TECHNICAL FIELD

The present invention relates to a method for determining operational information of a metering pump.

BACKGROUND

Metering or dosing pumps are used for feeding and dosing precise amounts of liquid. These metering pumps usually have a moveable displacement member for example in form of a membrane or piston driven by a drive motor via a drive system transferring the rotational movement of the motor into a linear movement of the displacement member.

For many applications it is desirable to determine operational information of the metering pump, e.g. to detect the presence of malfunctions or to compute an effective stroke length or other operational parameter of the pump.

EP 3591 226 discloses a metering pump that includes a control device that is designed in such a manner that it detects the current position of the displacement element, detects the torque of the electric drive motor at several positions of the displacement element and that monitors the torque in relation to the position of the displacement Element. This prior art metering pump includes an analyzing module that either compares torque or pressure curves detected over time or that compares a detected pressure or torque curve with a previously stored sample curve.

However, it remains desirable to provide a method for determining operational information of a metering pump that is applicable to different types of metering pumps or that can at least efficiently be adapted for use with different types of metering pumps. It is also generally desirable to provide such a method that is reliable and computationally efficient and that can be configured in a cost-efficient manner. SUMMARY

Thus, it remains desirable to provide a method for determining operational information of a metering pump that solves one or more of the above problems and/or that has other benefits, or that at least provides an alternative to existing solutions.

According to one aspect, disclosed herein are embodiments of a method for determining operational information of a metering pump, the metering pump comprising a dosing chamber, a displacement member and a drive motor for driving the displacement member. Various embodiments of the method comprise: a) receiving a plurality, in particular a sequence, of detected values of an indicator quantity indicative of a strength of activation of the displacement member at respective positions of the displacement member during operation of the metering pump; b) computing the operational information from a machine-learning model configured, in particular trained, to output said operational information responsive to receiving a plurality, in particular a sequence, of input values derived from detected values of the indicator quantity.

The inventors have realized that operational information of a metering pump can reliably be determined from detected values of an indicator quantity indicative of a strength of activation of the displacement member at respective positions of the displacement member during operation of the metering pump, by employing a trained machine-learning model. Various embodiments of the method can efficiently be adapted for use with different types of metering pumps.

During operation of various embodiments of the metering pump, the drive motor moves the displacement member, preferably in a reciprocating manner so that the displacement member by its movement increases and decreases the volume of the dosing chamber. When the dosing chamber is filled with an incompressible liquid, the change in volume of the dosing chamber defines the delivered liquid volume. In the presence of air or cavitation, the medium inside the dosing chamber is compressible. In that case, the change in volume is different from delivered liquid volume. Generally, the delivered volume is the change in volume times the effective stroke length.

During operation of the metering pump, a pressure inside the dosing chamber or a relating indicator quantity, e.g. a relating force or torque, may be detected and recorded, and used to determine an operational condition or other operational information of the metering pump. The pressure or related indicator quantity may be measured continuously or intermittently, in particular periodically, as a sequence, in particular a time series, of detected values. The drive motor may be an electric drive motor.

A relating force or torque may be a force acting on the displacement member, in particular by the drive motor. The force or torque acting on the displacement member are related, in particular substantially proportional, to the pressure inside the dosing chamber. Therefore, while some embodiments use a detected pressure in the dosing chamber as indicator quantity, other embodiments use a detected force or torque as an indicator quantity.

For the purpose of detecting the pressure, the metering pump may comprise a pressure sensor for detecting the pressure inside the dosing chamber. In an alternative solution, the pressure may be calculated on the basis of the drive torque or force provided by the drive motor. The calculation may be based on knowledge of the mechanical connection between the drive motor and the displacement member. The drive torque or force may for example be measured by a respective torque or force sensor, respectively, or may be derived from electric values of the drive motor. Generally, further examples of an indicator quantity include a motor current, a motor voltage or another quantity derivable from, or otherwise related to, the motor current and/or the motor voltage and/or another quantity related to the motor load.

The inventors have found that a suitably trained machine-learning model may determine a variety of operational information of a metering pump from a plurality of detected values of an indicator quantity. Examples of such operational information include the identification of discrete operational states as well as the prediction of continuous-valued operational parameters. Various embodiments of the method disclosed herein allow determination of such operational information without the need for expert knowledge being applied to identify how various operational states or parameter-values are derivable from the measured indicator quantity. Accordingly, various embodiments of the method disclosed herein use different types of machinelearning models.

In particular, the operational information may include a classification of an operational condition of the metering pump, in particular a classification of an error condition or malfunction of the metering pump. The operational condition may be a current condition or even a predicted future condition, e.g. a prediction of an imminent error condition likely to occur in the near future. To this end, the machine-learning model may include a classification model trained to output an identifier of one of a plurality of discrete classes. Examples of classes may include predetermined error conditions such as "cavitation", "air bubbles", "leak condition", etc. In some embodiments, the classification model may be trained to output an estimated likelihood that one or more error conditions and/or other classes of operational conditions are present or are likely to occur in the future, in particular in the near future.

Alternatively or additionally, in some embodiments, the operational information includes a value of an operational parameter. Accordingly, the machine-learning model may include a regression model trained to output a value of a continuous-valued operational parameter. Examples of operational parameters include: a discharge pressure, an effective stroke length, and a discharge flow. For example, one minus the effective stroke length is an indicator of the amount of air in the dosing chamber.

