Method for determining a solid fraction

Abstract

The invention relates to a computer-implemented method for determining a solid fraction in a flowable, in particular aqueous, medium, in particular by means of a microwave sensor, comprising: – receiving an input data set having at least one effective permittivity value ε e ƒƒ, in particular ascertained from measurement data, of the medium; – providing the input data set to a machine learning model, the machine learning model having an exchangeable, application-specific parameter set, the parameter set having at least one weighting factor wi and at least one bias bi; – determining a solid fraction value δ of the solid in the medium by means of the machine learning model; – generating an output data set comprising the solid fraction value δ. Furthermore, the invention relates to a machine learning model, a computer-implemented method for generating a training data set for a machine learning model and a microwave sensor.

Description

[0001] Method for determining a solids content

[0002] The invention relates to a computer-implemented method for determining a solid content, a machine learning model for determining a solid content, a computer-implemented method for generating a training data set for a machine learning model for determining a solid content, a microwave sensor for determining a solid content, a computer program with instructions which, when the program is executed by a computer, cause the computer to carry out the method according to the invention, a computer-readable medium which contains instructions which, when executed by a computer, cause the computer to carry out the method according to the invention and a method for configuring a microwave sensor according to the invention.

[0003] Microwaves can be used to determine the physical quantities of permittivity and loss factor of a medium in a process line. These two quantities – measured either at one or across many different frequencies – can be used to draw conclusions about application-specific parameters, such as the water content in a mixture of water and other non-polar or slightly polar components, or the solid content in a liquid medium.

[0004] The established transmission-Z-reflection measurement is described in LF Chen, CK Ong, CP Neo, VV Varadan, VK Varadan - “Microwave Electronics, Measurement and Materials Characterization”, John Wiley & Sons Ltd., 2004. For this purpose, the microwave signal is coupled to the medium in a container or measuring tube at two different positions, the scattering parameters (transmission and, if applicable, reflection) between these coupling structures are measured, and the measured scattering parameters are used to calculate the physical properties of the medium.

[0005] WO 2018 / 121927 A1 teaches a measuring arrangement for analyzing the properties of a flowing medium using microwaves. In addition to the microwave antennas, the measuring arrangement comprises an electrically insulating lining layer on the inner surface of the measuring tube. This lining layer forms a dielectric waveguide through which a microwave signal can be transmitted, at least in part, from a first microwave antenna to a second microwave antenna. One application for such a measuring arrangement is the determination of solids content in the liquid medium being conveyed. WO 2021 / 099152 A1 teaches a microwave antenna having a front section in contact with the medium, through which the excitation signal is emitted into the medium.

[0006] DE102017131269A1 discloses a method for continuously determining the fat content of milk with varying solids content using a Coriolis flowmeter and a microwave sensor. This type of measuring device can be integrated into milk processing processes. DE 42 27 727 A1 discloses a method for determining a state variable of a measuring liquid from its conductivity value α and its temperature value T using a neural network of the type mentioned above. Using this known method, the measuring liquid is classified as surfactant-containing or surfactant-free. The output of the neural network is either the value 1 (measurement liquid containing surfactant) or the value 0 (measurement liquid free of surfactant). To classify the measurement liquid, the concentration of non-ionic surfactants in the measurement liquid is determined.To determine the concentration of non-ionic surfactants, the neural network receives not only the conductivity and temperature of the measuring fluid, but also the pH and turbidity of the measuring fluid via two additional inputs in the input layer. Therefore, the known method does not use the neural network to determine specific concentration values ​​and output them via the output of the output layer.

[0007] DE 198 35 137 C2 discloses a method for determining a state variable of a measuring liquid from its conductivity value o and its temperature value T with the aid of a neural network with an input layer with at least two inputs for the conductivity value and the temperature value, with an output layer with an output for the state variable and with at least one hidden layer arranged between the input layer and the output layer with a plurality of neurons connected between the inputs and the output, characterized in that the concentration of the ions dissolved in the measuring liquid is determined as the state variable, wherein in each neuron the input values ​​of the inputs are each weighted with an input-side weighting factor, the weighted input values ​​are summed and the sum value is passed via an activation function.

[0008] The invention is based on the object of also providing a solution for determining the solid content in a medium based on the cited prior art.

[0009] The problem is solved by the method according to claim 1 and the subject matter of the adjacent claims.

