Method for detecting fluid parameters using a flow sensor configuration

Abstract

The invention concerns a method for detecting fluid (12) parameters (3) using a sensor configuration (2) of a flow meter (1), comprising: a) flowing training fluids (4) with known fluid parameters through the flow meter, wherein training measurement signals (5) from the sensor configuration are fed to a machine learning model (6); b) training the machine learning model to use real-time measurement signals (7) from the sensor configuration to detect real-time fluid (12) parameters (8); and c) using the trained machine learning model and real-time measurement signals fed to the trained machine learning model, to detect real-time fluid parameters, wherein feeding the real-time measurement signals from the sensor configuration to the machine learning model comprises processing (9) the real-time measurement signals by performing feature extraction or feature learning (10) thereon and feeding the processed real-time measurement signals (19) to the trained machine learning model.

Description

[0001] Title: Method for detecting fluid parameters using a flow sensor configuration Description FIELD OF THE INVENTION The present invention relates to a method for detecting one or more fluid parameters using a sensor configuration of a flow meter, as well as a flow meter for carrying out such a method. BACKGROUND OF THE INVENTION Accurate sensing of flow, pressure and other physical quantities of liquids and gases is crucial in a variety of applications that have high societal, environmental and economic impact. One could think of flow control in vaccine- producing process lines, which contain a sophisticated mixing / blending step of the vaccine with additives, or accurate deposition gas control for the production of semiconductors, but also more individual applications require accurate fluidic sensors, e.g. artificial respiration and accurate drug delivery to patients in hospitals. In many applications, mass flow meters are the preferred type of flow meters, but modern flow meters may comprise various additional functions. Nowadays, all the above mentioned applications rely on prior knowledge of the systems and of the physical relations between all parameters and e.g. the composition of a mixture. As an example, the viscosity of natural gas is a measure for its calorific value. The viscosity can be calculated from the Hagen- Poiseuille equation, using a predetermined flow rate and pressure drop. However, in the case of turbulent flow, the Hagen-Poiseuille equation is not valid, complicating determination of viscosity. Another problem is that the relation between viscosity and calorific value becomes less accurate when a high percentage of N2or CO2(as is the case in biogas) is present in the gas mixture. Furthermore, the possible addition of hydrogen gas in gas grids will also affect the relation between viscosity of the gas mix and its calorific value. Finding better-suited physical analytical models could provide a solution, but the model would have to be established for each individual circumstance, making this system time-consuming, inflexible and potentially inaccurate whenever conditions differ from those under which the model was derived. In the field of flow sensing, there is a need for multiparameter systems and methods that are able to extract more information from fluid mixtures than the flow rate. By making use of microfabrication technologies, in particular microelectromechanical systems (MEMS), multiple sensors - e.g. for flow rate, pressure and density - can be integrated into a single chip efficiently. To calculate e.g. the viscosity from the flow rate and pressure, conventional data processing methods involve filtering the raw sensor signals, calibration of the individual sensors and physical models. These methods are, however, time-consuming, not always applicable, and leave potentially relevant information undiscovered. OBJECT OF THE INVENTION It is therefore an object of the invention to provide a method and flow meter for detecting fluid parameters using a multiple-sensor configuration - i.e. multiple sensing structures together - that is less time-consuming, more broadly applicable, and allows for the discovery of potentially relevant information. SUMMARY OF THE INVENTION According to the invention, a method for detecting one or more fluid parameters using a sensor configuration of a flow meter is provided, comprising the consecutive steps of: a) flowing one or more training fluids with one or more known fluid parameters through the flow meter, wherein one or more training measurement signals from the sensor configuration are fed to a machine learning model; b) training the machine learning model, using the one or more training measurement signals, in order to use one or more real-time measurement signals from the sensor configuration to detect one or more real-time fluid parameters; and c) using the trained machine learning model, and one or more real-time measurement signals fed to the trained machine learning model from the sensor configuration, in order to detect one or more real-time fluid parameters, wherein feeding the one or more real-time measurement signals from the sensor configuration to the machine learning model comprises processing the one or more real-time measurement signals by performing feature extraction or feature learning on the one or more real-time measurement signals, and feeding the one or more processed real-time measurement signals to the trained machine learning model. The abovementioned method allows traditional physics rules to be combined with machine learning, e.g. modern feature learning, preferably deep learning techniques, to effectively deal with raw sensor data, verify physical constraints, understand the complex physical effects in (microfabricated) fluid channels, and improve future chip designs based on existing and novel (hidden) causal relations found through AI model interpretability. In addition to processing the one or more