阅读说明：本技术 用于操作风机系统的方法和具有后弯式离心风机的风机系统 (Method for operating a fan system and fan system having a back-curved centrifugal fan ) 是由 W·埃贝勒 A·劳 R·维施图普 R·纳泽 M·胡姆 于 2021-05-28 设计创作，主要内容包括：本发明涉及用于操作风机系统的方法和具有后弯式离心风机的风机系统。本发明涉及用于操作风机系统的方法及这样的风机系统。该风机系统具有控制装置,其具有人工神经网络。该控制装置控制后弯式离心风机的电动马达。该离心风机产生气流,其由实际流动值、特别是体积流率的实际值表征。该实际流动值不通过传感器装置检测,而是借助人工神经网络根据输入变量确定,且基于此电动马达借助控制装置来开环或闭环控制。马达电流和马达电压及其时间相关行为被提供给人工神经网络输入层,该行为可为时间微分(例如一阶梯度)或可为先前时间点的至少一先前值。特别有利的是人工神经网络确定内部或外部反馈的输出压力的实际值,从而形成输入层的输入变量。(The invention relates to a method for operating a fan system and a fan system with a back-curved centrifugal fan. The present invention relates to a method for operating a fan system and such a fan system. The fan system has a control device with an artificial neural network. The control device controls an electric motor of the backward-bending centrifugal fan. The centrifugal fan generates an air flow which is characterized by an actual flow value, in particular an actual value of the volume flow rate. The actual flow value is not detected by the sensor device, but is determined from the input variables by means of an artificial neural network, and on the basis of this the electric motor is controlled open-loop or closed-loop by means of the control device. The motor current and motor voltage and their time-dependent behavior, which may be a time differential (e.g., a first order gradient) or may be at least one previous value of a previous point in time, are provided to the artificial neural network input layer. It is particularly advantageous that the artificial neural network determines the actual value of the output pressure fed back internally or externally, thus forming the input variables of the input layer.)

1. A method for operating a fan system (10), the fan system (10) comprising a control device (11) and a back-curved centrifugal fan (12), the back-curved centrifugal fan (12) having a motor (13) and a rotor (14) driven by the motor (13), wherein the fan system (10) comprises a fan motor (13)The fan system (10) is configured to generate an air flow (G) consisting of at least one actual flow rate value (pa (kT), Q (kT), pa (t)_akt)、Q(t_akt) Characterized, wherein the method comprises the steps of:

-determining an operating parameter (U (kT); U (t)) forming a correcting variable and characterizing an operating state of the motor (13) of the centrifugal fan (12)_akt)），

-determining in a continuous or time-discrete manner at least one operating parameter (I (kT) forming at least one actual system variable and characterizing at least one operating state of the motor (13) of the centrifugal fan (12); I (t)_akt)），

-converting the correcting variable (U (kT); U (t)_akt) And the actual system variable (i (kt)); i (t)_akt) An artificial neural network (19) provided to the control device (11),

based on the correcting variable (U (kT); U (t)_akt) And the actual system variable (i (kt)); i (t)_akt) And an actual system variable (I ((k-1) T); time-dependent change in dI) (I ((k-1) T); dI) determining at least one actual flow value (q (kt) by means of the artificial neural network (19); q (t)_akt)），

-determining an actual flow value (Q), (kT), Q (t) based on at least one of the determined actual flow values (Q), (kT)_akt) To check the correcting variable (u (kt)); u (t)_akt) Whether modification is required.

2. Method according to claim 1, wherein the control device (11) comprises a regulator (20), a predetermined desired flow value (B) and the actual flow values (Q (kt), Q (t)_akt) Is submitted to the regulator (20).

3. The method according to claim 1 or 2, wherein the actual flow value of the at least one actual flow value determined by the artificial neural network (19) is an actual volumetric flow rate value (Q (kt), Q (t)_akt) And the desired flow value (B) is a desired volumetric flow rate value.

4. A method according to claims 2 and 3, wherein the desired volumetric flow rate value is kept constant during operation in order to obtain a constant volumetric flow rate.

5. The method according to any one of the preceding claims, wherein the artificial neural network (19) comprises an input layer (30), the actual system variables (I (kT) of an actual point in time (kT), I (t)_akt) And the correcting variable (U (kT)); u (t)_akt) And the previous values (I ((k-1) T)) of the actual system variables to the previous time point ((k-1) T) are submitted to the input layer (30).

6. The method according to claim 5, wherein a previous value (U ((k-1) T)) of a further correcting variable to a previous point in time ((k-1) T) is submitted to the input layer (30).

