Pump system and method for operating a pump system

Abstract

The invention relates to a pump system, wherein the pump system comprises a pump (10) comprising: a motor (26), and a controller (18) configured to control a speed of the motor (26). The pump system further comprises: at least two ultrasonic transducers (14, 16) which are configured to output raw data that depends on a velocity of a fluid which is pumped; and a processing unit (24) configured to execute a trained machine learning, ML, model (22) which is supplied with input data, wherein the input data comprises at least a portion of the raw data in digital form; wherein the ML model (22) is configured to output flow value data based on the input data. (Fig. 1)

Description

[0001] PUMP SYSTEM AND METHOD FOR OPERATING A PUMP SYSTEM

[0002] TECHNICAL FIELD OF THE INVENTION

[0003] The invention relates to a pump system with improved flow measurement capabilities and to a method for operating such a pump system .

[0004] BACKGROUND OF THE INVENTION

[0005] A centri fugal pump is a type of pump that raises or trans fers a fluid by converting rotational energy from an impeller into hydrodynamic energy of the fluid flow .

[0006] The flow rate provided by such a pump can be measured using ultrasonic waves . This measurement principle is often referred to as time-transit method, because a di f ference in transit-times between ultrasonic pulses propagating with and against a fluid flow is determined and converted to a flow value .

[0007] However, this approach requires a predictable and / or stable flow profile of the fluid to relate the transit-times to the volume flow across the whole flow channel . Turbulence and other instabilities in the flow channel may cause variations in the received signals . These instabilities are even more problematic in the vicinity of obstructions and rotational parts of a pump .

[0008] These problems are traditionally addressed by adding, e . g . , flow straighteners and / or more ultrasonic beams in order to probe larger regions of the flow profile , resulting in a more complex system and more advanced signal processing . Furthermore , at low flow rates the di f ference between the transit-times recorded with and against the flow becomes very small and sensitive to these instabilities and noise ef fects . For a centri fugal pump, these ef fects limit the lowest possible flow to be measured, due to the complex internal flow patterns which are related to the instabilities and noise .

[0009] Thus , it is an obj ective to provide an improved pump system and an improved method of operating a pump system which avoid the above-mentioned disadvantages . In particular, it is an obj ective to provide a pump system with improved flow measurement capabilities based on ultrasonic waves .

[0010] SUMMARY OF THE INVENTION

[0011] The obj ect of the present invention is achieved by the solution provided in the enclosed independent claims . Advantageous implementations of the present invention are further defined in the dependent claims .

[0012] According to a first aspect , the invention relates to a pump system, comprising a pump which comprises : a motor, and a controller configured to control a speed of the motor . The pump system further comprises : at least two ultrasonic transducers which are configured to output raw data that depends on a velocity of a fluid which is pumped; and a processing unit configured to execute a trained machine learning (ML ) model which is supplied with input data, wherein the input data comprises at least a portion of the raw data in digital form; wherein the ML model is configured to output flow value data based on the input data . This achieves the advantage that a pump is provided which can accurately determine a flow rate of a fluid which is moved by the pump .

[0013] In particular, the use of the ML model to calculate the flow rate results in a more stable measurement with less fluctuations and a larger dynamic range. Thereby, the accuracy of the flow measurements at low flow rates can be improved compared to conventional ultrasonic flow measurement methods. For instance, the ML model can be trained on the characteristics and / or the arrangement of the pump and / or the transducers. Furthermore, no complex pre-processing of the raw sensor data which requires memory and computing power is required.

[0014] The pump can be a centrifugal pump, e.g., comprising an impeller which is arranged in a pump body and which is connected to the motor by a shaft. The motor can be an electrical motor. Thus, the pump system can be a centrifugal pump system.

[0015] The flow value data can represent a flow rate, in particular a volumetric flow rate, of the fluid which is pumped by the pump. The fluid can comprise water or other media.