[0010] The computer-implemented method according to the invention for determining a solids content in a flowable, in particular aqueous, medium, in particular by means of a microwave sensor, comprises:

[0011] Receiving an input data set with at least one effective permittivity value, in particular determined from measurement data of the medium, and a temperature value T of the medium;

[0012] Providing the input data set to a machine learning model, wherein the machine learning model has an interchangeable, application-specific parameter set, wherein the parameter set includes at least one weighting factor and at least one bias b t has;

[0013] Determining a solid fraction value 8 of the solid in the medium using the machine learning model;

[0014] Generating an output data set including the solids content value 8;

[0015] Optional: Output the output data set, especially the solids content value 8.

[0016] Advantageous embodiments of the invention are the subject of the subclaims.

[0017] One embodiment provides that the input data set further comprises a frequency value f and / or an effective density value p eff of the medium.

[0018] One embodiment provides that the parameter set results from a mathematical model applied in a learning phase of the machine learning model to describe the properties of heterogeneous mixtures of substances consisting of an insoluble substance located in a carrier substance.

[0019] One embodiment provides that the mathematical model is selected from:

[0020] Maxwell-Garnett model

[0021] Parallel Wiener equation

[0022] Serial Wiener equation

[0023] The machine learning model according to the invention for determining a solids content in a flowable, in particular aqueous, medium, in particular for use in the method according to the invention, comprises: an input layer with at least two inputs for an effective permittivity value E of the medium and a temperature value T of the medium; an output layer with an output for a solids content value θ of the solid in the medium; at least two hidden intermediate layers present between the input layer and the output layer, each having a number N tswitched neurons, wherein the machine learning model is configured to weight input values ​​received in each neuron of a first intermediate layer of the at least two intermediate layers with a weighting factor, to sum the weighted input values ​​and a bias to a sum value, to convert the sum value into an output value via an activation function and to provide the output value to at least one neuron, in particular to all neurons, of a second intermediate layer of the at least two intermediate layers, wherein the weighting factor w t and the bias b t form an interchangeable, application-specific parameter set.

[0024] One embodiment provides that a hyperbolic tangent function and / or a ReLU (rectified linear unit) function is used as the activation function.

[0025] One embodiment provides that the input layer has at least one further input for a frequency value f and / or an effective density value p eff has.

[0026] One embodiment provides that the machine learning model has exactly three intermediate layers, wherein a total sum of the neurons of the three intermediate layers is selected from a range of 40 to 2000, in particular up to 1000 and preferably up to 200 neurons.

[0027] One embodiment provides that a first intermediate layer has a number of neurons selected from the range 20 to 80, wherein a second intermediate layer has a number of neurons selected from the range 10 to 80, wherein a third intermediate layer has a number of neurons selected from the range 10 to 60,

[0028] One embodiment provides that the machine learning model has exactly four intermediate layers, with a total sum of the neurons of the four intermediate layers being selected from a range of 100 to 250 neurons.

[0029] One embodiment provides that a first intermediate layer has a number of neurons selected from the range 20 to 60, wherein a second intermediate layer has a number of neurons selected from the range 50 to 80, wherein a third intermediate layer has a number of neurons selected from the range 20 to 60, wherein a fourth intermediate layer has a number of neurons selected from the range 10 to 50.

[0030] One embodiment provides that the machine learning model has a data size of less than or equal to 500 kB, with a deviation of less than 0.1% relative to the mathematical model serving as the basis for the learning phase.

[0031] The computer-implemented method according to the invention for generating a training data set for a machine learning model for determining a solid content 8 of a solid in a flowable, in particular aqueous, medium, in particular for the machine learning model according to the invention, comprises:

[0032] Creating an input training dataset, where the input training dataset has an effective permittivity value of the medium, a permittivity value of the solid, a permittivity value e m of the carrier medium, a solids content value 8 and a temperature T, wherein at least the effective permittivity value E eff and the permittivity value are determined by a random generator, whereby the solid content value 8 is calculated using a mathematical model depending on the effective permittivity value E eff , the permittivity value and the permittivity value sm .

[0033] One embodiment provides that the input training data set further comprises an additional parameter ß, which is selected from a number range from -1 to 1. The additional parameter ß determines whether the mathematical model is the parallel or serial Wiener equation.

