subsequent cycle, preferably several subsequent cycles.

[0074] The specified tolerance band can, for example, be entered by the operator and / or be based on desired quality requirements, data from the design phase of the molded part and / or the mold tool, or the like.

[0075] If the predictability criterion is not met, at least one of the following may be performed for at least one subsequent cycle - preferably several subsequent cycles: Activating a device for collecting quality data and / or requesting an operator to collect the quality data and / or controlling a reject switch and / or depositing at least one molded part produced in at least one further cycle in a test deposit and / or marking at least one molded part produced in at least one further cycle.

[0076] This means that it may be useful to request a sample of the quality of the manufactured molded parts if predictability is not guaranteed. This can be done for a defined number of cycles.

[0077] It should be noted that the first message and / or the second message and / or the third message can be issued as a human-understandable message (e.g. text, diagram and / or warning tone) or as a machine-understandable message (e.g. only controlling the reject switch).

[0078] The device for collecting quality data may, for example, be an optical or infrared system for automatically recording quality characteristics of a molded part, which are known per se in the prior art.

[0079] For example, if such a device is not available, an operator can simply be prompted to check the quality data, for example, by issuing a (visual or acoustic) signal. In certain embodiments, the signal can simply consist of activating the reject gate, thereby sending a molded part to be inspected for quality to an operator or another person for inspection of the molded part's quality.

[0080] It may be provided that an evaluation of all those operating points (cycle data) of the forming process and / or the quality data is output for which the predictability criterion is not met, and / or temporal changes in the relationships between the cycle data and the quality data are output.

[0081] It may be provided that the current cycle data are added to the cycle data and / or if current quality data - in particular those which are provided after activating a device for collecting quality data, after requesting an operator to collect the quality data and / or controlling a reject switch by an operator or the device for collecting quality data - are available for the at least one further cycle, the current quality data are added to the quality data.

[0082] This allows the current cycle data and / or current quality data for subsequent cycles to be already available in the data collection, enabling a better prediction of the quality of the molded parts produced. Shifts in the operating point of the molding process can thus be responded to easily, effectively, and quickly.

[0083] Alternatively or additionally, cycle data and / or quality data can also be removed from the data collection, for example if they are already of a certain age - ie a lot of time has passed since the data were collected or many cycles have been carried out - and therefore they can no longer be expected to be representative of the forming process in its current form.

[0084] In other words, the data collection can be adapted as follows: Continuous overwriting of old cycle data and / or quality data with current cycle data and / or quality data if the difference between the old and the current cycle data and / or quality data is less than one or more specified limit values ​​(if necessary separately for cycle data and quality data) and / or the possibility of resetting the entire system in the event that significant changes have been made to the forming system and / or forming machine.

[0085] The term "overwrite" is, of course, to be understood in a figurative sense here. Preferably, the data is not actually overwritten on the storage medium, but rather is removed from the data collection in some way and replaced with the current cycle and quality data.

[0086] In this way, effective monitoring of the molding process can be achieved without having to constantly retrain the model with complex machine learning techniques as the underlying data changes. Instead, predictions about the quality of the currently produced molded part are made using constantly updated data (so-called "lazy learning").

[0087] According to a preferred embodiment of the invention, training can be carried out simply by collecting data. At least one of the following can be provided: a training mode, whereby cycles are carried out and the measured and / or collected cycle data and / or quality data are added to the data collection, and / or if the distance between newly collected and / or measured cycle data and already available (historical) cycle data does not exceed a predetermined maximum distance, the newly collected and / or measured cycle data are added to the cycle data, and / or if random or other measurements or surveys of the quality are carried out, these are added to the quality data, preferably with assignment to the respective cycle (and of course the cycle data of the cycle are added to the cycle data).

[0088] It can be planned that the data collection consisting of cycle data and quality data, and if applicable, an associated mathematical model (e.g., a multilinear regression model), are saved for a molding machine configuration and are available with this configuration when production resumes. For injection molding machines, a machine configuration could, for example, be determined by a tool / material combination.

[0089] Examples of quantities that can be used as cycle data would be: maximum and average values ​​of injection pressure, torque, injection times, dosing times, injection pressure curve or the dosing torque.

[0090] In addition, cycle times, partial cycle times, parameters from peripheral devices (dosing devices, external hot runner, temperature control devices) and / or parameters from other process monitoring systems or process assistance systems can be part of the cycle data.

[0091] Examples of quality data would be: molded part weight, dimensions or surface finish.