[0016] The machine learning model can be a neural network model (or short: a neural network) . For example, the neural network model can be a convolutional neural network (CNN) model. However, other ML or neural network models are also possible, such as LSTM (Long Short-Term Memory) , IDConv, gradient boosting, decision trees, random-forest etc.

[0017] In an embodiment, the at least two ultrasonic transducers are mounted in or to a flow channel for the fluid. The ultrasonic transducers can be mounted downstream respectively upstream from each other . Each transducer can send and receive ultrasonic pulses to respectively from the other transducer ( s ) , wherein an ultrasonic pulse received by a transducer is converted to an electrical signal which is directly or indirectly forwarded to the ML model . The raw data can comprise these electrical signals .

[0018] The raw data depending on the velocity of the fluid which is pumped may refer to the timing and / or the frequency of the electrical signals ( that are output by the transducers ) depending on the velocity of the fluid through which the ultrasonic pulses are transmitted . The ML model can be trained to derive information on the fluid velocity and, more particular, on its flow rate from the raw data .

[0019] The flow channel can be arranged in the pump, in particular in a water carrying element of the pump ( e . g . , of the pump body) . Alternatively, the flow channel can be arranged outside of the pump . In the latter case , the flow channel can be formed by a pipe connected to a pump outlet .

[0020] In an embodiment , the raw data comprises electrical signals which are output by the at least two ultrasonic transducers , wherein the controller is configured to analyze a signal level of the electrical signals and to select a section of the respective signal after the signal level reaches a reference value ; wherein the controller is configured to supply said section to the ML model .

[0021] For instance , the electrical signals are digitali zed before being analyzed by the controller .

[0022] In an embodiment , the controller is configured to select each section to start a predefined time after the signal level reaches the reference value and to have a predefined length .

[0023] In an embodiment , the controller is configured to adapt a pump setting, in particular the speed o f the motor, based on the flow value data . This achieves the advantage that the pump can be precisely controlled based on the determined flow values , e . g . , to maintain a constant flow rate .

[0024] In an embodiment , the processing unit is comprised by the controller of the pump, or the processing unit is ( or is comprised by) an external computing device , in particular a central server, which is connected to the controller via a communication interface .

[0025] I f the ML model i s executed by an external computing device , the pump system can be configured to forward the raw data to the ML model via the communication interface and to receive the flow value data from the ML model via the communication interface .

[0026] In an embodiment , the pump system further comprises : an analog- to-digital converter (ADC ) configured to digitali ze the raw data which is output by the at least two ultrasonic transducers . In particular, the ADC is configured to digitali ze the electrical signals which are output by the ultrasonic transducers .

[0027] The ADC can be a component of the controller or of a measurement electronic of the transducers . In an embodiment , the input data which is supplied to the ML model further comprises at least one of : an electrical power intake of the motor, a speed of the motor, and a di f ferential pressure value . By using these additional input data, the pump measurement accuracy can be enhanced . At the same time , the ML model might require less computing power .

[0028] For example , the ML model is trained on individual characteristics of the pump . This achieves the advantage that the accuracy of the generated flow values can be enhanced .

[0029] In an embodiment , for training the ML model , the ML model is supplied with further input data, in particular further raw data, which is collected in determined flow conditions ; wherein the ML model is configured to use the collected further input data as training data .

[0030] The training of the ML model can be carried out using the pump of the pump system or using a separate training pump of a training system, wherein the respective pump is exposed to the determined flow conditions . In the latter case ( training pump ) , the ML model can be trans ferred to the pump system after training .

[0031] For example , the ML model is configured to correlate the further input data with the determined flow conditions and to adj ust its configuration based on said correlation .

[0032] In case of the ML model being a neural network, weights of the neural network can be adj usted based on said correlation . In a neural network, the weights can control the signal between two neurons ( i . e . , how much influence the input has on the output ) . In addition, the neural network can comprise biases which are constant and which are not influenced by a previous layer of the neural network . The biases are an additional input into a next neural network layer .

[0033] In an example , the trained ML model is configured to continuously or repeatedly adj ust its configuration over the li fetime of the pump based on the input data received over time .