[0034] One design provides that the mathematical model is selected from:

[0035] Maxwell-Garnett model

[0036] Parallel Wiener equation

[0037] Serial Wiener equation

[0038] One embodiment provides that the input training data set continues to have an effective density p eff of the medium. One embodiment provides that the input training data set is created for temperatures from a temperature range of 1 to 100°C, in particular 10 to 60°C, and / or wherein the effective permittivity value £ effand the permittivity value are selected so that a solids content value 8 between 0 and 35% is obtained.

[0039] The microwave sensor according to the invention for determining a solid content 8 of a solid in a flowable, in particular aqueous, medium, with means for carrying out the method according to the invention comprises: a transmitting antenna for generating a microwave signal passing through the medium; a receiving antenna for receiving the microwave signal passing through the medium; and evaluation electronics configured to determine an effective permittivity of the medium as a function of a provided propagation time and / or absorption of the microwave signal, wherein the application-specific parameter set is stored interchangeably in the evaluation electronics.

[0040] The computer program according to the invention with instructions which, when the program is executed by a computer, cause the computer to execute the computer-implemented method according to the invention for determining a solids content in a flowable, in particular aqueous, medium, in particular by means of a microwave sensor and / or the method according to the invention for generating a training data set for a machine learning model for determining a solids content 8 of a solid in a flowable, in particular aqueous, medium, in particular for the machine learning model according to the invention.

[0041] The computer-readable medium according to the invention, which contains instructions which, when executed by a computer, cause the computer to execute the computer-implemented method according to the invention for determining a solids content in a flowable, in particular aqueous, medium, in particular by means of a microwave sensor and / or the method according to the invention for generating a training data set for a machine learning model for determining a solids content 8 of a solid in a flowable, in particular aqueous, medium, in particular for the machine learning model according to the invention.

[0042] The method according to the invention for configuring the microwave sensor according to the invention comprises the following method steps:

[0043] Receiving a parameter set at the microwave sensor; and replacing a parameter set of the machine learning model already provided, in particular at the factory, at the microwave sensor with the received parameter set, or

[0044] Implementing the received parameter set on the parameter set-free machine learning model implemented in the microwave sensor.

[0045] Determine the solid fraction in the medium using the machine learning model and the newly provided parameter set.

[0046] The invention is explained in more detail with reference to the following figures. They show:

[0047] Fig. 1 : a microwave sensor for determining a solid content 8 of a solid in a flowable, in particular aqueous, medium;

[0048] Fig. 2: a process flow of an embodiment of the method according to the invention for determining a solid content in a flowable, in particular aqueous, medium, in particular by means of a microwave sensor;

[0049] Fig. 3: a neural network of a method according to the invention according to a first embodiment in a schematic representation; and

[0050] Fig. 4: a method sequence of an embodiment of the method according to the invention for generating a training data set for a machine learning model according to the invention for determining a solid content 8 of a solid in a flowable, in particular aqueous, medium.

[0051] Fig. 1 shows a microwave sensor 100 for determining a solid content 8 of a solid in a flowable, particularly aqueous, medium, with means for carrying out the method according to the invention for determining a solid content in a flowable, particularly aqueous, medium. When using the microwave sensor 100, the medium is guided through a measuring tube 101 in a process line.

[0052] The illustrated microwave sensor 100 itself comprises a transmitting antenna 116 for generating a microwave signal that passes through the medium when a medium is present in the measuring tube 101. The transmitting antenna 116 is arranged in an opening of the measuring tube 101. A suitable transmitting antenna 116 is disclosed in patent DE 10 2020 134 320 A1, to which reference is made in its entirety.

[0053] Furthermore, the microwave sensor 100 comprises a receiving antenna 118 for receiving the microwave signal generated by the transmitting antenna 116 and—if a medium is present in the measuring tube 101—passing through the medium. A suitable receiving antenna 118 is also disclosed in patent DE 10 2020 134 320 A1, to which reference is made in its entirety. Alternatively, the microwave sensor 100 can have only one antenna, which is a receiving antenna 118 and simultaneously a transmitting antenna 116. In this case, the microwave signal is generated by the antenna and also measured. For this purpose, the microwave signal is reflected at least once from an inner surface of the measuring tube back toward the transmitting antenna.