[0092] In accordance with the invention, the data collection can be stored locally on the forming machine. However, this is not mandatory. For example, the data collection can also be stored in a cloud storage.

[0093] Analogously, the processor unit intended to execute the computer program product according to the invention can also be arranged on the forming machine or implemented as a cloud computer.

[0094] Further advantages and details are shown in the figures and the accompanying figure descriptions. They show: Fig. 1 is a flow chart of an embodiment of a method according to the invention, Fig. 2a and 2b are two diagrams illustrating partial least squares regression, Fig. 3 is a diagram illustrating angle-based outlier detection, Fig. 4 is a diagram of a course of an actual process parameter and a process parameter predicted by lazy learning, Fig. 5 is a flow chart for commissioning and operating a method according to the invention, and Fig. 6 is a schematic representation of a molding system according to the invention.

[0095] Fig. 1 First, an overview flowchart of an embodiment of a method according to the invention is shown. As soon as a cycle or at least part of the cycle (subcycle) is completed, the method can begin.

[0096] It should be noted that the required current cycle data, which includes setting data and process data, can be available as measured values ​​or from a machine control 3 of the forming machine 1 (see Fig. 6 ) can be read out. The machine control 3 of the forming machine 1 can also serve as an operator interface.

[0097] First, within the framework of the "Input Analysis - Unknown Setting Parameters" (setting check), the setting data are checked to see whether they have a distance from the setting data that is available as part of the cycle data that is smaller than a predefined maximum setting distance.

[0098] If this is not the case (decision branch N), the third message can be output, which, for example, indicates that a new process setting has been detected.

[0099] If the setting check is positive (decision branch J), the next step is "Input Analysis - Anomaly Detection" (anomaly check). This checks whether the current cycle data is close to the cycle data contained in the data collection (see Fig. 2a and 2b ) . It should be noted that not only the setting data contained in the current cycle data and the cycle data are compared with each other, but if possible all existing (current) cycle data including all process data.

[0100] If the anomaly check fails, i.e., if an outlier is detected (decision branch N), a "detailed analysis of the anomaly" (the outlier) can optionally be performed. For example, if the procedure is still in a training phase, the subsequent second message can be suppressed or accompanied by an appropriate note.

[0101] Subsequently or in parallel to the detailed analysis of the anomaly, the second message can be output if an outlier is detected (decision branch N), in this embodiment in the form of a "result representation of [the] process anomaly".

[0102] This may, for example, include an analysis of the outlier, which is displayed on a user interface of the molding machine—in this embodiment, an injection molding machine (IMM)—or on a PC for operators, possibly via an internet-based customer portal of the manufacturer of the molding machine 1.

[0103] In addition, the second message can prompt operators to check the quality of the produced molded part.

[0104] If no outlier is detected (decision branch J), the predictability criterion can be tested.

[0105] If it turns out that the predictability criterion is not met, the first message is issued (in the flowchart from Fig. 1 not explicitly shown).

[0106] For example, a text message can simply be issued stating that the quality and / or quality date of the molded part produced with at least one further cycle cannot be reliably predicted (and therefore quality monitoring for this cycle is not available or is only available to a limited extent).

[0107] Alternatively or additionally, at least one of the following may be performed: Activating a device for collecting quality data and / or requesting an operator to collect the quality data and / or controlling a reject switch and / or depositing at least one molded part produced in at least one further cycle in a test deposit and / or marking at least one molded part produced in at least one further cycle.

[0108] If the predictability criterion is not met (decision branch N) and if the model used to predict the quality or the quality date or the corresponding data allow it, a (further) extrapolated prediction of the quality or the quality date could still be calculated and output.

[0109] If the predictability criterion is met (decision branch J), the quality forecast can be created. This can preferably be done using lazy learning.

[0110] It may also be provided that a "forecast tolerance", ie a confidence interval, is issued with the forecast.

[0111] In this embodiment, it is also provided that the current cycle data is added to the cycle data ("extension of the input database"). Note that the data collection could also be referred to as a "database." The cycle data could also be referred to as "inputs."

[0112] However, adding the current cycle data doesn't have to happen under all circumstances. For example, this could only happen if there is a deviation (of a certain magnitude) between the current cycle data and the cycle data. Alternatively, this could also happen if, for example, an outlier was detected, but the operating personnel confirmed that it was a "normal" cycle.

[0113] In addition, certain elements of the cycle data can optionally be removed from the data collection.