[0034] The trained ML model can further be configured to additionally output information on a status of the pump based on the input data . For example , the information on the status can comprise information on a wear of the pump and / or on changes of the pump performance over time .

[0035] According to a second aspect , the invention relates to a method for operating a pump system, wherein the pump system comprises a pump comprising a motor and a controller configured to control a speed of the motor, and wherein the method comprising the steps of : outputting raw data with at least two ultrasonic transducers , wherein the raw data depends on a velocity of a fluid which is pumped; supplying a trained machine learning (ML ) model with input data, wherein the input data comprises at least a portion of the raw data in digital form; and outputting flow value data based on the input data with the ML model .

[0036] The flow value data can represent a flow rate , in particular a current flow rate , of the fluid which is pumped by the pump system .

[0037] In an embodiment , the raw data comprises electrical signals which are output by the at least two ultrasonic transducers ; wherein the method further comprises : analyzing a signal level of the electrical signals ; selecting a section of the respective signal after the signal level reaches a reference value ; and supplying said section to the ML model .

[0038] In an embodiment , each section is selected to start a predefined time after the signal level reaches the reference value and to have a predefined length .

[0039] In an embodiment, the method comprises the further step of : adapting a pump setting, in particular the speed of the motor, based on the flow value data .

[0040] In an embodiment , the ML model is trained by : collecting further input data, in particular further raw data, in determined flow conditions ; supplying the ML model with the further input data ; and using the collected further input data as training data for training the ML model .

[0041] For example , the ML model is executed by a processing unit . The processing unit can be comprised by the controller of the pump or can be an external computing device , in particular a central server, which is connected to the controller via a communication interface .

[0042] The pump system can further comprise an analog-to-digital converter (ADC ) configured to digitali ze the raw data which is output by the at least two ultrasonic transducers .

[0043] For example , the input data which is supplied to the ML model further comprises at least one of : an electrical power intake of the motor, a speed of the motor, and a di f ferential pressure value .

[0044] The method according to the second aspect of the invention can be carried out by the pump system according to the first aspect of the invention .

[0045] BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The invention will be explained in the following together with the figures .

[0047] Fig . 1 shows a schematic diagram of a pump system according to an embodiment ;

[0048] Fig . 2 a schematic depiction of a measurement sequence for determining flow value data according to an embodiment ;

[0049] Fig . 3 an exemplary signal from an ultrasonic transducer ;

[0050] Fig . 4 a schematic depiction of a measurement sequence for determining flow value data according to an embodiment ;

[0051] Fig . 5 a flow chart of a method for operating a pump system according to an embodiment ; and

[0052] Fig . 6 a flow chart of a method for training an ML model according to an embodiment .

[0053] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Fig . 1 shows a schematic diagram of a pump system according to an embodiment .

[0054] The pump system comprises a pump 10 which comprises a motor 26 , and a controller 18 configured to control a speed of the motor 26 . The pump system further comprises : at least two ultrasonic transducers 14 , 16 which are configured to output raw data that depends on a velocity of a fluid which is pumped; and a processing unit 24 configured to execute a trained machine learning (ML ) model 22 , wherein the ML model 22 is supplied with input data, wherein the input data comprises at least a portion of the raw data in digital form; wherein the ML model 22 is configured to output flow value data based on the input data .

[0055] The ML model can be trained to determine the flow rate of the fluid based on the raw data, i . e . , unprocessed data provided by the ultrasonic transducers . This flow measurement calculation can be done using a structurally simple ML model 22 with low memory requirements and with low data processing loads . At the same time , the flow measurement can be highly accurate , as the ML model 22 can be trained on the individual characteristics and arrangement of the ultrasonic transducers 14 , 16 and / or the pump 10 .

[0056] The flow value data can represent a flow rate , in particular a volumetric flow rate , of the fluid . The fluid can comprise water or other media .

