Dual-Model Machine Learning for Process Control of Manufacturing Equipment and Rule Controllers

Abstract

Using a learning algorithm, train a machine learning model on a training dataset that describes input parameters to a manufacturing apparatus and corresponding output parameters from the manufacturing apparatus, obtain a physical model that describes the evolution of the state space of the manufacturing apparatus, generate a command for the physical model to change the setting of the simulation of the manufacturing apparatus by the physical model such that the physical model generates corresponding predicted output parameters in response to input data, and train a machine learning-based controller agent based on the setting and the corresponding predicted output parameters using another learning algorithm.

Description

Technical Field

[0001] (Cross - Reference to Related Applications) This application claims the benefit of U.S. Provisional Application No. 18 / 300,632, filed on April 14, 2023; U.S. Provisional Application No. 63 / 359,526, filed on July 8, 2022; and U.S. Provisional Application No. 63 / 391,065, filed on July 21, 2022. All of these are hereby incorporated by reference in their entirety into this specification.

[0002] The present disclosure relates to the control of manufacturing equipment.

Background Art

[0003] A manufacturing control system may generate an output signal in response to an input signal to operate a device under control in a specific manner.

Summary of the Invention

[0004] The method includes receiving, by a machine - learning model, a training data set that describes input parameters and corresponding output parameters of a manufacturing device; training the machine - learning model on the training data set using at least one learning algorithm to obtain a physical model that describes the evolution of the state space of the manufacturing device; changing the settings of the simulation of the manufacturing device by the physical model and configuring a machine - learning - based controller agent to generate a command to change the settings of the simulation of the manufacturing device by the physical model such that, in response to input data, the physical model generates corresponding predicted output parameters; and training the machine - learning - based controller agent with the settings and corresponding predicted output parameters using at least one other learning algorithm such that, in response to the input data, the machine - learning - based controller agent maintains the values of the predicted output parameters within their respective predefined ranges.

[0005] The method further includes configuring a machine learning-based controller agent to generate commands for the manufacturing apparatus in response to values of predicted output parameters from a physical model, and the manufacturing apparatus may execute those commands. The method may further include configuring a machine learning-based controller agent to generate commands for the manufacturing apparatus in response to input parameters to the manufacturing apparatus and corresponding output parameters from the manufacturing apparatus. Configuring a machine learning-based controller agent to generate commands for the physical model may include receiving, by the machine learning-based controller agent, one or more rules that define control actions for the manufacturing apparatus that are executed in response to a value of at least one output parameter from the manufacturing apparatus being outside a predefined range. The method further includes receiving time series data of the machine learning-based controller agent that describes the operating state of the manufacturing apparatus, and in response to an operating state indicating that a value of at least one operating parameter is outside a predefined range, the machine learning-based controller agent generates a command for the manufacturing apparatus to execute at least one control action. The one or more rules may be obtained from a rule controller. The method may further include receiving an input to the machine learning-based controller agent and modifying one or more rules in real time. The configuration may incorporate the rules. The method may further include operating the manufacturing apparatus using a rule controller and generating a training data set. The input parameters may include active control parameters, endogenous parameters, and exogenous parameters of the manufacturing apparatus. The output parameters may include characteristic parameters of components manufactured by the manufacturing apparatus. The physical model may be a sequence to sequence machine learning model. The sequence to sequence machine learning model may be an Encoder-Decoder model.The encoder-decoder model may include a long short-term memory model. At least one learning algorithm may be a supervised learning algorithm.

[0006] The method trains a machine learning model to obtain a physical model that describes the evolution of the state space of a manufacturing apparatus using at least one learning algorithm on a training data set that describes input parameters of the manufacturing apparatus and corresponding output parameters from the manufacturing apparatus, generates commands for the physical model to change the settings of a simulation of the manufacturing apparatus simulated by the physical model such that the physical model generates corresponding predicted output parameters in response to input data, configures a machine learning-based controller agent to simulate a simulation of the manufacturing apparatus by the physical model such that the physical model generates corresponding predicted output parameters in response to input data, trains the machine learning-based controller agent using at least one other learning algorithm with the settings and corresponding predicted output parameters, configures the machine learning-based controller agent such that the machine learning-based controller agent maintains the values of the predicted output parameters within their respective predefined ranges in response to input data, and configures the machine learning-based controller agent to generate commands for the manufacturing apparatus in response to the values of the predicted output parameters from the physical model.