real-time measurement signals by performing feature extraction or feature learning on said signals, the method may include additional pre-processing of (some of) the real-time measurement signals. In addition to processing the one or more real-time measurement signals by performing feature extraction or feature learning on said signals, the method may include representation learning or feature selection. The abovementioned method employs a flow meter containing multiple sensing structures and a trained neural network, which outperforms the state- of-the-art multiparameter systems in many applications, e.g. for real-time quality control of products made in chemical or pharmaceutical microreactors, or in the food industry. By embedding human knowledge of physical quantities into machine learning, e.g. deep neural networks, the abovementioned method is able to learn quicker and better how to use the sensing structures on the chip to give a physically relevant output signal. Deep neural networks may involve regression methods. Preferred simple regression methods include Linear Regression (LR), Support Vector Regression (SVR) and Gaussian Process Regression (GPR), though these methods should preferably not be used for unseen data (i.e. unseen condition and sensors). More advanced regression models are generally suitable for unseen data. These methods are described in e.g. “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-Coordinated 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 S. M. Nguyen, et al., in “2023 IEEE International Conference on Pervasive Computing and Communications (PerCom)”, 13-17 March 2023, ISBN: 978-1-6654-5378-3. The latter two papers are regression models for location prediction, which can be used for flow rate prediction as well. An example of a neural network type usable in the present invention includes a convolutional neural network (CNN), which is a network architecture for deep learning that learns directly from data. A CNN is particularly useful for finding patterns in data. Another example is a recurrent neural network (RNN), which is a class of artificial neural networks where connections between nodes can create a cycle, allowing output from some nodes to affect subsequent input to the same nodes. Yet another example is “long short-term memory” (LSTM), which is neural network that has feedback connections, unlike standard feedforward neural networks. CNN, RNN and LSTM are all more advanced regression models generally suitable for unseen data. CNN may preferably be CNN regression but also include other CNN- based models. Further examples of deep neural networks may involve regression methods include Attention-based models (i.e. Transformer). Algorithms based on statistics such as Auto Regressive Integrated Moving Average (ARIMA) may also be used. Machine learning in combination with a (microfabricated) multiparameter system additionally has the following further advantages: • more versatile: fluids with less distinctive, non-linear or other physically non-ideal parameters can also be classified. • faster: a new type of fluid does not require in-depth understanding of the fluid or reprogramming of the detection software, only retraining of the model. • more robust and oriented on the application: when trained well in many practical situations, the solution can be robust against exceptional situations. In the context of the present patent application, “deep symbolic learning” concerns all methods in AI that are based on high-level symbolic (human- readable) representations of problems, logic and search. “Fluid type” concerns the identity / identification of the fluid or fluid mixture, such as water or ethanol (or a mixture thereof). “Fluid parameters” means, in a broad sense, properties of a fluid, which can be substance-specific parameters, such as density or viscosity, but also non-substance specific parameters such fluid concentration. Viscosity includes both dynamic and kinematic viscosity. “Real-time fluid” is the fluid that is flowing through the channel and being measured. The real-time fluid can be a liquid, a gas, a vapour or a plasma, including various types of mixes, for example air (which is a mix of nitrogen, oxygen, argon, carbon dioxide and trace gasses, and optionally water). US 2021 / 010839 A1 discloses a method for server-side flow measurement using machine learning, more specifically multivariate machine learning, wherein ultrasonic flow meter data is processed in multiple steps. However, US 2021 / 010839 A1 merely employs a single sensor type (i.e. acoustic) and is limited to pipelines (i.e. flow tubes having relatively large diameters) and therefore e.g. unsuitable for microfabricated flow channels. US 2020 / 166398 A1 discloses a flow metering system with a flow meter coupled to a plurality of sensors and configured to measure volume of fluid flowing through the flow meter. The system also includes a metrology computer coupled to the flow meter and the sensors. The metrology computer is configured to receive live values from a plurality of sensors during a first time period, train an artificial intelligence engine based on the live values received during the first time period, and detect a sensor failure based on a deviation between a live value from the sensor and a predicted value for the sensor. US 2020 / 166398 A1, however, appears to exclusively concern metrology-related applications. US 2020 / 124461 A1 furthermore discloses a rather generic method for determining at least one process variable of a medium. The method includes steps of recording a first value for the process variable by means of a first method for determining the process variable and recording a second value for the process variable by means of a second method for determining the process variable. The method also includes steps of selecting at least one of the detected values for the process variable by means of a classifier and outputting the selected value for the process variable. US 2020 / 355073 A1 discloses an attenuated total