Method and apparatus for providing virtual sensor using machine learning

Abstract

The invention relates to a method for providing a virtual sensor which is configured to determine a state variable (x) which characterizes a state of a technical system (200, 300) on the basis of an operating variable profile (u) of the technical system (200, 300), in particular on the basis of measurements of the technical system (200, 300), in which first training data (D1) is provided, the first training data comprises a pair of a temporal curve (u (t)) of an operating variable, which is determined empirically, in particular as a function of measurements, and a curve (x (t)) of an associated state variable, a first model (h1) is determined from the data, said first model characterizing the evolution of the state variable (x (t)) as a function of the operating variable (u (t)) and the state variable (x (t)), and a second model (h2) characterizing the evolution of the state variable (x (t)) as a function of the operating variable (u (t)). Wherein second training data (D2) is generated by means of the first machine learning system, said second training data comprising a temporal curve (u (t)) of the operating variable and an associated curve (x (t)) of the state variable determined by means of the first model (h1), and wherein a second model (h2) is trained on the basis of the second training data (D2), the second model (h2) characterizes the evolution of the state variable (x (t)) with the operating variable (u (t)) (and of the state variable (x (t))) and has a lower complexity than the first model (h1), and wherein the virtual sensor is provided in such a way that it determines the operating variable (x) from the state variable (u) by means of the second model (h2).

Description

Technical Field

[0001] The present invention relates to a method for providing a virtual sensor, a virtual sensor, a method for operating a virtual sensor, a computer program, a machine-readable storage medium, and a training device. Background Technology

[0002] A method for providing an aging state model to determine the modeled state of aging (SOH) of a battery cell is known from DE 10 2022 200 022 A1, comprising the following steps: • Provide training datasets that correlate empirically determined aging states with the change curves of one or more operating variables at specific time points; • Provides a data-based aging model constructed using a set of neural differential equations to model the internal electrochemical state of an energy storage device, wherein the internal state can be mapped to an aging state, wherein at least one differential equation in the set of differential equations has a sum of deterministic model terms and data-based correction terms, the correction terms being formed by a data-based correction model; • The data-based aging model is trained using the training dataset to determine the model parameters of the data-based calibration model. It is also known that this aging model is used as a virtual sensor.

[0003] Therefore, a method for determining the modeled state of aging (SOH) of a battery cell at a specific time point is also known from the aforementioned literature, comprising the following steps: • Provides runtime variable change curves for one or more runtime variables; • Provides a trained, data-based aging state model that is based on a set of neural differential equations and is trained to correlate the curves of change of running variables with the aging state. • By evaluating the aging state model using the change curves of the running variables over time steps, the aging state of the model can be determined or predicted.

[0004] The neural differential equation set is known in Chen et al.'s "Neural Ordinary Differential Equations" (https: / / arxiv.org / pdf / 1806.07366.pdf). Summary of the Invention

[0005] Virtual sensors typically refer to the curves of changes in variables that characterize the state of a technical system, based on the curves of changes in operating variables (determined by measurement).

[0006] The present invention, having the features of independent claim 1, has the advantage over the prior art in that the virtual sensor has particularly powerful performance. In particular, it has robust extrapolation capabilities.

[0007] Other aspects of the invention are the subject of the independent claims. Advantageous improvements are the subject of the dependent claims.

[0008] In a first aspect, the present invention relates to a method for providing a virtual sensor configured to: determine a state variable change curve characterizing the state of a technical system based on a change curve of an operating variable (particularly a change curve of an operating variable determined by measurements of the technical system), wherein first training data is provided, the first training data comprising pairs of time-varying curves of operating variables determined empirically (i.e., particularly based on measurements) and associated change curves of state variables, wherein a first model is determined based on these data, the first model characterizing the evolution of state variables according to operating variables and state variables, wherein second training data is generated by means of the first machine learning system, the second training data comprising time-varying curves of operating variables and associated change curves of state variables determined by means of the first model, wherein a second model is trained based on the second training data, the second model characterizing the evolution of state variables according to operating variables (and state variables), and the second model having lower complexity (i.e., fewer adjustable parameters during training) compared to the first model, wherein a virtual sensor is provided such that the virtual sensor determines the operating variables according to the state variables using the second model.