The machine-learning model receives a plurality of input values representing the detected values of the indicator quantity and/or input values derived from the detected values of the indicator quantity. To this end, the machine-learning model may be configured to receive different types of representations of the plurality of input values and/or additional input values. In some embodiments, the machine-learning model is configured to receive a plurality, e.g. a sequence or array, of input values of the indicator quantity, each of the plurality of input values being associated with a respective position of the displacement member, and wherein the machine-learning model is configured to output said operational information responsive to receiving at least said plurality of input values, e.g. in the form of pairs of input data as described below, or otherwise. The input values may be the detected values or values derived therefrom, e.g. by a noise reduction process, an averaging over multiple detected values or the like.

In some embodiments, e.g. when the input values always represent values of the indicator quantity at the same respective predetermined positions of the displacement member, or at predetermined times during the cyclic movement of the displacement member, a one-dimensional representation of the indicator quantity may be used as input data for the machine-learning model, thus allowing for a memory-efficient representation, which may be particularly beneficial especially for embedded implementations. Accordingly, in some embodiments, each of the plurality of input values is associated with a respective predetermined position of the displacement member and/or with a respective predetermined time during the cyclic movement of the displacement member. Hence, the input values represent a sequence of values where the sequence has a predetermined phase-relationship with the cyclic movement of the displacement member.

In other embodiments, the machine-learning model is configured to receive a plurality of pairs of input data, each pair of input data comprising a position of the displacement member and a corresponding value of the indicator quantity at said position, and wherein the machine-learning model is configured to output said operational information responsive to receiving said plurality of pairs of input data. Accordingly, detected values at varying positions may be used.

In the above and other embodiments, the machine-learning model may thus receive a representation of a so-called pressure-stroke curve relating the pressure or similar indicator quantity with the current position of the displacement member. The pressure- stroke diagram may be represented as a closed curve in a pressure-position coordinate system. Alternatively the pressure-stroke diagram may be represented as an open curve representing the pressure (or other indicator quantity) as a function of the time or phase along the cyclic movement of the displacement member.

In some embodiments, the machine-learning model may receive a two-dimensional array of input values, the two-dimensional array representing a pressure-stroke diagram, e.g. an array of image pixels representing an image of a pressure-stroke curve. For example, the input may be represented as a raster image of a representation of the pressure-stroke diagram. Each pixel is represented as a number between 0 and 1 indicating black level of the pixel. Other embodiments may use a different representation of the pressure-stroke diagram, e.g. a more compact representation.

To this end, for the purpose of detecting the position of the displacement member, the metering pump may comprise a suitable mechanism for detecting the current position of the displacement member. For example, the metering pump may comprise a position sensor or the drive motor may be a stepper motor such that the position can be determined by counting the rotational angle of the drive motor.

Accordingly, in some embodiments, the method comprises:

- receiving position data indicative of monitored positions of the displacement member during operation of the metering pump, or computing position data from at least the received detected values of the indicator quantity;

- computing the plurality of input values, e.g. the sequence of values of the indicator quantity or the plurality of pairs of input data, from the received detected values of the indicator quantity and from the received or computed position data.

While some metering pumps allow determination of the position of the displacement member during operation, other types of metering pumps do not provide this information. It would thus be desirable to provide a method that is applicable to a wider range of metering pumps. To this end, in some embodiments, the machine-learning model is configured to receive a time series of detected values of the indicator quantity at respective points in time and to output said operational information responsive to receiving said time series of detected positions of the detected values of the indicator quantity. In particular, in some embodiments, the machine-learning model is configured to output said operational information based only on the received said time series of detected positions of the detected values of the indicator quantity. It will be appreciated that, when the cycle time of the movement of the displacement member, i.e. of the stroke cycle, is known and when the position of the displacement member is known for a reference time, the representation of the input data as a time series carries the same information as a representation of the indicator quantity as a function of position. However, as mentioned above, this information may not be readily available for all types of pumps. Nevertheless, the inventors have realized that a suitably trained machinelearning model may determine useful operational information from the indicator quantity alone, i.e. without the need of measuring the position of the displacement member.

In particular, in some embodiments, the machine-learning model includes a first machine-learning model and a second machine-learning model, the first machinelearning model being configured to compute, based on the received time series of detected values of the indicator quantity at respective points in time during the operation of the metering pump, an input representation of a plurality of input values of the indicator quantity, each input value being indicative of a value of the indicator quantity at a respective, in particular at a respective predetermined or otherwise known, position of the displacement member; the second machine-learning model may thus be configured to output the operational information responsive to receiving the computed input representation. In particular, the first machine-learning model may be trained to determine a phase and/or period of the cyclic motion of the displacement member from the time series of detected values of the indicator function. To this end, the first machine-learning model may receive pressure values or values of another indicator quantity obtained during a time window, which may have an unknown starting point relative to the stroke cycle of the pump and/or an unknown length relative to the stroke cycle of the pump. The first machine-learning model may output pressure values (or, if the first machine-learning model receives values of another indicator quantity, values of said another indicator quantity) for a time window corresponding to one stroke of the pump, the time window starting at a predetermined point of the stroke cycle, e.g. the bottom dead point. The second machine-learning model may then receive the pressure (or other indicator quantity) values corresponding to a single stroke as its input. Alternatively, the first machine-learning model may output pressure values (or, if the first machine-learning model receives values of another indicator quantity, values of said another indicator quantity) for a time window corresponding to a predetermined number of strokes or otherwise of a predetermined duration relative to the duration of a stroke cycle. The second machine-learning model may thus receive the pressure values or other indicator quantity values corresponding to said predetermined duration.