internal reflection optical sensor for obtaining downhole fluid properties. The sensor is thus limited to merely optical measurements and furthermore is unsuitable for micro-system applications. US 2022 / 244157 A1 discloses a method for determining and identifying anomalies in fork meters. The method merely employs “mechanical methods” and is also unsuitable for micro-system applications. US 2022 / 128388 A1 discloses a water metering system that also appears unsuitable for micro-system applications. CN 101900589 A discloses an air-entrainment liquid flow measuring method based on mass flow meter. CN 101900589 A, however, appears to merely relate to two-phase flows. CN 115 355 959 A discloses a gas-liquid two-phase flow measurement method and system based on machine learning and physical constraint. CN 115355959 A also appears to relate only to two-phase flows. An embodiment relates to an aforementioned method, wherein the machine learning model employs deep symbolic learning. The Applicant expects that using deep symbolic learning on e.g. microfluidic sensor data may cause a breakthrough in the design and use of cutting-edge multiparameter sensing systems. Deep symbolic learning can combine traditional physics rules with modern deep learning techniques to effectively deal with raw sensor data, verify physical constraints, understand the complex physical effects in e.g. microfabricated fluid channels, and improve future chip designs based on existing and novel (hidden) causal relations found through AI model interpretability. Thus, Deep Symbolic Learning is a framework that can jointly (i.e. parallel instead of sequentially) learn perception and symbolic functions, also see e.g. “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 methods such as deep symbolic Artificial Intelligence is a unique way to analyse the physics of the sensors, and the information derived from the sensors, from both a scientific and an engineering point of view. This also holds for the emerging class of AI. Currently, many machine learning algorithms are focused on e.g. object recognition from images, language processing or interpretation of the results from sensor networks. Also noise reduction and improvement of single MEMS devices has been performed. However, these are all commercially available calibrated off-the-shelf (digital) sensors. The physics of the proposed microfluidic sensors produce data in a completely different (time dependent) format. The Applicant submits that only novel insights in machine learning can lead to algorithms that can uncover unknown correlations between signals and fluid properties (e.g. end-to-end learning), which might result in new symbolic functions, or that has better interpretability along with competing accuracy (symbolic reasoning). Especially in the field where the Applicant intends to control flow by means of other parameters such as pressure or viscosity (as in the case of gas chromatography) the Applicant expects to find these so far undiscovered correlations and to understand the learnt policy. Furthermore, the Applicant submits that the opportunity to design and fabricate sensing structures and optimize them for machine learning is unique. An embodiment relates to an aforementioned method, wherein the machine learning model uses, analyses or explores temporal information or time- series data, e.g. in sequential data. The sensors obviously provide this information and exploring the temporal information can improve classification accuracy in at least some situations. An embodiment relates to an aforementioned method, wherein the machine learning model identifies time ranges that are particularly relevant for training the machine learning model. For example, the machine learning model may select relevant time ranges representing about 50% of the total running time, or less than 50% of the total running time, or less than 40% of the total running time, or less than 30% of the total running time, or less than 20% of the total running time, or less than 10% of the total running time. An embodiment relates to an aforementioned method, comprising retraining the machine learning model, i.e. constantly refining the machine learning model, for instance in case of a new fluid being introduced or in case of drifting populations. An embodiment (thus) relates to an aforementioned method, wherein the retraining is a continuous process. An embodiment relates to an aforementioned method, wherein the one or more training measurement signals comprise raw, uncalibrated measurement signals, i.e. raw, unfiltered (electrical) measurement signals with noise, distortion and non-ideal effects. Such measurement data can be used to better understand the complex physical effects in the (e.g. microfabricated) fluid channels. An embodiment relates to an aforementioned method, wherein flow meter is optimized for a certain class of fluids and the training fluids from that class are selected for training the model: for example, if a flow meter is optimized for measuring gas, the real-time fluid will be a gas or a mixture of gasses, then the machine learning model would be trained on training gasses, whereas a liquid flow meter would be trained on training liquids. Preferred training gasses include air, N2and CO2, whereas preferred training liquids include water, isopropanol and ethanol. The skilled person can determine if a flow meter is safe and suitable for a class of fluids. This embodiment is especially relevant for flow measurements where fluids that require additional safety measures could be present. Thus, the method could in further addition include the step of a safety check, optionally using other sensors, for example a pH meter sensor. An embodiment relates to an aforementioned method, wherein the one or more fluid parameters comprise 2, 3, 4 or 5 fluid parameters. An embodiment relates to an aforementioned method, wherein the detected fluid parameters comprise fluid type and / or fluid mixture, such that the fluid (mixture) can be identified or classified. An