Boundary layer detection using a machine learning algorithm

Description

[0001] The invention relates to radar-based boundary layer detection based on machine learning algorithms.

[0002] In automation technology, particularly in process automation, field devices are frequently used to measure various process variables. These variables can include, for example, fill level, flow rate, pressure, temperature, pH value, redox potential, conductivity, or dielectric constant of a medium in a process plant. To acquire these measurements, field devices incorporate suitable sensors or are based on appropriate measurement principles. The Endress+Hauser Group manufactures and distributes a wide range of field device types.

[0003] Radar-based measurement methods have become established for measuring the fill level of contents in containers because they are robust and require little maintenance. A key advantage of radar-based measurement methods is their ability to measure the fill level virtually continuously and without contact. In the context of this patent application, the terms " radar " or " High frequency "On electromagnetic waves with frequencies between 0.03 GHz and 300 GHz. Common frequency bands used for level measurement are 2 GHz, 26 GHz, 79 GHz, or 120 GHz. The two common measurement principles are the pulse transit-time principle (also known as " Pulse radar "known) as well as the FMCW principle (“Frequency Modulated Continuous Wave”)Based on the pulse transit-time method and the FMCW method, level measuring devices can be implemented with comparatively low circuit complexity and high level resolution in the sub-millimeter range. Radar-based measuring principles are described in more detail, for example, in "Radar Level Detection, Peter Devine, 2000".

[0004] In addition to free-radiating radar measurement, where high-frequency signals are transmitted and received via an antenna, there is also the guided radar variant. Here, instead of an antenna, an electrically conductive measuring probe (e.g., a coaxial waveguide or a metal rod) is used, which is lowered into the container to guide the high-frequency signals. Similar to free-radiating radar, the high-frequency signal is reflected in the measuring probe at the level of the contents and guided back along the probe to the level gauge. This variant of radar-based level measurement is also known as... "TDR" ( Time Domain Reflectometry (). An advantage of this variant is that, due to the guided signal transmission, less power is required to operate the level gauge. A TDR-based level gauge is described, for example, in US patent 10,07,743 B2.

[0005] Regardless of the measurement principle (pulse transit time, FMCW, TDR), the fill level is determined by calculating the signal maximum resulting from the surface of the contents and its corresponding distance value from the reflected received signal. The various radar-based measurement principles are generally described, for example, in " Radar Level Detection, Peter Devine, 2000 " .

[0006] Depending on the type of material whose fill level is to be determined, it may be permeated with impurities, additives, or foreign substances, which may form an additive layer above or below the actual material, such as foam, sediment, or an oily or aqueous phase. In this case, a more or less pronounced boundary layer forms between the actual material and the additive layer, often also referred to as a " Separating layer"Understood, out. Accordingly, the level measurement can be disrupted or distorted by any additional layers of fill material, as the corresponding signal maximum of the fill material surface cannot be reliably identified. For this reason, it is of interest to be able to reliably detect not only the fill level but also the presence or even the thickness of an additive fill material layer. A TDR-based level measuring device, which determines any additive fill material layer by comparing the amplitude with a theoretical target value, is described in European patent application 2 722 655 A1. However, detection by identifying the corresponding signal maximum is also at least uncertain in this case, since the additive fill material layer or its surface may only reflect the corresponding signal very weakly."

[0007] US 2012 / 299768 A1 discloses a level gauge for determining the position of the level of a substance in a container and / or a separating layer between two substances. The level gauge comprises an echo curve detection device for detecting one or more echo curves, an echo identification device for evaluating the at least one echo curve, a multiple echo detection device for classifying one or more echoes of a multiple reflection at a substance surface and / or at the bottom of the container as multiple echoes, and a self-learning device designed to automatically determine the length of the container's manhole using the multiple echoes classified by the multiple echo detection device.

[0008] The invention is therefore based on the objective of being able to reliably determine any boundary layer.

[0009] The invention solves this problem by means of a corresponding measuring system for boundary layer detection on a product located in a container. For this purpose, the measuring system comprises the following components: A radar-based measuring device comprising: a measuring probe extending into the container, by means of which high-frequency signals are sent towards the contents and, after reflection at the surface of the contents, are received as receiving signals; a signal generation unit designed to generate the high-frequency signal to be sent according to the TDR method; a receiving unit designed to record the received signal; and an evaluation unit in which, according to the invention, a machine learning algorithm is implemented to detect the mass or volume fraction of the contents in the boundary layer based on the received signal, by determining a distribution of the attenuation coefficient and / or the dielectric value in the container along the measuring probe.