Providing separate first and second machine-learning models may result in a more compact, memory-efficient representation of the overall model. In particular, when the machine-learning model includes multiple models, e.g. for detecting respective operational conditions or for estimating respective parameters, the machine-learning model may include a single first machine-learning model performing the phase detection of the time series, and a plurality of second machine-learning models, each receiving the output of the first machine-learning model as an input. However, it will be appreciated that, in other embodiments, the first and second machine-learning models may be combined into a single machine-learning model, which may thus receive a time series of values of the indicator quantity where the time series has an unknown phase shift and/or an unknown duration relative to the cyclic motion of the displacement member. The combined machine-learning model may be trained to output the operational information directly from said time series of unknown phase and/or duration. Accordingly, the training set for such a combined machine-learning model may include input time series of different relative phase shifts and/or durations relative to the cyclic motion of the displacement member.

The present disclosure relates to different aspects including the method described above and in the following, corresponding apparatus, systems, methods, and/or products, each yielding one or more of the benefits and advantages described in connection with one or more of the other aspects, and each having one or more embodiments corresponding to the embodiments described in connection with one or more of the other aspects and/or disclosed in the appended claims.

In particular, according to one aspect, disclosed herein are embodiments of a computer- implemented method for creating a trained machine-learning model. The method comprises: a) obtaining a set of training data items, each training data item including a plurality of input values and a corresponding target output, the plurality of input values being indicative of an indicator quantity indicative of a strength of activation of a displacement member of a metering pump at respective positions of the displacement member during operation of said metering pump, the corresponding target output being indicative of operational information observable during said operation of said metering pump; b) training a machine-learning model from the obtained set of training data to output operational information responsive to receiving a plurality of input values.

Generally, for the purpose of the present disclosure, the term "trained machine-learning model" refers to a machine-learning model having a set of parameters, such as weights, that have been adapted based on a set of training data using a suitable training algorithm, such as an unsupervised or a supervised training algorithm. Similarly, the term "training a machine-learning model" refers to the process of adapting the machine-learning model based on the training data. In particular the training, in particular the adaptation of the model parameters of the machine-learning model, may be performed using supervised learning based on a set of training data where each training data item is labelled by a corresponding target output.

Various embodiment of the method disclosed herein may be computer-implemented. Accordingly, disclosed herein are embodiments of a data processing system configured to perform the steps of the method described herein. In particular, the data processing system may have stored thereon program code adapted to cause, when executed by the data processing system, the data processing system to perform the steps of the method described herein. The data processing system may be embodied as a single computer or other data processing device, or as a distributed system including multiple computers and/or other data processing devices, e.g. a client-server system, a cloud based system, etc. The data processing system may include a data storage device for storing the computer program and detector data. The data processing system may include a communications interface for receiving the detected values and/or other types of sensor data. In some embodiments, the data processing system may partly or completely be embodied as a suitably programmed or otherwise configured processing unit, e.g. a control device for controlling operation of a metering pump. Accordingly, a part of the data processing system or the whole data processing system may be accommodated in a housing of the metering pump, e.g. as part of the control device for controlling operation of a metering pump or as a separate processing unit. Alternatively or additionally, the data processing system may include one or more data processing apparatus external to the metering pump. The data processing system may receive the detected values of an indicator quantity from the metering pump or otherwise, e.g. from a separate pressure sensor.

According to one aspect, disclosed herein are embodiments of a metering pump. Various embodiments of the metering pump comprise a displacement member, a drive motor for driving the displacement member, and a data processing system as disclosed above and in the following. In particular, the metering pump may comprise a processing unit, which may be integrated into or separate from a control device configured to control operation of the metering pump; the processing unit may be configured to perform the steps of the method described herein. The processing unit of the pump may perform an embodiment of the process described herein alone as a stand-alone device or as part of a distributed data processing system, e.g. in cooperation with an external data processing system such as with a portable data processing device and/or with a remote host computer and/or with a cloud-based architecture. The processing unit may be separate from or partially or completely be integrated into a control device for controlling operation of the pump. The pump may further include an integrated sensor configured to measure the indicator quantity or a quantity from which the indicator quantity can be derived. According to another aspect, disclosed herein are embodiments of a system, the system comprising a metering pump and a data processing system as disclosed herein, wherein the metering pump comprises a dosing chamber, a displacement member, a drive motor for driving the displacement member. In some embodiments, the data processing system is separate from the metering pump and comprises an interface for receiving a plurality of detected values of an indicator quantity indicative of a strength of activation of the displacement member at respective positions of the displacement member during operation of the metering pump. Alternatively, the metering pump comprises the data processing system.

Yet another aspect disclosed herein relates to embodiments of a computer program configured to cause a data processing system to perform the acts of the method described above and in the following. A computer program may comprise program code means adapted to cause a data processing system to perform the acts of the method disclosed above and in the following when the program code means are executed on the data processing system. The computer program may be stored on a computer-readable storage medium, in particular a non-transient storage medium, or embodied as a data signal. The non-transient storage medium may comprise any suitable circuitry or device for storing data, such as a RAM, a ROM, an EPROM, EEPROM, flash memory, magnetic or optical storage device, such as a CD ROM, a DVD, a hard disk, and/or the like.

BRIEF DESCRIPTION OF THE DRAWINGS

Preferred embodiments will be described in more detail in connection with the appended drawings, where

FIG. 1 schematically shows an example of a dosing or metering pump.

FIG. 2 schematically shows an example of a pressure-stroke diagram showing a pressurestroke curve of a metering pump.

FIG. 3 illustrates examples of pressure-stroke curves of dosing pumps in the presence of different malfunctions or other operational conditions. FIG. 4 schematically illustrates a flow diagram of a training method for creating a trained machine-learning model for subsequent use in a method for determining operational information of a metering pump.

FIG. 5 schematically illustrates a flow diagram of a method for determining operational information of a metering pump.

FIGs. 6 through 8 schematically illustrate examples of machine-learning models.