embodiment relates to an aforementioned method, wherein the detected fluid parameters comprise substance-specific fluid parameters, such as density (especially liquid density), viscosity (both dynamic and kinematic as described above), heat capacity, heat conductivity or electrical conductivity. The combination of sensing structures in the sensor configuration, i.e. a Coriolis mass flow and additional sensors density sensor and pressure sensors, can be used to estimate such substance-specific parameters of the fluid. An embodiment relates to an aforementioned method, wherein the detected fluid parameters comprise fluid concentrations – as an advantageous example of non-substance specific fluid property. An embodiment relates to an aforementioned method, wherein the one or more fluids flow through a sensor configuration that is part of a microfluidic chip, e.g. a MEMS-chip or microfabricated chip, of the flow meter (with associated flow rates). Alternatively, the one or more fluids may flow through a sensor configuration that is part of a “macro” or traditional flow meter, such as a traditional thermal mass flow sensor, a Coriolis flow sensor or an ultrasonic flow sensor. An embodiment relates to an aforementioned method, wherein the machine learning model employs a neural network. The neural network may be trained to e.g. use the microfluidic chip’s sensing structures to classify fluids and find their fluid parameters. E.g. deep symbolic learning allows for real-time fluid data processing by using a combination of such deep neural networks and physics in flow sensing. It should be noted that deep neural networks (DNN) have been widely applied in smart sensing, as DNN can provide inferencing models enabling automation of the data processing. However, this black-box approach fails to explain the learnings in a way that can be understood by humans, since DNN does not take the physics of the sensors into account. For example, a neural network does not understand that the density of a fluid cannot be negative. This is where the aforementioned deep symbolic learning comes in, which includes physical constraints that limit unmeaningful output signals and decisions, to improve transparency and interpretability. An embodiment relates to an aforementioned method, wherein processing the one or more real-time measurement signals comprises hidden feature detection, which may reveal hidden correlations between the output signals of the sensor configuration / sensing structures (on the chip) and physical quantities. This could lead to finding previously unknown correlations, opening new opportunities for the design of sensing structures. E.g. relations between pressure and temperature, i.e. unknown relations, may thus be established. It is also possible to take advantage of (typically low level) sensitivities of a sensor to physical quantities other than the quantity they were designed to measure. An embodiment relates to an aforementioned method, wherein the machine learning model employs unsupervised machine learning. An embodiment relates to an aforementioned method, wherein the machine learning model employs self-supervised learning. An embodiment relates to an aforementioned method, wherein the machine learning model employs model compression, pruning, and neural architecture search. Another aspect of the invention concerns a flow meter for carrying out the aforementioned method, configured for communicative connection with a computer system to be discussed hereafter, comprising a fluid inlet, a fluid outlet, and one or more fluid channels connecting the fluid inlet to the fluid outlet, such that one or more fluids can flow through the one or more fluid channels from the fluid inlet to the fluid outlet, and a sensor configuration for generating one or more fluid-related measurement signals. An embodiment relates to an aforementioned flow meter, wherein the flow meter comprises a Coriolis mass flow meter, a thermal flow meter, an ultrasonic flow meter or any other type of flow meter / flow controller. When detecting the fluid parameter viscosity, it is known that Coriolis mass flow sensors are particularly suitable for detecting dynamic viscosity and thermal flow meter sensors are able to detect 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). Of course, measurement of pressure and flow is also required. Coriolis mass flow sensors are suitable for directly determining density of the fluid and thus can be used to determine dynamic viscosity. An embodiment relates to an aforementioned flow meter, wherein the sensor configuration comprises one or more additional sensors, such as flow rate sensors, pressure sensors, density sensors and / or viscosity sensors, providing an intelligent, (single-chip) multi-parameter sensor system. If the flow meter is a Coriolis flow meter, and an additional flow rate sensor is desired, that would typically be a sensor using another principle, such as a thermal flow sensor or the reverse. The pressure sensors could, for example, be MEMS capacitive pressure sensors or MEMS piezoresistive strain gauge sensors; use of MEMS additional sensor is especially desirable if the flow meter sensor configuration is MEMS. An embodiment relates to an aforementioned flow meter, wherein the flow meter comprises a microfluidic chip, wherein the fluid inlet is connected to the fluid outlet with one or more microfabricated fluid channels. In addition to the engineering advantages that come with combining microfluidic sensors with machine learning, modern data processing also opens up many scientific possibilities. Machine learning can help in gaining more knowledge on what exactly happens in the microfabricated fluid channels. Another aspect of the invention concerns a computer system, comprising: - one or more computer system processors; and - memory storing instructions that are executable by the one or more computer system processors to perform at least steps b) and c), and preferably also the feeding of the one or more training measurement signals from the