[0007] The configuration may include receiving, by a machine learning-based controller agent, one or more rules that define control actions for a manufacturing apparatus to be taken in response to a value of at least one output parameter from the manufacturing apparatus being outside a predefined range. The method may further include receiving time series data of a machine learning-based controller agent that describes an operating state of the manufacturing apparatus, and generating, by the machine learning-based controller agent, a command for the manufacturing apparatus to perform at least one control action in response to an operating state indicating that a value of at least one operating parameter is outside a predefined range. One or more rules may be obtained from a rule controller. The method may further include operating the manufacturing apparatus using the rule controller to generate a training data set.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3

Figure 4

Embodiments for Carrying Out the Invention

[0009] Embodiments are described herein. However, it should be understood that the disclosed embodiments are merely examples, and other embodiments may take various alternative forms. The figures are not necessarily to scale. Some functions may be exaggerated or minimized to show details of specific components. Therefore, the specific structures and function details disclosed herein should not be construed in a limiting sense, but rather should be construed as a representative basis for teaching those skilled in the art.

[0010] The various features illustrated or described with reference to any example can be combined with features illustrated or described in one or more other examples to produce embodiments that are not explicitly illustrated or described. The combinations of features shown provide representative embodiments for typical applications. However, various combinations and modifications of features consistent with the teachings of this disclosure may be desirable for a particular application or implementation.

[0011] Machines used in mass production often have control parameters that affect the measurable characteristics of the manufactured parts. To give a simple example, a stamping machine may apply a certain pressure for a certain period of time to form metal into a desired shape. The ability of the stamping machine to repeatedly produce the same desired shape depends on this pressure and time. From experience or otherwise, the operator of the stamping machine may determine that certain ranges of values for pressure and time are acceptable when the stamping machine manufactures parts of specified dimensions. If the values of these control parameters change over time, the parts manufactured one hour ago may have a slightly different shape from the parts manufactured one hour later, resulting in a decrease in consistency between parts.

[0012] In this example, the actual pressurization may be a function of the power supplied to the stamping machine with respect to a given pressure setting. Therefore, fluctuations in the supplied power can cause fluctuations in pressurization even when the pressure setting does not change. The fluctuations in the supplied power may be related to fluctuations in the part shape, although there is a time lag. That is, considering the processing time associated with the stamping machine, a change in the supplied power at time zero may appear as a deviation from the desired shape 42 seconds later. If it is possible to predict the impact of a sudden change in the supplied power on the part shape at a later time, the pressure setting can be strategically changed to offset such a change. Specifically, when a decrease in power occurs, the pressure setting can be increased accordingly. When an increase in power is expected, the pressure setting can be decreased accordingly, and so on.

[0013] Statistical techniques such as statistical process control are commonly used to monitor and control manufacturing processes in order to produce fewer waste and more specification-compliant products. In the context of complex manufacturing processes, these techniques may have limitations in effectiveness. Machines used in mass production may have hundreds (or even thousands in some cases) of active control parameters (as well as internal and external parameters) that affect the measurable characteristics of the manufactured parts. In this context, an active control parameter is a parameter that is actively controlled and / or managed (e.g., speed setting, pressure setting, etc.), an internal parameter is a parameter that appears in the machine but is not necessarily controlled (e.g., machine vibration, machine temperature, etc.), and an external parameter is a parameter related to ambient conditions (e.g., humidity, ambient temperature, etc.). The ability to predict the impact of changes in active control parameters, internal parameters, and / or external parameters, and external parameters are parameters related to ambient conditions (e.g., humidity, ambient temperature, etc.). Therefore, the ability to predict the impact of changes in active control parameters, internal parameters, and / or external parameters on the measurable characteristics of parts is a complex endeavor.

[0014] As will be described in more detail below, in some embodiments, supervised machine learning techniques can be used to establish a physical model that describes the input-output behavior of a machine. This physical model can ultimately be used to train an intelligent (machine learning-based) controller that monitors the operation of the machine in real time.