[0054] The microwave sensor 100 also includes a temperature sensor 104, which is arranged in a specially provided opening in the measuring tube 101. The temperature sensor is configured to determine a current temperature value of the guided medium and provide it to an evaluation electronics unit 102. Alternatively, a temperature can be specified by the customer, or the current temperature value can be determined using an external temperature sensor and provided to the evaluation electronics unit.

[0055] The evaluation electronics 102 is provided, which is configured to determine an effective permittivity of the medium as a function of a provided propagation time and / or absorption of the microwave signal. The absorption results, for example, from the difference between a signal strength value of the received microwave signal and a signal strength value of the generated microwave signal. The receiving antenna 118 is configured to provide the received microwave signal to the evaluation electronics 102. The microwave signal generated and emitted by the transmitting antenna 116 is also provided to the evaluation electronics 102. The evaluation electronics 102 can be arranged in a microwave sensor housing (not shown), in which the transmitting antenna 116 and the receiving antenna 118 are also arranged. An application-specific parameter set 103, which is linked to a machine learning model, is stored in the evaluation electronics 102.This parameter set is stored interchangeably, meaning that the original or factory-stored parameter set 103 can be replaced by a different parameter set 103*. The evaluation electronics 102 is also configured to determine the effective permittivity value E. eff of the medium and a temperature value to a machine learning model 105 stored in the evaluation electronics 102. The machine learning model 105 has corresponding inputs for the provided effective permittivity value E eff and the provided temperature. Furthermore, the machine learning model 105 has an interchangeable, application-specific parameter set 103. This can be replaced, for example, by the customer with a parameter set 103* that differs from the parameter set 103 and is provided by the manufacturer. The parameter set 103 has at least one weighting factor w t and at least one bias b tThe machine learning model 105 is configured to determine a solids content value 8 of the solids in the medium and output this value. The output can be provided via a display unit (with a display).

[0056] The evaluation electronics 102 is capable of receiving a parameter set 103* and replacing a parameter set 103 of the machine learning model 105 already provided, in particular at the factory, on the microwave sensor 100 with the received parameter set 103*. Alternatively, whenever no parameter set 103 has yet been provided, the received parameter set 103* can be implemented on the parameter-free machine learning model 105 permanently implemented in the microwave sensor 100, in particular in the evaluation electronics 102. The received parameter set 103* can be provided via a cloud or handheld device (tablet, smartphone, laptop) connected to the microwave sensor 100. Furthermore, the evaluation electronics 102 is configured to determine the solids content 8 in the medium using the machine learning model 105 and the newly provided parameter set 103*.

[0057] Fig. 2 shows a process flow of an embodiment of the computer-implemented method according to the invention for determining a solids content in a flowable, particularly aqueous, medium, in particular using a microwave sensor. The method comprises the following process step:

[0058] Receiving an input data set with at least one effective permittivity value, in particular determined from measurement data of the medium, and a temperature value T of the medium.

[0059] In addition, the input data set may further contain a frequency value f for the microwave signal used and / or an effective density value p eff of the medium.

[0060] Furthermore, the procedure includes the following step:

[0061] Providing the input dataset to a machine learning model.

[0062] Machine learning is a general term for the "artificial" creation of knowledge from experience. An artificial system (machine learning model) learns from provided examples and cases and can generalize them after the learning phase is complete. To achieve this, machine learning algorithms build a statistical model based on the input data sets and test it against them. The machine learning model thus recognizes patterns and regularities in the input data sets.

[0063] The machine learning model has a replaceable, application-specific parameter set with at least one weighting factor w t and at least one bias b^. The input values ​​entering the respective neuron are weighted with the corresponding weighting factor w tweighted (e.g., by multiplication). All weighted input values ​​are then summed in the neuron. A bias b assigned to the neuron is then applied to the sum. t summed. The result is then passed through an activation function in the next step. The activation function can be, for example, a hyperbolic tangent function. Alternatively, the ReLu function can be used. The resulting output value is then provided to all neurons in the next intermediate layer or to one neuron in the output layer. The output value y t can be described by the following equation where gt(z) stands for the activation function and P for the number of input values ​​over which the summation is carried out.

[0064] The parameter set results from a mathematical model applied during a learning phase of the machine learning model to describe the properties of heterogeneous mixtures. In this case, these consist of an insoluble substance in a carrier substance (e.g., water).