[0114] Individual steps in connection with Fig. 1 The procedure described will be discussed in more detail below.

[0115] Optionally, as mentioned, the reject switch could be triggered if, for example, a quality parameter or a quality indicator (as a prediction of a quality date) and / or the corresponding confidence range lies outside the tolerance range for the quality of the molded part (could also be referred to as a component).

[0116] It should be noted that the first message, the second message, and the third message need not be output in chronological order. In fact, in preferred embodiments, at most one of these messages is output.

[0117] The in connection with Fig. 1The process steps described are described in more detail in the exemplary embodiment described below. This is an injection molding process as an example of a molding process, with one or more molded parts 2 being produced per cycle. The molded part 2 produced in this way or the entirety of the molded parts 2 produced in a cycle can also be referred to as a unit or product.

[0118] The following data is collected: The injection molding machine records cycle-based individual values ​​and time series (cycle data). Furthermore, cycle-based component characteristics, such as component length or weight, are measured for quality prediction (quality data).

[0119] The type and number of recorded signals remains as unchanged as possible over the lifetime of the forecasting system.

[0120] In practice, component quality measurements are often performed manually and then only on a random basis. This is, on the one hand, a key reason for the relevance of quality forecasting. On the other hand, the system described here can also help determine suitable sampling cycles. After all, the calculations already include a pure cycle data analysis, which can also be used to evaluate the process without the need for quality measurements.

[0121] In this example, per cycle p various cycle-based values ​​recorded ( u 1 , ... up ). ∘ e.g. maximum and average values ​​of injection pressure, torque, injection times, dosing times, etc. q different time series with s samples each ( W 1.1: s , ... W q ,1: s ) . ∘ e.g. the injection pressure curve, the dosing torque, ... as well as 1Value of the quality measurement (e.g. the component weight) is taken.

[0122] Because the sampling rate of the time series is very high, the data is downsampled by 1 / f. Alternatively, only certain features from the time series can be used, such as rise times, minimums, maximums, or averages.

[0123] This results in m = p + q ∗ s f Cycle values ​​and 1 Quality value.

[0124] The system requires an initial database (i.e., the data collection) before anomalies can be detected and quality characteristics predicted. This database can be loaded, for example, from a previous production batch of the same molded part 2, or collected during production.

[0125] With n samples (cycles) in the database, this results in a total data set of n*m values ​​plus n target values ​​(quality values ​​of a quality parameter). Data preparation & modeling for dimensionality reduction

[0126] Since the absolute values ​​of the individual signals are included in the model in all subsequent calculations, it is advantageous to normalize the cycle data. This ensures that fluctuations of all signals, regardless of their unit, are equally included in the model. Common normalization methods can be used for normalization, such as ZScore, or normalizing the initial database to the range [0, 1] or [-1, 1].

[0127] To subsequently apply the system to new data (the current cycles), this current cycle data must be normalized using the same normalization parameters as the initial database. The normalization parameters are therefore also saved. For a normalization method with, for example, two parameters, this would result in an additional 2*(p+q) values.

[0128] In the next step, the cycle data are transformed in such a way that the number of its dimensions is greatly reduced and its most meaningful elements in terms of target variance are emphasized.

[0129] In this example, this transformation is achieved by a well-known Partial Least Squares (PLS) regression model, the basic properties of which are briefly discussed below: PLS is, similar to Principal Component Analysis (PCA), a linear model t ( s ) = θ T< m ( s ) to describe fundamental relationships between two matrices (here cycle and target data), with t ( s )Quality characteristic (target) of the cycle θ T< Parameter vector m ( s )Cycle data of the cycle s .

[0130] An advantage over classical linear regressions is that the calculation of θ T<is also possible with linearly dependent variables (cycle data).

[0131] The parameter vector θ T< is calculated as follows: 1. Transformation into new vector space ( t , m ): In this, the first entries of the vector describe m the largest variance of the transformed target variable t 2. Selection of the most important m PLS Entries (= components) of m̃. 3. Calculating the regression vector θ̃ = ( M r T< M r ) -1< M r T< t r , M r ... data matrix ( M r = [ m r ( s_ 1), m r ( s _2), m r ( s _3) ... ]) 4. Inverse transformation of θ̃ → θ

[0132] Steps 1 and 2 are essential for the dimension reduction, whereby the reduced number of dimensions m PLS < m is a hyperparameter of the system that can be specified. The index r denotes the dimension-reduced quantities in the transformed space. The original data set M with m * n Values ​​were therefore reduced to M r with m PLS * n values.