[0015] Put simply, in supervised learning, a function that maps inputs to outputs is discovered based on examples of pairs of inputs and outputs. This infers a function from labeled training data that includes a series of learning examples. Each example can be a pair defined by an input object (e.g., a vector) and a desired output value (a supervision signal). The algorithm analyzes the training data and generates the inferred function. This function can be used to map new examples. Under optimal circumstances, the algorithm accurately predicts the output values for unseen inputs. For this, the algorithm may need to generalize from the training data.

[0016] To establish the above physical model, a number of supervised learning algorithms can be used, such as support vector machines, linear regression, logistic regression, naive Bayes, linear discriminant analysis, decision trees, k-nearest neighbors, similarity learning, neural networks, etc. In this specification, for the purpose of discussion, neural networks, particularly recurrent neural networks, are taken as an example. However, this discussion is generally applicable to other learning techniques including supervised learning techniques.

[0017] In certain agreements, the control device of a machine can be programmed to automatically maintain control parameters within a specified range. Continuing with the above example, the control device associated with a stamping machine can be programmed to monitor the sensed values of pressure, temperature, and time, and if any of these values go outside the corresponding specified range, the control device can take automatic measures to return the value to the specified range. For example, if the specified range of the temperature value is from 80°C to 85°C and the sensed value of the temperature exceeds that range and reaches 86°C, the controller may instruct the heating element of the stamping machine to lower the temperature so that the sensed value returns to the range of 80°C to 85°C. Similarly, if the specified pressure range is from 300 Pa to 325 Pa and the sensed value of the pressure falls below that range and becomes 294 Pa, the controller may issue an instruction to increase the pressure on the pressure element of the stamping machine so that the sensed value returns to the range of 300 Pa to 325 Pa. For example, due to natural variations in raw materials and the environment, the sensed values and the output of the manufacturing process may vary. There are also other factors that can cause variations, such as wear of the device.

[0018] There may be a time lag, for example related to the sensing device, between the point when the sensing value and the output of the manufacturing process actually go outside the desired range and the point when the control device senses such a situation. There may also be a further time lag between the control device sensing such a situation and the control parameter being brought back within the specified range by the action commanded by the control device. During such time, the parts manufactured by the machine may not have their specified dimensions or other characteristics.

[0019] Therefore, consider a controller and machine learning techniques that enable the manufacturing device to predict that the sensed values and the output of the manufacturing process will go outside the specified range and take corrective measures before that happens to maintain the control parameters within the specified range. This avoids the time during which the control parameter is outside the specified range, and as a result, avoids the time during which the factory system may produce undesirable results (e.g., parts that do not have the desired dimensions or other characteristics).

[0020] In one example, a process expert uses an interface to describe to a rule controller control inputs that change to correct a situation where a desired output or outputs deviate from a target in a manufacturing process. Each combination of a desired output and a control input is called a "rule". Each output can have multiple rules that are used to maintain it within specifications and / or targets, and each may have a different mathematical weighting for execution order or importance (for example, if the dimensions of a manufactured component are too large, open the upstream cooling water valve and the temperature of the cooling water flowing from the valve must be within a predefined range before the downstream cooling water valve opens. The value of parameter A can range from 0 to 2, the target value is 1, the value of B can range from 0 to 10, and the target value can be 5. If the value of A is set higher than the value of B and a situation occurs where A and B do not reach their target values, the control operation is executed so that A approaches its target value and B remains within the range but moves away from the target value. Within a rule, the process expert has several options. That is, a target value can be defined for each desired output, and each desired output can have a different weighting (for example, ranking by importance). Also, a proportionality constant can be defined and combined with an error function to modify the control input. This expert advice can take the form of rules stored in a database. When the manufacturing process is operating and generating outputs, software running on a processor (for example, a rule controller or engine) accesses the corresponding rules from the database, extracts real-time data from the process (that is, the current value of the desired output), and executes the rules. This automates the human decision-making process and provides several advantages. That is, human bias is eliminated from the decision-making process, immediate action can be taken if a deviation in the desired output is detected, and / or the need for human correction is eliminated.