[0009] This offers several advantages: it requires less real training data while improving extrapolation capabilities. Furthermore, the reduced complexity allows for more resource-efficient virtual sensors.

[0010] In one improved approach, it can be proposed that the first model and / or the second model include a physical model and a machine learning system, such as an artificial neural network or a Gaussian process.

[0011] Physical models offer greater interpretability in this context, which is particularly important for safety-critical applications of virtual sensors. In this modeling approach, machine learning systems compensate for potential inaccuracies in the physical model. However, this data-driven aspect of the modeling can sometimes be difficult to interpret or explain clearly.

[0012] Therefore, in one improved approach, it can be proposed that the first and second models comprise the same physical model. This means that the machine learning system of the second model has reduced complexity compared to the machine learning system of the first model. This makes the virtual sensor more interpretable because the data-based terms of machine learning are transformed into structured terms of the physical model. Furthermore, this enables more robust predictions from the virtual sensor, i.e., improved extrapolation capabilities.

[0013] In one improvement in this regard, it can be proposed that the second model consists only of the physical model, that is, the machine learning system is reduced to a "model" without free parameters.

[0014] Alternatively or additionally, it can be proposed that the physical model of the second model further reduce its complexity using model lordnungsreduktion methods to achieve a better relationship between accuracy and model complexity. For this purpose, standard model lordnungsreduktion methods, such as the Krylov subspace method, can be used.

[0015] Advantageously, in these aspects, it can be proposed that the first and / or second models describe the evolution of state variables as neural differential equations. Such neural differential equations are known, for example, in Chen et al.'s "Neural Ordinary Differential Equations" (2019, https: / / arxiv.org / pdf / 1806.07366.pdf). Due to their high model accuracy and robustness, these neural differential equations are particularly suitable for modeling dynamic systems, and thus also well-suited for virtual sensors. Subsequently, the state variables can be advantageously determined by integrating the neural differential equations.

[0016] An improved approach proposes selecting a second model from multiple candidate functions with varying complexities (i.e., different numbers of adjustable parameters). This selection is Pareto-optimal in terms of both model complexity and performance among the candidate functions. This approach effectively resolves the conflict between model quality and the number of parameters.

[0017] In a specific exemplary application, the technology system is an energy storage device, especially a battery, and the state variable particularly characterizes the aging state of the energy storage device.

[0018] In another exemplary application, the technical system is an electric motor, and the state variable specifically characterizes the temperature of the motor rotor.

[0019] In another aspect, the present invention relates to a virtual sensor configured to determine state variables characterizing the state of a technical system based on operating variable change curves (especially operating variable change curves determined based on measurements of the technical system); the virtual sensor is provided using one of the methods described above.

[0020] In another aspect, the present invention relates to a method for operating the virtual sensor, wherein the operating variable change curve of the technical system is determined (in particular, measured), and the state variable characterizing the state of the technical system is determined. Attached Figure Description

[0021] The embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. In the drawings: Figure 1 The illustration shows a first exemplary application of the virtual sensor; Figure 2 This illustration shows a second exemplary application of the virtual sensor; Figure 3 An implementation of the method for providing a virtual sensor is illustrated in a flowchart, along with the subsequent optional operation of the virtual sensor; Figure 4 An apparatus for performing the method of providing the virtual sensor is illustrated schematically. Detailed Implementation

[0022] Figure 1 A motor vehicle (100) is shown, which has a sensor (110) that, in this embodiment, measures the voltage applied to an electrical energy storage device (200) (particularly a battery, or fuel cell, which may contain at least one battery cell). The electrical energy storage device (200) can, for example, be used to supply energy to the drive unit (130) of the motor vehicle (100). The measurements from the sensor (110) are transmitted to a computer (120), which, via... Figure 3 The method shown provides a virtual sensor from which variables are determined, in some embodiments, that determine the State of Health (SOH) of the energy storage device (200). Based on the determined SOH, the drive unit (130) can be manipulated accordingly. In some embodiments, the computer (120) may be located inside the vehicle (100). In alternative embodiments, the computer (120) may be located outside the vehicle (100).