[0010] Under the term " Unit"Within the scope of the invention, any electronic circuit suitable for the intended purpose is understood to be an electronic circuit. Depending on the requirements, this could be an analog circuit for generating or processing corresponding analog signals. However, it could also be a digital circuit such as a microcontroller or a storage medium in conjunction with a program. The program is designed to execute the corresponding process steps or to apply the necessary arithmetic operations of the respective unit. In this context, various electronic units of the measurement system within the meaning of the invention can potentially also access a common physical memory or be physically operated by means of the same digital circuit."

[0011] The inventive implementation of a machine learning algorithm circumvents the problem that the boundary layer or the additive filler layer in the received signal often does not generate a clearly identifiable signal maximum, which classical distance measurement methods according to the prior art, such as TDR, pulse transit time, or FMCW methods, require for its detection. Instead, received signals recorded under complex measurement conditions, such as multiple reflections, can be interpreted much better using machine learning algorithms.

[0012] In addition to the detection of the boundary layer according to the invention, it is also conceivable that the evaluation unit or the receiving unit of the measuring device is also designed to additionally determine the fill level of the contents in the container based on classical distance measurement methods using the received signal.

[0013] Within the scope of the invention, the evaluation unit can in particular be designed to use the machine learning algorithm to analyze the received signal. The system aims to determine the vertical thickness and / or vertical position of the boundary layer relative to its height above the container brine, the mass or volume fraction of the fill material in the boundary layer as a function of its height above the container brine, and / or a dielectric constant value relative to its height above the container brine. In this case, the mass or volume fraction of the fill material in the boundary layer is detected, for example, by the evaluation unit using a machine learning algorithm to determine the distribution of the damping coefficient, conductivity, and / or dielectric constant within the container along the measuring probe.

[0014] As " machine learning algorithmWithin the scope of the invention, "a machine learning algorithm" is defined in principle as any universal mathematical function approximator that maps at least one of its input variables to at least one of its output variables. Internal parameters of the corresponding mapping functions are adjusted during a training phase using known training data. In particular, supervised learning is implemented as the algorithmic approach. In this context, the invention does not specify the concrete form in which the machine learning algorithm is implemented. For example, the machine learning algorithm can be implemented in the form of " Decision Trees ", " Support Vector Machine ", " Naive Bayes ", or " k Nearest Neighbour"can be implemented. However, the boundary layer can be detected particularly effectively if the machine learning algorithm is based on a non-symbolic approach, such as an artificial neural network, especially in the form of a deep learning method (better known as "deep learning"). Machine learning algorithms are described in more detail, for example, in " Introduction to Artificial Intelligence " (Wolfgang Ertel, 2017).

[0015] Furthermore, it is not essential within the scope of the invention how or where the evaluation unit is implemented. For example, the evaluation unit can be designed as an integral part of the measuring device, or as a component of a higher-level network such as a cloud or a server, or as a component of a plant-specific process control system.

[0016] The measuring device is designed based on the TDR method, so that the transmission unit is implemented as a measuring probe extending into the container. Correspondingly, the signal generation unit generates the transmitted high-frequency signal, in this case pulse-wise, according to the TDR method. An advantage of the TDR method in relation to the invention is that the received signal has a higher overall signal amplitude compared to free-radiating radar, which in principle makes it easier for the machine learning algorithm to identify the boundary layer.

[0017] Corresponding to the measuring system according to the invention, the problem underlying the invention is also solved by a corresponding measuring method for detecting a boundary layer of a fill material in a container using the measuring system according to one of the previously described embodiments. The method comprises the following process steps: Emitting a high-frequency signal via the measuring probe into the container, recording the received signal after reflection of the high-frequency signal there, and evaluating the received signal using the machine learning algorithm, so that the mass or volume fraction of the filling material in the boundary layer is detected by the machine learning algorithm determining a distribution of the damping coefficient and / or the dielectric value in the container along the measuring probe.

[0018] In the training phase, which is required before the actual measurement operation, the machine learning algorithm can be trained, for example, using experimentally obtained and / or simulation-generated received signals, such as those obtained through " CST Microwave Studio "to be trained".