sensor configuration to a machine learning model as mentioned under step a), of the aforementioned method. The computer system may be communicatively connected to (the sensor configuration of) the flow meter. Another aspect of the invention may thus relate to an assembly of such a computer system and such a sensor configuration of a flow meter, wherein the flow meter is communicatively connected to the computer system. Another aspect of the invention concerns a non-transitory computer- readable medium comprising instructions that are executable by one or more computer system processors of a computer system to perform at least steps b) and c), and preferably also the feeding of the one or more training measurement signals from the sensor configuration to a machine learning model as mentioned under step a), of the aforementioned method. BRIEF DESCRIPTION OF THE DRAWINGS The invention will be explained by means of the exemplary embodiment depicted in the accompanying Figure and the detailed description of the Figure below. Figure 1 shows a schematic example embodiment of a microfluidic chip with a sensor configuration comprising different sensing structures, wherein measurement signals are fed to a neural network that can be trained for different fluids. DETAILED DESCRIPTION As will be shown with respect to Figure 1, and as discussed in the foregoing, the Applicant has come up with a method for detecting one or more fluid 12 parameters 3 using a sensor configuration 2 of a flow meter 1. The sensor configuration 2 comprises multiple sensors 17, 20. The sensor may be a chip. The method comprises the consecutive steps of: a) flowing one or more training fluids 4 with one or more known fluid parameters 3 through the flow meter 1, wherein one or more training measurement signals 5 from the sensor configuration 2 are fed to a machine learning model 6; b) training the machine learning model 6, using the one or more training measurement signals 5, in order to use one or more real-time measurement signals 7 from the sensor configuration 2 to detect one or more real-time fluid 12 parameters 8; and c) using the trained machine learning model 6, and one or more real- time measurement signals 7 fed to the trained machine learning model 6 from the sensor configuration 2, in order to detect one or more real-time fluid 12 parameters 8. Feeding the one or more real-time measurement signals 7 from the sensor configuration 2 to the machine learning model 6 comprises processing 9 the one or more real-time measurement signals 7 by performing feature extraction or feature learning 10 on the one or more real-time measurement signals 7, and feeding the one or more processed real-time measurement signals 19 to the trained machine learning model 6. Raw electrical measurement signals 11 with noise, distortion and non-ideal effects are thus fed via an optional pre-processing 9a, a processing 9 and feature extraction or feature learning unit 10 (to e.g. keep voltages within the range of A to D converters) to a machine learning model 6. As discussed in the foregoing, the method may be used for real-time fluid classification, i.e. to sense and recognize a substance or even its concentration in mixtures, which has potential for many applications. E.g. detection of concentrations during drug delivery and caloric value estimation of natural gas. Preferably, the machine learning model 6 employs deep symbolic learning. The method may also comprise retraining the machine learning model 6. The retraining may be a continuous process. The one or more training measurement signals 5 may comprise raw, uncalibrated measurement signals 11. The machine learning model identifies relevant time ranges for training the machine learning model by detecting the resonance and selecting appropriate data. Preferably, the one or more fluid parameters 3, 8 comprise 2, 3, 4 or 5 fluid parameters 3, 8. The detected fluid parameters 3, 8 may comprise fluid type and / or fluid mixture. The detected fluid parameters 3, 8 may also comprise substance- specific fluid parameters, such as density, viscosity, heat capacity, heat conductivity or electrical conductivity, which may be considered additional sensors 18. The detected fluid parameters 3, 8 may for instance also comprise fluid concentrations. Preferably, the one or more fluids 4, 12 flow through a microfluidic chip 13 of the flow meter 1. The machine learning model 6 preferably employs a neural network. Processing the one or more real-time measurement signals 7 may furthermore comprise hidden feature detection. Figure 1 shows a flow meter 1 for carrying out the aforementioned method. 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 to the fluid outlet 15, such that one or more fluids 4, 12 can flow through the one or more fluid channels 16 from the fluid inlet 14 to the fluid outlet 15. The flow meter 1 also comprises a sensor configuration 2 for generating one or more fluid-related measurement signals 5, 7, 11. The flow meter 1 may comprise a Coriolis mass flow sensor 17, a thermal flow meter (not shown), an ultrasonic flow meter (not shown) or any other type of flow meter. The sensor configuration 2 may comprise one or more flow rate sensors, pressure sensors 20, density sensors and / or viscosity sensors. The flow meter 1 may comprise an additional sensor 18, such as a temperature sensor, a humidity sensor, a density sensor, a viscosity sensor, a heat capacity sensor, a heat conductivity sensor or an electrical conductivity sensor. The additional sensor 18 may also be a second flow sensor; this second flow sensor would preferably work on a different principle (e.g. a sensor configuration comprising a Coriolis flow sensor would be combined with a thermal flow meter or an ultrasonic flow meter). As shown in Figure 1, the flow meter 1 comprises a microfluidic chip 13, wherein the fluid inlet 14 is connected to the fluid outlet 15 with one or more microfabricated fluid channels 16. The flow meter 1 as shown in Figure 1 provides a fully integrated sensing system, preferably