7. The method according to any one of the preceding claims, wherein the artificial neural network (19) comprises an input layer (30), the actual values of the actual points in time of the actual system variables ((I (t))_akt) And an actual point in time (t) of the actual system variable_akt) Is submitted to the input layer (30).

8. Method according to claim 7, wherein the actual point in time (t) of the further correcting variable_akt) Is submitted to the input layer (30).

9. The method according to any one of the preceding claims, wherein an actual flow value of the at least one actual flow value determined by the artificial neural network (19) is an actual output pressure value (pa (kt), pa (t)_akt)）。

10. The method according to any one of the preceding claims, wherein the at least one actual flow value (pa (kt), pa (t) determined by the artificial neural network (19)_akt) Actual flow values (pa (kT)), pa (t)_akt) Is fed back to an input layer (30) of the artificial neural network (19).

11. Method according to claims 9 and 10, wherein said actual output pressure values (pa (kt), pa (t)_akt) Is fed back to the input layer (30).

12. The method according to any one of the preceding claims, wherein the artificial neural network (19) comprises neurons (33), and wherein each neuron (33) comprises an activation function (F).

13. Method according to claim 12, wherein the activation function (F) is formed by a rectifier.

14. Method according to claim 12 or 13, wherein the activation function (F) is limited to a maximum value (F)_max）。

15. Method according to any of the preceding claims, wherein at least one of the at least one actual system variable depends on or is the fan speed (n), and wherein the fan speed (n) is determined indirectly or is detected directly by means of a speed sensor (21).

16. A fan system (10) comprising a control device (11) and a backward curved centrifugal fan (12), the backward curved centrifugal fan (12) having a motor (13) and a rotor (14) driven by the motor (13), wherein the control device (11) is configured to perform the method according to any of the preceding claims.

Technical Field

The present invention relates to a method for operating a fan system and such a fan system. The fan system has a control device and a backward-curved centrifugal fan. In this back-curved centrifugal fan, curved fan blades are arranged, which extend from a radially inward side to a radially outward side obliquely to the radial plane opposite to the direction of rotation.

Background

The centrifugal fan has a rotor with fan blades and a motor configured to drive the rotor.

Open or closed loop control of the centrifugal fan will generate an air flow, in particular an air flow. The flow is characterized by at least one flow parameter, which may be described by one or more actual flow values, such as pressure or volumetric flow rate. In many cases, it is desirable to control the volumetric flow rate of the airflow generated by the centrifugal fan in an open or closed loop manner.

If there is no sensor for measuring the volume flow rate, the back-bent centrifugal ventilator has the problem that the correlation between the motor current of the motor of the centrifugal ventilator and the resulting volume flow rate at each operating point is undefined. Thus, the resulting volumetric flow rate cannot be easily inferred from the motor current and other operating parameters of the centrifugal fan. In addition, due to the large number of external influences and disturbance variables, closed-loop control of the volume flow rate is difficult without the actual value of the volume flow being detected by the sensor device.

In the past, experiments have been conducted to solve this problem mathematically or algorithmically, wherein the backward-curved centrifugal fan is described by a model and can therefore be controlled on the basis of this model. However, it has been shown that the adjustment of the desired volume flow rate based on this method is relatively inaccurate.

Disclosure of Invention

It is therefore an object of the present invention to provide a method and a fan system which can be implemented without measuring flow parameters used as open-loop or closed-loop control variables, but which still allows for accurate open-loop or closed-loop control.

The invention is solved by a method having the features of claim 1 and by a fan system having the features of claim 16.

The fan system comprises a control device and a backward-bending centrifugal fan. The centrifugal fan has a motor and a rotor drivable by the motor. The back-curved centrifugal fan has a rotor with fan blades extending from a radially inward side to a radially outward side in a curved manner inclined with respect to a radial plane and thereby having an extension opposite to the direction of rotation of the rotor. Thus, the radially inward edge of the fan blade is arranged before the radially outward edge of the fan blade, seen in the direction of rotation. The side of each fan blade facing the direction of rotation is convex, while the side facing the opposite direction of rotation is concave.

With a rotating rotor, a gas flow is generated, which can be characterized by one or more flow parameters or actual flow values (flow values). For example, a volume flow actual value and/or an output pressure actual value may be used as the at least one actual flow value.