FIGs. 9 through 11 illustrate examples of prediction results obtained by a method for determining operational information of a metering pump.

DETAILED DESCRIPTION

As an example of a dosing or metering pump, FIG. 1 schematically shows a membrane pump. It has to be understood that the invention may be carried out in a similar manner with other types of dosing pumps, for example with a metering or dosing pump using a piston as a displacement member instead of a membrane. The pump as shown in FIG. 1 has a pump or dosing chamber 2, a side wall of which is formed by a membrane 4. This membrane 4 is a displacement member. By displacement of the membrane 4 the volume inside the dosing chamber 2 can be increased for filling the dosing chamber 2 and decreased for discharging the liquid from the dosing chamber 2. At the lower side of the dosing chamber 2 there is arranged a suction valve 6 whereas on the opposite side there is arranged a pressure valve 8. Both valves are designed as check valves. In this example, ball shaped valve elements are closing the valve by gravity. However, additionally a biasing element, such as a spring, can be provided. During operation, liquid is sucked from a liquid container 3 via a suction line 5 through the suction valve 6 into the dosing chamber 2 and discharged out of the dosing chamber 2 through the pressure valve 8. From the pressure valve 8 the liquid is discharged via a pressure line 9 and a pressure loading valve 7 for example into a pipe 11 of a facility. The pressure loading valve 7 in the pressure line 9 defines the pressure in the pressure line 9, i.e. maintains the pressure on the outlet side of the pressure valve 8 at a predefined pressure. This pressure is set by the pressure loading valve 7. Connected to the supply line 9 is a pulsation damper 13 for equalizing a pressure pulsation occurring in the outlet or pressure line 9. The membrane 4 is moved in reciprocating manner via the connection rod 10. For driving the connection rod 10 in reciprocating manner there is provided a drive motor, in particular an electric drive in form of an electric drive motor 12, for example a stepper motor. The rotating drive motor 12 moves the connection rod 10 via an eccentric drive 14 transferring the rotational movement into a linear reciprocating movement. The eccentric drive 14 is coupled to the electric drive motor 12 via a gear drive 16. The connection rod 10 is connected to the eccentric drive 14 at a connection point 18 which is distanced from the rotational axis x of the eccentric drive 14 by the eccentricity e. This causes the linear movement of the connection rod 10 into direction S if the eccentric drive 14 is rotated in the rotational direction R. In this example, furthermore, a spring 20 is arranged in the drive. The spring 20 is a compression spring connected to the connection rod 10 such that the spring 20 is compressed when the connection rod 10 is moved backwards into direction SI moving the membrane 4 in the retracted position. The spring 20 can accumulate energy during the suction stroke. This energy is released during the pressure stroke when the connection rod 10 together with the membrane 4 is moved in the forward, i.e. advanced position in the direction S2. By this the spring 20 smooths the torque to be applied by the electric drive motor 12 during the entire stroke. It has to be understood that it is also possible to arrange a spring being compressed during the pressure stroke and acting as a return spring. Furthermore, the invention may also be realized without a spring 20.

The dosing pump has a control device 22 controlling the electric drive motor 12. The control device 22 comprises a monitoring module 24 for monitoring the operation of the dosing pump. The control device 22 may comprise usual electronic components like, in particular, a CPU or other processing unit, a storage device and one or more software applications for control of the dosing pump. The software applications may be stored on the storage device and be for execution on the CPU. The monitoring module 24 may preferably be realized as a software module. In this example, the monitoring module 24 is integrated into the control device 22. However, it would be possible to transfer information to an external computing or monitoring device, in particular a cloud device acting as a monitoring module 24. For this the control device 22 may comprise a communication interface 26 for wired and/or wireless communication. The monitoring module 24 is configured to continuously or intermittently record a pressure P inside the dosing chamber 2 and the position of the displacement member. The pressure inside the dosing chamber 2 and the position of the displacement member, e.g. the membrane 4 of the membrane pump of FIG. 1, may be recorded as a representation of a pressure-stroke curve in a pressure-stroke diagram. For detecting the position of the membrane 4 along the direction S, in this example, an encoder 28 detecting the angular position of the rotor of the drive motor 12 is used. Furthermore, it is possible to detect certain positions of the drive or the displacement member, for example by a single sensor and to calculate the further positions on basis of the known velocity of the displacement member and the time past. Furthermore, instead of a special encoder, a stepper motor may be used. In knowledge of the transmission ratio of the gear drive 16 and the geometrical design of the eccentric drive 14 based on the angular position, the position in direction S can be calculated. The pressure P inside the dosing chamber 2 may either be detected by a pressure sensor 30 or indirectly by detecting the torque of the drive motor 12 or a force acting in the drive and calculating the pressure P on the basis of the force F acting onto membrane 4, or otherwise. In this example, a pressure sensor 30 is arranged at the dosing chamber 2 and connected to the control device 22. In case that a force or torque is detected as an indicator quantity instead of the pressure, it is possible to continuously record this force or torque over the position of the displacement member instead of recording the pressure, as the pressure is related to, in particular proportional to, the force or related to the torque, in particular to the torque multiplied by a term that depends on the position of the eccentric drive.

