Fluid parameter detection method using flow sensor configuration

Abstract

The present invention relates to a method for detecting fluid (12) parameters (3) using a sensor configuration (2) of a flow meter (1), comprising the steps of: a) flowing a training fluid (4) having known fluid parameters through the flow meter and supplying a training measurement signal (5) from the sensor configuration to a machine learning model (6); b) training the machine learning model and detecting real-time fluid (12) parameters (8) using a real-time measurement signal (7) from the sensor configuration; and c) detecting real-time fluid parameters using the trained machine learning model and the real-time measurement signal supplied to the trained machine learning model, wherein the step of supplying the real-time measurement signal from the sensor configuration to the machine learning model includes processing the real-time measurement signal (9) by performing feature extraction or feature learning (10) and supplying the processed real-time measurement signal (19) to the trained machine learning model. Includes.

Description

[Technical Field]

[0001] The present invention relates to a method for detecting one or more fluid parameters using a flow meter sensor configuration, and to a flow meter for performing such a method. [Background technology]

[0002] Accurately detecting the flow rate, pressure, and other physical quantities of liquids or gases is crucial in a wide range of applications with significant social, environmental, and economic impacts. Examples include flow control in vaccine manufacturing process lines (including the sophisticated mixing / formulation process of vaccines and additives) or precise film deposition gas control for semiconductor manufacturing. More specific applications also require accurate fluid sensors, such as mechanical ventilation and precise drug administration to patients in hospitals. While mass flowmeters are the preferred type of flowmeter in many applications, modern flowmeters often offer a variety of additional features.

[0003] Currently, all of the above applications depend on prior knowledge of the system and the physical relationships between all parameters (e.g., the composition of the mixture). For example, the viscosity of natural gas is an indicator of its calorific value. Viscosity can be calculated from Hagen-Poiseuille's equation using a given flow rate and pressure drop. However, in the case of turbulence, Hagen-Poiseuille's equation is not applicable, and determining viscosity becomes complicated. Another problem is that the presence of high concentrations of N2 or CO2 in the gas mixture (as in the case of biogas) reduces the accuracy of the relationship between viscosity and calorific value. Furthermore, the possibility of adding hydrogen gas to the gas network also affects the relationship between the viscosity and calorific value of the gas mixture. Finding a more appropriate physical analysis model may be a solution, but the model needs to be established for each individual situation, and this system is time-consuming, lacks flexibility, and is potentially inaccurate under conditions different from those under which the model was derived.

[0004] In the field of flow measurement, there is a need for multi-parameter systems and methods that can extract information other than flow rate from fluid mixtures. By utilizing microfabrication technology, particularly MEMS (Micro-Electro-Mechanical Systems), multiple sensors such as flow rate, pressure, and density can be efficiently integrated onto a single chip. For example, when calculating viscosity from flow rate and pressure, conventional data processing methods require filtering of raw sensor signals, calibration of individual sensors, and physical modeling.

[0005] However, these methods are time-consuming, not always applicable, and can leave potentially important information undiscovered.

[0006] Therefore, an object of the present invention is to provide a time-efficient, more widely applicable method and flow meter for detecting fluid parameters using multiple sensor configurations (i.e., combinations of multiple sensing structures), which enables the discovery of potentially relevant information. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] U.S. Patent Application Publication No. 2021 / 010839 [Patent Document 2] U.S. Patent Application Publication No. 2020 / 166398 [Patent Document 3] U.S. Patent Application Publication No. 2020 / 124461 [Patent Document 4] U.S. Patent Application Publication No. 2020 / 355073 [Patent Document 5] U.S. Patent Application Publication No. 2022 / 244157 [Patent Document 6] U.S. Patent Application Publication No. 2022 / 128388 [Patent Document 7] Chinese Patent Application Publication No. 101900589 [Patent Document 8] Chinese Patent Application Publication No. 115355959 [Overview of the project] [Means for solving the problem]

[0008] The present invention provides a method for detecting one or more fluid parameters using a flow meter sensor configuration, comprising the following steps: a) flowing one or more training fluids having one or more known fluid parameters through the flow meter and supplying one or more training measurement signals from the sensor configuration to a machine learning model; b) training the machine learning model using the training measurement signals and detecting real-time fluid parameters using real-time measurement signals from the sensor configuration; and c) detecting real-time fluid parameters using the trained machine learning model and real-time measurement signals supplied to the trained machine learning model from the sensor configuration, the step of supplying one or more real-time measurement signals from the sensor configuration to the machine learning model comprising processing the one or more real-time measurement signals by performing feature extraction or feature learning, and supplying the processed one or more real-time measurement signals to the trained machine learning model.

[0009] The methods described above combine conventional physical laws with machine learning (e.g., modern feature learning, preferably deep learning techniques in particular) to enable effective processing of raw sensor data, verification of physical constraints, understanding of complex physical effects within (microfabricated) fluid channels, and improvement of future chip designs based on existing and novel (hidden) causal relationships discovered through the interpretability of AI models.

[0010] In addition to processing one or more real-time measurement signals by performing feature extraction or feature learning on the aforementioned signals, this method may also include additional preprocessing of (a portion of) the real-time measurement signals.