The control device of the fan system comprises an artificial neural network for open-loop or closed-loop control of the at least one actual flow value. For open-loop or closed-loop control of the at least one actual flow value, operating parameters of the motor and/or the centrifugal fan are used. The operating parameters forming the correcting variable are set by the control device. At least one operating parameter is detected, forming at least one system actual variable, and may be directly sensed, measured or determined by calculation. The correcting variable and the at least one actual system variable are in particular an electrical operating parameter of the motor, for example the motor current, the motor voltage or the frequency of the motor current or the motor voltage.

The operating parameters of the motor used as actual system variables can be determined continuously or in a time-discrete manner during operation. Thus, the actual system variable may be detected continuously or repeatedly at time points predetermined by time intervals.

If, for example, the motor current or the motor voltage is modified, the fan system reacts to a change in the actual system variable, which also depends on the fan speed. If the motor voltage is used as the set correcting variable, the motor current can be used as the actual system variable.

The fan speed and hence the actual system variables also depend on the output pressure or volumetric flow rate of the airflow. Since the fan speed and in particular its time-dependent variation also depends on the output pressure. Due to the compressibility of the gas and the pressure-dependent density and inertia associated therewith, the change in the fan speed depends not only on the change in the correcting variable but also on the current output pressure. The time-dependent progression of the pressure-dependent change in the fan speed from the at least one actual system variable is evident. This relationship can be used to train or learn an artificial neural network. The actual system variables and the time-dependent changes of the actual system variables as well as the correcting variables can be transmitted to the artificial neural network and can be used in the case of a change of the actual system variables for determining and outputting the modified correcting variables such that a desired output pressure and/or volume flow rate of the gas flow is obtained. The correlation between the actual system variable and the output pressure and/or the volume flow rate is taken into account with sufficient resolution (resolution) since time-dependent changes of the actual system variable are taken into account.

The rotational speed of the motor or the rotor supporting the fan blades may be used as the fan rotational speed.

The rotor supporting the fan blades of the centrifugal fan is in particular connected in a torque-proof manner to the rotor of the motor, so that the motor speed corresponds to the rotor speed.

By means of the artificial neural network, the at least one actual flow value of the generated gas flow or air flow can be determined. The determination is based on the operating parameters (in particular the motor current and the motor voltage) used as correcting variable and at least one actual system variable, and also on the time-dependent variation of the operating parameters forming the actual system variable. Furthermore, preferably, the time-dependent changes of the correcting variable are also transmitted to the neural network and taken into account during the open-loop or closed-loop control of the air flow.

The operating point can be determined since time-dependent variations are taken into account. At each operating point of the back-curved centrifugal fan, the relationship between the operating parameter and the resulting volumetric flow rate is ambiguous, so that based solely on the actual values of the correcting variable and the actual system variable, it is not possible to determine the actual flow value. An unambiguous association is possible if additionally time-dependent changes of the actual system variables are taken into account. In this way, open-loop or closed-loop control of the generated air flow or air flow is possible without the use of flow sensors, in particular output pressure sensors or volumetric flow rate sensors.

Preferably, the motor current or the motor voltage is transmitted as a correcting variable to the neural network. In addition, the fan speed or an actual system variable (e.g., motor voltage of the motor current) related to the fan speed that is not used as a correcting variable may be transmitted to the artificial neural network. In addition, at least one additional value is transmitted to the artificial neural network so that the time-dependent behavior of the actual system variable can be determined. For example, the additional value can be at least the value of a temporally preceding actual system variable or a temporal differentiation (preferably at least a first order temporal differentiation) of the actual system variable. Alternatively, additional values may be transmitted to the artificial neural network so that the time-dependent behavior of the correcting variable may be determined. The additional value can be, for example, at least one temporally preceding value of the correcting variable or a temporal differentiation thereof (preferably a temporal differentiation of at least one order).

Based on the artificial neural network, a number of additional parameters may be considered which affect the generated airflow and may vary and may affect the actual system variables. Such changes may arise due to increased flow resistance in the system, for example, in the event that filters arranged upstream or downstream through which the gas flows are increasingly loaded and thus clogged. In the case of air conditioning control of buildings, the flow characteristics may change, for example, as a result of a door to a room being opened or closed. For general and similar systems, a change in the volume of the intake volume or the output flow volume of the ventilator system may cause a changed state. Additional disturbance or influencing variables may be leaks in the flow channel, blockages due to contamination in the flow channel, etc. Furthermore, temperature changes in the surrounding area, air humidity in the surrounding area, pressure in the surrounding area or other conditions in the surrounding area also have an influence on the air flow generated by the fan system and can be taken into account very well by means of an artificial neural network. The complexity based on the plurality of influencing parameters can be mapped very well by means of an artificial neural network.

