Tempering device and tempering method for tempering a tempering material using a tempering model, and machine learning process and computer program product for forming the tempering model
Patent Information
- Application Number
- PCT/EP2026/054246
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-02-21
- Filing Date
- 2026-02-17
- Publication Date
- 2026-08-27
Smart Images
Figure EP2026054246_27082026_PF_FP_ABST
Abstract
Description
[0001] 202501276 Foreign version
[0002] 1
[0003] Description
[0004] Tempering device and tempering method for tempering a temperable material using a tempering model, as well as machine learning methods and computer program product for creating the tempering model.
[0005] The invention relates to a tempering device and a tempering method for tempering a material to be tempered using a tempering model. In addition, a machine learning method and a computer program for creating the tempering model are described.
[0006] The material to be tempered can be any solid and / or liquid. Tempering refers to a thermal treatment (tempering) of the material. This thermal treatment adds thermal energy to the material. However, thermal treatment can also remove thermal energy from the material.
[0007] The material to be tempered is, for example, a thermoplastic granulate. During or after tempering the granulate, a plastic forming process is carried out. This process is, for example, extrusion, in which the thermoplastic is heated and continuously forced through a die under pressure.
[0008] The entire process (tempering and extrusion) is controlled and regulated to produce the most uniform extrudate (extruded product) possible. The temperature of the plastic throughout the entire process has a significant influence on this. To regulate the plastic's temperature, a heating band is arranged around a cylinder of an extruder's piston (extrusion device, such as a piston or screw extruder). The temperature of the thermoplastic is controlled via the heating band. A temperature controller is attached to the heating band to regulate the tempering process.
[0009] The majority of the heat absorbed by the plastic granules throughout the entire process does not pass through the heating belt. The majority of the heat is generated by friction and shearing of the plastic granules within the cylinder. 202501276 Foreign version
[0010] 2
[0011] The heating band is used to adjust the temperature of the plastic granules or the temperature of the plastic that has been plasticized by the input of thermal energy into the plastic granules, so that a target temperature that is as homogeneous as possible results in the plasticized plastic.
[0012] An extrusion machine is typically equipped with several heating belts. The temperatures for each heating belt must be set individually. This can be complex and time-consuming.
[0013] The object of the present invention is to provide a way in which the tempering of a material to be tempered, for example by setting the temperature of a heating band as described above, can be controlled simply and flexibly.
[0014] To solve the problem, a tempering device for tempering a temperable material is required, comprising
[0015] - at least one condition monitoring unit for recording the current state of the tempered material and
[0016] - at least one control unit with at least one control parameter for controlling the tempering of the material to be tempered, wherein
[0017] - the control unit has a proportional integral differential controller designed to create a data-driven tempering model with the control parameter based on the actual state of the tempering material using a machine learning method and
[0018] - the data-driven tempering model can be used with the control parameter to control the tempering of the tempering material.
[0019] To solve the problem, a tempering process for tempering a temperable material using the tempering device is also specified, whereby the following process steps are carried out:
[0020] a) Recording the current state of the tempered material using the state recording unit,
[0021] b) Forming the data-driven temperature model with the control parameter based on the recorded actual state variable using the
[0022] Proportional-Integral-Differential Controller and 202501276 Foreign Version
[0023] 3
[0024] c) Controlling the tempering of the material to be tempered using the tempering model with the control parameter, whereby the following further process steps are carried out to create the tempering model:
[0025] ba) Providing an initial model of the Temper model and
[0026] bb) Applying a machine learning method to the initial model, taking into account the recorded state variable.
[0027] To provide the output model of the Temper model, input data is used in particular, which is selected at least from the group of proportional-integral-differential data of the proportional-integral-differential controller, device data of the Temper device and environmental data of the Temper device.
[0028] For the machine learning process, a CGI (artificial intelligence) model is preferably used. Furthermore, a neural network is specifically employed for the CGI model. In addition to the tempering device and the tempering process, a machine learning method for generating the tempering model, which is used for the tempering device and / or for the tempering process, and a computer program for executing the machine learning process are also provided to solve the problem.