[0019] The invention is explained in more detail using the following figures. They show: Fig. 1 : A TDR-based measuring device for boundary layer detection according to the invention in a container, and Fig. 2 : a method for determining the boundary layer according to the invention.

[0020] To understand the boundary layer detection according to the invention, it is necessary to refer to the following: Fig. 1 A TDR-based measuring device 1 is shown, which can be used to determine the fill level L of a fill material 2, 2' located in a container 3.

[0021] According to the TDR method, the measuring device 1 comprises a measuring probe 11 as a transmission unit 11 for high-frequency signals S HF , R HF to the filling material 2, 2' or after reflection in the container 3. Not according to the invention, contrary to the in Fig. 1 In the illustrated embodiment, the measuring device 1, in the case of free-radiating radar, comprises at least one antenna as the transmission unit 11. To determine the fill level L, the measuring probe 11 extends inside the container from the measuring device 1 at the top of the container 3 to virtually the bottom of the container 3. The installation height h of the measuring device 1 above the bottom of the container is known and stored in the level measuring device 1 or in an external evaluation unit 4, such as a decentralized server or a central process control system.

[0022] The high-frequency signal SHF to be emitted is directed via the measuring probe 11, either as a pulse or with a frequency ramp, towards the material 2, 2' according to the TDR method. Due to the change in the dielectric constant DK at the surface of the material 2, 2', the emitted high-frequency signal SHF is subsequently reflected at the level of the material surface by the measuring probe 11 and, after a corresponding signal propagation time t, is received in the measuring device 1 as a received signal RHF. The signal propagation time t of the signal SHF, RHF depends according to... t = 2 d c from the distance d d = h − L the container top to the surface of the contents. Here, c is the propagation speed of the high-frequency signal S HF , R HF along the measuring probe 11, which is in the range of the speed of light c.

[0023] To generate the high-frequency signal SHF, the level measuring device 1 includes a correspondingly designed signal generation unit: In the case of the TDR method, this unit can, for example, be based on a capacitor that is discharged to generate the pulse lasting approximately 100 ps to 1 ns. In the case of free-radiating radar according to the pulse-time-of-flight or FMCW method (not according to the invention), the signal generation unit can, for example, include a frequency-controlled high-frequency resonant circuit or a crystal oscillator. To enable the signal generation unit to generate the high-frequency signal SHF in the required clock rate, either in a pulsed or ramped manner, the capacitor or crystal oscillator is driven accordingly by clocking or modulation.

[0024] A transmit / receive switch is interposed between the signal generation unit of the measuring device 1 and the measuring probe 11 to feed the received signal R HF, after reflection in the container 3, to a receiving unit in which the received signal R HF is digitized or recorded. The design of the transmit / receive switch is not, in principle, fixed. In the case of the TDR method, as used in the Fig. 1 In the illustrated embodiment, the transmit / receive switch can, for example, be designed as a purely electrical node. Particularly in the case of non-inventive, freely radiating radar, the transmit / receive switch can, for example, be implemented as a duplexer.

[0025] To determine the fill level L, the received signal R HF can be recorded in the receiving unit, for example, by undersampling the received signal R HF according to the pulse-time delay principle, so that the received signal R HF is stretched in time by a defined factor. The time-stretching factor depends on the sampling rate. For this to achieve sufficient time stretching, the corresponding sampling rate must be chosen such that it differs from the clock rate of the outgoing signal pulses S HF only in the per mille range. From a circuit design perspective, this time stretching simplifies the determination of the fill level L based on the received signal R HF. In contrast to the method described in Fig. 1In the embodiment shown, the received signal R HF is rectified in addition to time dilation in the case of non-inventive, freely radiating radar, so that the recorded received signal R HF has only one polarity - plus or minus - with respect to a fixed reference potential.

[0026] The determination of the fill level L using the possibly time-dilated received signal R HF is shown on the left curve in Fig. 2 To illustrate this in more detail: The curve shown there depicts the temporal amplitude profile of the potentially time-stretched received signal R HF. The distance d between the top of the container and the surface of the contents is proportional to the signal propagation time axis t shown in the diagram. Fig. 2The curves shown. Ideally, i.e., without any external interference, the received signal R HF includes at least one signal maximum, which is caused at the surface of the material 2, 2'. The fill level L is then determined based on the propagation time t corresponding to the signal maximum and the formulas mentioned above.