[0023] For example, the optimal operating strategy can be used to manipulate the drive unit (130).

[0024] Alternatively or additionally, maintenance measures for the energy storage device (200) can be predicted based on the predicted aging condition.

[0025] In other implementations of the energy storage device (200) provided by a battery, the charging strategy of the battery can be optimally controlled.

[0026] Figure 2 Other embodiments for applying virtual sensors are described. A motor vehicle (100) has a sensor (111) that, in this embodiment, measures the voltage applied to a motor (300), particularly a generator or electric motor, especially the voltage applied to the stator windings or rotor (310) windings of the motor (300). The motor (300) may, for example, be... Figure 1 Part of the drive unit (130) shown, and from the same in Figure 1 The energy storage device (200) shown receives energy. Measurements from the sensor (111) are transmitted to the computer (120), which, via... Figure 3 The method shown provides a virtual sensor from which to determine a variable characterizing the temperature within the rotor (310) of the motor (300). Based on the determined temperature, the actuators (135) of the vehicle (100) can be manipulated accordingly. In some embodiments, a computer (120) may be located within the vehicle (100). In alternative embodiments, the computer (120) may be located outside the vehicle (100).

[0027] In some implementations, the cooling strategy of the motor (300) can be adjusted according to the determined temperature.

[0028] Figure 3 The flowchart describes the process for providing and subsequently operating the virtual sensor.

[0029] First, a virtual sensor is provided. For this purpose, (1000) initial training data is provided: .

[0030] These training data include time-varying curves of the operating variables (u) of the technical system, and associated state variables (x) characterizing the state of the technical system. Figure 1 and Figure 2 In the embodiment shown, the operating variable (u) corresponds to the variable determined by the sensors (110, 111). These associated values, consisting of the change curves of the operating variable (u) and the state variable (x), can be determined, for example, on a test bench.

[0031] Then, the first model (1100) is provided: .

[0032] In some implementations, the first model (h1) includes a physical model (f), such as a 1-node or 2-node model (e.g., a rotor temperature model) or an electrochemical cell model. This physical model (f) determines an approximation of the state vector x based on the operating variable (u) (for the battery shown in Figure 1, this could be, for example, a curve showing the variation of the operating variable (u) during cycling with respect to the current intensity required to charge or discharge the battery). This physical model is parameterized using a parameter θ.

[0033] The first model (h1) also includes a machine learning system (g), such as an artificial neural network or a Gaussian process. This machine learning system (g) describes the data-based portion of the modeling, which is used to improve the purely physical modeling. The machine learning system utilizes parameters... To parameterize.

[0034] Now, in some implementations, the first model (h1) is trained as a set of neural differential equations in a known manner, using the first training data D1, i.e., according to the following equations: (1) In some implementations, the state variable (x) describes the state vector and includes physically measurable or interpretable variables, such as the temperature of the rotor (310) of the motor (300), or characterizing the thickness of the SEI or the amount of cyclic lithium in the battery (200).

[0035] Then, artificial data (1200) are generated using the first model (h1) trained in this way. For this purpose, curves of change of the running variable (u) are provided, for example based on load distribution maps, which in some implementations are derived from real data (e.g., from real driving or test driving).

[0036] Then, using the first model (h1), the associated state variable (x) change curves are determined and provided from the change curves of these respective operating variables (u). This can be interpreted as artificial labeling. The pairs thus generated, consisting of the change curves of the operating variable (u) and the state variable (x), are provided as the second training data: In some implementations, the variation curve of the running variable (u) covers a very wide range, thus forcing the first model (h1) to extrapolate. This allows for the simulation to generate a very large second training dataset (D2).

[0037] In some implementations, it can be proposed to generate the change curve of the state variable (x) by sampling the change curve of the state variable (x) measured in step (1000) with a higher time resolution.