[0133] The ( m PLS ) Principal components describe the largest part of the covariance between the two original spaces of input and target parameters. This is Fig. 2a and Fig. 2b illustrated (Image source: learnche.org). The first two components are intended to explain the points in the input space (left) and the target space (right) as well as their relationship as accurately as possible. Furthermore, they are orthogonal in the input space.

[0134] The PLS components do not correspond to real signals from the machine, but rather to a linear combination of the machine signals and form the basis of the new dimension-reduced space.

[0135] The number of components m pLS is one of the main parameters of the system, which can either be set or determined automatically.

[0136] While the number of cycle-based individual values p typically contains an order of magnitude of 50-500 cyclic parameters and the total number of cycle data m with included time series data at high sampling rates can reach several 10,000, the number of PLS ​​components m PLS typically in the range 10-20.

[0137] This step can therefore achieve a significant reduction in dimension with moderate loss of information.

[0138] The calculated PLS model is then used to project the input data onto the (dimensionally) reduced space of the PLS components. All subsequent analyses are performed in this space.

[0139] To transform the original input data into the dimension-reduced space, a projection matrix with m * m PLS Values ​​are required and therefore stored. Summary of saved parameters

[0140] For a database that includes n cycles and 1 quality characteristic (target), the following values ​​(all real decimal numbers) are required and therefore stored: 1. The transformed (projected onto PLS components) and normalized cycle data basis: A matrix with n * m Values. 2. The target values ​​that belong to the cycles of the database A vector with n values ​​(per target). 3. The untransformed database without normalization A matrix with n* m Values. 4. The normalization parameters A matrix with 2 * ( p + q ) values ​​for the common methods zScore or MinMax normalization. 5. The projection matrix of the PLS model A matrix with m * m PLS values, where m PLS is a model parameter. Outlier detection (anomaly testing)

[0141] During system operation, a new set of cycle data is acquired for each cycle from the injection molding machine and connected peripherals. This data is first normalized by the system using the stored normalization parameters. The normalized cycle data is then projected onto the space of the PLS components using the PLS model.

[0142] Now begins an analysis that essentially aims to distinguish two cases: 1) Does the point of the new cycle data in PLS space lie near already known points from the cycle database? 2) Does the point of the new cycle data in PLS space lie outside of already known regions?

[0143] In the first case, a prediction of the quality characteristic can be made based on the current cycle data of the component.

[0144] The second case corresponds to an outlier in the cycle data, and in this case, a quality control of the component in this cycle is recommended. The recorded quality value can then be added to the existing database along with the cycle data in order to perform a quality forecast in the future based on similar cycle data.

[0145] For anomaly testing and thus to distinguish between these two cases, angle-based outlier detection (ABOD) can be used.

[0146] Two straight lines are drawn from the new point to all possible pairs of elements in the database, and the angle between them is determined. If the distribution of these angles is random and uniform, the new point lies in the middle of the already known data points (case 1).

[0147] If all these angles are rather small, this means that the new point lies far outside the cluster of known points (case 2).

[0148] Angle-based outlier detection is in Fig. 3 (Image source: researchgate.net). While a point in the center of the cluster exhibits an angular distribution of up to 180° and still reaches a certain width even in the limiting case (beta), the outlier is characterized by a narrow-band angular distribution. No outlier: Checking the predictability criterion and possible quality prediction using k nearest neighbors.

[0149] If the cycle data of a new cycle lie within a range in the space of the PLS components in which there is already a sufficient amount of (historical) cycle data, the predictability criterion is checked and the quality value belonging to the current cycle (the current quality date) can be forecast.

