Machine controller and method for configuring and using a machine controller

Abstract

For configuring a machine controller (CTL), geometry data (GD) and measured physical data (PD) of a machine (M) are received. The geometry data (GD) and the physical data (PD) are input to a machine learning module (NN) and a simulation module (SIM) of the machine controller (CTL). By means of the input data, the simulation module (SIM) generates first values (K, F) of a first physical property of components (CP) of the machine (M) on a discretized grid (GR). Further, an evaluator (EV) is provided for evaluating a physical compatibility of the first values (K, F) with second values of a second physical property of the components (CP) and for generating residuals (D) quantifying the compatibility. The evaluator (EV) evaluates a compatibility of the first values (K, F) with output data (U) of the machine learning module (NN) and generates resulting residuals (D). Furthermore, the machine learning module (NN) is trained to minimize the resulting residuals (D), thus configuring the machine controller (CTL) for controlling the machine (M) by means of the output data (U) of the trained machine learning module (NN).

Description

[0001] Complex machines such as motors, turbines, robots, power plants, manufacturing plants, internal combustion engines, or vehicles typically require sophisticated control or monitoring strategies to operate within production and stable operating ranges and to predict potential failures. To implement such control or monitoring strategies, modern machine controllers frequently use numerical simulators to simulate the dynamic state and behavior of the corresponding machine during operation. In particular, mechanical failures caused by fatigue or external influences can be effectively reduced through preventative and predictive maintenance based on machine runtime simulations. Such runtime simulations of machines are often denoted as digital twins of the machine.

[0002] In many cases, runtime simulation provides a specification of the machine's actual state and predictions of future states, such as degradation or failure. However, it typically requires significant computational resources to run detailed simulations in real time.

[0003] To meet computational demands, several strategies are currently known:

[0004] According to the first known method, so-called reduced-order models can be used. These are approximate models obtained by reducing the machine's state space. Reduced-order models are useful for real-time simulations and / or in situations with limited computational resources. However, compared to full-order models, the reduction of the state space often incurs higher simulation errors.

[0005] Another approach involves using machine learning-based surrogate models. These surrogate models are trained to reproduce the simulation results of a simulator. After training, the surrogate model can replace the simulator. However, the black-box nature of such machine learning models makes simulations risky. In particular, if the machine or digital twin reaches a state with very little coverage from the training data, the predicted output of the surrogate model may be incorrect. This is because machine learning models typically learn input-output combinations without considering the underlying physics.

[0006] According to another approach, so-called physics-informed neural networks can be used. Such networks are described in the publication: “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential Equations,” by Raissi, Maziar, Paris Perdikaris, and George E. Karniadakis, in the Journal of Computational Physics 378 (2019), pp. 686–707. However, it remains unclear whether physics-informed neural networks are competitive for real-time simulations. Furthermore, they must undergo costly retraining when physical constraints change.

[0007] The purpose of this invention is to provide a method for configuring a machine controller, a method for controlling a machine, and a corresponding machine controller, which allows for higher accuracy for predictive control and / or requires less computational effort.

[0008] This objective is achieved by the method according to claim 1, the method according to claim 8, the machine controller according to claim 11, the computer program product according to claim 12, and the computer-readable storage medium according to claim 13.

[0009] According to a first aspect of the invention, a machine controller is configured including a machine learning module and a simulation module. For this purpose, geometric data and measured physical data of a machine to be controlled are received. Specifically, the machine may be a motor, turbine, robot, power plant, manufacturing plant, internal combustion engine, or vehicle. The geometric data and physical data are input as input data to the machine learning module and the simulation module. Using the input data, the simulation module generates a first value for at least one first physical characteristic of a component of the machine on a discretized grid. Furthermore, an evaluator is provided for evaluating the physical compatibility of the first value with a second value of at least one second physical characteristic of the component that differs from the at least one first physical characteristic, and for generating a residual that quantifies the compatibility. The evaluator evaluates the compatibility of the first value with the output data of the machine learning module and generates the resulting residual. Furthermore, the machine learning module is trained to minimize the resulting residual, thus configuring the machine controller to control the machine using the output data of the trained machine learning module.