[0011] In addition to processing one or more real-time measurement signals by performing feature extraction or feature learning on the aforementioned signals, this method may also include representation learning or feature selection.

[0012] The aforementioned method employs a flow meter with multiple detection structures and a trained neural network, outperforming state-of-the-art multi-parameter systems in many application areas, such as real-time quality control of products in chemical or pharmaceutical microreactors or the food industry.

[0013] By incorporating human knowledge about physical quantities into machine learning (such as deep neural networks), this method can learn more quickly and effectively how to utilize sensing structures on a chip to obtain physically meaningful output signals.

[0014] Deep neural networks may include regression methods. Preferred simple regression methods include linear regression (LR), support vector regression (SVR), and Gaussian process regression (GPR), but these methods should not be used with unknown data (i.e., unknown conditions and sensors). More advanced regression models are generally suitable for unknown data. These methods are, for example, “Short-term temperature forecasts using a convolutional neural network-An application to different weather stations in Germany” by D. Kreuzer et al.,in “Machine Learning with Applications”,Volume 2,15 December 2020, Article 100007, “SomBe:Self-Organizing Map for Unstructured and Non-CoCoCo iBeacon Constellations” by Duc.V.Le,et al.,in “2018 IEEE International Conference on Pervasive Computing and Communications (PerCom)”,19-23 March 2018,ISBN:978-1-5386-3224-6,or “Learning the world from its words:Anchor-agnostic Transformers for Fingerprint-based Indoor Localization”,by SMNguyen,et al.,in “2023 IEEE International Conference on Pervasive Computing and Communications Communications(PerCom)”,13-17 March This is explained in 2023, ISBN: 978-1-6654-5378-3. The latter two papers are regression models for location prediction and can also be applied to flow rate prediction.

[0015] As an example of a neural network that can be used in the present invention, a convolutional neural network (CNN), which is a deep learning network architecture that learns directly from data, can be mentioned. CNN is particularly useful for discovering patterns in data. As another example, a recurrent neural network (RNN) can be mentioned. This is a type of artificial neural network where the connections between nodes may form a cycle, and thus the output from some nodes may potentially affect the input to the same nodes later. As yet another example, there is "long short-term memory (LSTM)", which is a neural network with feedback connections, different from a standard feedforward neural network. CNN, RNN, and LSTM are all more advanced regression models and are generally suitable for unknown data.

[0016] Regression CNN is preferred for CNN, but other CNN-based models are also included.

[0017] Further examples of deep neural networks include those that include regression methods, and models based on attention mechanisms (e.g., transformers), etc. Algorithms based on statistics such as autoregressive integrated moving average (ARIMA) can also be used.

[0018] By combining machine learning with (microfabricated) multi-parameter systems, the following additional advantages can be obtained: Improved versatility: Fluids with difficult-to-identify, non-linear, or physically non-ideal parameters can also be classified. Speeding up: For new types of fluids, in-depth understanding of the fluid or reprogramming of the detection software is not required, and only re-learning of the model is necessary. More robust and application-oriented: When sufficiently learned in a variety of situations, robust solutions can be obtained for exceptional situations.

[0019] In the context of this patent application, "deep symbolic learning" refers to all methods of AI based on high-order symbolic (human-readable) representations of problems, logic, and search.

[0020] "Fluid type" refers to the identity / identification of a fluid or fluid mixture, such as water or ethanol (or a mixture thereof).

[0021] "Fluid parameters" in a broad sense refer to the properties of a fluid, including not only material-specific parameters such as density or viscosity, but also non-material-specific parameters such as fluid concentration. Viscosity includes both kinematic viscosity and kinematic viscosity. "Real-time fluid" refers to the fluid flowing through a channel and being measured.

[0022] The real-time fluid can be a liquid, gas, vapor, or plasma, and may include various mixtures (e.g., nitrogen, oxygen, argon, carbon dioxide, trace gases, and optionally air containing water).

[0023] U.S. Patent Application Publication 2021 / 010839 discloses a server-side flow measurement method using machine learning (more specifically, multivariate machine learning) that processes ultrasonic flow meter data in multiple stages. However, U.S. Patent Application Publication 2021 / 010839 employs only a single sensor type (i.e., acoustic) and is limited to pipelines (i.e., flow channels with relatively large diameters), making it unsuitable for, for example, microfabricated channels.

[0024] U.S. Patent Application Publication 2020 / 166398 discloses a flow measurement system comprising a flow meter connected to multiple sensors and configured to measure the volume of fluid passing through the flow meter. The system also includes a measuring computer connected to the flow meter and sensors. The measuring computer is configured to receive real-time values ​​from the multiple sensors during a first time period, to train an artificial intelligence engine based on the real-time values ​​received during the first time period, and to detect sensor failures based on the deviation between the real-time values ​​from the sensors and the sensor's predicted values. However, U.S. Patent Application Publication 2020 / 166398 appears to be limited to measurement-related applications.

[0025] U.S. Patent Application Publication No. 2020 / 124461 further discloses a fairly general method for determining at least one process variable of a medium. This method includes the steps of recording a first value of a process variable by a first method for determining a process variable, and recording a second value of a process variable by a second method for determining a process variable. The method also includes the steps of selecting at least one of the detected values ​​of a process variable using a classifier, and outputting the value of the selected process variable.