[0150] This is done using the following method: 1. In the space of the PLS components, the distance of the current cycle data point to all points in the database is calculated. Different distance metrics, such as the Manhattan metric or the Euclidean metric, can be used for this purpose. 2. The k data points with the shortest distance to the new data point, referred to as the k nearest neighbors (kNN), are selected. 3. The predictability criterion is then checked. a. The quotient V Z = max dist Z kNN quant 95 dist Z is calculated using the cycle data, where max( dist ( Z kNN )) the maximum distance of the knearest neighbors to each other and quantum 95 ( dist ( Z kNN )) the 95% quantile value of the distances of the k nearest neighbor to all other cycle data. b. In the second step, the quotient V Q = max Q kNN − min Q kNN quan t 95 Q − quan t 5 Q calculated using the quality data. Here, min( Q kNN ) and max( Q kNN ) the minimum and maximum quality value of the k nearest neighbors and quan t 5 ( Q fcNN ) and quan t 95 ( Q kNN ) the 5% or 95% quantile of all quality values ​​in the database. If the quotient V V = V Z V Q in a defined interval around 1 , the predictability criterion is met. V v < 1 indicates a relatively large dispersion of the quality data of the nearest neighbors and thus an inadmissible predictability, while V v> 1 indicates a relatively high scatter of the cycle data of the nearest neighbors and thus an insufficient correlation between quality and cycle data in this area. The interval around 1 is determined by the parameter ε to [1 - ε ,1 + ε ] is defined. 4. If the predictability criterion is not met, a message is issued stating that the quality datum cannot be reliably predicted at the current operating point of the forming process. 5. If the predictability criterion is met, the predicted quality characteristic can be calculated, for example, as the distance-weighted mean of the quality values ​​of the k nearest neighbors: Q praed = ∑ i = 1 k Q i dist i 2 ∑ i = 1 k 1 dist i 2

[0151] The parameter k , the number of nearest neighbors, is a main parameter of the algorithm, along with m PLS , the number of PLS ​​components.

[0152] It is expected that the parameter ε can be chosen universally. In the present example, very good results were achieved with 2 = 0.1. Outliers: Adding one or more new samples to the database

[0153] If an outlier is found in the data of a new cycle, it can be added to the database as described below. This allows the algorithm to learn new operating points of the modeled injection molding process.

[0154] In the event of an outlier in the cycle data or a data point in an area that is not yet sufficiently described, the system requests a quality assessment of the component. This can be done either as a notification to the user or as a trigger for an automatic quality measurement system or a reject switch.

[0155] The algorithm thus supports the selection of suitable samples, as it recommends quality measurements especially when the cycle data is not yet known from history.

[0156] The complete data set of the cycle from m + 1 raw values ​​(m cycle values ​​and 1 quality value) are stored in the same way as the m PLS Values ​​that describe the cycle data point in the transformed space.

[0157] Already in the next cycle, the system has the expanded database available for the analysis of the cycle data and the quality forecast.

[0158] Note: Assuming that the cycle data and quality values ​​do not change abruptly in the injection molding process, some new data points can be added to the database before the normalization parameters and the PLS model need to be recalculated.

[0159] Furthermore, it is advisable to perform several quality measurements in succession after an outlier and to expand the database with the new cycle data and quality values. This serves, on the one hand, to ensure that the component quality is satisfactory (even outside the historically known operating points of the process) and, on the other hand, to quickly and comprehensively map a new operating point of the cycle data into the database. The minimum required number of new quality measurements can be, for example, the number of nearest neighbors. k be used. Removing an element of the data collection / database

[0160] To keep the system's computing power and storage space to a minimum, it's advisable to limit the number of data points in the database to a minimum. To achieve this, the current data point can replace an old one from the database, with the goal of evenly representing all the process's work areas.

[0161] This is ensured if the following criteria (1)-(3) are met, where in the first step the k nearest neighbors must be determined. (1) If the current quality measurement value lies between the minimum and maximum value of the quality measurement values ​​of the k nearest neighbor? (2) Is the ratio max distance Z i Z kNN max distance Z i Z less than a set limit G z ∈ [0,1]? Here, max( distance ( Z i , Z kNN )) the maximum distance of the current cycle data from the k nearest neighbors in the space of the PLS components and max( distance ( Z i , Z )) the maximum distance of the current cycle data from the cycle data in the space of the PLS components. (3) If the ratio max distance Q i Q kNN max distance Q i Q less than a set limit GQ ∈ [0,1]? Here, max( distance ( Q i , Q kNN)) the maximum distance of the current quality value from the quality values ​​of the k nearest neighbors and max( distance ( Q i , Q )) the maximum distance of the current quality value from the quality data.

[0162] If all these criteria are met, the nearest neighbor with the smallest distance from the current cycle values ​​in the PLS component space is replaced by the new one. Instead of using the maximum values, the 95% quantile values ​​can also be used to suppress the influence of erroneous measurements. Further samples

[0163] If, in addition to the cycles suggested by the system based on outliers, samples of the quality values ​​are taken (e.g. based on a defined test plan), these cycles can be added to the database, taking into account the described method for keeping the number of data points to a minimum.