Prior to the start-up of the wind turbine system, the artificial neural network may be trained based on expert knowledge and/or empirically determined data. In one embodiment, the artificial neural network may also be configured to train and learn at the installation location if additional training data is available. An artificial neural network may also be configured to update if an updated version is available.

In addition, artificial neural networks are also capable of determining and indicating changes. For example, a comparison model characterizing the state of the artificial neural network may be stored and compared to the actual state of the artificial neural network. From this, the actual condition or actual state may be determined, e.g. whether the filter is clogged, whether the door in the room is open, etc. It is therefore also possible to infer external disturbance variables or influencing variables which are actually present in open-loop or closed-loop control.

For learning of the artificial neural network, a gradient-based learning algorithm may be used. The size of the artificial neural network may vary. For example, an artificial neural network may have an input layer, an output layer, and one or more hidden layers. The number of neurons in each layer may be equal or different.

It is advantageous if the control device comprises a regulator to which a control deviation between a predetermined desired flow value and an actual flow value is transmitted. For example, the desired flow value may be a desired volumetric flow rate value and the actual flow value may be an actual volumetric flow rate value. The regulator is configured to adjust one or more operating parameters of the motor in accordance with the control deviation. This volume flow rate can be used as a control variable. In particular, the open-loop control or the closed-loop control of the control device is configured to maintain a constant volume flow rate.

In a preferred embodiment, the artificial neural network may comprise at least one feedback, wherein the neuron output values of one layer in the artificial neural network are fed back into a previous layer, in particular the input layer, in the form of neuron input values. The feedback may be implemented within the artificial neural network or may be implemented external to the artificial neural network. The artificial neural network may be constructed as a recurrent neural network. For example, actual flow values, e.g., actual output pressure values and/or actual volumetric flow rate values, determined by the artificial neural network may be fed back to previous layers of the artificial neural network, e.g., input layers. This feedback is particularly useful to allow the condition of the fan system to remain sufficiently stable. In addition, in doing so, a floating average determination may be achieved with changes based on external disturbances and influencing variables.

In one embodiment, the artificial neural network comprises an input layer to which the actual values of the correcting variable at the actual point in time, the actual values of the actual system variables at the actual point in time and at least one previous value of the actual system variables, and optionally at least one previous value of the correcting variable, can be transmitted at a previous point in time. If the actual value of the actual point in time and the previous value of the previous point in time are known, the time-dependent change of the respective operating parameter (actual system variable or correcting variable) can be determined by means of differentiation. Thus, for example, the actual motor voltage, the actual motor current and at least one previous value of the motor voltage and/or the motor current can be transmitted to an input layer of the neural network, in order to be able to take into account time-dependent motor voltage variations or motor current variations. Based on these values, the artificial neural network may determine an actual volumetric flow rate value and/or an actual output pressure value of the generated airflow.

In a further embodiment, in addition to the actual value of the correcting variable at the actual point in time and the actual value of the at least one actual system variable at the actual point in time, at least one time-dependent differentiation of the at least one actual system variable and, optionally, a time-dependent differentiation of a further correcting variable may also be transmitted to the input layer of the artificial neural network. For example, in the case of continuous measurement of the at least one actual system variable, the time derivation can be carried out by means of a differentiator and can be provided to the input layer. The at least one time differential preferably comprises at least a first order time differential.

Each neuron in each layer of the artificial neural network preferably has a nonlinear activation function. In a preferred embodiment, the activation function may be formed by a so-called rectifier (ReLU). Other activation functions may be used instead, such as a threshold function or a sigmoid function.

It is advantageous if the activation function is limited to a maximum value. In this case, for example, rare characteristics may be better trained or learned. For example, the maximum value of the activation function may be equal to 6.

Drawings

Preferred embodiments of the invention can be gathered from the dependent claims, the description and the drawings. Hereinafter, preferred embodiments of the present invention are explained in detail based on the drawings. The figures show:

figure 1 shows a highly schematic block diagram of a fan system with a control device and a back-curved centrifugal fan,

figure 2 shows a block diagram of an artificial neural network of the control means of the wind turbine system of figure 1,

figure 3 shows a block diagram of another embodiment of the artificial neural network of the control device of figure 1,

figure 4 shows a block diagram of a neuron of the artificial neural network of figures 1-3,

figure 5 shows the activation function for the neuron of figure 4,

fig. 6 illustrates an exemplary relationship between volumetric flow rate, motor current, and fan speed for the fan system of fig. 1, wherein equal volumetric flow rate lines intersecting equal output pressure lines are illustrated,

figure 7 illustrates an exemplary relationship between fan speed, motor current and control level of motor voltage for the fan system of figure 1,

FIG. 8 illustrates an exemplary relationship between fan speed, volumetric flow rate, and control level or motor voltage for the fan system of FIG. 1, an

Fig. 9 shows a characteristic curve of a set of characteristic curves, which illustrates, in an exemplary manner, a relationship between a volume flow rate and a motor current of the fan system of fig. 1.