The control device 22 further comprises a processing module 50 configured to implement a trained machine-learning model. The processing module 50 may e.g. be implemented as a software module executed by the CPU of the control device 22, or otherwise. The trained machine-learning module may comprise a suitable representation of the model structure and of parameter values of the model parameters, e.g. of the weights of a neural network model. In this example the processing module 50, including the trained machine-learning module, is integrated into the control device 22. However, it would be possible to implement the processing module 50 and/or the machine-learning module on a data processing system external to the control device of the pump. To this end, the control device 22 may exchange information with an external data processing system, e.g. a cloud computing architecture, that functions as a processing module and/or implements a machinelearning module. During operation, the processing module 50 receives recorded pressure values and associated positions of the displacement member from the monitoring module 24. The processing module 50 optionally processes the received information and feeds it into the trained machine-learning module, which in turn returns corresponding operational information as described herein. The processing module may be configured to display the information on a display of the control unit and/or raise an alarm in case of a detected fault condition and/or forward the information and/or any raised alarms to an external device or system via the communication interface 26, and/or the like. An example of a process performed by the processing module 50 and the machine-learning module will be described in more detail below. The trained machine-learning module may be commissioned with the control device 22 during installation or it may subsequently be loaded onto the control device 22, e.g. via communication interface 26.

It will be appreciated that embodiments of the method disclosed herein may also be implemented to compute operational information of other types of metering pumps. The machine-learning model may be integrated into such a pump or implemented by an external computing device, e.g. by a separate control unit that may be communicatively coupled to the pump and/or to a sensor configured to measure an indicator quantity. Yet further, the machine-learning model may be implemented by a remote data processing system, e.g. as a cloud service, configured to receive the indicator quantity or quantities from the pump and/or from a separate sensor, and compute the operational information as described herein.

As will further be discussed below, some embodiments of a metering pump may not be capable of monitoring the position of the displacement member. Such pumps may only be capable of monitoring the pressure in the dosing chamber or another, related indicator quantity. Accordingly, in such embodiments the machine-learning model may receive the pressure or other indicator quantity as its only input.

FIG. 2 schematically shows an example of a pressure-stroke diagram depicting a pressure-stroke curve as can be detected by the monitoring module 24 in general, or otherwise. The abscissa shows the stroke lengths S in percent, i.e. the linear movement of the membrane 4 between its position representing the minimum volume of the dosing chamber 2 and the position defining the maximum volume of the dosing chamber 2. The ordinate shows the pressure P as detected by the pressure sensor 30. A stroke of 0 percent corresponds to the lower dead center 32 and the stroke length of 100 percent corresponds to the upper dead center 34. The curve illustrates four phases of the membrane movement. The lower portion of the curve represents the suction phase 36, the portion with rapidly increasing pressure on the left side represents the compression phase 38, the upper portion represents the discharge phase 40 and the right portion with rapidly decreasing pressure represents an expansion phase 42, in which the internal pump volume is expanded. The expansion phase 42 together with the suction phase 36 corresponds to a movement of the membrane 4 in the direction SI, whereas the compression phase 38 and the discharge phase 40 form the pressure stroke in direction S2.

When the monitoring module 24 of the control device 22 continuously or intermittently records or monitors the pressure and associated displacement values, changes in the pressure-stroke curve over time or over several strokes can be detected by the monitoring device. Different problems or malfunctions which may occur in the dosing pump have different effects on the course of the curve in the pressure-stroke diagram.

FIG. 3 illustrates examples of pressure-stroke curves of dosing pumps in the presence of different malfunctions or other operational conditions, such as cavitation, the presence of air bubbles, etc. It will be appreciated that other indicator quantities, such as force or torque, may be represented in dependence of stroke length in a similar manner, thus relating in a different form of indicator-stroke diagrams, which may be used to detect operational conditions and estimate operational parameters in a similar manner. Various embodiments of the method disclosed herein provide an efficient way of reliably detecting such problems or malfunctions from the recorded pressure, or other indicator quantity, and displacement values.

FIG. 4 schematically illustrates a flow diagram of a training method for creating a trained machine-learning model for subsequent use in a method for determining operational information of a metering pump.

Initially, the process obtains a set of training data items. To this end, in step SI, the process obtains a set of input sequences. Each input sequence represents values of an indicator quantity indicative of a strength of activation of a displacement member of a metering pump at respective positions of the displacement member during operation of said metering pump. For the purpose of the following description, embodiments of the methods and apparatus disclosed herein will mainly be described with reference to pressure as indicator quantity. However, it will be appreciated that the various methods and apparatus may use other indicator quantities instead or in addition to pressure.

The input sequence may represent a time series of measured pressure values P(to), P ), P(t _n ) where the times to, ..., t _n are respective times during at least one stroke cycle of a metering pump. The number n, n>l, of recorded values may depend on the rate at which the metering pump records the values. Alternatively, the input sequence may represent a sequence of pairs of pressure and position data: Pi, S , (P2,S2),...,(Pn,S _n ), where each pair (Pi, Si) represents a pressure and a corresponding position, in particular the position of the displacement member at which the pressure was measured. It will be appreciated that, e.g. when the measured pressure values always represent pressure values at respective predetermined positions, the positions Si may not need to be explicitly included in the input sequence. Instead, the input sequent may be represented as a pressure vector p[i], i = where the index / enumerates the predetermined positions S,. It will further be appreciated that other embodiments may use other representations of an input sequence, e.g. use a measured torque or force instead of the pressure. In one embodiment, the input sequence may be represented as an array of values/data pairs. In other embodiments, the input sequence may be represented in a different manner, e.g. as an image of a pressure-stroke curve where the pressure-stroke curve represents a sequence of pairs of pressure and associated position data.

The input sequences may be obtained by operating a plurality of metering pumps under various operational conditions and by recording pressure and/or position data from the corresponding sensors of the pump. For example, the recorded pressure and/or position data may be obtained by a monitoring module of a metering pump, e.g. as described in connection with FIG. 1 above, or otherwise. The recorded data may be received from the pump via a suitable communications interface of the pump or they may be stored locally and later retrieved from a data storage device of the pump or a separate monitoring unit.