[0164] It may be advisable to analyse the new quality value, namely whether it lies in the interval [(1 - a ) * min( Q kNN ), (1 + a ) * max( Q kNN )], where min( Q kNN ) and max( Q kNN ) describe the minimum and maximum quality value of the k nearest neighbors of the cycle values ​​of the cycle corresponding to the quality value. The parameter a is introduced to compensate for measurement inaccuracies.

[0165] If the quality value lies outside this interval, a new sample should be requested and then checked to see whether this new quality value lies within the interval, if necessary for a new set of k nearest neighbors. If this is the case, the cycle data and the quality value of the new cycle are added to the database, and those of the first sample are discarded. Otherwise, it can be assumed that the relationships between cycle data and quality values ​​have shifted during the process and the model must be adjusted. This can be done, for example, by repeatedly taking samples and overwriting the elements from the database with the new elements until the quality value of the sample is again within the interval [(1 - a ) * min( Q kNN ), (1 + a ) * max( Q kNN )] lies. Parameter determination of m PLS and k

[0166] The two main parameters of the algorithm, the number of PLS ​​components m PLS and the number of nearest neighbors k can be set by the operator or determined automatically.

[0167] However, the automatic determination requires a greater computational effort and is performed using a grid search across both parameters and cross-validation. For the grid search, minimum and maximum values ​​of the parameters ( m PLS min and m PLS max and k min and kmax ) and the number of values ​​to be examined per parameter.

[0168] In cross-validation, the existing database is n CV equally sized and randomly distributed blocks of samples. Now, in n CVRounds, each of these blocks is treated as a set of new cycle data and a quality forecast is created. Since the quality data of these samples is already known, the deviation of the forecast from the actual value can be calculated. The mean absolute error (MAE) from these n CV Runs is an indicator of the forecast quality of the current database-parameter combination.

[0169] Grid-Search now means that this calculation is repeated for each combination of k and m PLS is carried out within the specified parameter range.

[0170] That parameter set ( m PLS and k ) with the lowest MAE is finally selected.

[0171] In order to maintain a high model quality during operation even with major shifts in the process operating points, the parameters m PLS and kafter specified periods of time (e.g., every 6 hours), after a specified number of cycles, or after a specified number of added samples (quality samples) using the current database. However, this only makes sense if additional quality samples have been added to the database since the last parameter determination. Since this calculation is generally computationally intensive due to the grid search, it may be advisable to perform it on an external computing device (e.g., on a dedicated computing cluster in the cloud) and to determine the optimal parameters. m PLS and k to the quality control system.

[0172] However, the method presented here shows a very robust behavior against shifts in the working point of a forming process, which is based on Fig. 4It shows an actual quality parameter (average height of molded part 2) and a quality parameter predicted using the above system plotted against the number of cycles.

[0173] As can be seen, a relatively strong change in the operating point of the forming process was almost completely compensated after only about 100 cycles (i.e., the new operating point was essentially "learned").

[0174] Fig. 5 shows a flowchart relating to the commissioning and ongoing operation of a method according to an embodiment of the invention.

[0175] In principle, it is preferable to first define the possible quality defects ("component defects") and the relevant process parameters ("process variables"), whereby Fig. 5 examples are given in each case.

[0176] An initial training base can then be generated as a basic data set. After that, the model training can begin (see above, Parameter Determination of m PLS and k ) .

[0177] The actual process can then be carried out while the forming machine is in operation.

[0178] In parallel, the adaptation of the data collection and / or the model can be carried out.

[0179] Fig. 6 shows schematically an embodiment of a forming machine 1 according to the invention as part of a forming system.

[0180] In this embodiment, the molding machine 1 is an injection molding machine. A mold 8 is mounted on the injection molding machine, which is essentially separate from the injection molding machine and can be replaced.

[0181] In each cycle, a schematically shown molded part 2 is produced.

[0182] A reject switch 5 is provided in this embodiment.

[0183] A device 4 for collecting quality data is also provided. This could, for example, be implemented as a scale and / or an optical camera or an infrared camera.

[0184] A computing unit with a memory 6 and a processor unit 7 is also provided. The computing unit can be configured, for example, as a cloud computer. In preferred embodiments, however, the computing unit—i.e., the memory 6 and the processor unit 7—is part of the machine control system 3 of the molding machine.

[0185] During operation - i.e. while cycles are repeatedly carried out - for example, the Fig. 1 described procedures were carried out.