[0026] U.S. Patent Application Publication 2020 / 355073 discloses a damped total reflection optical sensor for acquiring downhole fluid properties. Therefore, this sensor is limited to simple optical measurements and is unsuitable for microsystem applications.

[0027] U.S. Patent Application Publication No. 2022 / 244157 discloses a method for determining and identifying anomalies in a fork meter. This method employs only a "mechanical method" and is unsuitable for application to microsystems.

[0028] U.S. Patent Application Publication No. 2022 / 128388 discloses a water volume measurement system, but this also appears unsuitable for application to microsystems.

[0029] Chinese Patent Application Publication No. 101 900 589 discloses a method for measuring the flow rate of an air-entrained liquid based on a mass flow meter. However, Chinese Patent Application Publication No. 101 900 589 appears to be related only to two-phase flow.

[0030] Chinese Patent Application Publication No. 115 355 959 discloses a method and system for measuring gas-liquid two-phase flow based on machine learning and physical constraints. Chinese Patent Application Publication No. 115 355 959 also appears to be related only to two-phase flow.

[0031] One embodiment relates to the method described above, wherein the machine learning model employs deep symbolic learning. The applicant expects that, for example, the application of deep symbolic learning to microfluidic sensor data will bring about groundbreaking advancements in the design and use of state-of-the-art multi-parameter sensing systems. By combining conventional physical laws with modern deep learning techniques, deep symbolic learning can enable effective processing of raw sensor data, verification of physical constraints, understanding of complex physical effects such as microfabricated fluid channels, and improvement of future chip designs based on existing and novel (hidden) causal relationships discovered through the interpretability of AI models. Thus, deep symbolic learning is a framework that allows perceptual and symbolic functions to be learned in parallel (i.e., simultaneously, not sequentially). For further details, see also “Deep Symbolic Learning: Discovering Symbols and Rules from Perceptions” by A. Daniele et al., in “Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence”, Macao, SAR, 19-25 August 2023. Using modern data processing techniques such as Deep Symbolic AI provides a unique way to analyze the physical properties of sensors and the information they provide from both scientific and technical perspectives. This also applies to the emerging field of AI. Currently, many machine learning algorithms focus on object recognition from images, language processing, and interpretation of sensor network results. Noise reduction and performance improvements for single MEMS devices have also been implemented. However, these are all commercially available, calibrated, off-the-shelf (digital) sensors. The physical properties of the proposed microfluidic sensor generate data in a completely different (time-dependent) format. The applicant argues that only new insights in machine learning can lead to algorithms (e.g., end-to-end learning) that elucidate the unknown correlation between signals and fluid properties. This may result in new symbolic functions or improved interpretability (symbolic inference) along with competing accuracy.In particular, in fields where the applicant attempts to control flow using other parameters such as pressure or viscosity (such as in gas chromatography), the applicant expects to discover previously undiscovered correlations and understand learned policies. Furthermore, the applicant argues that the opportunity to design and manufacture sensing structures and optimize them for machine learning is unique.

[0032] One embodiment relates to the method described above, in which a machine learning model utilizes, analyzes, and explores time-series information or time-series data (e.g., continuous data). Sensors obviously provide this information, and exploring time-series information can improve classification accuracy in at least some situations.

[0033] In one embodiment, relating to the method described above, the machine learning model identifies a time range that is particularly important for training the machine learning model. For example, the machine learning model may select a relevant time range that accounts for about 50% of the total operating time, less than 50% of the total operating time, less than 40% of the total operating time, less than 30% of the total operating time, less than 20% of the total operating time, or less than 10% of the total operating time.

[0034] One embodiment, relating to the method described above, includes retraining the machine learning model, i.e., continuous improvement of the machine learning model. For example, this may occur when a new fluid is introduced or when there is a drifting population.

[0035] (Therefore) one embodiment relating to the method described above is one in which retraining is a continuous process.

[0036] One embodiment relates to the method described above, wherein one or more training measurement signals include raw data, uncalibrated measurement signals, i.e., raw data including noise, distortion, and non-ideal effects, and unfiltered (electrical) measurement signals. Such measurement data can be used to better understand the complex physical effects within a fluid channel (e.g., a microfabricated one).

[0037] One embodiment relates to the method described above, wherein the flow meter is optimized for a specific fluid class, and a training fluid is selected from that class for model training. For example, if the flow meter is optimized for gas measurement, the measured fluid will be a gas or a mixture of gases, and the machine learning model will be trained with the training gas. On the other hand, a liquid flow meter will be trained with a training liquid. Preferred training gases include air, N2, and CO2, and preferred training liquids include water, isopropanol, and ethanol. Those skilled in the art can determine whether the flow meter is safe and appropriate for a specific fluid class. This embodiment is particularly important in flow measurements where there may be fluids that require additional safety measures. Therefore, the method may further include a safety check step that optionally uses other sensors (e.g., a pH meter sensor).

[0038] One embodiment relates to the method described above, wherein one or more fluid parameters include two, three, four, or five fluid parameters.

[0039] One embodiment relates to the method described above, wherein the detected fluid parameters include a fluid type and / or a fluid mixture, thereby enabling the identification or classification of the fluid (mixture).