[0010] According to a second aspect of the invention, the machine is controlled by a machine controller configured as specified above. For this purpose, operational data is received from the machine and input as input data to a trained machine learning module and a simulation module. The simulation module generates a first value for the at least one first physical characteristic using the input data. Furthermore, the trained machine learning module outputs output data as a predicted second value for the at least one second physical characteristic. Thus, an evaluator generates a residual that quantifies the compatibility between the first value and the predicted second value. The machine is controlled using the predicted second value and the obtained residual.

[0011] To implement the method of the present invention, a machine controller, a computer program product, and a preferably non-transitory computer-readable storage medium are provided.

[0012] The method and / or the machine controller of the present invention can be implemented by means of one or more processors, computers, application-specific integrated circuits (ASICs), digital signal processors (DSPs), and / or field-programmable gate arrays (FPGAs).

[0013] This invention allows for the efficient configuration and use of predictive machine controllers. Since the machine learning module can be used to predict second values, these second values ​​no longer need to be calculated by the simulation module. This often allows for avoiding a full simulation of all relevant physical characteristics of the machine. In particular, those physical characteristics that require expensive simulation can be left to the machine learning module, thus significantly reducing computational workload. Furthermore, this invention also allows for considering the physical and technical interrelationships of the machine by evaluating the physical compatibility of the machine learning module's predictions with the physical characteristics of the first physical feature. This often leads to higher prediction accuracy even in machine states with poor training data coverage. Therefore, in many cases, successful training requires less training data. Moreover, when the configured machine controller is used to control the machine, the resulting residuals can be used to estimate the uncertainty or accuracy of the predicted second values, thus enhancing control reliability.

[0014] Specific embodiments of the invention are specified by the dependent claims.

[0015] According to an advantageous embodiment of the invention, the at least one first physical characteristic may be the stiffness of the component, the force acting on the component, the thermal input to the component, and / or the electromagnetic input to the component. In particular, such a first physical characteristic may be a spatially resolved characteristic. Typically, the at least one first physical characteristic may include given, known, measured, or easily derived physical characteristics of the component. Therefore, the simulation module typically does not require a large computational workload to determine the first value from the input data. In particular, it typically does not require a full simulation involving solving one or more partial differential equations, thus significantly reducing the computational workload.

[0016] According to another advantageous embodiment of the invention, the at least one second physical property may be strain, deformation, displacement, temperature, fluid properties, vibration, or electromagnetic properties of the component. In particular, such a second physical property may be a spatially resolved property. Typically, the at least one second physical property may include properties that would require a significant amount of computation if classically simulated. Specifically, the at least one second physical property may include unknowns in physical partial differential equations. Such unknowns are those quantities that are typically determined by solving partial differential equations. However, according to embodiments of the invention, these unknowns can instead be determined by a machine learning module, thus avoiding the costly solution of partial differential equations.

[0017] Furthermore, the simulation module can use the finite element method to generate the discretized mesh and / or first values. Alternatively or additionally, the finite difference method (FDM) or the finite volume method (FVM) can be used. Several efficient implementations of these methods are available for this purpose.

[0018] Furthermore, the training of the machine learning module can be performed using reinforcement learning methods, gradient descent methods, particle swarm optimization methods, and / or genetic algorithms. Several efficient implementations are available for applying these training methods.

[0019] According to a preferred embodiment of the invention, the evaluator can determine the residual as a quantification of the non-satisfaction of a discretized physical partial differential equation, which associates at least one first physical property as a known quantity with at least one second physical property as an unknown quantity. In many cases, the trial solutions of the unknowns that satisfy or do not satisfy the partial differential equation can be determined and quantified with far less effort than actually solving the equation.

[0020] According to another embodiment of the invention, the predicted second value can be accepted or discarded for use in controlling the machine based on the obtained residual. Specifically, the obtained residual can be compared with one or more thresholds specifying a tolerance range. Then, if the corresponding predicted second value is within that range, the corresponding predicted second value can be accepted; if the corresponding predicted second value is outside that range, the corresponding predicted second value is discarded.