[0029] The invention employs a model-based approach. First, based on input data, including data on the detected state of the material being tempered, a tempering model (a surrogate model for the temperature treatment of the material) is created (and trained) using a machine learning method. This model is then used to control the tempering process. For this purpose, the control unit of the tempering device preferably includes at least one temperature controller for regulating the temperature of the material being tempered. Accordingly, a temperature controller is used for the control unit in the described tempering process, forming the control parameter. The temperature controller can be used to heat or cool the material being tempered. The temperature controller, or its parameters, are optimized using the machine learning method. The temperature controller is implemented as a PID controller (proportional-integral-derivative controller).It is particularly advantageous to use multiple temperature controllers. The temperature control device therefore has several temperature controllers. These can be controlled simultaneously. Mutual influences between the temperature controllers can thus be modeled. 202501276 Foreign version.
[0030] 4
[0031] In one embodiment, the current state of the tempering material can be described using at least one physical or chemical quantity, and the state monitoring device includes at least one physical or chemical sensor for detecting this quantity. In particular, the physical quantity is the temperature of the tempering material, and the physical sensor includes a temperature sensor for detecting this temperature. In another embodiment, the physical quantity is the shape of the tempering material, and the physical sensor includes a shape sensor for detecting this shape. The shape of the tempering material includes, among other things, its dimensions. For example, the tempering material is a plastic granulate. The shape sensor detects the shapes or dimensions of individual granules of the plastic granulate.
[0032] Any input data can be used for the machine learning process. Preferably, the input data used for the machine learning process is selected from the group consisting of state data of the tempering material, tempering device data of the tempering device, and environmental data of the tempering device's environment. Tempering device data includes, for example, parameters of the temperature controllers used. Environmental data of the tempering device includes, for example, the ambient temperature or humidity. It is also conceivable to use comparison state data of a comparison tempering material as input data. The comparison tempering material is, for example, a tempered product from a previous tempering process.
[0033] From the multitude of possible tempering device data, these are preferably selected from the group of real data from an existing tempering device and simulation data from a virtual tempering device. The real data incorporates empirical values from an existing, comparable tempering device. The virtual data from a virtual tempering device uses non-real data from a non-real tempering device. The virtual tempering device is, for example, a simulated tempering device. The non-real data consists of simulation data derived from a simulation of the tempering device.
[0034] The tempering process can be carried out over any desired period of time. Preferably, the tempering process, the process steps of the tempering process, and / or the further process steps of the tempering process are carried out in a time of less than 120 minutes, and particularly in a time of less than 60 minutes. Short time intervals are also possible.
[0035] 5
[0036] possible. Preferably, the time is less than 1 s and particularly less than 100 ms. This applies especially to individual process steps, for example, the formation of the tempering model or controlling the tempering of the material to be tempered.
[0037] The tempering process can be used for tempering a wide variety of materials. Such materials include inorganic materials like glass, ceramics, or metals. In these cases, tempering would be referred to as sintering. It is particularly advantageous to use at least one thermoplastic feedstock as the material to be tempered, whereby the tempering process at least supports a physical and / or chemical transformation of the feedstock into a plastic product. The plastic product can be an intermediate product in a multi-stage production chain for manufacturing a final plastic product. When using a plastic feedstock, input data selected from the groups of plastic feedstock data and / or plastic product data is used for the machine learning process. The input data relates to corresponding input variables.
[0038] With thermoplastic materials, the malleability of the material can be influenced by its temperature. It is advantageous to combine the tempering process with a shaping process. In a specific embodiment, at least one plastic shaping process selected from the group of extrusion, injection molding, and blow molding is applied during and / or after the chemical and / or physical transformation of the plastic feedstock into the plastic product.
[0039] The machine learning method used can be extended, in particular for sensitivity analysis. This involves determining the influence of the input data of the machine learning method on the output data of the machine learning model.
[0040] In summary, the following advantages of the invention with the model-based approach should be highlighted:
[0041] - The sintering process can be quickly and flexibly adapted to changes in the material being tempered.