[0040] One embodiment relates to the method described above, wherein the detected fluid parameters include material-specific fluid parameters (such as density (particularly liquid density), viscosity (kinematic viscosity and kinematic viscosity as described above), heat capacity, thermal conductivity, or electrical conductivity). Such material-specific parameters of the fluid can be estimated using a combination of sensing structures in the sensor configuration, namely a Coriolis mass flow sensor and additional density and pressure sensors.

[0041] One embodiment relates to the method described above, wherein the detected fluid parameters include the fluid concentration. This is an advantageous example of non-material specific fluid properties.

[0042] One embodiment relates to the method described above, wherein one or more fluids flow through a sensor configuration which is part of a microfluidic chip (e.g., a MEMS chip or a microfabricated chip) of a flow meter (along with associated flow values).

[0043] Alternatively, one or more fluids may flow through a sensor configuration that is part of a "macro" or conventional flow meter (such as a conventional thermal mass flow sensor, Coriolis flow sensor, or ultrasonic flow sensor).

[0044] One embodiment, related to the method described above, employs a machine learning model that utilizes a neural network. The neural network is trained to classify fluids and identify their fluid parameters, for example, using a sensing structure on a microfluidic chip. For example, deep symbolic learning, combined with physics in fluid sensing, enables real-time fluid data processing. It should be noted that deep neural networks (DNNs) are widely applied in smart sensing because they can provide inference models that enable automated data processing. However, this black-box approach has the drawback that DNNs do not take into account the physical characteristics of the sensor, making it difficult to explain what they have learned in a way that is understandable to humans. For example, neural networks do not understand that fluid density cannot take negative values. This is where the aforementioned deep symbolic learning becomes useful. It improves transparency and interpretability by including physical constraints and limiting meaningless output signals and judgments.

[0045] One embodiment relates to the aforementioned method and includes latent feature detection in the processing of one or more real-time measurement signals, which may reveal hidden correlations between the output signals of the sensor configuration / sensing structure (on the chip) and physical quantities. This can lead to the discovery of previously unknown correlations and open up new opportunities in the design of sensing structures. For example, a previously unknown relationship, such as the relationship between pressure and temperature, may be established. It is also possible to leverage the (usually low-level) sensitivity of the sensor to quantities other than the physical quantities for which it was designed to measure.

[0046] One embodiment, relating to the method described above, employs unsupervised machine learning as the machine learning model.

[0047] One embodiment of the method described above involves a machine learning model employing self-supervised learning.

[0048] One embodiment, relating to the method described above, involves a machine learning model employing model compression, pruning, and neural architecture search.

[0049] Another aspect of the present invention relates to a flow meter for carrying out the method described above, configured for communication connection with a computer system described later, and comprising a fluid inlet, a fluid outlet, and one or more fluid channels connecting the fluid inlet and the fluid outlet, thereby allowing one or more fluids to flow from the fluid inlet to the fluid outlet through the one or more fluid channels, and comprising a sensor configuration for generating one or more fluid-related measurement signals.

[0050] One embodiment relates to the aforementioned flow meter, which includes a Coriolis mass flow meter, a thermal flow meter, an ultrasonic flow meter, or other types of flow meter / flow control devices. When detecting viscosity, a fluid parameter, Coriolis mass flow sensors are particularly suitable for detecting kinematic viscosity, and thermal flow meter sensors are known to be capable of detecting kinematic viscosity (Schut et al., FULLY INTEGRATED MASS FLOW, PRESSURE, DENSITY AND VISCOSITY SENSOR FOR BOTH LIQUIDS AND GASES MEMS 2018, ISBN:978-1-5386-4782-0 / 18). Naturally, measurement of pressure and flow rate is also necessary. Coriolis mass flow sensors are suitable for directly measuring the density of a fluid and can therefore be used to measure kinematic viscosity.

[0051] One embodiment relates to the aforementioned flow meter and provides an intelligent (single-chip) multi-parameter sensor system in which the sensor configuration includes one or more additional sensors such as a flow sensor, a pressure sensor, a density sensor, and / or a viscosity sensor. If the flow meter is a Coriolis flow meter and an additional flow sensor is desired, it is usually a sensor using a different principle, such as a thermal flow sensor, or vice versa. Examples of pressure sensors include MEMS capacitive pressure sensors or MEMS piezoresistive strain gauge sensors. If the sensor configuration of the flow meter is MEMS, the use of MEMS additional sensors is particularly desirable.

[0052] One embodiment relates to the aforementioned flow meter, wherein the flow meter comprises a microfluidic chip, and the fluid inlet is connected to the fluid outlet via one or more microfabricated fluid channels. In addition to the engineering advantages of combining microfluidic sensors with machine learning, modern data processing technologies also open up many scientific possibilities. Machine learning can contribute to gaining insights into what is actually happening within the microfabricated fluid channels.

[0053] Another aspect of the present invention relates to a computer system having - one or more computer system processors, and - a memory that stores instructions for performing at least steps b) and c) which are executable by the one or more computer system processors, and preferably also capable of supplying one or more training measurement signals from a sensor configuration to a machine learning model, as described in step a) of the method. The computer system may be communicably connected to a flow meter (sensor configuration).