[0021] In addition to the above, rules created to automate the process can also accelerate the construction of machine learning control models. This can be achieved by extracting rule parameters (e.g., control inputs and desired outputs) from the rule controller database and using them in the training algorithm of machine learning. This reduces the amount of data required for accurate prediction and correction. For example, rule parameters can be used to configure a machine learning-based controller agent and generate commands for a physical model that describes the evolution of the state space of a manufacturing device. The machine learning-based controller agent can change the settings of the manufacturing device simulated by the physical model that may incorporate the rule parameters so that the physical model generates the corresponding predicted output parameters. Further, by operating the manufacturing device under the control of the rule controller for the purpose of creating a dataset with which the machine learning model can be trained, the corresponding rule parameters implicitly represented by the dataset can be introduced into the machine learning model.

[0022] Briefly reviewing, an artificial neural network that can be used to create a machine learning control model may include four parts: nodes, activations, connections, and connection weights. An artificial neural network usually consists of a large number of nodes. There are generally two types of network connections: input connections and output connections. An input connection is a path through which a node receives information, and an output connection is a path through which a node sends information. A connection can be both an input connection and an output connection. For example, when a connection is used to move information from a first node to a second node, that connection is an output connection for the first node and an input connection for the second node.

[0023] Generally, a recurrent neural network can store inputs via internal memory and process sequential data such as time series data indicating ambient conditions, control inputs to a manufacturing apparatus, and measurable characteristics of parts manufactured by the manufacturing apparatus. With this internal memory, the recurrent neural network can track information regarding the received inputs and then predict what will happen next. Therefore, the recurrent neural network has two inputs: the current and the most recent past. Weights are applied to the current input and the past input. These weights may be adjusted for the purposes of gradient descent and backpropagation of error. Furthermore, the mapping from input to output does not have to be one-to-one.

[0024] The long short-term memory network is an extension of the recurrent neural network. With long short-term memory, the recurrent neural network can store inputs in a so-called memory in a form that can be read, written, and deleted over a longer period of time. This memory can determine whether to save or delete information based on the importance assigned to the information. The importance of specific information may be learned over time by long short-term memory. Typical long short-term memory has sigmoid-type input, forget, and output gates. These gates determine whether to accept a new input, delete it, or allow the new input to affect the output at the current time step.

[0025] A sequence-to-sequence model can be constructed using a recurrent neural network. A general sequence-to-sequence architecture is an encoder-decoder architecture with two main components: an encoder and a decoder. The encoder and decoder can be, for example, LSTM (long short-term memory) models, although other models are also envisioned. The encoder reads the input sequence and summarizes its information into an internal state or context vector. The output of the encoder is discarded, but the internal state is saved and helps the decoder make accurate predictions.

[0026] The initial state of the decoder is initialized to the final state of the encoder. That is, the internal state vector of the final cell of the encoder is input to the first cell of the decoder. With this initial state, the decoder can start generating the output sequence.

[0027] The above and other concepts can be adapted for use within the context of the manufacturing environment described above. As an example, long-term short-term encoder-decoder models, transformers (e.g., bidirectional encoder representations from transformers, generative pre-trained transformers 3s, etc.) can form the basis of a sequence-to-sequence that is trained to interpret time series data describing ambient conditions and manufacturing operations and predict corresponding component characteristics. Subsequently, an appropriate controller can take corrective actions to maintain parameters within a specified range before such a situation occurs.

[0028] The time series data can include actual control parameter values (e.g., current, machine revolutions per minute, machine pressure, machine temperature, etc.), endogenous control parameter values (e.g., machine vibration, etc.), exogenous parameter values (e.g., ambient temperature, humidity, etc.), changes in these values over a predefined duration, and other related data, and can be preprocessed using various digital signal processing techniques (e.g., Fourier signal processing techniques). Environmental temperature, humidity, etc., changes in these values over a predefined duration, and other related data are collected, preprocessed using various digital signal processing techniques (Fourier analysis, wavelet analysis, etc.), and a feature set can be generated to describe the evolution of the state space (the set of all possible configurations) of the manufacturing apparatus in the frequency domain and / or time domain. For a given application, a specific set of digital signal processing techniques can be determined using standard methodologies including simulation, trial and error, etc.