[0038] In some implementations, it may be proposed to interpolate the measured curve of the state variable (x) to determine the curve of the simulated state variable (x). In particular, this interpolation can be performed at an abstract level and, in some implementations, may include a certain number of fast charging cycles and / or charging current of the battery (200), and / or the driving performance and / or operating strategy of the vehicle (100), or intervention limits (e.g., derating temperature). This interpolation can be based on a single parameter or on a combination of at least two parameters.

[0039] These artificially generated data can now also be provided as second training data (D2), or added to second training data (D2). In some implementations, it is proposed that the second training data (D2) contains more data than the first training data (D1). This is because: the curves of change of the state variable (x) determined by the first model (h1) are easier to generate than the curves of change of the state variable (x) measured in reality; and the first model (h1) has a physical model f(x(t),u(t),t,θ) adapted to the domain, and therefore it can have better externalities than a pure machine learning system for certain operating variables.

[0040] Subsequently, a second model (1300) is provided: Like the first model (h1), the second model (h2) also includes a physical term (f2) and a second machine learning system (g2). In some implementations, the second machine learning system may be given as an artificial neural network, while in others it may be given as a Gaussian process. In some implementations, the physical term (f2) is the same as the physical term (f) of the first model (h1).

[0041] In some implementations, it can be proposed that the second machine learning system (g2) is always zero, i.e., g2(x(t),u(t),t, The value is set to 0. This ensures that as much information encoded in the first model (h1) is transferred to the physical model (f) as possible when training the second model (h2) subsequently. This guarantees the highest possible interpretability and better extrapolation capability of the second model (h2).

[0042] In alternative implementations, the second machine learning system (g2) can be proposed as a neural network with a smaller capacity than the machine learning system (g) given as a neural network. This can mean that the architecture of the second machine learning system (g2) has fewer parameters compared to the architecture of the machine learning system (g), such as fewer neurons per layer and / or fewer layers. In some implementations, the nonlinear activation function of the machine learning system (g) can be replaced with a linear activation function in the second machine learning system (g2), which also results in fewer parameters. (In other words, in some implementations this might mean that the dimension of ξ is smaller than...) ).

[0043] In an alternative implementation, it can be proposed that the second machine learning system (g2) is a Gaussian process with a smaller capacity than the machine learning system (g) given as a Gaussian process. This could mean that the second machine learning system (g2) has fewer inducing points than the first machine learning system (g).

[0044] When training the second model (h2) subsequently, this results in the second machine learning system (g2) having a smaller model capacity than the machine learning system (g) of the first model (h1). This allows the physical component (f2) in the second model (h2) to interpret as much information as possible from the machine learning system (g) of the first model (h1). Furthermore, the (now simplified) second machine learning system (g2) can still help compensate for any potential modeling errors, achieving higher accuracy.

[0045] In some implementations, the second machine learning system (g2) may include an interpretable, data-based model, such as symbolic regression, linear feature combination, xNN known from https: / / arxiv.org / pdf / 1806.01933.pdf, or a (preferably linear) combination of basis functions. This provides high accuracy while maintaining good interpretability during the subsequent training of the second model (h2), thus achieving good verifiability.

[0046] Then, the second model (h2) is trained (1400) using the second training data (D2), which, in some implementations, is similar to the description of the first model (h1) above, and is trained as a system of neural differential equations: (2) Now, a virtual sensor comprising the second model (h2) is provided (1500). In this case, the virtual sensor is configured to: feed the change curve of the operating variable (u) input to it to the second model, determine the change curve of the state variable (x) by integrating equation (2) (in some embodiments, numerical integration), and provide it at its output.

[0047] In some implementations, it may be proposed to further compress the second model (h2) before providing the virtual sensor to use the virtual sensor as resource-efficiently as possible. For example, this can be achieved by performing the Krylov subspace method with torque balancing.