fabricated using microtechnology. Multiple sensors 17, 20 – e.g., for flow rate, pressure and density – can be integrated into a single microfluidic chip 13 to provide a multiparameter system allowing for real-time fluid data processing. Furthermore, the data can be used to better understand the complex physical effects in the microfabricated fluid channels 16 and improve the microfluidic chip 13 design. The microfluidic chip 13 may e.g. be fabricated by the surface channel technology (SCT) fabrication process. Example – fluid classification using a Coriolis flow meter In an example setup for fluid classification, e.g. as shown in Figure 1, the sensor configuration 2 may comprise a Coriolis mass flow sensor 17 and at least two pressure sensors 20, 22. The pressure sensors 20, 22 are preferably positioned in the channels upstream and downstream of the Coriolis mass flow sensor 17, respectively. The Coriolis mass flow sensor 17 is preferably made of silicon nitride and comprises free-hanging channels which typically have a semi-circular cross-section. Rectangular, square or circular cross-sections are possible too. The frame of the free- hanging channels may be rectangular. If the frame is rectangular, it is preferably fixed in the center at one of its sides, typically one of the longer sides. The channel can be actuated magnetically. Lorentz actuators are particularly suitable, i.e. actuation is achieved in a twist mode. The channel is activated in twist mode by an alternating current through the wiring on top of the channel in a magnetic field causes the channel to move in so called twist mode. Once a fluid flows through the vibrating channel, a fictitious Coriolis force is induced which forces the channel to vibrate in a second mode, called a swing mode, at the same frequency. This mode is a measure for the mass flow rate through the channel. Two electrodes 21, for example gold or platinum electrodes, may be positioned at both sides of the twist mode axis. Without mass flow, the phase shift between the electrodes is 180o. If the vibrating channel enters swing mode, due to a mass flow, this causes the phase shift of the electrodes for capacitive readout of the Coriolis flow meter 17 (for example consisting of two comb-shaped structures) to change to below 180oshifted from each other. The phase shift ∆φ can be derived from the ratio between the swing mode amplitude ẑswingand the twist mode amplitude ẑtwistat the electrode’s 21 position xe: which is approximately proportional to the mass flow Φ for small phase shifts, with small shifts corresponding to low mass flows. The mass of the fluid in the vibrating channel has a significant effect on the resonance frequency. with keff the effective spring constant of the channel for the twist mode, Vc the volume of the channel, Vfthe volume of the fluid which corresponds to the internal volume of the channel, ρcthe density of the channel walls and ρfthe density of the fluid. The density of the fluid can therefore be measured with the same structure and in parallel with the mass flow. Twist can also be measured by, for example, optical means. Alternatively, the actuation causes the channel to move in a swinging mode and the twist is measured. The pressure sensors 20, 22 may comprise semi-circular silicon nitride channels, and preferably the cross-section is same as the cross-section of the Coriolis mass flow sensor 17. The pressure sensors 20, 22 are not free-hanging, but are fixed in the silicon bulk. The flat silicon nitride ceiling of the channels forms a diaphragm or membrane that deforms due to a pressure inside the channel that is different than the pressure outside the channel. The membrane deformation causes meandering electrodes (for example gold or platinum electrodes) on top of the channel to elongate and compress, i.e. a strain gauge. A Wheatstone bridge configuration is preferably used for resistive readout (not shown). The following model approximates the output voltage Vbridge: with Vsupplythe supply voltage of the Wheatstone bridge 21 and ∆P the gauge pressure . The combination of sensing structures, i.e. the Coriolis mass flow sensor 17 and the pressure sensors 20, 22 as well as a density function of the Coriolis mass flow sensor 17, can be used to estimate the viscosity of the fluid. The Hagen-Poiseuille law gives the pressure drop ∆P as: with η the dynamic viscosity, L the length of the channel, ρ the density of the fluid, reffthe effective radius and the mass flow. This is a very rough estimation, which assumes an incompressible Newtonian fluid with uniform density, no acceleration of the fluid and a highly laminar flow. Nevertheless, in non-ideal cases as well as the ideal case, a higher the viscosity still has always leads to a higher a positive effect on the pressure drop. Cases are considered non-ideal if the fluid is a.o. compressible. Thus, the sensing structures of the sensor configuration 2 are together able to measure the fluid’s mass flow, pressure drop, density and viscosity. Although rough linear estimations of the physical sensing principles are known from the prior art, the actual relations are more complex. This is especially the case for non-ideal fluid, such as compressible fluids or fluids with a very low density (e.g. gases). Furthermore, other complex physical effects might give interesting information about the fluid, including fluid identification. The magnitudes of the higher order harmonics of the Coriolis mass flow sensor 17 are pressure-dependent. Such a dependence could be used to improve the accuracy of the integrated pressure sensors 20, 22. Deep learning can be used as an automated way to find correlations between sensor outputs and actual fluid parameters 3, 8 – being a very sophisticated and automated calibration routine. In the sensor configuration 2 described above four raw measurement signals are generated: two capacitive (from capacitive readout 21) and two from pressure sensors 20, 22 e.g.: 1st: upstream pressure sensor measurement signal, 2nd: downstream pressure