[0057] The pump 10 can comprise a pump body 11 having an inlet and outlet for connecting pipes 12 . For instance , the pump 10 is a centri fugal pump . The motor 26 can be an electric motor. The motor 26 can be configured to drive an impeller of the pump 10. The impeller can be arranged in a section of the pump body 11, wherein the motor 26 is connected to the impeller via a shaft.

[0058] The ML model 22 can be configured to output the flow value data to the controller 18. The controller 18 can then adapt one or more pump settings, in particular the speed of the motor, based on the flow value data.

[0059] The at least two ultrasonic transducers 14, 16 can be mounted in or to a flow channel for the fluid. The ultrasonic transducers 14, 16 can be components of an ultrasonic flow sensor.

[0060] In the example shown in Fig. 1, a part of this flow channel is formed by the pipe 12 which is connected to the pump 10. In particular, the pipe 12 is connected to the pump body 11, which e.g. comprises a centrifugal stack. The ultrasonic transducers 14, 16 are thereby mounted to a section of the pipe 12 which is outside of the pump 10.

[0061] However, the transducers 14, 16 could also be mounted to a section of the flow channel which is in the pump 10, e.g. in a water carrying element of the pump body.

[0062] For instance, the ultrasonic transducers 14, 16 can be clamp-on sensors (e.g., clamped on the outside of a pipe as shown in Fig. 1) , or in-media sensors (e.g., arranged inside the flow channel and in contact with the fluid) .

[0063] The controller 18 can comprise a processing unit 24 and a memory 20. For example, the processing unit 24 can be configured to execute the ML model , which can be stored in the memory 20 in the form of an ML algorithm .

[0064] Alternatively, the ML model 22 can be executed by an external computing device , in particular a command center server, which is connected to the pump controller 18 via a communication interface .

[0065] The controller 18 can be comprised by an edge module of the pump

[0066] 10 , e . g . an electronic module which comprises the pump electronics and which can be mounted to a side of the pump body

[0067] 11 .

[0068] Fig . 2 schematically depicts a pos sible measurement sequence for determining the flow value data with the pump system 10 .

[0069] The transducers 14 , 16 are typically mounted downstream respectively upstream from each other along the flow channel ( as indicated by the arrow in Fig . 1 ) . For instance , the upstream transducer 14 transmits a first ultrasonic pulse through the media ( i . e . , fluid) to the downstream transducer 16 ( downstream US signal ) , and the downstream transducer 16 then transmit a second ultrasonic pulse through the media to the upstream transducer 14 (upstream US signal ) .

[0070] The analog signals from the ultrasonic transducers 14 , 16 can be converted to digital signals that are sent to the ML model 22 . How the transducers 14 and 16 are in communication with the ML model 22 is not limited, and may be achieved via a wireless or wired connection For instance , the transducers 14 , 16 can be connected (wired) to the pump controller 18 , in particular, to a dedicated processor of the pump controller 18 configured to do a sampling and / or conversion (analog-to-digital) of the signal and then send the digitalized signal to the ML model.

[0071] For example, the transducers 14, 16 can be driven by an electrical circuit. The ultrasonic pulses received by each transducer 14, 16 can be converted to a respective electrical signal (raw signal) which is in-turn converted to a digital signal by an analog-to-digital converter (ADC) . The ADC can be a component of the pump controller 18 or of a measurement electronic of the transducers 14, 16. Subsequently, the digitalized raw signals can be transmitted to the ML model 22. For example, the timing information (time of emission and time of reception) of the signals is also transmitted to the ML model 22. The ML model 22 then outputs flow value data based on the received signals. For instance, the ML model 22 can be trained to determine a difference in transit-times between the ultrasonic pulses propagating in and against the flow based on the received signals and to derive a flow rate from said difference. In particular, the difference in transit-times between the ultrasonic pulses depends on a velocity of the fluid and, thus, on a flow rate of the fluid.

[0072] In an example, feature engineering can be applied to reduce the memory and processing requirements of the ML model 22. For example, a reduced sample of the raw signal (s) can be selected as input for the ML model.