Detailed Description

In FIG. 1, a wind turbine system 10 is illustrated highly schematically based on a block diagram. The fan system 10 comprises a control device 11 for open-loop or closed-loop control of the backward curved centrifugal fan 12. The backward curved centrifugal fan 12 has a motor, for example an electric motor 13, which is operatively connected to a rotor 14 of the centrifugal fan. During rotation, the rotor 14 generates an air flow G, e.g., an air flow, having an output pressure pa and a volumetric flow rate Q downstream of the centrifugal fan 12 or downstream of the fan system 10.

The backward-curved centrifugal fan 12 or the rotor 14 has fan blades 15 extending in a curved manner between a radially inner edge and a radially outer edge, which fan blades 15 are arranged so as to be evenly distributed over the rotor 14 in the direction of rotation. The radially inner edge is arranged before the radially outer edge, viewed in the direction of rotation. Each fan blade 15 has a convex shape on one side facing the direction of rotation and a concave shape on the opposite side.

The gas flow G can be characterized by at least one flow parameter, for example by the output pressure pa or the volume flow rate Q. At least one of these flow parameters, and the volumetric flow rate Q according to this example, may be open-loop or closed-loop controlled. Open-loop or closed-loop control of the flow parameters of the gas flow G should take place without detection by means of a sensor according to the example, and in particular without the use of a volume flow rate sensor and/or a pressure sensor.

In this embodiment, the flow parameters of the gas flow G should be closed-loop controlled by means of the control device 11 to values preset by the desired flow value B. The desired flow value B is a desired volumetric flow rate value and, therefore, the flow rate value Q forms the control variable according to this example. The actual volumetric flow rate value is not detected by the sensor device but is determined in the control device 11 by means of the use of an artificial neural network 19. Based on the actual flow value determined by the artificial neural network 19, which in the present case is the actual volumetric flow rate value, the control deviation D can be calculated by calculating the difference from the desired flow value B and can be submitted to the regulator 20 of the control device 11. The regulator 20 may then modify one or more operating parameters of the electric motor 13, such as the motor current I or the motor voltage U, in order to minimize and, if desired, eliminate the control deviation D.

Due to the profile of the fan blades 15, a characteristic curve of the backward curved centrifugal fan 12 is created, describing the relationship between the motor current I and the volume flow rate Q, as shown in fig. 9. The progression is non-linear and includes a maximum. With a constant volume flow rate Q, the characteristic curve shifts with increasing fan speed in the direction of higher values of the motor current I. Due to the curved parabolic progression, the determination of the motor current I of the electric motor 13 and the fan speed n (speed of the electric motor 13 or the rotor 14) cannot be made, and an unambiguous mapping to the volume flow rate Q of the generated air flow G cannot be made, since the operating point is not unambiguous. This ambiguous relationship is illustrated by the dashed lines in fig. 9, which each depict a constant motor current value with two intersections of the associated characteristic curve.

To further illustrate the operational behavior of the backward curved centrifugal fan 12, additional spatial characterization areas are illustrated in FIGS. 6-8. Fig. 6 shows the relationship between the generated volume flow rate Q, the output pressure pa, the motor current I and the fan speed n. In fig. 6, a line of constant output pressure pa intersects a line of constant volumetric flow rate Q. Fig. 7 illustrates the relationship between the fan speed n, the motor current I, and the control level (control level) of the electric motor 13 between 0% and 100%, where the control level increases as the motor voltage U increases. Fig. 8 shows the relationship between the fan speed n, the volume flow rate Q of the air flow G, and the control level of the electric motor 13.

For determining the fan speed n, a speed sensor 21 may alternatively be provided, however, this speed sensor 21 is preferably omitted. Alternatively, at least one additional sensor 22 can also be provided in order to detect ambient conditions or other influencing parameters, for example the air humidity h and/or the input pressure pe, which can correspond, for example, to the air pressure in the surrounding area. Instead of detecting the input pressure pe by means of a sensor device, it may also be determined in another way, for example by means of a calculation which is dependent on the geographical location of the installation of the fan system 10, in particular on the geographical height above sea level.