[0029] Referring to FIG. 1, as an example, a manufacturing system 10 may include a manufacturing apparatus 12 (e.g., an extruder, a press, a stamper, etc.) that physically or virtually manufactures (e.g., assembles, creates, etc.) a manufactured component 14 (e.g., a tube, a panel, etc.). The manufacturing system 10 may also include one or more environmental condition (exogenous) sensors 16, a current sensor 18, a voltage sensor 20, one or more additional sensors 22, one or more characteristic sensors 24, and a database 26. The environmental condition sensors 16 measure one or more environmental conditions (e.g., humidity, temperature, etc.) in the vicinity of the manufacturing apparatus 12. The current and voltage sensors 18, 20 measure the current and voltage supplied to the manufacturing apparatus 12. The additional sensors 22 measure other active control and endogenous parameters of the manufacturing apparatus 12. The characteristic sensors 24 measure various characteristic parameters (e.g., length, rigidity, thickness, etc.) of the manufactured component 14.

[0030] These sensed values may be sequentially reported to the database 26. That is, at time t0, each of the sensors 16, 18, 20, 22, 24 detects their values and reports them to the database 26, and at time t1, each of the sensors 16, 18, 20, 22, 24 detects their values and reports them to the database 26, and so on. Thereby, the database 26 receives time-series data that describes ambient conditions, endogenous properties, control parameter values related to the operation of the manufacturing apparatus 12, and characteristic parameter values related to the manufactured component 14 manufactured by the manufacturing apparatus 12. Such a configuration can be used to collect a huge amount of data for training purposes.

[0031] By performing various transformations (e.g., data cleansing, band-pass filtering, convolution operation, principal component analysis, wavelet transform, etc.) on the time-series data held in the database 26, a streaming feature set can be generated that covers the associated state space that describes the evolution of the manufacturing process related to the manufacturing apparatus 12. The associated state space can be iteratively identified during model training and evaluation.

[0032] Referring to FIG. 2, one or more processors 28 may implement a long short-term memory Encoder-Decoder model 30 (or another suitable model) trained with at least a portion of the streaming feature set from the database 26. For example, 60 minutes, 600 minutes, or 6000 minutes, etc. of the streaming feature set can be used to train the long short-term encoder-decoder model 30 to recognize the relationship between the sensed environmental conditions and endogenous and control parameter values of the sensors 16, 18, 20, 22, and the sensed characteristic parameter values of the characteristic sensor 24. The model 30 (physical model), if properly trained, can predict future characteristic parameter values of the component 14 manufactured from the streaming feature set.

[0033] A pattern of specific ambient conditions and / or control parameter values may precede other control parameter values that are outside the specified range. As a simple example, the model 30 can recognize that when the humidity is less than 30% and the supply voltage to the manufacturing apparatus 12 (for example, the aforementioned stamping machine) exceeds 225 V, the pressure of the manufacturing apparatus 12 may start to increase at a rate of 1 Pa per second. If the current pressure is 323 Pascals and the specified range of pressure is from 300 Pascals to 325 Pascals, the pressure is considered to be outside the specified range within 3 seconds. As another example, when the measured dimension is increasing at a certain rate and is likely to exceed its limit, countermeasures can be taken before the limit is reached. Of course, in the actual scenario, there can be hundreds or thousands of patterns of ambient conditions and / or control parameter conditions before a certain control parameter value falls outside its specified range, which can be more complex.

[0034] Referring to FIGS. 1 and 3, one or more processors 28 can further implement a controller agent 32 trained based on the streaming feature sets from the model 30 and the database 26. Before the training of the controller agent 32, the model 30 (or other source) can notify the controller agent 32 of the control limits of the manufacturing apparatus 12 that can be simulated by the model 30. The control limits can include, for example, the operating pressure range of the press (300 psi to 500 psi), the operating temperature range of the drying oven (50°C to 80°C), etc. Further, the controller agent 32 can receive the target feature parameter values of the manufactured part 14 (e.g., target length = 3 cm, target stiffness = 5 N / m, etc.). During the training of the controller agent 32, the model 30 and the controller agent 32 can each synchronously receive the same portion of the streaming feature set from the database 26 to simulate the feedback from the sensors 16, 18, 20, 22, 24 during manufacturing execution. Thereby, the model 30 can generate the predicted feature parameter values of the simulated manufactured part and report them to the controller agent 32. Then, the controller agent 32 can instruct the model 30 to perform control actions to change the control settings within the control limits. In the first iteration, assuming that the operating pressure of the press simulated by the model 30 is 310 psi and the operating temperature of the drying oven simulated by the model 30 is 62°C, the controller agent 32 can increase one by a certain amount and decrease the other by a certain amount, and learn how such changes affect the predicted feature parameter values from the model 30 with respect to the target feature parameter values. The amount of change can be arbitrary or governed by a predetermined rule.The controller agent 32 can self-learn using one or more learning algorithms regarding how to change the control settings of the manufacturing apparatus 12 in order to maintain the predicted characteristic parameter values, and thus the actual characteristic parameter values, at or near the target characteristic parameter values, even if the values from the sensors 16, 18, 20, 22, 24 change, by executing such iterations not millions of times but at least thousands of times in a relatively short period of time.