[0048] In some implementations, in this case and / or when providing the second model (h2), it may be proposed to: identify multiple candidates for the second model (h2) or the compressed second model (h2), and then, in order to provide the virtual sensor, select the Pareto-optimal model, especially in terms of resource consumption and performance.

[0049] This concludes the method for providing virtual sensors.

[0050] Subsequently, in some embodiments, it may be proposed that the virtual sensor (1600) is now running, as shown in the embodiments of Figures 1 and 2, for example.

[0051] Steps (1000) to (1500) can be repeated, for example, at preset time intervals. Thus, the first model (h1) is initially set to be equal to the already trained second model (h2), and advantageously, the unchanged version of the first model (h1) initialized in this way is offered as an option for Pareto optimal selection.

[0052] Figure 4 illustrates a training device (400) configured to perform the method shown in Figure 3 (excluding optional step 1600). In some embodiments, the method is executed as a computer program stored on a computer-readable storage medium (420) and contains instructions that, when the computer (410) runs the computer program, cause the computer (410) of the training device (400) to execute the method.

Claims

1. A method for providing a virtual sensor, the virtual sensor being configured to: determine a state variable (x) characterizing the state of a technical system (200, 300) based on a change curve (u) of an operating variable, the change curve being determined, in particular, based on measurements of the technical system (200, 300), wherein, First training data (D1) is provided, comprising pairs of time-varying curves (u(t)) of the running variable determined empirically (i.e., particularly based on measurement) and the associated state variable's changing curves (x(t)). A first model (h1) is determined based on this data, characterizing the evolution of the state variable (x(t)) with respect to the running variable (u(t)) (and the state variable (x(t))). Second training data (D2) is generated using the first machine learning system, comprising the time-varying curves (u(t)) of the running variable (u(t) ... (u The second model (h2) is trained based on the second training data (D2), and the second model characterizes the evolution of the state variable (x(t)) according to the running variable (u(t)) (and the state variable (x(t))), and the second model has lower complexity than the first model (h1), wherein the virtual sensor is provided such that the virtual sensor determines the running variable (x) according to the state variable (u) using the second model (h2).

2. The method according to claim 1, wherein, The first model (h1) and / or the second model (h2) include a physical model (f) and a machine learning system (g).

3. The method according to claim 2, wherein, The first model (h1) and the second model (h2) include the same physical model (f).

4. The method according to any one of the preceding claims, wherein, The second model (h2) includes only the physical model (f).

5. The method according to any one of the preceding claims, wherein, The first model (h1) and / or the second model (h2) describe the evolution of the state variable (x(t)) as a neural differential equation.

6. The method according to any one of the preceding claims, wherein, The second model (h2) is selected from a plurality of candidate functions having different complexities, i.e., different numbers of adjustable parameters, wherein the second model (h2) selected from the plurality of candidate functions is Pareto optimal in terms of the model complexity and performance of the candidate functions.

7. The method according to any one of the preceding claims, wherein, The technical system is an energy storage device (200), especially a battery, and the state variable (x(t)) particularly characterizes the aging state (SOH) of the energy storage device (200).

8. The method according to any one of claims 1 to 6, wherein, The technical system is a motor (300), and the state variable (x(t)) characterizes the temperature of the rotor (310) of the motor (300).

9. A virtual sensor configured to determine a state variable (x) characterizing the state of a technical system (200, 300) based on a change curve (u) of an operating variable of the technical system (200, 300), the change curve of the operating variable being determined, in particular, based on measurements of the technical system (200, 300), the virtual sensor being provided using the method according to any one of claims 1 to 6.

10. A method for operating a virtual sensor according to claim 9, wherein the method determines, in particular measures, the operating variable change curve (u) of the technical system (200, 300), and determines the state variable (x) characterizing the state of the technical system (200, 300).

11. A computer program comprising instructions that, when executed by a computer (410), cause the computer to perform the method according to any one of claims 1 to 8 or 10.

12. A machine-readable storage medium (420) having a computer program as claimed in claim 11 stored thereon.

13. A training apparatus (400) comprising a computer (410) configured to perform the method according to any one of claims 1 to 8.

Images