[0186] If the predictability criterion is not met in a particular cycle, the reject gate can be activated, for example, so that the manufactured molded part 2 reaches the device 4 for collecting quality data via path A. Specifically, the molded part 2 could be deposited in a test tray 9 of the device 4 for collecting quality data. It is then conceivable that the device 4 for collecting quality data is automatically activated as soon as a molded part 2 is deposited in the test tray 9, or that the device 4 for collecting quality data is activated via a separate activation signal.

[0187] If the predictability criterion is already met, the molded part 2 can continue to run normally along path G as an accepted good part.

[0188] Paths A and G are labeled based on the terms "reject" and "good part." However, it is of course possible that the predictability criterion is not met and molded part 2 still meets the quality requirements, i.e., is a good part. This molded part 2 can then simply be fed to path G.

[0189] In this embodiment, it is not necessary for the device 4 for collecting quality data to always be active and to check each molded part 2. This also makes it possible to use devices 4 that require more time than the duration of a cycle to collect the quality data.

[0190] Of course, other configurations are also conceivable in principle. For example, instead of a device 4 for collecting quality data, a person could also assess the quality of the molded part 2. List of reference symbols:

[0191] forming machine 1 molded part 2 Machine control 3 Device for collecting quality data 4 Reject switch 5 memory 6 Processor unit 7 mold tool 8 Test filing 9

Claims

1. Method for monitoring a molding process carried out in cycles, wherein a data collection is provided which contains at least the following data: - in each case cycle data of the molding process carried out on a molding machine (1) for past cycles and - quality data from molded parts (2) produced using the molding machine (1) for at least some of the past cycles, and wherein, with the molding machine (1), at least one further cycle of the molding process is carried out as well as current cycle data of the at least one further cycle are gathered, characterized in that the following further steps are carried out: - determining at least two nearest neighbors in the form of the cycle data from at least two of the past cycles, such that the cycle data of the at least two nearest neighbors lie nearer to the current cycle data than the cycle data which do not pertain to the at least two nearest neighbors, wherein only those past cycles for which quality data are contained in the data collection are used for the determination of the at least two nearest neighbors, - checking a predictability criterion, wherein it is checked whether a quality variation of the quality data of the cycles of the at least two nearest neighbors is smaller than a maximum variation and / or larger than a minimum variation and, - if the predictability criterion is not met, issuing a first notification that a quality and / or a quality datum of the molded part (2) produced with the at least one further cycle is not reliably predictable.

2. Method according to claim 1, characterized in that the quality variation is calculated as a - preferably normalized - difference between a maximum quality value of the at least two nearest neighbors and a minimum quality value of the at least two nearest neighbors.

3. Method according to one of the preceding claims, characterized in that the quality variation of the quality data is used in the checking of the predictability criterion in a form wherein the quality variation is correlated with a cycle data variation of the cycle data of the at least two nearest neighbors.

4. Method according to claim 3, characterized in that the cycle data variation is calculated as a - preferably normalized - maximum distance between the at least two nearest neighbors.

5. Method according to one of the preceding claims, characterized in that, if the prediction criterion is met, a quality prognosis based on the cycle data and the quality data is calculated and issued.

6. Method according to claim 5, characterized in that the quality prognosis is calculated by means of at least one of the following: - arithmetic mean of the quality data of the cycles of the at least two nearest neighbors, - median of the quality data of the cycles of the at least two nearest neighbors, - multilinear regression model.

7. Method according to claim 5 or 6, characterized in that, if the quality prognosis lies outside a predefined tolerance range, at least one of the following steps is carried out for at least one subsequent cycle - preferably several subsequent cycles: - activating a device (4) for gathering quality data and / or - prompting an operator to gather the quality data and / or - actuating a reject gate (5) and / or - depositing at least one molded part (2) produced in the at least one further cycle in an inspection tray (9) and / or marking at least one molded part (2) produced in the at least one further cycle.

8. Method according to one of the preceding claims, characterized in that, if the predictability criterion is not met, at least one of the following steps is carried out for at least one subsequent cycle - preferably several subsequent cycles: - activating a device (4) for gathering quality data and / or - prompting an operator to gather the quality data and / or - actuating a reject gate (5) and / or - depositing at least one molded part (2) produced in the at least one further cycle in an inspection tray (9) and / or marking at least one molded part (2) produced in the at least one further cycle.

9. Method according to one of the preceding claims, characterized in that - in the context of an anomaly check, it is checked whether the current cycle data constitute an outlier and, - if this is the case, a second notification is issued.