[0073] Therefore, the controller 18 (or a dedicated processor of the controller) can be configured to analyze a signal level of the electrical signals from the transducers 14, 16 and to select a section of each signal after the signal level reaches (or exceeds) a reference value. Said section is then supplied to the ML model 22 (instead of the full raw signal) .

[0074] Fig. 3 shows an example of an electrical signal provided by a transducer 14, 16. For instance, a conical section of this raw signal can be sufficient to identify a phase shift between the upstream and downstream signals and, thus, determine the flow rate. In addition, the reference value can be included in the vector sent to the ML model 22 to allow compensation for the absolute speed of sound.

[0075] The section highlighted in Fig. 3 can be selected to start at a predefined time after the signal level reaches the reference value and to have a predefined length (e.g., length in time) .

[0076] The selection of the signal section can be carried out in a preprocessing step before sending the signal to the ML model 22. For instance, the pre-processing step includes identifying the signal reaching the reference value (e.g., signal is above 0.5) . The reference value may function as a representation of the speed of sound. A conically shaped section of the raw signal is then identified at a predetermined (e.g., temporal) distance from the reference value. In other words, the conically shaped section is a range where its upper and lower boundaries are at predetermined distance from the reference point. How the predetermined distance is determined is not limited, and may be determined by empirical trials.

[0077] For example, an additional ML model (e.g., a neural network) could be used to upscale the electrical signal provided by the transducers. The thus upscaled signals can be preprocessed and / or forwarded to the ML model for determining the flow value data . Fig . 4 schematically depicts a further possible measurement sequence for determining the flow value data with the pump system 10 .

[0078] Thereby, additional input data is supplied to the ML model 22 . The additional input data can comprise pump operational data, such as rotational speed of the motor (RPM) and an electrical power intake of the motor .

[0079] For instance , the input data comprises further parameters , such as : a temperature , a density, and / or a viscosity of a fluid to be pumped . For instance , the temperature will impact the speed of sound in the medium . Thus , by taking into account the medium temperature the ML model can determine the flow rate more accurately .

[0080] The input data can comprise further sensor and / or measurement data, such as system pressure of the pump or di f ferential pressure value ( i f the pump comprises a di f ferential pressure sensor ) .

[0081] For instance , i f the pressure at some location in the pump 10 is nearing the steam pressure of the medium ( often water ) , local boiling followed by implosion may occur . This phenomenon is called cavitation . Cavitation can influence the characteristics of the pump, including the flow . In order to take this into account the system pressure can be added as input data .

[0082] In an example , the input data further comprises a geometric parameter of the pump 10 and / or information on its installation . By taking into account additional input data, the accuracy of the determined flow values can be enhanced . For instance , the ML model 22 is trained on the relationship between the di f ferent input parameters and their influence on the flow rate .

[0083] In summary, the ML model 22 is configured to receive raw data from the transducers 14 and 16 as input , and in result output a flow measurement of the medium which is pumped . The ML model 22 can be trained to determine the flow in real-time based on the raw data / signals from the transducers 14 and 16 .

[0084] How the ML model 22 is implemented is not limited . In some nonlimiting embodiments , the ML model 22 is implemented as a neural network . However, it is contemplated that the ML model 22 is implemented in other manners , such as LSTM ( Long Short-Term Memory) or IDConv, gradient boosting, decision trees , randomforest , etc .

[0085] In the following, a possible training routine for training the ML model 22 will be discussed :

[0086] For training the ML model 22 , the ML model can be supplied with further input data, in particular further raw data from ultrasonic transducers , which is collected in determined flow conditions . The ML model 22 can be configured to use the collected further input data as training data .

[0087] For instance , the training of the ML model 22 may be carried out within the pump controller 18 , e . g . by exposing the pump 10 to the determined flow conditions , and directly training the ML which is executed by the processing unit 24 of the controller 18 . Alternatively, the ML model 22 can be trained remotely, e . g . in a test center with a different pump. The thus trained ML model 22 can then be loaded into the pump system after training.