[0027] As in Fig. 1 As depicted, an additive layer 2' of the fill material, such as an aqueous phase of an otherwise oil-based fill material 2, is formed above the fill material 2. The additive layer 2' is vertically separated from the fill material 2 by a boundary layer in which both components, i.e., oil and water, are present as an emulsion in a defined mass or volume fraction %.

[0028] The exemplary embodiment in Fig. 1 or Fig. 2This clarifies that the signal maximum of the received signal RHF in this case is not caused by the actual fill material 2, but by the additive fill material layer 2'. At the level of the boundary layer or below, the received signal RHF also exhibits an increased signal amplitude due to reflection in the boundary layer or at the surface of the actual fill material 2; however, a specific signal maximum cannot be determined from this in case of doubt. Therefore, regardless of whether it is based on the pulse transit time, FMCW, or TDR principle, the measuring device 1 can only determine the cumulative fill level of the fill material 2 and the additive fill material layer 2'. Depending on the application or process plant, it is essential to know the fill level L of the pure fill material 2, as shown in Fig. 1This is shown. In this context, it may also be necessary to know the presence or thickness of the additive fill layer 2' as a parameter. The thickness of the boundary layer or the predominant mass or volume fraction % of the fill material 2 or the additive 2' there may also be relevant parameters. The mass or volume fraction % should be determined with as much spatial resolution as possible, i.e., as a function of the height h above the container base.

[0029] Thus, the term " Boundary layer detection "within the scope of the invention, the determination of one or more of these previously presented parameters.

[0030] According to the invention, parameters related to the boundary layer can be determined by applying a machine learning algorithm (MLA) to the potentially time-dilated received signal RHF. The specific MLA algorithm used is not predetermined. However, artificial neural networks, especially deep learning, prove to be particularly effective in this regard. This inventive concept circumvents the problem that the boundary layer or the additive filler layer often does not generate a clearly identifiable signal maximum in the received signal, which, according to the prior art, would be necessary for its detection.

[0031] Since the application of the machine learning algorithm MLA may require high computing power, it is advantageous if the MLA algorithm is not implemented in the measuring device 1 itself, but in the external evaluation unit 4, as the measuring device 1 in the process plant may be subject to a limited power supply. The measuring device 1, or its receiving unit, can then receive the recorded received signal R HF via a suitable interface, such as... PROFIBUS ", " HARD ", " Wireless HART ", " 4 - 20mA ", " Bluetooth " or " Ethernet "Transferred to the evaluation unit 4. Thus, the evaluation unit 4 together with the measuring device 1 forms a corresponding measuring system for boundary layer detection according to the invention in the container 3. If sufficient computing power is also available in the measuring device 1, the machine learning algorithm MLA for boundary layer detection can also be implemented in the measuring device 4 itself."

[0032] The application of machine learning algorithms (MLA) requires training the measuring system according to the invention under known conditions. Therefore, before regular measurement operation, the measuring device 1 must be subjected to various training situations with known parameters, such as the fill level L of the fill material 2, the thickness of the boundary layer, the thickness of the additive fill material layer, a known fill material type or its dielectric value DK, the dielectric value DK of the additive fill material layer to suspend. The underlying received signals R can be affected. HF not only experimentally, but also through simulation, such as via " CST Microwave Studio " . This avoids a time-consuming training phase in the process plant.

[0033] Fig. 2This illustrates an advantageous way in which the measuring system according to the invention, after completion of the learning phase, can detect the boundary layer or the aforementioned parameters in regular measurement operation based on the received signal R HF and the machine learning algorithm MLA: Using the machine learning algorithm MLA, the dielectric constant DK in the container 3 is first determined from the signal propagation time t dependent on the received signal R HF as a function of the signal propagation time t or as a function of the height h, as shown by the middle curve in Fig. 2 This represents the following. Based on this, the individual parameters can be deduced in evaluation unit 4 using a correspondingly programmed logic or evaluation function f: The constant dielectric constant DK in the upper transit time region must be assigned to the atmosphere above the fill material or above the additive fill material layer 2'. If the dielectric constant DK changes over a defined range of the signal transit time t, the boundary layer must be present in this range. The constant dielectric constant DK below the range belonging to the boundary layer must be assigned to the fill material 2, whereby the dielectric constant DK allows conclusions to be drawn about the fill material type (a dielectric constant in the range of 80 can be assigned to a 100% aqueous medium; see also the right-hand curve in [reference]). Fig. 2The constant dielectric constant DK above the signal propagation time region belonging to the boundary layer must be assigned to the additive fill layer 2', whereby the corresponding dielectric constant DK in turn allows conclusions to be drawn about the type of additive fill layer 2' (a dielectric constant in the range of two can, for example, be assigned to a 100% oil-containing medium). The fill level L can be determined from the position of the boundary layer. The thickness of the additive fill layer 2' can be determined from the extent of the signal propagation time region assigned to the additive fill layer 2'. Depending on how the dielectric constant DK changes within the boundary layer, the mass / volume fraction % of the fill material 2 or the additive fill material 2' within the boundary layer also changes, as shown from the middle and right curves in Fig. 2It follows that since the dielectric constant DK increases continuously with increasing signal propagation time t or with decreasing height h, the proportion % of the fill material 2 must increase continuously (from 0% to 100%) with decreasing height h. Conversely, the proportion % of the additive fill material 2' must also decrease continuously (from 100% to 0%) with increasing distance d.