[0021] According to a preferred embodiment of the invention, a digital twin of a machine component can be constructed using an analog module and a trained machine learning module. The digital twin can be continuously provided with operational data of the machine, thus simulating or representing the state or behavior of the component in real time. The digital twin can also provide unrecorded or unmeasured operational or physical data of the machine, difficult-to-measure data, or data relating to the future state or behavior of the machine, for use in optimized or predictive control.

[0022] Specific embodiments of the invention are described in the following paragraphs with reference to the figures. The figures are illustrated in a schematic form:

[0023] Figure 1 The machine controller of the present invention in the configuration phase, and

[0024] Figure 2 : A configured machine controller that controls the machine.

[0025] Figure 1 A machine controller (CTL) in the configuration phase is illustrated schematically. The machine controller (CTL) includes one or more processors (PROCs) for executing method steps of the corresponding inventive method, and a memory (MEM) for storing processed data. The memory (MEM) also includes non-transitory memory for storing programs that implement the corresponding inventive method.

[0026] A machine controller (CTL) is designed to control a machine M, such as a motor, turbine, robot, power plant, manufacturing plant, internal combustion engine, vehicle, or various technical systems. For the purposes of this invention, the term "controlling a machine" should generally also cover monitoring the machine or other measures or actions taken by the machine controller (CTL) to facilitate the operation of the machine M. In this embodiment, it is assumed that the machine M is a motor.

[0027] To effectively control machine M, the machine controller CTL should be configured to determine or predict at least one physical, control-related characteristic of one or more components of machine M. For the purposes of this invention, the term "physical characteristic of a component" should also cover the physical behavior of that component. Such physical behavior can be considered as dynamic physical characteristics. For simplicity, only one component of machine M is explicitly considered in this embodiment. This component may be, for example, the rotor of motor M.

[0028] According to the present invention, configuration is performed by training a machine learning module NN using a simulation module SIM. In this embodiment, the machine learning module NN and the simulation module SIM are part of a machine controller CTL. The machine learning module NN is preferably implemented as an artificial neural network.

[0029] To provide training data for training the machine learning module NN, a database DB is coupled to the machine controller CTL. The database DB contains training data TD for machine M. The training data TD includes geometric data GD of machine M or its components and measured physical data PD of machine M or its components.

[0030] Geometric data GD may include data about the geometry, positioning, or orientation of the machine M or its components, and / or data about the geometric boundary conditions of the machine M or its components.

[0031] The physical data PD specifically includes data about the physical state of the machine M or its components, or data about the environment of the machine M. This may include data about the stiffness of the machine M or its components, applied or exerted forces, temperature, vibration, thermal input, electromagnetic input, applied voltage, flowing current, fluid characteristics, rotational speed, velocity, and / or other physical characteristics. Preferably, at least some of the physical data PD is spatially resolved.

[0032] According to this embodiment, training data TD, including geometric data GD and physical data PD, is transferred from the database DB to the machine controller CTL. Such training using stored training data from the database is often labeled as offline training. Alternatively or additionally, training data can also be received from a running machine M. Such training using real-time data from a running machine is often labeled as online training.

[0033] Within the machine controller CTL, geometric data GD and physical data PD are input as input data to both the simulation module SIM and the machine learning module NN.

[0034] The simulation module (SIM) can be a conventional simulator or a module of such a simulator. In particular, the simulation module can include or be part of a so-called FEM solver (FEM: Finite Element Method), FDM solver, or FVM solver, as provided, for example, by a conventional FEM, FDM, or FVM package.

[0035] The Simulation Module (SIM) generates a discretized mesh (GR) using input data. This discretized mesh (GR) specifies the geometry of the component in a preferably spatially discretized form. Mesh generation (GD) is typically based on geometric data (GD). However, physical data (PD) can also be taken into account during this generation, for example, influencing mesh refinement at specific locations. Such mesh generation is a standard task in many conventional FEM packages.

[0036] Furthermore, according to this embodiment, the simulation module SIM derives a stiffness matrix K and a vector of applied or exerted forces F from the physical data PD and geometric data GD. The stiffness matrix K specifies the stiffness of the components at the nodes of the mesh GR. Therefore, the vector F specifies the force acting on or exerted by the components at the nodes of the mesh GR. Typically, matrix K and vector F specify the values ​​of a first physical property of the components on the mesh GR. In the following text, the values ​​of the first physical property of the components are designated as first values.