[0054] Therefore, another aspect of the present invention may relate to an assembly of such a computer system and such a sensor configuration of a flow meter, wherein the flow meter is communicably connected to the computer system.

[0055] Another aspect of the present invention relates to a non-temporary computer-readable medium that includes instructions executable by one or more computer system processors of a computer system to perform at least steps b) and c) and preferably also to perform the supply of one or more training measurement signals from the sensor configuration to a machine learning model, as described in step a). [Brief explanation of the drawing]

[0056] The present invention will be described by the embodiments shown in the accompanying drawings and by the detailed description of the drawings below.

[0057] [Figure 1] Figure 1 shows a schematic embodiment of a microfluidic chip with sensor configurations having different detection structures, where the measurement signals are fed to a neural network trainable for different fluids. [Modes for carrying out the invention]

[0058] As shown in Figure 1 and as previously stated, the applicant has found a method for detecting one or more parameters 3 of a fluid 12 using a sensor configuration 2 of a flow meter 1. The sensor configuration 2 includes a plurality of sensors 17, 20. The sensors may be chips. The method includes the following steps: a) flowing one or more training fluids 4 having one or more known fluid parameters 3 through the flow meter 1 and supplying one or more training measurement signals 5 from the sensor configuration 2 to a machine learning model 6; b) training the machine learning model 6 with one or more training measurement signals 5 so that it can detect one or more real-time fluid 12 parameters 8 using one or more real-time measurement signals 7 obtained from the sensor configuration 2; and c) detecting one or more real-time fluid 12 parameters 8 using the trained machine learning model 6 and one or more real-time measurement signals 7 supplied to the trained machine learning model 6 from the sensor configuration 2.

[0059] Supplying one or more real-time measurement signals 7 from sensor configuration 2 to machine learning model 6 includes processing one or more real-time measurement signals 7 by performing feature extraction or feature learning 10 on one or more real-time measurement signals 7, and supplying one or more processed real-time measurement signals 19 to trained machine learning model 6. Bioelectric measurement signals 11, including noise, distortion, and non-ideal effects, are supplied to machine learning model 6 via optional preprocessing 9a, processing 9, and feature extraction or feature learning units 10 (e.g., to keep voltages within the range of an A / D converter).

[0060] As mentioned above, this method can be used for real-time fluid classification, i.e., detection and recognition of substances, as well as concentration measurement in mixtures, and is expected to have a wide range of applications. Examples: concentration detection during drug delivery, estimation of the calorific value of natural gas.

[0061] Preferably, the machine learning model 6 employs deep symbolic learning.

[0062] This method may involve retraining machine learning model 6. Retraining may be an ongoing process.

[0063] One or more training measurement signals 5 may include uncalibrated raw measurement signals 11.

[0064] Machine learning models identify the relevant time range for training by detecting resonances and selecting appropriate data.

[0065] Preferably, one or more fluid parameters 3,8 include 2, 3, 4, or 5 fluid parameters 3,8. The detected fluid parameters 3,8 may include fluid type and / or fluid mixture. The detected fluid parameters 3,8 may also include material-specific fluid parameters such as density, viscosity, heat capacity, thermal conductivity, or electrical conductivity, which can be considered additional sensors 18. The detected fluid parameters 3,8 may also include, for example, fluid concentration.

[0066] Preferably, one or more fluids 4, 12 flow through the microfluidic chip 13 of the flow meter 1.

[0067] Machine learning model 6 preferably employs a neural network.

[0068] Processing of one or more real-time measurement signals 7 may further include the detection of latent features.

[0069] Figure 1 shows a flow meter 1 for implementing the method described above. The flow meter 1 comprises a fluid inlet 14, a fluid outlet 15, and one or more fluid channels 16 connecting the fluid inlet 14 and the fluid outlet 15, allowing one or more fluids 4, 12 to flow from the fluid inlet 14 to the fluid outlet 15 through one or more fluid channels 16.

[0070] The flowmeter 1 also includes a sensor configuration 2 for generating one or more fluid-related measurement signals 5, 7, and 11. The flowmeter 1 may include a Coriolis mass flow sensor 17, a thermal flow meter (not shown), an ultrasonic flow meter (not shown), or other types of flow meters. The sensor configuration 2 may include one or more flow sensors, a pressure sensor 20, a density sensor, and / or a viscosity sensor.

[0071] The flow meter 1 may include additional sensors 18 such as a temperature sensor, humidity sensor, density sensor, viscosity sensor, heat capacity sensor, thermal conductivity sensor, or electrical conductivity sensor. The additional sensors 18 may also be a second flow sensor, which preferably operates on a different principle (for example, a sensor configuration including a Coriolis flow sensor may be combined with a thermal or ultrasonic flow meter).

[0072] As shown in Figure 1, the flow meter 1 is equipped with a microfluidic chip 13, and the fluid inlet 14 is connected to the fluid outlet 15 via one or more microfabricated fluid channels 16.

[0073] The flowmeter 1 shown in Figure 1 provides a fully integrated sensing system, preferably manufactured using microtechnology. By integrating multiple sensors 17, 20 (e.g., for flow rate, pressure, and density) onto a single microfluidic chip 13, a multi-parameter system enabling real-time fluid data processing can be provided. Furthermore, this data can be used to deepen the understanding of complex physical effects within the microfabricated fluid channel 16 and to improve the design of the microfluidic chip 13.