[0088] In the latter case (training in a different pump) , the training of the ML model 22 may comprise mounting a pump into a test rig (comprising transducers) which provides a controlled flow rate, motor speed and power, as well as the raw data used as further input data for training.

[0089] The training of the ML model can be based on a supervised learning procedure. The training data can depend on the input data which the ML model 22 is intended to receive during later operation to determine the flow value data. The following table shows possible training data for different input data together with possible training labels:

[0090] For instance, when the training data is inputted into the ML model 22, the model can carry out a training logic to determine a set of features associated with the training label and the training data.

[0091] In other words, raw data from transducers can be used as training dataset and the controlled (i.e. known) flow rate can be used as training labels associated with the training dataset. Accordingly, the ML model 22 can be configured to generate an inferred function which is capable of outputting the flow rate during the in-use phase, based on the in-use input.

[0092] For instance, the further input data used for training the ML model 22 can comprise additional parameters, such as: electrical power intake of the pump motor, speed of the pump motor, and / or differential pressure values. In general, the further input data can comprise all types of data which are later used by the ML model 22 to determine the flow values.

[0093] In additional or alternatively, the ML model 22 can be continuously trained during normal operation of the pump 10. For instance, the ML model 22, e.g. a neural network, can be configured to further adjust its configuration (e.g., the weights of the neural network) over its lifetime based on the input data received over time.

[0094] For instance, after changing a component of the pump (e.g., an inlet) , information on this change can be input in the pump, and the ML model 22 can be trained on a new model or can be adjusted to take this change into account.

[0095] Besides providing flow information, the trained ML model 22 can be configured to provide additional data. For instance, the ML model 22 can be configured to further output information on a status of the pump based on the input data.

[0096] In an alternative embodiment to the above, the absolute time- of-flight (aToF, transit-time ) of the transducers signals are derived from the raw / full signals or directly from a timing circuit in the electronics, and the difference (dToF, delta ToF) in transit-time can either be obtained from the transit-time or from correlation between the two signals. These derived values (and not the raw signals) can then be used, e.g. by a specifically trained ML model, to determine the flow value data.

[0097] For a flow measurement in a pump, this alternative embodiment comprises providing one or more motor parameters for a current operation point of the pump, i.e., the pump speed (rpm) and power. These parameters are included in the calculation for the flow rate (Q) , e.g. according to Q = K * f (aToF, dToF, rpm, power, .... ) . In one example, the function f, is estimated by an n-order linear regression of the parameter to the actual flow, but other analytical functions may also be possible to use.

[0098] Fig. 5 shows a flow chart of a method 50 for operating a pump system according to an embodiment. The pump system comprises a pump 10, e.g. a centrifugal pump, comprising a motor 26 and a controller 18 configured to control a speed of the motor 26. The pump system further comprises at least two ultrasonic transducers 14, 16. For example, the method 50 can be carried out with the pump system as shown in Fig. 1.

[0099] The method 50 comprising the steps of: outputting 51 raw data with the at least two ultrasonic transducers 14, 16, wherein the raw data depend on a velocity of a fluid which is pumped; supplying 54 a trained ML model 22 with input data, wherein the input data comprises at least a portion of the raw data in digital form; and outputting 55 flow value data based on the input data with the ML model.

[0100] For instance, the raw data comprise electrical signals which are output by at least two ultrasonic transducers 14, 16. The method 50 can further comprise : analyzing 52 a signal level of the electrical signals ; selecting 53 a section of a respective signal after the signal level reaches a reference value , wherein said section is supplied to the ML model 22 .

[0101] For example , each section is selected 53 to start a predefined time after the signal level reaches the reference value and to have a predefined length .

[0102] The method 50 may comprise the further step of : adapting a pump setting, in particular the speed o f the motor, based on the flow value data .

[0103] Fig . 6 shows a flow chart of a method 60 for training the ML model 22 according to an embodiment . The method 60 comprises the steps of : collecting 61 further input data, in particular further raw data, in determined flow conditions ; supplying 62 the ML model with the further input data ; and using 63 the collected further input data as training data for training the ML model .