[0042] - Materials can be saved. For example, no material (e.g.,
[0043] (plastic granules) are consumed. This enables many optimization steps that would be costly without the invention due to material costs. 202501276 Foreign version
[0044] Time is saved: The Temper model can be evaluated within milliseconds, significantly accelerating the optimization of the control device (e.g., the Temper controller). Of course, the Temper model also requires a training period, which can easily last several hours. However, this training period is a one-time event. Subsequent optimization of the Temper controller can then be performed very quickly.
[0045] - A sensitivity analysis can be performed using the tempering model. This provides an indication of, for example, the extent to which the properties of a starting material of the tempering material (e.g., plastic granules) influence the behavior of the tempering device.
[0046] - When using multiple temperature controllers, they can be controlled simultaneously. The Temper model considers all temperature controllers at once, so that mutual influences are already modeled.
[0047] - During the training of the Temper model as part of the Temper model formation process, unknown process fluctuations can be modeled through cost-effective retraining.
[0048] The invention is described in more detail below with reference to an exemplary embodiment and the accompanying figures. The figures are schematic and not to scale.
[0049] Figure 1 shows a tempering device.
[0050] Figure 2 shows a tempering process.
[0051] Figure 3 shows the construction of a temper model.
[0052] Figure 4 shows the output of the temper model.
[0053] Given is a tempering device 1 for tempering a temperable good 11, comprising - at least one state recording unit 12 for recording an actual state of the temperable good and 202501276 foreign version
[0054] 7
[0055] - at least one control unit 13 with at least one control parameter for controlling the tempering of the material to be tempered, wherein
[0056] - the control unit has a proportional-integral differential controller designed to create a data-driven tempering model with the control parameter based on the actual state of the tempering material using a machine learning method and
[0057] - the data-driven tempering model can be used with the control parameter to control the tempering of the tempering material.
[0058] Using the tempering device, a tempering process 100 is carried out with the following process steps:
[0059] a) Recording 101 the actual state parameter of the tempered good using the state recording unit,
[0060] b) Form 102 a temper model with the control parameter based on the recorded actual state variable using the proportional-integral-differential controller and
[0061] c) Controlling the tempering of the material to be tempered using the tempering model, wherein the following further process steps are carried out to create the tempering model: ba) Providing an initial model of the tempering model and
[0062] bb) Apply 1022 a machine learning method to the initial model, taking into account the recorded state variable.
[0063] According to the exemplary embodiment, the material to be tempered is a thermoplastic granulate. The thermoplastic is used as the starting material. During tempering, thermal energy is introduced into the granulate, causing it to heat up. This results in a plastically deformable product.
[0064] The tempering device is an integral part of an extrusion device in the form of a screw extruder with an extruder cylinder 14 and an extruder screw 15 arranged in the extruder cylinder 14. A plastic forming process is carried out during or after the tempering of the plastic.
[0065] The plastic granules 11 are fed into the extruder cylinder via an extruder filling hopper 16. Inside the extruder cylinder, the plastic is plasticized by increasing the temperature. The plasticized plastic is then conveyed along the extruder screw by the extruder screw.
[0066] 8
[0067] The extruder cylinder is advanced. Through an opening 18 at the end of the extruder cylinder, the plasticized plastic exits the extruder cylinder as extrudate 19.
[0068] The control unit 13 of the tempering device includes a PID controller. The P, l, and D values of the PID controller are used as control parameters. A multitude of state detection units 12 in the form of temperature sensors are arranged along the extrusion cylinder 1.
[0069] The tempering device consists of several partial tempering devices (sections) 10 arranged along an extrusion cylinder 14. A plurality of temperature sensors 17 are arranged along the extrusion cylinder to detect the temperature of the material being tempered in the respective section of the extrusion cylinder. Each section is associated with a heating belt 141.