[0065] The mathematical model is selected from one of the listed models / equations:

[0066] Maxwell-Garnett model

[0067] Parallel Wiener equation

[0068] Serial Wiener equation

[0069] The Maxwell-Garnett model describes the dielectric properties of bodies made of different materials with different dielectric properties. The model assumes the following relationship with a solids content value of 8, an effective permittivity value of the medium, the permittivity value of the solid and the permittivity value sm the carrier medium (e.g. water).

[0070] The following relationship applies to the Wiener equation where ß = -1 is assumed for the serial Wiener equation and ß = +1 for the parallel Wiener equation.

[0071] Furthermore, the method comprises the method step: determining a solid fraction value 5 of the solid in the medium by means of the machine learning model.

[0072] Furthermore, the procedure includes the following step:

[0073] Generate an output data set that includes the solids content value 8.

[0074] Optionally, the method may also include the following step:

[0075] Output the output data set, especially the solids content value 8.

[0076] Fig. 3 shows a neural network of a method according to the invention according to a first embodiment in a schematic representation. A suitable machine learning model would be, for example, MLPRegressor. However, other machine learning models are also suitable for the method according to the invention. The depicted machine learning model for determining a solids content in a flowable, in particular aqueous, medium, in particular for use in the computer-implemented method according to the invention, comprises an input layer ES with at least two inputs E1, E2. An effective permittivity value is passed via input E1. of the medium into the machine learning model, and a temperature value T of the medium is input to the machine learning model via input E2. The machine learning model also has an output layer AS with an output A1 for a solid fraction value 8 of the solid in the medium. Between the input layer and the output layer, there are at least two hidden intermediate layers, each with a number N t switched neurons.

[0077] Alternatively (not shown), the input layer ES can have at least one further input for a frequency value f and / or one further input for an effective density value p ff The input is or the inputs are connected to at least one of the neurons, in particular to all neurons, and represents or represents the corresponding density value p at the at least one neuron, in particular at all neurons. eff and / or the frequency value.

[0078] In the embodiment shown, all neurons in an intermediate layer ZS1, ZS2 are exclusively networked with neurons in a neighboring intermediate layer. This means that there are no further connections to neurons in a non-neighboring intermediate layer. Figure 3 shows 5 neurons (circle) per intermediate layer. However, the dotted lines indicate that more neurons can be provided. The connections are shown by the arrows. For reasons of clarity, only the connections between the first two neurons and the neurons in the second intermediate layer ZS2 have been shown. However, it goes without saying that the other neurons are also networked with one another. The embodiment shown has exactly two intermediate layers. Alternatively, the machine learning model can also have more than two, in particular exactly three, intermediate layers.The machine learning model is selected so that the total number of neurons in the three intermediate layers is between 40 and 200. To optimally utilize the required memory, it is advantageous if the number of neurons in the first intermediate layer is selected from the range of 20 to 80, the number of neurons in the second intermediate layer is selected from the range of 10 to 80, and the number of neurons in the third intermediate layer is selected from the range of 10 to 60.

[0079] In an alternative and equally advantageous embodiment, the machine learning model has exactly four intermediate layers with a total of 100 to 250 neurons. In this case, the first intermediate layer has a number of neurons selected from the range of 20 to 60, the second intermediate layer has a number of neurons selected from the range of 50 to 80, the third intermediate layer has a number of neurons selected from the range of 20 to 60, and the fourth intermediate layer has a number of neurons selected from the range of 10 to 50.

[0080] The design results in a machine learning model with a data size of less than or equal to 500 kB, with a deviation of less than 0.1% relative to the mathematical model used as the basis for the learning phase. If more than 500 kB of memory is available, up to 1500 neurons can be provided with two fixed inputs (effective permittivity and temperature).

[0081] The machine learning model shown is designed to weight input values ​​received in each neuron of a first intermediate layer of the at least two intermediate layers with a weighting factor, the weighted input values ​​and a bias b tto sum to a sum value, to convert the sum value into an output value via an activation function and to provide the output value to at least one neuron, in particular to all neurons, a second intermediate layer of the at least two intermediate layers or to the output layer AS. The weighting factors w t and the bias b t are part of the interchangeable, application-specific parameter set. The activation function can be a hyperbolic tangent function and / or a ReLU (rectified linear unit) function. Alternatively, the activation function can also be an exponential linear unit (ELU) or a sigmoid function.