[0074] The microfluidic chip 13 is manufactured, for example, by a surface channel technology (SCT) manufacturing process.

[0075] Example 1 - Fluid classification using Coriolis flowmeters In an exemplary configuration for fluid classification (see Figure 1), sensor configuration 2 includes a Coriolis mass flow sensor 17 and at least two pressure sensors 20, 22. The pressure sensors 20, 22 are preferably located in the upstream and downstream channels of the Coriolis mass flow sensor 17, respectively. The Coriolis mass flow sensor 17 is preferably made of silicon nitride and typically comprises a free-hanging channel having a semicircular cross-section. Rectangular, square, or circular cross-sections are also possible. The frame of the free-hanging channel may be rectangular. If the frame is rectangular, it is preferably fixed in the center of one side (typically one of the longer sides). The channel is magnetically drivable. A Lorentz actuator is particularly suitable, i.e., driving in torsional mode is achieved. The channel is activated in torsional mode by an alternating current through wiring above the channel, and a magnetic field moves the channel into the so-called torsional mode. As the fluid flows through the vibrating channel, a pseudo-Coriolis force is induced, and the channel enters a second vibrating mode called the "swing mode" at the same frequency. This mode serves as an indicator of the mass flow rate through the channel.

[0076] For example, two electrodes 21, which may be gold or platinum electrodes, may be positioned on either side of the torsional mode axis. In the absence of mass flow, the phase difference between the electrodes is 180°. When the vibration channel enters swing mode due to mass flow, the phase difference of the capacitive reading electrodes of the Coriolis flowmeter 17 (e.g., consisting of two comb-like structures) shifts to less than 180° from each other. The phase difference Δφ is given by the electrode 21 position x e Swing mode amplitude in JPEG2026512841000002.jpg813 and twist mode amplitude This can be derived from the ratio with JPEG2026512841000003.jpg812.

[0077]

number

[0078] This is approximately proportional to the mass flow rate Φ for small phase differences, and small phase differences correspond to low mass flow rates. The mass of the fluid within the vibration channel has a significant effect on the resonant frequency.

[0079]

number

[0080] In the formula, k eff V is the effective spring constant of the channel in torsional mode. c V is the volume of the channel. f ρ is the volume of fluid corresponding to the internal volume of the channel. c ρ is the density of the channel wall. f This is the density of the fluid. Therefore, the density of the fluid can be measured in parallel with the mass flow rate using the same structure. Torsion can also be measured, for example, by optical means. Alternatively, the channel can be moved in an oscillating mode by a drive, and its torsion can be measured.

[0081] The pressure sensors 20, 22 are composed of semicircular silicon nitride channels, and it is desirable that the cross-sectional shape be the same as that of the Coriolis mass flow sensor 17. The pressure sensors 20, 22 are not free-hanging but are fixed within the silicon bulk. The flat silicon nitride ceiling portion of the channel forms a diaphragm or membrane that deforms when a pressure difference occurs between the inside and outside of the channel. Due to this membrane deformation, the serpentine electrodes (e.g., gold or platinum electrodes) arranged above the channel are stretched and compressed and function as strain gauges. For resistive readout, a Wheatstone bridge configuration is preferred (not shown). The output voltage V bridge is approximated by the following model:

[0082]

Number

[0083] where V supply is the supply voltage of the Wheatstone bridge 21, and ΔP is the gauge pressure. Using the combination of the sensing structures, i.e., the Coriolis mass flow sensor 17 and the pressure sensors 20, 22, and the density function of the Coriolis mass flow sensor 17, the viscosity of the fluid can be estimated. According to Hagen-Poiseuille's law, the pressure drop ΔP is given as follows:

[0084]

Number

[0085] where η is the kinematic viscosity, L is the flow path length, ρ is the fluid density, r effΦ is the effective radius, and Φ is the mass flow rate. This is a very rough estimate assuming an incompressible Newtonian fluid, uniform density, no fluid acceleration, and highly laminar flow. However, even in non-ideal cases, as in ideal cases, there is always a tendency for higher viscosity to have a positive effect on the pressure drop. Non-ideal cases include when the fluid is compressible. Therefore, the sensing structure of sensor configuration 2 is capable of comprehensively measuring the mass flow rate, pressure drop, density, and viscosity of the fluid. While a rough linear estimation of the physical sensing principle is known from prior art, the actual relationship is more complex. This is particularly noticeable in non-ideal fluids such as compressible fluids or very low-density fluids (e.g., gases). Furthermore, there are other complex physical effects that can provide interesting information about the fluid, including fluid identification. The magnitude of the higher harmonics of the Coriolis mass flow sensor 17 depends on the pressure. This dependency can be used to improve the accuracy of the built-in pressure sensors 20 and 22. Deep learning can be used as a method to automatically discover the correlation between the sensor output and the actual fluid parameters 3 and 8, resulting in a highly automated calibration routine.