[0035] Referring to FIG. 4, when the controller agent 32 is properly trained (for example, the error between the predicted characteristic parameter value and the target characteristic parameter value falls within a predetermined range), one or more processors 28 may be arranged within the manufacturing system 10 to receive the live data output by the sensors 16, 18, 20, 22, 24, preprocess the data using the various transformations described above, and generate live streaming functional coordinates spanning the associated state space that describes the evolution of the manufacturing process associated with the manufacturing apparatus 12. Similarly to the above, the trained controller agent 32 can then instruct the manufacturing apparatus 12 to change the control settings within the control limits to maintain the predicted functional parameter values, and thus the actual functional parameter values, at or near the target functional parameter values based on the live streaming functional set and the corresponding predicted functional parameter values. For example, if the current pressure is 323 Pa, the specified range is 300 Pa to 325 Pa as described above, and the model 30 recognizes the state of the surroundings and / or control parameters and indicates that the pressure is rising at a rate of 1 Pa per second, the model 30 can provide the controller agent 32 with a predicted control parameter value indicating that the pressure will exceed the upper limit within 3 seconds. In response to this data, the controller agent 32 can begin to decrease the pressure before it exceeds the upper limit value and generate a control action for the manufacturing apparatus 12 to maintain the pressure within a predetermined range.

[0036] The algorithms, methods, or processes disclosed herein can be provided to, or implemented by, a computer, a controller, or a processing device, which may include any dedicated electronic control unit or programmable electronic control unit. Similarly, the algorithms, methods, or processes can be stored in many forms as data and instructions executable by a computer or a controller. This includes, but is not limited to, information permanently stored on a non-writable storage medium such as a read-only memory device, and information variably stored on a writable storage medium such as a compact disc, a random access memory device, or other magnetic and optical media. The algorithms, methods, or processes can also be implemented as executable software objects. Alternatively, the algorithms, methods, or processes can be embodied, in whole or in part, using suitable hardware components such as application-specific integrated circuits, field-programmable gate arrays, state machines, or other hardware components or devices, or combinations of firmware, hardware, and software components.

[0037] Although exemplary embodiments have been described above, these embodiments are not intended to describe all possible forms encompassed by the claims. The terms used herein are terms for description rather than limitation, and it is understood that various changes may be made without departing from the spirit and scope of the disclosure. For example, the terms "controller" and "controller" may be used interchangeably herein, and the same is true for the terms "processor" and "processor".

[0038] As described above, the features of the various embodiments can be combined to form additional embodiments that are not explicitly described or illustrated. The various embodiments can be described as providing advantages or preferences over other embodiments or prior art implementations with respect to one or more desired characteristics, but one of ordinary skill in the art will recognize that one or more functions or characteristics may be compromised to achieve the desired overall system attributes that depend on the particular application and implementation. These attributes include, but are not limited to, cost, strength, durability, life cycle cost, marketability, appearance, packaging, size, ease of repair, weight, ease of manufacture, ease of assembly, and the like. Thus, an embodiment that is less desirable than other embodiments or prior art implementations with respect to one or more characteristics may not be outside the scope of the disclosure and may be desirable in a particular application.