[0082] Fig. 4 shows a method flow of an embodiment of the method according to the invention for generating a training data set for a machine learning model according to the invention for determining a solids content 8 of a solid in a flowable, in particular aqueous, medium. The method comprises the following method step: determining at least the effective permittivity value £ eff of the medium and the permittivity value of the solid using a random generator.

[0083] The random generator is subject to a probability distribution, such as the normal distribution. The randomly generated permittivity values ​​result in a data set that flows into the next process step:

[0084] Calculation of the solid content value 8 using a mathematical model depending on the effective permittivity value c e ^, the permittivity value « and the permittivity value s m of the carrier medium.

[0085] The mathematical model is selected from one of the listed models:

[0086] Maxwell-Garnett model

[0087] Parallel Wiener equation

[0088] Serial Wiener equation.

[0089] For the training dataset, for example, the Maxwell-Garnett model can be used. For this, the Maxwell-Garnett equation is rearranged so that the effective permittivity can be calculated as a function of the remaining parameters. Furthermore, the permittivity of the solid can be fixed. For example, a value between 2 and 5 can be assumed for the permittivity of the solid. The permittivity of the carrier medium (e.g., water) is temperature-dependent. The temperature is therefore included in the calculation of the permittivity of the carrier medium (e.g., using the Debye model). For the solid content, a value from a range that is reasonable for the application (e.g., 0-35% dry matter) is assumed.

[0090] The next step in the process includes:

[0091] Creating an input training dataset comprising an effective permittivity value c e ^, a permittivity value a permittivity value s m , a solids content value 8, a temperature T and an additional parameter ß, which is selected from a numerical range from -1 to 1.

[0092] For this process step, the Wiener equation is used as a model to describe the medium, therefore the additional parameter ß is used. Alternatively, the Maxwell-Garnett model can also be used.

[0093] Alternatively, the input training dataset can also contain an effective density p of the medium. The temperatures are preferably selected from a temperature range of 1 to 100°C, in particular 10 to 60°C. The effective permittivity value and the permittivity value £ fare selected such that a solids content value 8 between (and including) 0 and 50% results – as is to be expected, for example, in wastewater applications. However, the limit can also be higher. For example, applications are known in which a solids content of more than 50% is possible. In this case, the input training dataset would have solids content values ​​8 between 0 and the upper limit relevant for the application.

Claims

PATENT CLAIMS 1 . A computer-implemented method for determining a solids content in a flowable, in particular aqueous, medium, in particular by means of a microwave sensor, comprising: Receiving an input data set with at least one effective permittivity value s, in particular determined from measurement data eff of the medium, and a temperature value T of the medium; Providing the input data set to a machine learning model, wherein the machine learning model has an interchangeable, application-specific parameter set, wherein the parameter set includes at least one weighting factor and at least one bias b t has; Determining a solid fraction value ö of the solid in the medium using the machine learning model; Generating an output data set including the solids content value 5; Optional: Output the output data set, especially the solids content value S.

2. The method according to claim 1, wherein the input data set further comprises a frequency value and / or an effective density value p eff of the medium.

3. Method according to claim 1 or 2, wherein the parameter set results from a mathematical model applied in a learning phase of the machine model for describing the properties of heterogeneous mixtures of substances consisting of an insoluble substance which is located in a carrier substance. REVISED SHEET (RULE 91) ISA / EP 4. The method according to claim 3, wherein the mathematical model is selected from: Maxwell-Garnett model Parallel Wiener equation Serial Wiener equation.

5. Machine learning model for determining a solids content in a flowable, in particular aqueous, medium, in particular for use in the method according to one of claims 1 to 4, comprising: an input layer with at least two inputs for an effective permittivity value E of the medium and a temperature value T of the medium; an output layer with an output for a solids content value θ of the solid in the medium; at least two hidden intermediate layers located between the input layer and the output layer, each having a number of connected neurons, wherein the machine learning model is configured to weight incoming input values in each neuron of a first intermediate layer of the at least two intermediate layers with a weighting factor, the weighted input values and a bias to sum up to a sum value, to convert the sum value into an output value via an activation function and to provide the output value to at least one neuron, in particular to all neurons, of a second intermediate layer of the at least two intermediate layers, wherein the weighting factor w t and the bias b t form an interchangeable, application-specific parameter set.