In step S2, each of the obtained input sequences are labelled with one or more target output values indicative of the operational information of the pump operating when said input sequence has been obtained.

The target output values may be obtained in a variety of ways, e.g. by manually determining an operational condition, e.g. an operational malfunction, of the pump and by assigning the target value a class identifier representing the determined operational condition. For example, some operational conditions, such as cavitation may be induced, e.g. by choking the inlet of the pump. In other embodiments, the target values may be obtained by performing reference measurements of e.g. the output pressure or effective stroke length of the metering pump. For example, target values of the effective stroke length may be determined from a flow measurement and from the pump speed and pump volume.

In any event, the process may store the obtained target output values associated with the corresponding input sequences to which they are related, thus creating a set of labelled training data items. Each training data item includes an input sequence and a corresponding target output, the input sequence being indicative of an indicator quantity indicative of a strength of activation of a displacement member of a metering pump at respective positions of the displacement member during operation of said metering pump, the corresponding target output being indicative of operational information observable during said operation of said metering pump. It will be appreciated that the training data may include multiple target values for each input sequence, e.g. an identifier for classifying the operational condition and/or respective values of one or more operational parameters. Accordingly, the training data may be used to train a single machine-learning model or different machine-learning models to output respective types of operational information.

In step S3, the obtained training data is used to train a machine-learning model to output operational information responsive to receiving a plurality of input values, in particular responsive to receiving an input sequence of input values. The machinelearning model may be a feed-forward neural network, such as a convolutional neural network.

For example, the convolutional neural network may receive an input indicative of a pressure-stroke curve. The pressure-stroke curve may be represented as a sequence of ID data points, as a sequence of 2D data points or as 2-dimensional input representation of a pressure-stroke diagram depicting a pressure-stroke curve, or otherwise. In this case, the machine-learning model may be a feed-forward neural network, such as a convolutional neural network.

When the input values are represented as a time series, the machine-learning model may receive a one-dimensional sequence of inputs. In this case, the machine-learning model may be a feed-forward neural network, a recurrent neural network or another suitable type of machine-learning model.

Generally, when the input to the machine-learning model is a sequence or onedimensional array, the network may be a dense, fully-connected network, a onedimensional convolutional network and/or a recurring network, such as a long shortterm memory (LSTM) network, or a combination thereof. For example the first layer(s) of the network may be convolutional, and the following layers fully connected. In one example, convolutional and LSTM layers may be combined, e.g. such that the first layer(s) is/are ID convolutional, the subsequent layer(s) LSTM, and further layer(s) dense.

The training of the machine-learning model may be based on a suitable training algorithm known as such in the art, e.g. a backpropagation algorithm.

The training process results in a trained machine-learning model which may then be used to predict operational information about a metering pump responsive to receiving a plurality of input values as described herein. To this end, a representation of the trained machine-learning model may be stored in the processing or control device of a metering pump or on another data processing system external to the pump, e.g. as a computer-program module. For example, the representation of the trained machinelearning model may be loaded or otherwise implemented as a processing module 50 onto a metering pump, e.g. as described in connection with FIG. 1 above, or otherwise.

FIG. 5 schematically illustrates a flow diagram of a method for determining operational information of a metering pump.

In initial step S4, the process records pressure and/or position data from the corresponding sensors of the metering pump for which operational information is to be determined. The process may record the data during normal operation of the pump and represent the recorded data as an input sequence of recorded values or in another suitable form. It will be appreciated that, in other embodiments, the process may obtain detected values of another type of indicator quantity that is indicative of a strength of activation of the displacement member at respective positions of the displacement member during operation of the metering pump. Moreover, in some embodiments, the process may only obtain the values of the indicator function; in other embodiments, the process may record the values of the indicator function as well as the corresponding position data. The process feeds the obtained data to a suitable processing unit which receives the indicator function values and, optionally, the corresponding position data as input. In subsequent step S5, the process derives a plurality, in particular a sequence, of input values from the detected values of the indicator quantity and feeds the input values into a trained machine-learning model, e.g. trained according to the process described in connection with FIG. 4. For example, deriving the input values may involve using the detected values directly or it may involve a suitable scaling of the input values and/or other pre-processing steps, such as a suitable filtering or the like. The specific preprocessing steps may depend on the representation of the input values to be fed into the machine-learning model, e.g. whether the trained machine-learning model is configured to receive a 2D representation of a pressure-stroke curve, a time series of, optionally normalized, pressure values and/or the like.

The trained machine-learning model outputs one or more corresponding output values in response to receiving the sequence of input values. Depending on the type of operational information represented by the output of the trained machine-learning model, the output value may have different forms. For example, when the operational information represents a type of operational condition, e.g. a type of malfunction, the machine-learning model may be a classification model and the output value may represent a class identifier representing an operational condition of the metering pump. When the operational information represents a continuous quantity, the trained machine-learning model may be a regression model and the output value may be a continuous value representing the continuous quantity. In some embodiments, the machine-learning model may output multiple output values. To this end the machinelearning model may include multiple part-models, each trained to output a particular type of operational information. The different part-models may receive the same or different representations of the input values.

In step S6, the process may then return the output value from the trained machinelearning model directly or compute a derived output value from the output of the machine-learning model. In some embodiments, the process may perform one or more actions responsive to the output value, e.g. issue an alarm in response to a predicted error condition, or even perform one or more control action for controlling the pump of another component of a system in response to the output of the machine-learning model.

FIGs. 6 through 8 schematically illustrate examples of machine-learning models for use in embodiments of the method disclosed herein.