Claims

1. In a machine learning model, receiving a training data set that describes input parameters to a manufacturing apparatus and corresponding output parameters; training the machine learning model with the training data set using at least one learning algorithm to obtain a physical model that describes the evolution of the state space of the manufacturing apparatus; configuring a machine learning-based controller agent to generate commands for the physical model to change the configuration of a simulation of the manufacturing apparatus by the physical model such that the physical model generates corresponding predicted output parameters in response to input data; training the machine learning-based controller agent with the configuration and corresponding predicted output parameters using at least one other learning algorithm such that the machine learning-based controller agent maintains the values of the predicted output parameters within respective predetermined ranges in response to input data. A method comprising.

2. The method of claim 1, further comprising configuring a machine learning-based controller agent to generate commands for the manufacturing apparatus in response to the values of the predicted output parameters from the physical model such that the manufacturing apparatus executes commands for the manufacturing apparatus.

3. The method of claim 2, further comprising configuring the machine learning-based controller agent to generate commands for the manufacturing apparatus in response to the input parameters to the manufacturing apparatus and the corresponding output parameters from the manufacturing apparatus.

4. In the machine learning-based controller agent, the configuring step includes receiving one or more rules that define control actions for the manufacturing apparatus that are executed in response to the value of at least one output parameter from the manufacturing apparatus being outside a predetermined range. The method of claim 1, comprising.

5. Receiving time series data of a machine learning-based controller agent that describes the operating state of the manufacturing apparatus; In response to the operating state indicating that the value of at least one operating parameter is outside a predetermined range, the manufacturing apparatus is caused by the machine learning-based controller agent to execute at least one of the control operations by generating a command for the manufacturing apparatus to execute at least one of the control operations, the method according to claim 4 further comprising the step of.

6. The method according to claim 4, wherein the one or more rules are obtained from a rule controller.

7. The method according to claim 4, further comprising the step of receiving an input of a machine learning-based controller agent that changes the one or more rules in real time.

8. The method according to claim 4, wherein the setting includes the rule.

9. Furthermore, the method according to claim 1, wherein the manufacturing apparatus is operated by a rule controller to generate the training data set.

10. The method according to claim 1, wherein the input parameters include active control parameters, endogenous parameters, and exogenous parameters of the manufacturing apparatus.

11. The method according to claim 1, wherein the output parameters include characteristic parameters of parts manufactured by the manufacturing apparatus.

12. The method according to claim 1, wherein the physical model is a sequence to sequence machine learning model.

13. The method according to claim 12, wherein the sequence to sequence machine learning model is an Encoder-Decoder model.

14. The method according to claim 13, wherein the Encoder-Decoder model includes a Long short-term memory model.

15. The method according to claim 1, wherein at least one of the learning algorithms is a supervised learning algorithm.

16. Training a machine learning model using at least one learning algorithm on a training data set that describes input parameters to the manufacturing apparatus and corresponding output parameters from the manufacturing apparatus to obtain a physical model that describes the evolution of the state space of the manufacturing apparatus; Configuring a machine learning-based controller agent to generate a command for the physical model to change the simulation settings of the manufacturing apparatus by the physical model so as to generate predicted output parameters corresponding to the input data. Training the machine learning-based controller agent on the settings and corresponding predicted output parameters using at least one other learning algorithm such that, in response to the input data, the machine learning-based controller agent maintains the values of the predicted output parameters within their respective predefined ranges. Configuring a machine learning-based controller agent to generate a command for the manufacturing apparatus in response to the values of the predicted output parameters from the physical model so that the manufacturing apparatus executes the command for the manufacturing apparatus. A method comprising.

17. The configuring step includes receiving, by the machine learning-based controller agent, one or more rules that define control actions for the manufacturing apparatus that are executed in response to the value of at least one output parameter from the manufacturing apparatus being outside a predefined range. The method according to claim 16.

18. Receiving machine learning-based controller agent time series data that describes the operating state of the manufacturing apparatus. Generating, by the machine learning-based controller agent, a command for the manufacturing apparatus to execute at least one of the control actions in response to the operating state indicating that the value of at least one operating parameter is outside a predefined range. The method according to claim 17, further comprising.

19. The one or more rules are obtained from a rule controller. The method according to claim 17.

20. The method according to claim 16, further comprising operating the manufacturing apparatus with a rule controller to generate the training data set.

Images