[0086] In the sensor configuration 2 described above, four raw measurement signals are generated: two capacitive signals (from capacitive readout 21) and signals from two pressure sensors 20 and 22, for example: 1st: Upstream pressure sensor measurement signal, Second: Downstream pressure sensor, Third: Left side capacitance measurement signal, 4th: Right-side capacitance measurement signal However, the input to the classifier consists of four features: 1st: Average pressure, Second: Pressure drop, Third: Amplitude spectrum of capacitance signal, Fourth: Phase spectrum of a capacitance signal.

[0087] Table 1 shows the predictive accuracy of multiple classification methods applied to four types of liquids (water, isopropanol, ethanol, and acetone) using the aforementioned apparatus equipped with seals for non-corrosive liquids:

[0088] [Table 1]

[0089] From Table 1, it can be deduced that the neural network performs extremely well in fluid classification using the example apparatus described above.

[0090] The applicant claims that the method and flow meter according to the present invention have many applications. For example, if the flow rate and composition of a drug mixture to be administered can be accurately recorded in real time, drug delivery will be safer and more effective. Similarly, composition measurement in the food industry is important for the quality and safety of food, as precise control of flow rate, pressure, and composition is crucial. The intelligent microfluidic sensor system described above is a non-limiting example, and other configurations employing other measurement principles are also possible.

[0091] Other suitable applications include: Respiratory systems for patients with lung disease: Real-time monitoring of inhaled and exhaled gas composition to promote recovery (especially in the treatment of lung disease using gases) - Medical professionals are familiar with numerous case studies, including the use of Heliox in COPD. Fault-free nutritional delivery and excretion management in organ-on-chip systems: Accelerating biomedical research Measure the calorific value of fuel gases (including environmentally friendly alternative fuels such as biogas and hydrogen) and reduce CO2 emissions; To improve specialty nutritional foods by adding additives (such as vitamins) in controlled amounts to foods and beverages to enhance nutritional value, taste, and other consumer characteristics; To prevent waste and food poisoning, (food) packaging should be effectively cleaned (preferably by administering H2O2); Improved safety, predictability, and efficiency through expanded applications in pharmaceutical and vaccine production; and Promoting the safe and efficient expansion of solar panel production using CVD and ALD processes.

[0092] Example 2 In this method, a MEMS Coriolis chip 17 is used in a measurement device of the type shown in Figure 2.

[0093] Measuring device The sensor chip is placed inside an incubator for ambient temperature control. Optionally, an additional temperature sensor 18 is placed near the device for local temperature measurement at the chip level. The pressure controller and flow reference with control valve are placed outside the incubator to control the pressure input and flow rate to the device. Finally, a multifunction I / O device NI PCI-6143 is used for data acquisition. Six voltage signals are acquired by time sampling: P1 and P2 from upstream and downstream pressure sensors 20, 22; C1 and C2 from the capacitive electrodes of the Coriolis mass flow sensor 21; signal D for driving Coriolis torsion; and temperature-dependent resistance track R on the Coriolis chip.

[0094] Data processing Eight features can be obtained from these measurement signals. Three of these eight features are calculated using a Fast Fourier Transform (FFT) based on the resonant frequency (f0). These three features represent the phase difference between the driving electrode and the Coriolis electrode (ΔΦC1, ΔΦC2) and the phase difference between the two electrodes (ΔΦ). Furthermore, the pressure sensors (P1 and P2) and the additional temperature sensor 18 use the final measured values ​​without applying the FFT. These features are normalized using a minimum-maximum normalization technique and converted to a range of 0 to 1.

[0095] Machine Learning Data splitting is performed in a 30 / 70 ratio based on random sampling from all fluids, with 30% of the data used for model training and 70% for performance evaluation. Mean absolute percentage error (MAPE) is used as the evaluation metric. Density estimation uses the features Rtrack, P1, P2, f0, and ΔΦC1. Viscosity estimation uses these features plus (P1-P2) / ΔΦ, and ΔΦC1 is replaced with ΔΦ.

[0096] Explainability The Shapley Additive Explanation (SHAP) is based on Shapley values. The calculation of these values ​​for features is based on the marginal contributions in any possible association between various feature input variables ("players"). A SHAP diagram is used to explain the feature impact on the model, showing the correspondence between feature values ​​(0-1 in this example) and viscosity and density determinations.

[0097] Three types of statistical machine learning (ML) methods are employed: linear regression (LR), support vector regression (SVR), and Gaussian process regression (GPR). For details, see T. Hofmann et al: “A review of kernel methods in machine learning”, Mac-Planck-Institute Technical Report, 156, 2006 and / or Vladimir Vapnik: “The nature of statistical learning theory”, Springer Science & Business Media, 1999.

[0098] result Using the experimental setup shown in Figure 1, three types of liquids (ethanol, water, and isopropanol) were measured. A total of 3870 experiments were conducted, of which 951 random measurements were used to train a machine learning model, and the remainder were used for testing. Measurements were taken using various parameters (temperature 15-35°C, pressure 4-6 bar, mass flow rate 0-5 gh). -1 The measurements were conducted using the following method. Measurements without mass flow rate were excluded because no pressure drop occurred. Estimated density and viscosity were analyzed as functions of temperature. As summarized in Table 2 below, GPR demonstrated the best performance with mean absolute percentage errors of less than 0.01% and less than 1% for density and viscosity, respectively, demonstrating its potential as a state-of-the-art viscosity and density estimation technique.