In particular, FIG. 6 illustrates a machine-learning model 120 that receives a representation of a pressure-stroke curve as input 110, i.e. a representation of a series of data points, (Pi, Si), i = 1, n, each data point representing a pair (Pi, Si) of values representing a pressure P, and a corresponding displacement position S,-. The machinelearning model may receive the series of data points as sequence of duplets (Pi, Si), as a 2D image of the pressure stroke diagram or in another suitable form. The machinelearning model outputs the operational information 130 it is trained to compute. Examples of operational information include a classification of an operational condition of the metering pump, in particular a classification of an error condition of the metering pump. Other examples include a value of an operational parameter, in particular a continuous-valued parameter, such as a discharge pressure, an effective stroke length, and a discharge flow or the like.

FIG. 7 illustrates a machine-learning model 150 that receives a ID representation of pressure values P, as input 140 and is trained to output an output value 130 indicative of the operational information. The ID representation may represent a time series of detected pressure values (or pressure values derived from the detected values) or a series of detected or derived pressure values at respective positions during the stroke cycle. In some embodiments, the ID representation may represent a single stroke cycle, starting at a predetermined position during the stroke cycle, i.e. the ID representation may be a representation of a pressure-stroke curve. In other embodiments, the ID representation may represent another duration, e.g. more or less than a single stroke cycle; alternatively or additionally, the ID representation may represent a time window starting at an unknown position during the stroke cycle. To this end, in some embodiments, the machine-learning model 150 may be, or include, a recurrent network. Alternatively, the machine-learning model may be or include a feed-forward network configured to receive a fixed-length one-dimensional input sequence representing pressure values sampled over a time period of a predetermined length.

In any event, the machine-learning model 150 outputs an output value 130 in response to receiving the time series as input. As described in connection with FIG. 6, the output value may be indicative of a classification of the operational state of the metering pump or indicative of an operational parameter of the metering pump.

It is generally desirable to provide a machine-learning model that can easily be adapted to different types of pumps, in particular without the need of obtaining large training data sets for each individual type of pump.

For example, some metering pumps are not capable of outputting position information about the displacement of the displacement member. They may only be capable of outputting pressure values or another type of indicator quantity. Accordingly, it is desirable to provide a method that can also compute operational information for this type of metering pump.

In particular, the process may receive a time series of detected pressure values without information about how the time series relates to the movement of the displacement member. In particular, the process may not know the relative phase of the time series relative to the stroke cycle and/or the process may not know the duration of the stroke cycle relative to the time series, i.e. how many cycles the time series covers.

Accordingly, FIG. 8 illustrates a machine-learning model 850 that receives a representation 840 of a time series as input, similar to what was described in connection with FIG. 7. However, the time series 840 may cover more or less than a single stroke cycle and the start of the time series relative to the stroke cycle may also be unknown. The machine-learning model 850 of FIG. 8 comprises two part models 851 and 853, respectively. The first part model 851 is trained to receive the time series 840 and to output a representation of a pressure-stroke curve 852 which may then serve as input to the second part model 853. The representation of the pressure-stroke curve 852 may be in the form of a ID representation of pressure values P, covering a single stroke cycle and starting from a predetermined position during the stroke cycle. The ID representation may then serve as an input to the second part model 853, e.g. as the input 140 to the model of FIG. 7. Alternatively, the representation of the pressure-stroke curve 852 may be a representation of a series of data points, (Pi, Si), i = 1, ..., n, each data point representing a pair (Pi, S) of values representing a pressure P, and a corresponding displacement position Si, as described in connection with FIG. 6. In any event, the second part model 853 may be similar to the model 120 of FIG. 6 or to the model 150 of FIG. 7 and trained to output an output value 130 indicative of operational information responsive to receiving a representation of the pressure stroke diagram 852, e.g. as described in connection with FIG. 6 or FIG. 7. It will be appreciated that the representation of the pressure-stroke curve does not need to be a ID representation of pressure values corresponding to a cycle of the displacement member as illustrated in FIG. 8. Instead, it may be a 2D representation as illustrated in FIG. 7.

The first part model 851 thus only needs to be trained to translate a time series 840 of pressure values into a corresponding representation of a pressure stroke curve 852. As the pressure evolves periodically with the stroke cycle time as a period, the translation task corresponds to detecting a phase shift of the pressure time series relative to the stroke cycle and/or a period of the stroke cycle relative to the time series 840. The inventors have realized that a machine-learning model can be trained to perform this task. Accordingly, the second part model 853 may efficiently be re-used for, or at least adapted to, different types of metering pumps, e.g. by a suitable transfer learning process based on relatively few additional training examples. In particular, the second part model may even be used or adapted for use with metering pumps that do provide position information about the displacement member. Similarly, the output of the first part model 851 may be used as input to multiple second part models for detecting respective operational conditions or estimating respective operational parameters. It will be appreciated that the first part model may also be divided into multiple submodels, e.g. a first sub-model that receives a time series of pressure values and outputs a, preferably fixed length time series having a length corresponding to a single stroke cycle. A second part model may receive the output of the first part model and output a phase-shifted time series covering a single stroke cycle starting at a predetermined position of the displacement member, e.g. the "bottom dead point". Depending on the type of pump, in particular depending on the position / cycle information available from a given pump, the first sub-model, the second sub-model or both sub-models may be omitted.

It will be appreciated, that, in some embodiments, the machine learning model 850 may be implemented as a single model.

FIGs. 9 through 11 illustrate examples of prediction results obtained by a method for determining operational information of a metering pump as disclosed herein.