6. Machine learning model according to claim 5, wherein a hyperbolic tangent function and / or a ReLU (rectified linear unit) function is used as the activation function. REVISED SHEET (RULE 91) ISA / EP 7. Machine learning model according to claim 5 or 6, wherein the input layer has at least one further input for a frequency value f and / or an effective density value p eff has.

8. Machine learning model according to one of claims 5 to 7, wherein the machine learning model has exactly three intermediate layers, wherein a total sum of the neurons of the three intermediate layers is selected from a range of 40 to 2000, in particular up to 1000 and preferably up to 200 neurons.

9. The machine learning model of claim 8, wherein a first intermediate layer has a number of neurons selected from the range 20 to 80, wherein a second intermediate layer has a number of neurons selected from the range 10 to 80, wherein a third intermediate layer has a number of neurons selected from the range 10 to 60.

10. Machine learning model according to one of claims 5 to 8, wherein the machine learning model has exactly four intermediate layers, wherein a total sum of the neurons of the four intermediate layers is selected from a range of 100 to 250 neurons.

11. Machine learning model according to claim 10, wherein a first intermediate layer has a number of neurons selected from the range 20 to 60, wherein a second intermediate layer has a number of neurons selected from the range 50 to 80, REVISED SHEET (RULE 91) ISA / EP wherein a third intermediate layer has a number of neurons selected from the range 20 to 60, wherein a fourth intermediate layer has a number of neurons selected from the range 10 to 50.

12. Machine learning model according to one of claims 5 to 11, wherein the machine learning model has a data size of less than or equal to 500 kB, with a deviation of less than 0.1% relative to the mathematical model serving as the basis for the learning phase.

13. Computer-implemented method for generating a training data set for a machine learning model for determining a solid content 8 of a solid in a flowable, in particular aqueous, medium, in particular for the machine learning model according to one of claims 5 to 12, comprising: Creating an input training dataset, where the input training dataset has an effective permittivity value s eff of the medium, a permittivity value s f of the solid, a permittivity value s m of the carrier medium, a solids content value 8 and a temperature ?, wherein at least the effective permittivity value s eff and the permittivity value are determined by a random generator, whereby the solid content value 8 is calculated using a mathematical model depending on the effective permittivity value s eff , the permittivity value and the permittivity value s m .

14. The computer-implemented method of claim 13, wherein the input training data set further comprises an additional parameter / ? selected from a number range of -1 to 1. REVISED SHEET (RULE 91) ISA / EP 15. A computer-implemented method according to 13 or 14, wherein the mathematical model is selected from: Maxwell-Garnett model Parallel Wiener equation Serial Wiener equation.

16. The computer-implemented method according to 13 to 15, wherein the input training data set further has an effective density p eff of the medium.

17. Computer-implemented method according to 13 to 16, wherein the input training data set is created for temperatures from a temperature range of 1 to 100°C, in particular 10 to 60°C, and / or wherein the effective permittivity value s eff and the permittivity value are selected so that a solids content value 5 between 0 and 35% is obtained.

18. A microwave sensor for determining a solid content 8 of a solid in a flowable, in particular aqueous, medium, with means for carrying out the method according to one of claims 1 to 4, comprising: a transmitting antenna for generating a microwave signal passing through the medium; a receiving antenna for receiving the microwave signal passing through the medium; evaluation electronics configured to determine an effective permittivity of the medium as a function of a provided propagation time and / or absorption of the microwave signal, wherein the application-specific parameter set is stored interchangeably in the evaluation electronics. REVISED SHEET (RULE 91) ISA / EP 19. A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method according to any one of claims 1 to 4 and / or 13 to 17.

20. A computer-readable medium containing instructions which, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 4 and / or 13 to 17.

21. A method for configuring a microwave sensor according to claim 18, comprising the steps of: Receiving a parameter set at the microwave sensor; and Replacing a parameter set of the machine learning model already provided, in particular at the factory, on the microwave sensor with the received parameter set, or Implementing the received parameter set on the parameter set-free machine learning model implemented in the microwave sensor. Determine the solid fraction in the medium using the machine learning model and the newly provided parameter set. REVISED SHEET (RULE 91) ISA / EP

Images