[0037] In this embodiment, the first physical properties are spatially resolved stiffness and spatially resolved force. Additionally or alternatively, the first physical properties may include thermal or electromagnetic inputs to the component and / or various given, known, measured, or easily derived properties of the component. In particular, the first physical properties may be selected in such a way that a corresponding first value can be derived on the mesh GR without requiring excessive computation, particularly without requiring a full simulation of partial differential equations for the component.

[0038] In the current context, the component's behavior follows physical partial differential equations, which relate a first physical property to one or more second physical properties of the component. Specifically, the first physical property is a known variable or quantity in the partial differential equation, while the second physical property is an unknown variable or quantity in the partial differential equation. This means that to determine the value of the second physical property, a solution to the partial differential equation must be found.

[0039] According to this embodiment, the partial differential equations relate a given stiffness and force as a first physical property to an unknown deformation or strain of the component as a second physical property. To numerically determine the solution, i.e., the deformation, the partial differential equations are discretized on a mesh GR, resulting in a discretized differential equation DPDE. The unknown deformation or strain is also discretized on the mesh GR, resulting in a vector U representing the unknown deformation values ​​at the nodes of the mesh GR. Such discretization is a standard task in many conventional FEM packages.

[0040] Then, the discretized partial differential equation DPDE can be written as F = K * U, that is, matrix K multiplied by vector U should give vector F. In order to solve this equation, the transformed vector U must be determined such that K * UF = 0 is satisfied.

[0041] Instead of numerically solving the discretized differential equation DPDE using a classical linear algebra solver, the aim—according to the present invention—is to determine the deformation vector U from its input data GD and PD using a machine learning module NN. For this purpose, the machine learning module NN should be trained such that its output, as a result of processing the input data GD and PD, reproduces the solution U of the discretized differential equation DPDE as accurately as possible. In other words, the trained machine learning module NN should function as a predictor of the unknown variable U of the discretized differential equation DPDE. During training, the output data of the machine learning module NN can be considered as a trial solution to the discretized differential equation DPDE. Therefore, the output data of the machine learning module NN is also designated as U in the figure.

[0042] As a measure of the accuracy of solving the discretized differential equation DPDE using the output data U of the machine learning module NN, the residuals D quantifying the unsatisfiedness of the equation are calculated by the evaluator EV. For this purpose, the evaluator EV receives the grid GR, matrix K, and vector F from the simulation module SIM, and the output data U from the machine learning module NN.

[0043] The evaluator (EV) is part of the machine controller's control chain (CTL) and is typically designed to evaluate and quantify the physical compatibility of a first value of one or more first physical characteristics of a component with a second value of one or more second physical characteristics. In the current case, the evaluator (EV) is, for example, based on D = (K * UF). 2 Alternatively, the residual D, calculated from matrix K, vector F, and output data U using D = |K*UF|, serves as the preferred Euclidean distance between vector K*U and vector F. The residual D quantifies the preceding physical compatibility. In a sense, the residual D can be viewed as the prediction loss or solution error of the machine learning module NN. It should be noted that determining whether a discretized partial differential equation is satisfied or not typically requires far less computation than actually solving the equation using classical methods.

[0044] According to the present invention, the residual D is minimized by training a machine learning module NN. Here, the term "training" generally means optimizing the mapping from the input data of the machine learning module to its output data during the training phase, based on pre-determined and / or learned criteria. In the present case, this criterion includes the non-satisfaction of the output data U of the machine learning module NN with the discretized partial differential equation DPDE, which is quantified by the residual D. Several machine learning methods, particularly those for reinforcement learning, are available for minimizing the residual D. The mapping can be optimized by adjusting the mapping parameters of the machine learning module NN. In the case of artificial neural networks, the connection structure of its neurons and / or the weights of the connections between neurons can be varied to optimize the mapping. For such optimization, various numerical standard methods, such as gradient descent, particle swarm optimization, or genetic algorithms, are available.