10. Method according to one of the preceding claims, characterized in that, in the context of an anomaly check, it is checked whether the current cycle data constitute an outlier and, if this is the case, at least one of the following steps is carried out for at least one subsequent cycle - preferably several subsequent cycles: - activating a device (4) for gathering quality data and / or - prompting an operator to gather the quality data and / or - actuating a reject gate (5) and / or - depositing at least one molded part (2) produced in the at least one further cycle in an inspection tray (9) and / or - marking at least one molded part (2) produced in the at least one further cycle.

11. Method according to claim 9 or 10, characterized in that the anomaly check is carried out by means of angle-based outlier detection.

12. Method according to one of the preceding claims, characterized in that the quality data of the past cycles are gathered by measurement on the molded parts (2) and / or by human assessment of the molded parts (2).

13. Method according to one of the preceding claims, characterized in that the cycle data and / or the current cycle data contain setting data for the molding machine (1), which are assigned to the past cycles or to the at least one further cycle, respectively.

14. Method according to claim 13, wherein the cycle data and the current cycle data contain setting data of the molding machine (1), characterized in that, - in the context of a setting check, it is checked whether a setting distance between the setting data assigned to the at least one further cycle and the setting data of the cycle data is smaller than a predetermined maximum setting distance and, - if this is not the case, a third notification is issued.

15. Method according to one of the preceding claims, characterized in that the cycle data and / or the quality data are transformed and dimension-reduced and in that the transformed and dimension-reduced cycle data and / or the transformed and dimension-reduced quality data are used for the determination of the at least two nearest neighbors and / or for the checking of the predictability criterion and / or the anomaly check.

16. Method according to claim 15, characterized in that the cycle data and / or the quality data are transformed and dimension-reduced by means of a multilinear regression model, in particular a partial least squares regression and / or a principal component regression.

17. Method according to one of the preceding claims, characterized in that the cycle data and / or the current cycle data contain process data, which are gathered during the past cycles and / or during the at least one further cycle by process measurement and / or reading out of a machine control system (3), wherein the process measurement and / or the reading out of the machine control system (3) is carried out on the molding machine (1) itself or a molding system which contains the molding machine (1).

18. Method according to one of the preceding claims, characterized in that - the current cycle data are added to the cycle data and / or, - if current quality data are present for the at least one further cycle, the current quality data are added to the quality data.

19. Computer program product for monitoring a molding process carried out in cycles, in particular for carrying out a method according to one of the preceding claims, wherein the computer program product contains instructions for accessing a data collection and the data collection contains at least the following data: - in each case cycle data of the molding process carried out on a molding machine (1) for past cycles, - quality data from molded parts (2) produced using the molding machine (1) for at least some of the past cycles, and - current cycle data from at least one further cycle of the molding process carried out using the molding machine (1), characterized in that the computer program product contains further instructions for carrying out the following steps: - determining at least two nearest neighbors in the form of the cycle data from at least two of the past cycles, such that the cycle data of the at least two nearest neighbors lie nearer to the current cycle data than the cycle data which do not pertain to the at least two nearest neighbors, wherein only those past cycles for which quality data are contained in the data collection are used for the determination of the at least two nearest neighbors, - checking a predictability criterion, wherein it is checked whether a quality variation of the quality data of the cycles of the at least two nearest neighbors is smaller than a maximum variation and / or larger than a minimum variation and, - if the predictability criterion is not met, issuing a first notification that a quality and / or a quality datum of the molded part (2) produced with the at least one further cycle is not reliably predictable.

20. Computer program product according to claim 19, characterized in that the computer program product contains further instructions for carrying out the following steps: - carrying out an anomaly check, wherein it is checked whether the current cycle data constitute an outlier and, - if this is the case, initiating at least one of the following steps for at least one subsequent cycle - preferably several subsequent cycles: - activating a device (4) for gathering quality data and / or - prompting an operator to gather the quality data and / or - actuating a reject gate (5) and / or - depositing at least one molded part (2) produced in the at least one further cycle in an inspection tray (9) and / or - marking at least one molded part (2) produced in the at least one further cycle.

21. Computer-readable storage medium in which a computer program product according to claim 19 and / or 20 is stored.

22. Molding machine, in particular suitable for carrying out a method according to one of claims 1 to 18, which molding machine (1) is set up for carrying out a molding process to be carried out in cycles and contains at least one memory (6) and at least one processor unit (7), wherein a computer program product according to claim 19 and / or 20 is stored in the at least one memory (6) and the at least one processor unit (7) is set up for executing the computer program product according to claim 19 and / or 20.