[0070] To control the tempering process, a data-driven tempering model is created and trained (Figure 3). Input data 104 relating to the following quantities are used to create the tempering model:
[0071] - Controller parameters of the temperature controller
[0072] - Machine data of the screw extruder
[0073] - Properties of the plastic granules
[0074] - Environmental data of the screw extruder
[0075] The machine data includes, for example, actual and / or target temperatures of the heating strips. A tempering model with a neural network is used to train the tempering model. A neural network is applied to the input data (reference 105). This generates output data 106 of the tempering model. The output data includes, among other things, quality data for tempering control. This is illustrated in Figure 4, where time 109 in seconds is plotted against a specific control variable 110. For example, the time 107 at which a stable tempering state is reached can be used as a quality criterion (see Figure 4). A stable state is reached, for example, when the control variable, such as the temperature of a heating strip or the temperature of the material being tempered, lies within a specific tolerance range.Overshoot magnitude (the extent of a deviation between the actual temperature of the item being tempered and the target temperature of the item being tempered) can also be used as a quality criterion. 202501276 Foreign version.
[0076] 9
[0077] To train the tempering model, a correspondingly large amount of data (or simulation steps) is used, which includes a high degree of variability. This data can originate from the actual operation of an existing tempering device (real data) or from a simulation of the device (simulation data). It is also conceivable to use data from the operation of an identical tempering device.
[0078] The (PID) controller to be optimized is subsequently not optimized against real operating data or (expensive) simulation data, but against the temperature model. Controller parameters are determined that allow the target temperatures of the heating elements to be reached as quickly as possible without significant overshoot. This can be achieved, for example, with the objective function: costFunction = a*(t_stabil) + b*(Überschwingen_max) -> min
[0079] reach, where a and b are weights with 0<=a,b<=1 and a+b=1, t_stabil is the time until a stable state is reached, and Überschwingen_max is the overshoot magnitude.
[0080] The optimization can be carried out using conventional methods for temperature controller optimization, such as the method according to Ziegler and Nichols, the method according to Chien Hrones Reswick, or empirical optimization.
[0081] To optimize the controller parameters, statistical methods such as reinforcement learning and Bayesian optimization or methods from classical mathematical optimization such as the Nelder-Mead method can be used instead of traditional optimization methods.
[0082] The runtime for determining stable temperatures is thus reduced from several minutes to just a few milliseconds. Optimization is performed using the same methods as traditional controller optimization. However, the temperature model is now evaluated in each iteration step. This eliminates additional costs that would arise during conventional extrusion process optimization, such as costs associated with running the extrusion process, costs for measurements during extrusion, or costs for necessary optimizations. This saves not only time but also reduces scrap costs.
[0083] In one variation of the implementation example, the tempering model is generalized through retraining. For this purpose, parameters of the plastic granules are provided to the tempering model. If the tempering model is then trained with different batches of granules, a generally valid tempering model for different batches of granules can be developed.
[0084] 10
[0085] This eliminates the need to re-optimize the controllers when using another batch of granules. Furthermore, the model becomes more robust against process fluctuations.
[0086] In reality, however, there will always be process fluctuations, such as changes to plastic granules with new properties, for which no data is yet available during the training phase. In such a case, according to another variation of the implementation example, the Temper model is adapted without retraining. For this purpose, the entire Temper model is trained with the available training data during the training phase. If input parameters change compared to the training data, pre-trained weights of the Temper model are retained, and additional weights are introduced. Only the additional weights are then trained.
[0087] A sensitivity analysis is then performed on the Temper model used. This determines which input data of the Temper model have a particular influence on the output data of the Temper model.
[0088] In sensitivity analysis, a small perturbation term is added to the output data of the Temper model based on a previously trained neural network. The resulting error is then propagated back to the input data of the Temper model. Input data with large changes are highly sensitive, meaning they play a significant role in the output data of the Temper model.
[0089] This makes it possible to identify the input variables that are relevant quality drivers. For example, the results might show that a specific mixing ratio of the plastic is the most sensitive input variable. This provides more detailed information about the underlying process.
[0090] Using the tempering model, it's possible, for example, to determine the contribution of the plastic granule's properties to the temperature and the contribution of the extruder itself. This information can then be used to better guide and thus accelerate the optimization process, for example, by eliminating irrelevant or less relevant input features.