[0045] The residual D determined by the evaluator EV is returned to the machine learning module NN—as Figure 1 The dashed arrow in the diagram indicates that this is to minimize the error by adjusting the neural weights, for example, those of a machine learning module (NN). In this way, the NN learns, on the one hand, the correlation between its input data GD and PD, and on the other hand, the solution U of the discretized partial differential equation DPDE. After successful training, the machine learning module NN may be able to determine the corresponding solution U of the discretized partial differential equation DPDE with sufficient accuracy from the input data GD and PD (here, deformation or strain).

[0046] In this manner, the trained machine learning module NN constitutes a surrogate model for solving the discretized partial differential equations DPDE, thus eliminating the need to invoke a full simulation. In fact, according to this embodiment, the simulation module SIM only needs to provide the specifications of the simulation model, namely the discretized grid, the first discretized values, and optionally one or more physical relationships between them. Furthermore, it has been proven that the adherence to physical laws via the physical differential equations DPDE significantly enhances the robustness, accuracy, and effectiveness of the method of this invention.

[0047] After training, the machine controller CTL can use the machine learning module NN to control the machine M, such as... Figure 2 As shown in the illustration. Figure 2 Zhongyu Figure 1 The same reference numerals in the text indicate the same or corresponding entities, which are preferably as shown above.

[0048] The machine controller CTL is coupled to the running machine M, which includes a component CP. As mentioned above, the component CP can be the rotor of a motor.

[0049] The machine controller (CTL) receives operational data (OD) from the operating machine M, which includes measured physical and geometric data of the machine M. Instead of directly obtaining geometric data from the machine M, it can be obtained at least partially from the machine M's specifications. The operational data (OD) can further specify the state, control actions, performance, emissions, and / or resource consumption of the machine M or its environment. For this purpose, the machine M can be equipped with several sensors for measuring and providing process data, machine data, and / or environmental data of the machine M.

[0050] The operational data OD is fed as input data into the simulation module SIM and the trained machine learning module NN.

[0051] As described above, the simulation module SIM generates a vector of discretized mesh GR, stiffness matrix K, and applied or exerted forces F from the input data. The mesh GR, stiffness matrix K, and vector F are transferred from the simulation module SIM to the evaluator EV. A trained machine learning module NN is executed along with the operational data OD and derives output data U from them as a prediction of the solution to the discretized partial differential equation DPDE. According to this embodiment, the output vector U includes the values ​​of deformation or strain of component CP. It should be noted that executing a trained machine learning model typically requires far less computational effort than a full simulation. Therefore, in many cases, the solution to the differential equation DPDE can be predicted in real time.

[0052] The predicted output vector U is fed from the trained machine learning module NN to the evaluator EV. The evaluator EV then determines the residual D obtained as described above from the mesh GR, stiffness matrix K, vector F, and output vector U.

[0053] The resulting residual D and output vector U are then transmitted, respectively, from the evaluator EV or from the trained machine learning module NN to the control signal generator CG of the machine controller CTL. The control signal generator CG generates a control signal CS for controlling the machine M based on the predicted output vector U and the obtained residual D. The residual D is advantageously used by the control signal generator CG to evaluate the uncertainty or accuracy of the predicted solution U. Depending on the residual D, the predicted solution U is accepted or discarded for controlling the machine M. Specifically, if the residual D is small, thus indicating a reliable prediction U, the predicted output vector U is used to control the machine M. Otherwise, the prediction U is ignored or its weight is reduced.

[0054] The generated control signal CS is transmitted from the control signal generator CG and applied to machine M to control machine M.

[0055] The simulation module SIM, the trained machine learning module NN, and the evaluator EV constitute a hybrid simulator HS for simulating one or more second physical properties, particularly the deformation of component CP, without requiring classical solution of the fundamental partial differential equations. Therefore, in many cases, the computational workload is significantly reduced. Furthermore, the hybrid simulator HS also provides residuals D as a measure of possible simulation error. This measure is particularly valuable in regions where the available training data TD poorly covers the state space.