FIG. 9 shows a result of a classification model trained to predict five different classes of error conditions of a metering pump from a representation of the pressure-stroke curves. The operational conditions included pump operating normally and pumps operating under different types of malfunctions, such as air bubbles, cavitation or the like. FIG. 9 shows how the predicted classes correlated with the known actual operational condition of the pumps. It is noted that the individual points are intentionally displaced slightly from the classification values, so as to allow them to be visually distinguished from each other. As can be seen from FIG. 9, the classification results in a clear distinction of the various operational conditions with only few misclassifications, i.e. the trained machine-learning model was found to perform an accurate and reliable classification.

FIGs. 10 and 11 show respective correlations of two trained regression models. The model of FIG. 10 was trained to predict the outlet pressure of a metering pump from a pressure-stroke curve, while the model of FIG. 11 was trained to predict the effective stroke length of the metering pump. As can be seen from the correlations to respective reference measurements, the predicted values by the method disclosed herein provides an accurate estimate of the operational parameters of the pump.

Embodiments of the method described herein may be computer-implemented. In particular, embodiments of the method may be implemented by means of hardware comprising several distinct elements, and/or at least in part by means of a suitably programmed data processing system. In the apparatus claims enumerating several means, several of these means can be embodied by one and the same element, component or item of hardware. The mere fact that certain measures are recited in mutually different dependent claims or described in different embodiments does not indicate that a combination of these measures cannot be used to advantage.

In summary, various aspects disclosed herein may be summarized as follows:

Embodiment 1: A method for determining operational information of a metering pump, the metering pump comprising a dosing chamber, a displacement member and a drive motor for driving the displacement member, wherein the method comprises: a) receiving a plurality of detected values of an indicator quantity indicative of a strength of activation of the displacement member at respective positions of the displacement member during operation of the metering pump; b) computing the operational information from a machine-learning model trained to output said operational information responsive to receiving a plurality of input values derived from detected values of the indicator quantity.

Embodiment 2: A method according to embodiment 1, wherein the indicator quantity comprises a pressure inside the dosing chamber and/or a torque of the drive motor.

Embodiment 3: A method according to any one of the preceding embodiments, wherein the operational information includes a classification of an operational condition of the metering pump, in particular classification of an error condition of the metering pump. 1

Embodiment 4: A method according to embodiment 3, wherein the machine-learning model includes a classification model trained to output an identifier of one of a plurality of discrete classes.

Embodiment 5: A method according to any one of the preceding embodiments, wherein the operational information includes a value of an operational parameter.

Embodiment 6: A method according to embodiment 5, wherein the machine-learning model includes a regression model trained to output a value of a continuous-valued operational parameter.

Embodiment 7: A method according to embodiment 6, wherein the operational parameter is indicative of one or more of the following operational parameters: a discharge pressure, an effective stroke length, and a discharge flow.

Embodiment 8: A method according to any one of the preceding embodiments, wherein the machine-learning model is configured to receive a plurality of input values of the indicator quantity, each of the plurality of input values being associated with a respective position of the displacement member, and wherein the machine-learning model is configured to output said operational information responsive to receiving said plurality of pairs of input data.

Embodiment 9: A method according to embodiment 8, comprising:

- receiving position data indicative of monitored positions of the displacement member during operation of the metering pump, or computing position data from at least the received detected values of the indicator quantity;

- computing the plurality of input values from the received detected values of the indicator quantity and from the received or computed position data.

Embodiment 10: A method according to any one of embodiments 1 through 7, wherein the machine-learning model is configured to receive a time series of detected values of the indicator quantity at respective points in time and to output said operational information responsive to receiving said time series of detected positions of the detected values of the indicator quantity.

Embodiment 11: A method according to embodiment 10, wherein the machine-learning model includes a first machine-learning model and a second machine-learning model, the first machine-learning model being configured to compute a plurality of input values of the indicator quantity based on the received time series of detected values of the indicator quantity at respective points in time during the operation of the metering pump, each input value being indicative of a value of the indicator quantity at a respective position of the displacement member; the second machine-learning model being configured to output the operational information responsive to receiving the computed plurality of input values.

Embodiment 12: A computer-implemented method for creating a trained machinelearning model for use in a method according to embodiment 1, the training method comprising: a) obtaining a set of training data items, each training data item including a plurality of input values and a corresponding target output, the plurality of input values being indicative of an indicator quantity indicative of a strength of activation of a displacement member of a metering pump at respective positions of the displacement member during operation of said metering pump, the corresponding target output being indicative of operational information observable during said operation of said metering pump; b) training a machine-learning model from the obtained set of training data to output operational information responsive to receiving a plurality of input values.

Embodiment 13: A data processing system configured to perform the steps of the method defined in any one of embodiments 1 through 12. Embodiment 14: A metering pump comprising a dosing chamber, a displacement member, a drive motor for driving the displacement member, and a data processing system as defined in embodiment 13.

Embodiment 15: A system comprising a metering pump and a data processing system as defined in embodiment 13; wherein the metering pump comprises a dosing chamber, a displacement member, a drive motor for driving the displacement member.

Embodiment 16: A system according to embodiment 15, wherein the data processing system is separate from the metering pump and comprises an interface for receiving a plurality of detected values of an indicator quantity indicative of a strength of activation of the displacement member at respective positions of the displacement member during operation of the metering pump.

Embodiment 17: A system according to embodiment 15, wherein the metering pump further comprises the data processing system.

Embodiment 18: A computer program comprising computer program code configured, when executed by a data processing system, to cause the data processing system to perform the steps of the method according to any one of embodiments 1 through 12.

It should be emphasized that the term "comprises/comprising" when used in this specification is taken to specify the presence of stated features, elements, steps or components but does not preclude the presence or addition of one or more other features, elements, steps, components or groups thereof.