Claims
202501276 Foreign version 11 Patent claims 1. Tempering device (1) for tempering a temperable material (11), comprising - at least one condition monitoring unit (12) for recording the actual state of the tempered material and - at least one control unit (13) with at least one control parameter for controlling the tempering of the material to be tempered, wherein - the control unit has a proportional-integral differential controller designed to create a data-driven tempering model with the control parameter based on the actual state of the tempering material using a machine learning method and - the data-driven tempering model can be used with the control parameter to control the tempering of the tempering material.
2. Tempering device according to claim 1, wherein - the control unit has at least one temperature controller for regulating the temperature of the material being tempered and - the control parameter is formed by the temperature controller.
3. Tempering device according to claim 1 or 2, wherein - the current state of the tempered material can be described using at least one physical or chemical quantity and - the condition monitoring device has at least one physical or chemical sensor for detecting the physical or chemical quantity.
4. Tempering device according to claim 3, wherein - the physical quantity is the temperature of the tempered material and - the physical sensor has a temperature sensor for detecting the temperature of the material being tempered.
5. Tempering device according to claim 3 or 4, wherein - the physical quantity is a form of the tempered material and - the physical sensor includes a shape sensor for detecting the shape of the tempered material. 202501276 Foreign version 12 6. Tempering method for tempering a temperable material using a tempering device according to one of claims 1 to 5, wherein the following process steps are carried out: a) Detecting (101) the actual state parameter of the temperable material using the state detection unit, b) Forming (102) the data-driven temper model with the control parameter based on the recorded actual state variable using the Proportional-integral differential controller and c) Controlling (103) the tempering of the material to be tempered using the tempering model with the control parameter, wherein the following further process steps are carried out to form the tempering model: ba) Providing (1021) an initial model of the Temper model and bb) Applying (1022) a machine learning procedure to the initial model, taking into account the acquired state variable.
7. Temper method according to claim 6, wherein input data are used to provide the output model of the temper model, which are selected at least from the group consisting of proportional-integral-differential data of the proportional-integral-differential controller, device data of the temper device and environmental data of the temper device.
8. Tempering method according to claim 6 or 7, wherein a temperature controller is used for the control unit, which forms the control parameter.
9. Temper method according to one of claims 6 to 8, wherein a computer model is used for the machine learning method.
10. Temper method according to claim 9, wherein a neural network is used for the Kl model.
11. Tempering method according to any one of claims 6 to 10, wherein input data for the machine learning method are used which are selected from the group consisting of state data of the material to be tempered, tempering device data of the tempering device and environmental data of an environment of the tempering device. 202501276 Foreign version 13 12. Tempering method according to claim 11, wherein the tempering device data of the tempering device are selected from the group consisting of real data of an existing tempering device and simulation data of a virtual tempering device.
13. Tempering process according to any one of claims 6 to 12, wherein the tempering process, the process steps of the tempering process and / or the further process steps of the tempering process are carried out in a time of less than 120 min and in particular in a time of less than 60 min.
14. Tempering method according to claim 13, wherein the time is less than 1 s and in particular less than 100 ms.
15. Tempering process according to any one of claims 6 to 14, wherein - at least one thermoplastic polymer feedstock is used as the material to be tempered, and - a physical and / or chemical conversion of the polymer feedstock into a polymer product is at least supported by the tempering process.
16. Tempering method according to one of claims 15, wherein input data selected from the group plastic feedstock data and / or plastic product data are used for the machine learning method.
17. Tempering process according to claim 14 or 15, wherein during and / or after the chemical and / or physical conversion of the plastic feedstock into the plastic product, at least one plastic forming process selected from the group consisting of extrusion, injection molding and blow molding is applied.
18. Machine learning method for forming a tempering model that can be used for a tempering device according to any one of claims 1 to 5 and / or for a tempering method according to any one of claims 6 to 17.
19. Machine learning method, wherein an influence of the input data of the machine learning method on output data of the machine learning model is determined.
20. Computer program product for performing the machine learning method according to claim 18 or 19.