[0056] In many cases, the reduced computational workload allows for real-time simulation. In such cases, a hybrid simulator (HS) can be used to implement a digital twin (DT) of the component (CP). The hybrid simulator (HS) or digital twin (DT) can accordingly be continuously provided with operational data (OD) of the running machine (M), thus simulating or representing the state or behavior of the component (CP) in real time and in parallel with the actual operation of the running machine (M). The hybrid simulator (HS) or digital twin (DT) can also correspondingly provide physical data of the machine (M), which is not directly measured or refers to the future state or behavior of the machine (M). This physical data can then be used in real time for optimized or predictive control of the machine (M) or for predictive or preventative maintenance.

Claims

1. A computer-implemented method for configuring a machine controller (CTL) comprising a machine learning module (NN) and a simulation module (SIM), the method comprising: a) receiving geometric data (GD) and measured physical data (PD) of a machine (M) to be controlled, b) inputting the geometric data (GD) and the physical data (PD) as input data to the machine learning module (NN) and the simulation module (SIM), c) generating, by the simulation module (SIM), first values (K, F) of at least one first physical property of a component (CP) of the machine (M) on a discretized grid (GR) by means of the input data, d) providing an evaluator (EV) for evaluating a physical compatibility of the first values (K, F) with second values of at least one second physical property of the component (CP) different from the at least one first physical property and for generating residuals (D) quantifying the compatibility, e) evaluating, by the evaluator (EV), a compatibility of the first values (K, F) with output data (U) of the machine learning module (NN) and generating resulting residuals (D), and f) training the machine learning module (NN) to minimize the resulting residuals (D), thus configuring the machine controller (CTL) for controlling the machine (M) by means of output data (U) of the trained machine learning module (NN).

2. The method according to claim 1, wherein the at least one first physical property is a stiffness of the component (CP), a force acting on the component (CP), a thermal input to the component (CP) and / or an electromagnetic input to the component (CP).

3. The method according to claim 1 or 2, wherein the at least one second physical property is a strain, a deformation, a displacement, a temperature, a fluid property, a vibration or an electromagnetic property of the component (CP).

4. The method according to claim 1 or 2, wherein the simulation module (SIM) uses a finite element method for generating the discretized grid (GR) and / or the first values (K, F).

5. The method according to claim 1 or 2, wherein the machine learning module (NN) comprises an artificial neural network, a recurrent neural network, a convolutional neural network, a reinforcement learning model, a Bayesian neural network, an autoencoder, a deep learning architecture, a support vector machine, a data-driven trainable regression model, a k-nearest neighbor classifier, a physical model and / or a decision tree.

6. The method according to claim 1 or 2, wherein the training of the machine learning module (NN) is performed by means of a reinforcement learning method, a gradient descent method, a particle swarm optimization method and / or a genetic algorithm.

7. The method according to claim 1 or 2, wherein the evaluator (EV) determines the residual (D) as a quantified measure of non-fulfillment of a discretized physical partial differential equation (DPDE) relating the at least one first physical property as a known quantity to the at least one second physical property as an unknown quantity.

8. A computer-implemented method for controlling a machine (M) by a machine controller (CTL) configured according to one of the preceding claims, the method comprising: - receiving operational data (OD) from the machine (M) and inputting them as input data to a trained machine learning module (NN) and a simulation module (SIM), - generating first values (K, F) of at least one first physical property by the simulation module (SIM) by means of the input data, - outputting output data (U) by the trained machine learning module (NN) as predicted second values of at least one second physical property, - generating a resulting residual (D) of quantifying a compatibility of the first values (K, F) with the predicted second values (U) by an evaluator (EV), and - controlling the machine (M) by means of the predicted second values (U) and the resulting residual (D).

9. The method according to claim 8, wherein the predicted second values (U) are accepted or discarded for controlling the machine (M) depending on the resulting residual (D).

10. The method according to claim 8 or 9, wherein - a digital twin (DT) of a component (CP) constituting the machine (M) is constituted by means of the simulation module (SIM) and the trained machine learning module (NN), and - the digital twin (DT) is continuously provided with operational data (OD) of the machine (M) so as to simulate or represent a state or behavior of the component (CP) in real time.

11. A machine controller (CTL) for controlling a machine (M), adapted to perform the method according to one of claims 1 to 10.

12. A computer program product for controlling a machine (M), adapted to perform the method according to one of claims 1 to 10.

13. A non-transitory computer-readable storage medium storing the computer program product according to claim 12.

Images