[0104] For instance , the pump 10 can be mounted in a test rig which provides the determined flow conditions . Alternatively, the training of the ML model 22 can be carried out with a di f ferent pump, wherein the trained ML model 22 can be trans ferred to the pump system after said training .

[0105] The ML model 22 can be configured to correlate the collected further input data with the determined flow conditions , and to adj ust its configuration based on said correlation .

[0106] The method 60 can comprise the further step of : continuously or repeatedly adj usting the configuration of the ML model 22 , e . g . a neural network) over the lifetime of the pump 10 based on the input data received over time.

Claims

Claims1. A pump system, comprising: a pump (10) comprising: a motor (26) , and a controller (18) configured to control a speed of the motor (26) ; at least two ultrasonic transducers (14, 16) which are configured to output raw data that depends on a velocity of a fluid which is pumped; and a processing unit (24) configured to execute a trained machine learning, ML, model (22) which is supplied with input data, wherein the input data comprises at least a portion of the raw data in digital form; wherein the ML model (22) is configured to output flow value data based on the input data.

2. The pump system of claim 1, wherein the at least two ultrasonic transducers (14, 16) are mounted in or to a flow channel for the fluid.

3. The pump system of claim 1 or 2, wherein the raw data comprises electrical signals which are output by the at least two ultrasonic transducers (14, 16) ; wherein the controller (18) is configured to analyze a signal level of the electrical signals and to select a section of each signal after the signal level reaches a reference value; and wherein the controller (18) is configured to supply said section to the ML model (22) .

4. The pump system of claim 3,wherein the controller (18) is configured to select each section to start a predefined time after the signal level reaches the reference value and to have a predefined length.

5. The pump system of any one of the preceding claims, wherein the controller (18) is configured to adapt a pump setting, in particular the speed of the motor (26) , based on the flow value data.

6. The pump system of any one of the preceding claims, wherein the processing unit (24) is comprised by the controller (18) of the pump (10) ; or wherein the processing unit is an external computing device, in particular a central server, which is connected to the controller (18) via a communication interface.

7. The pump system of any one of the preceding claims, further comprising : an analog-to-digital converter, ADC, configured to digitalize the raw data which is output by the at least two ultrasonic transducers (14, 16) .

8. The pump system of any one of the preceding claims, wherein the input data which is supplied to the ML model(22) further comprises at least one of: an electrical power intake of the motor (26) , a speed of the motor (26) , and a differential pressure value.

9. The pump system of any one of the preceding claims, wherein, for training the ML model (22) , the ML model is supplied with further input data, in particular further raw data,which is collected in determined flow conditions; wherein the ML model (22) is configured to use the collected further input data as training data.

10. A method for operating a pump system, wherein the pump system comprises a pump (10) comprising a motor (26) and a controller (18) configured to control a speed of the motor (26) , the method comprising the steps of: outputting (51) raw data with at least two ultrasonic transducers (14, 16) , wherein the raw data depends on a velocity of a fluid which is pumped; supplying (54) a trained machine learning, ML, model (22) with input data, wherein the input data comprises at least a portion of the raw data in digital form; and outputting (55) flow value data based on the input data with the ML model (22) .

11. The method of claim 10, wherein the raw data comprises electrical signals which are output by the at least two ultrasonic transducers (14, 16) ; and wherein the method further comprises: analyzing (52) a signal level of the electrical signals; and selecting (53) a section of the respective signal after the signal level reaches a reference value, wherein said section is supplied to the ML model (22) .

12. The method of claim 11, wherein each section is selected (53) to start a predefined time after the signal level reaches the reference value and to have a predefined length.

13. The method of any one of claims 10 to 12, comprising the further step of: adapting a pump setting, in particular the speed of the motor (26) , based on the flow value data.

14. The method of any one of claims 10 to 13, wherein the ML model (22) is trained by: collecting (61) further input data, in particular further raw data, in determined flow conditions; supplying (62) the ML model (22) with the further input data; and using (63) the collected further input data as training data for training the ML model (22) .

Images