[0099] [Table 2] Table 2: Comparison of Mean Absolute Error Rate (MAPE) and Worst-Case Error in Linear Regression (LR), Support Vector Machines (SVR), and Gaussian Process Regression (GPR) [Explanation of Symbols]

[0100] 1 Flowmeter 2 Sensor Configuration 3 Fluid Parameters 4 Training fluid 5 Training measurement signal 6. Machine Learning Models 7. Real-time measurement signal 8. Real-time fluid parameters 9. Real-time measurement signal processing 9a Optional preprocessing 10. Feature extraction or feature learning 11 Uncalibrated measurement signal 12 Real-time fluid 13 Microfluidic Chips 14 Fluid inlet 15 Fluid outlet 16 fluid channels 17. Coriolis mass flow sensor 18 Additional sensors 19 Processed measurement signal 20 Upstream pressure sensor 21 Capacitive electrode of Coriolis mass flow sensor 22 Downstream pressure sensor

Claims

1. A method for detecting one or more fluid (12) parameters (3) using the sensor configuration (2) of a flow meter (1), a) A step of flowing one or more training fluids (4) having one or more known fluid parameters through the flow meter, and supplying one or more training measurement signals (5) from the sensor configuration to a machine learning model (6); b) A step of training a machine learning model using one or more training measurement signals so that it can detect one or more real-time fluid (12) parameters (8) using one or more real-time measurement signals (7) from the sensor configuration; c) A step of detecting one or more real-time fluid parameters using a trained machine learning model and one or more real-time measurement signals supplied to the trained machine learning model from the sensor configuration, wherein supplying one or more real-time measurement signals from the sensor configuration to the machine learning model includes processing one or more real-time measurement signals (9) by performing feature extraction or feature learning (10) on one or more real-time measurement signals, and supplying the one or more processed real-time measurement signals (19) to the trained machine learning model. A method that involves a series of steps.

2. The method according to claim 1, wherein the machine learning model (6) employs deep symbolic learning.

3. The method according to claim 1 or 2, wherein the machine learning model (6) uses, analyzes, or explores time information or time series data.

4. The method according to claim 3, wherein the machine learning model (6) identifies a relevant time range for training the machine learning model.

5. The method according to any one of claims 1 to 4, comprising retraining the machine learning model (6).

6. The method according to claim 5, wherein the retraining is a continuous process.

7. The method according to any one of claims 1 to 6, wherein the one or more training measurement signals (5) include raw, uncalibrated measurement signals (11).

8. The method according to any one of claims 1 to 7, wherein the one or more fluid parameters (3, 8) include 2, 3, 4, or 5 fluid parameters.

9. The method according to any one of claims 1 to 8, wherein the detected fluid parameters (3, 8) include a fluid type and / or a fluid mixture.

10. The method according to any one of claims 1 to 9, wherein the detected fluid parameters (3, 8) include material-specific fluid parameters such as density, viscosity, heat capacity, thermal conductivity, or electrical conductivity.

11. The method according to any one of claims 1 to 10, wherein the detected fluid parameters (3, 8) include the fluid concentration.

12. The method according to any one of claims 1 to 11, wherein the sensor configuration of the flow meter (1) is part of a microfluidic chip (13).

13. The method according to any one of claims 1 to 12, wherein the step of processing one or more real-time measurement signals (7) includes detecting latent features.

14. The method according to any one of claims 1 to 13, wherein the machine learning model (6) employs a regression method.

15. The method according to any one of claims 1 to 14, wherein the machine learning model (6) employs linear regression (LR), support vector regression (SVR), Gaussian process regression (GPR), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), attention mechanism-based models including transformers, or autoregressive integrated moving average models (ARIMA).

16. - One or more computer system processors; and - A computer system executable by one or more computer system processors, including a memory for storing instructions for performing at least step b) and step c) of the method according to any one of claims 1 to 15.

17. A non-temporary computer-readable medium, which includes instructions for performing at least steps b) and c) of the method according to any one of claims 1 to 15, and which can be executed by one or more computer system processors of a computer system.

18. A flow meter (1) for carrying out the method according to any one of claims 1 to 15, configured for communication connection with the computer system described in claim 16, comprising a fluid inlet (14), a fluid outlet (15), and one or more fluid passages (16) connecting the fluid inlet and the fluid outlet, thereby enabling one or more fluids (4, 12) to flow from the fluid inlet to the fluid outlet through the one or more fluid passages, A flow meter (1) comprising a sensor configuration (2) for generating one or more fluid-related measurement signals (5, 7, 11).

19. The flow meter (1) according to claim 18, wherein the flow meter includes a Coriolis mass flow meter (17), a thermal flow meter, an ultrasonic flow meter, or other types of flow meters.

20. The flowmeter (1) according to claim 18 or 19, wherein the sensor configuration (2) includes one or more flow sensors, pressure sensors (20, 22), density sensors, viscosity sensors, thermal conductivity sensors, electrical conductivity sensors, additional Coriolis flow sensors, additional thermal flow sensors, and / or other additional sensors (18).

21. An assembly comprising a flow meter (1) according to any one of claims 18 to 20 and a computer system according to claim 16, wherein the flow meter (1) is connected to the computer system in a manner that allows communication.

Images