sensor, 3rd: left capacitive measurement signal, 4th: right capacitive measurement signal However, the input for classifier comprises four features: 1st: average pressure, 2nd: pressure drop, 3rd: magnitude spectrum of capacitive signals, 4th: phase spectrum of capacitive signals. Table 1 shows the percentage of correct predictions for several classification methods used – for four different types of liquids (water, isopropanol, ethanol and acetone) using the above-mentioned set-up, which has sealing suitable for non-corrosive liquids: Table 1 From Table 1 it can be derived that neural network performance proved quite satisfactory for fluid classification using the abovementioned example setup. The Applicant submits that the method and a flow meter according to the present invention has many applications. Drug delivery, for example, would be safer and more effective if the flow and the composition of the administered medicine mixture are accurately recorded in real-time. The same holds for composition measurements in the food industry, in which accurate control of flow, pressure and composition is important to food quality and safety. The abovementioned intelligent microfluidic sensor system is a non-limiting example; other set-ups, optionally using other measurement principles are possible. Further suitable applications include: • respiratory systems for lung disease patients that can real-time monitor the inhaled and exhaled gas composition to accelerating recovery, especially treatment of lung conditions with gas – the skilled person in the medical field will be familiar with the use of Heliox in COPD and many other examples; • provide failure-free nutrition supply and waste drainage in organ-on- a-chip systems to accelerate biomedical research; • determine the calorific value of fuel gases (including greener alternatives such as biogas and hydrogen) to reduce CO2footprint; • improving specialty nutrition by dosing controlled amounts of additives (such as vitamins) to food and beverages to increase nutritional value, taste or other consumer properties; • effective cleaning of (food) packaging to prevent waste and food poisoning, preferably by H2O2dosing; • upscaling applications in pharmaceutical and vaccine production to increase safety, predictability and efficiency; and • facilitating upscaling of safe and efficient production of solar cell panels with CVD and ALD processes. Example 2 The method uses a MEMS Coriolis chip 17 in a measurement setup of the type illustrated in Figure 2. Measurement setup The sensor chip is placed inside an incubator for ambient temperature control. Optionally, an additional temperature sensor 18 is placed near the device to measure the temperature locally at chip-level. A pressure controller and flow reference with a control valve are situated outside the incubator to control pressure input and flow to the device. Lastly, for data acquisition, a Multifunction I / O Device NI PCI-6143 is used. Six voltage signals are acquired by sampling in time: P1 and P2 from the upstream and downstream pressure sensors 20, 22, C1 and C2 from the capacitive electrodes of the Coriolis mass flow sensor 21, D the driving signal applied to actuate the Coriolis twist and R the temperature-dependent resistive tracks on the Coriolis chip. Data processing From these measured signals, eight features can be acquired. Three of these eight features are computed using the Fast Fourier Transform (FFT) based on resonance frequency (f0). These three features represent phase differences between actuation and Coriolis electrodes (ΔΦC1, ΔΦC2) as well as the difference between these electrodes (ΔΦ). Additionally, pressure sensors (P1 and P2) as well as an additional temperature sensor 18 using the last reading in time without FFT. These features are normalized using a min-max normalization technique, which transforms the features to the range of 0 and 1. Machine The data split is based on random sampling across all fluids and is 30 / 70, where 30% of the data is used to train the models, and 70% of the data is used to test the performance. The used metric for evaluation is the Mean Absolute Percentage Error (MAPE). To determine density, features Rtrack, P1, P2, f0, and ΔΦC1are used. Viscosity estimation in addition to these features uses (P1-P2) / ΔΦ, and replaces ΔΦC1with ΔΦ. Shapley Additive explanations (SHAP) are based on Shapley values. These value calculations for a feature are based on marginal contributions across all possible coalitions, where the coalitions are between various feature input variables (‘players’). A SHAP figure will be used to explain the feature impact on the model, showcasing the feature value (which in this case is between 0 and 1), and its corresponding impact on viscosity and density determination. Three different statistical machine learning (ML) approaches are used: Linear Regression (LR), Support Vector Regression (SVR) and Gaussian Process Regression (GPR) as explained in 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. Results Three different liquids (ethanol, water and isopropanol) were measured with the experimental setup of Figure 1. In total, 3870 experiments were performed, of which 951 random measurements were used for training the ML models, and the rest for testing. The measurements were performed with varying parameters: temperatures between 15 and 35◦C, pressures between 4 and 6 bar and mass flows between 0 and 5 g h−1. Measurements with no mass flow were removed since these do not yield a pressure drop. The estimated density and viscosity as functions of the temperature were analyzed. As summarized in Table 2 below, GPR performed best with a mean absolute percentage error of < 0.01% and < 1% for density and viscosity respectively, showcasing the potential for state-of-the-art viscosity and density estimation. Table 2: Comparison of the Mean Absolute Percentage Error (MAPE) and worst-case error for the applied methods: Linear Regression (LR), Support Vector Machine (SVR) and Gaussian Process Regression (GPR)