[0034] Instead of determining the respective percentage indirectly via the dielectric constant, it is also conceivable within the scope of the invention to determine the percentage directly from the received signal R HF using the machine learning algorithm MLA, as shown in Fig. 2 is also illustrated.

[0035] In the one based on Fig. 1 and Fig. 2 In the illustrated embodiment, only an additive layer of filler material 2' is formed alongside the actual filler material 2. It goes without saying that the inventive idea can also be applied when several additive layers of filler material 2' are present. Reference symbol list

[0036] 1 Level gauge 2 Filling material 2'Additive filling layer 3 Container 4 Evaluation unit DK Dielectric constant d Distance f Logic function h Height above container bottom L Fill level MLA Machine learning algorithm R RF received signal S RF high-frequency signal t Signal propagation time % Mass or volume fraction of the filling material

Claims

1. Measuring system for boundary layer detection on a filling material (2, 2') in a container (3), comprising: - A radar-based measuring device (1), with ∘ a measuring probe (11) extending into the container (3), by means of which high-frequency signals (SHF) can be transmitted to the filling material (2, 2') and, after reflection at the filling material surface, can be received as reception signals (RHF), ∘ a signal generation unit designed to generate the high-frequency signal (SHF) to be transmitted in accordance with the TDR method, ∘ a receiving unit designed to record the received signal (RHF), and - an evaluation unit (4) in which a machine learning algorithm (MLA) is designed to detect the mass or volume fraction (%) of the filling material (2, 2') in the boundary layer on the basis of the received signal (RHF), by the machine learning algorithm (MLA) determining a distribution of the attenuation coefficient and / or the dielectric constant (DK) in the container (3) along the measuring probe (11).

2. Measuring system according to claim 1, wherein the evaluation unit (4) is designed to use the machine learning algorithm (MLA) - a thickness and / or a vertical position of the boundary layer (2') in relation to the height (h) above the brine in the container (3), and / or - the mass or volume fraction (%) of the filling material (2, 2') in the boundary layer as a function of the height (h) above the brine in the container (3), and / or - a dielectric constant (DK) in relation to the height (h) above the brine in the container (3),3. Measuring system according to one of the preceding claims, wherein the machine learning algorithm (MLA) is designed as an artificial neural network, in particular in the form of a deep learning method.

4. Measuring system according to one of the preceding claims, wherein the evaluation unit (4) is designed to determine the fill level (L) of the filling material (2) in the container (3) based on the received signal (RHF).

5. Measuring system according to one of the preceding claims, wherein the evaluation unit (4) is designed as an integral part of the measuring device (1).

6. Measuring system according to one of claims 1 to 5, wherein the evaluation unit (4) is designed as part of a higher-level network.

7. Method for detecting a boundary layer (2') of a filling material (2) in a container (3) using the measuring system according to one of the preceding claims, comprising the following method steps: - Transmitting a high-frequency signal (SHF) via the measuring probe (11) into the container (3), - recording the received signal (RHF) after reflection of the high-frequency signal (SHF), - evaluating the received signal (RHF) using machine learning algorithm (MLA) so that the mass or volume fraction (%) of the filling material (2, 2') in the boundary layer is detected by the machine learning algorithm (MLA) determining a distribution of the attenuation coefficient and / or the dielectric constant (DK) in the container (3) along the measuring probe (11).

8. Method according to claim 7, wherein the machine learning algorithm (MLA) is trained using experimentally obtained and / or simulation-generated reception signals (RHF).