[0002] LIST OF REFERENCE NUMERALS 1. Flow meter 2. Sensor configuration 3. Fluid parameters 4. Training fluid(s) 5. Training measurement signals 6. Machine learning model 7. Real-time measurement signals 8. Real-time fluid parameters 9. Real-time measurement signals processing 9a. Optional pre-processing 10. Feature extraction or feature learning 11. Uncalibrated measurement signals 12. Real-time fluid 13. Microfluidic chip 14. Fluid inlet 15. Fluid outlet 16. Fluid channel 17. Coriolis mass flow sensor 18. Additional sensor 19. Processed measurement signals 20. Upstream pressure sensor 21. Capacitive electrodes of the Coriolis mass flow sensor 22. Downstream pressure sensor

Claims

CLAIMS 1. Method for detecting one or more fluid (12) parameters (3) using a sensor configuration (2) of a flow meter (1), comprising the consecutive steps of: a) flowing one or more training fluids (4) with one or more known fluid parameters through the flow meter, wherein one or more training measurement signals (5) from the sensor configuration are fed to a machine learning model (6); b) training the machine learning model, using the one or more training measurement signals, in order to use one or more real-time measurement signals (7) from the sensor configuration to detect one or more real-time fluid (12) parameters (8); and c) using the trained machine learning model, and one or more real-time measurement signals fed to the trained machine learning model from the sensor configuration, in order to detect one or more real-time fluid parameters, wherein feeding the one or more real-time measurement signals from the sensor configuration to the machine learning model comprises processing (9) the one or more real-time measurement signals by performing feature extraction or feature learning (10) on the one or more real-time measurement signals, and feeding the one or more processed real-time measurement signals (19) to the trained machine learning model.

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

3. Method according to claim 1 or 2, wherein the machine learning model (6) uses, analyses or explores temporal information or time-series data.

4. Method according to claim 3, wherein the machine learning model (6) identifies relevant time ranges for training the machine learning model.

5. Method according to any one of the preceding claims, comprising retraining the machine learning model (6).

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

7. Method according to any one of the preceding claims, wherein the one or more training measurement signals (5) comprise raw, uncalibrated measurement signals (11).

8. Method according to any one of the preceding claims, wherein the one or more fluid parameters (3, 8) comprise 2, 3, 4 or 5 fluid parameters.

9. Method according to any one of the preceding claims, wherein the detected fluid parameters (3, 8) comprise fluid type and / or fluid mixture.

10. Method according to any one of the preceding claims, wherein the detected fluid parameters (3, 8) comprise substance-specific fluid parameters, such as density, viscosity, heat capacity, heat conductivity or electrical conductivity.

11. Method according to any one of the preceding claims, wherein the detected fluid parameters (3, 8) comprise fluid concentrations.

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

13. Method according to any one of the preceding claims, wherein processing the one or more real-time measurement signals (7) comprises hidden feature detection.

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

15. Method according to any one of the preceding claims, 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-based models including Transformer, or Auto Regressive Integrated Moving Average (ARIMA).

16. Computer system, comprising: - one or more computer system processors; and - memory storing instructions that are executable by the one or more computer system processors to perform at least steps b) and c) of the method according to any one of the claims 1 - 15.

17. Non-transitory computer-readable medium comprising instructions that are executable by one or more computer system processors of a computer system to perform at least steps b) and c) of the method according to any one of the claims 1 - 15.

18. Flow meter (1) for carrying out the method according to any one of the claims 1 - 15, configured for communicative connection to a computer system according to claim 16, comprising a fluid inlet (14), a fluid outlet (15), and one or more fluid channels (16) connecting the fluid inlet to the fluid outlet, such that one or more fluids (4, 12) can flow through the one or more fluid channels from the fluid inlet to the fluid outlet, and a sensor configuration (2) for generating one or more fluid-related measurement signals (5, 7, 11).

19. Flow meter (1) according to claim 18, wherein the flow meter comprises a Coriolis mass flow meter (17), a thermal flow meter, an ultrasonic flow meter or any other type of flow meter.

20. Flow meter (1) according to any one of the preceding claims 18 - 19, wherein the sensor configuration (2) comprises one or more flow rate sensors, pressure sensors (20, 22), density sensors, viscosity sensors, heat conductivity sensors, electrical conductivity sensors, additional Coriolis flow sensors, additional thermal flow sensor and / or other additional sensors (18).

21. Assembly of a flow meter (1) according to any one of the claims 18 - 20 and a computer system according to claim 16, wherein the flow meter (1) is communicatively connected to the computer system.

Images