The operating parameters of the centrifugal fan 12 and in particular of the electric motor 13 are transmitted to an artificial neural network 19. One of the operating parameters forms the correcting variable and the other operating parameter forms the actual system variable. The motor voltage U may be used as a correcting variable and the motor current I may be used as the actual system variable, or alternatively the reverse. The actual system value may be calculated, estimated or measured. The desired motor voltage value output by means of the control device 11 can be used as the motor voltage U, so that the measurement of the actual motor voltage value can be omitted. According to an example, the fan speed n or a change thereof is monitored indirectly, for example by means of the motor current I. Alternatively or additionally, the fan speed n can also be detected by means of a speed sensor 21 and can be supplied to the control device 11.

Alternatively, the input pressure pe or the air humidity h may be additional input variables for the artificial neural network 19.

In the control device 11 or the artificial neural network 19, not only the respective actual values of the motor voltage U and the motor current I are taken into account, but also their time-dependent behavior (time-dependent behavior) or time-dependent progression. To this end, for example, a plurality of values of the motor voltage U and the motor current I detected at different points in time can be input as input variables into the artificial neural network 19 (fig. 2). The difference of these values divided by the detected temporal distance is, at least approximately, a slope value characterizing the time-dependent change.

Alternatively, a time differential, for example a motor current change dI and, in addition, optionally a fan speed change dn, can be generated by means of the differentiator 23, which accordingly represents a first time differential of the motor current I or of the fan speed n (fig. 3). Additionally or alternatively, higher order time differentials may also be detected and considered. These configurations are particularly suitable if the operating parameters are detected time-continuously. In fig. 3, the actual value is determined by means of the actual measurement time t_aktTo illustrate.

Fig. 2 schematically illustrates the possibility of detecting the operating parameters in a time-discrete manner at time intervals T, respectively. The actual point in time is represented by the time interval kT. The previous time interval is characterized by (k-1) T. The time-dependent behavior is characterized for the input variable. For the actual time interval, the motor current I and the motor voltage U are correspondingly transmitted to the artificial neural network 19 in the form of the actual motor current value I (kt) and in the form of the actual motor voltage value U (kt). In addition, at least one previous motor current I ((k-1) T) and at least one previous motor voltage value U ((k-1) T) are transmitted to the artificial neural network 19 in order to take into account time-dependent behavior.

As already explained, additional input variables can also be submitted to the artificial neural network, as shown in fig. 2 and 3. In addition to the transmission of the actual values of interest, all input variables can also be transmitted with regard to their time-dependent behavior, for example in that a previous value of a previous point in time is submitted (fig. 2) or a time differential is determined by means of the differentiator 23 (fig. 3). In both cases, knowledge about the time-dependent behavior is present in the artificial neural network 19.

The artificial neural network 19 is only schematically presented in fig. 2 and 3. It comprises an input layer 30, at least one hidden layer 31 and an output layer 32. Each of these layers 30, 31, 32 of the artificial neural network 19 may include any number of neurons 33 (fig. 4), depending on the particular architecture. In this input layer, the number of neurons 33 may correspond to the number of input variables transmitted to the input layer 30, for example. In this embodiment, one neuron 33 is provided for each actual value of the input variable and for each submitted previous value (fig. 2) or for each submitted time differential (fig. 3), additional neurons 33 being provided in the input layer 30. According to this example, at least four or five neurons are provided in the input layer 30.

At least one of the following input variables is transmitted to each neuron 33 in the input layer 30: actual values of the motor currents I (kT), I (t)_akt) And/or the actual values n (kT), n (t) of the fan speed_akt) Actual values of the motor voltage U (kT), U (t)_akt) The actual value pa of the output pressure, the previous value I ((k-1) T) or time derivative dI of the motor current I and/or the previous value n ((k-1) T) or time derivative dn of the fan speed n. According to this example, the previous value U ((k-1) T) or the time differential dU of the motor voltage U is also transmitted to the input layer 30. Alternatively, the previous value pa ((k-1) T) (fig. 2) of the output pressure or the time differential dpa (fig. 3) of the output pressure pa may also be transmitted to the input layer 30. The input pressure pe and the delivery of the air humidity h are also optional.

As shown in FIG. 4, at least one value of another layer and/or at least one of the input variables is input at the neuron input with a neuron input value x₁To x_nIs submitted to each neuron 33. Neuron input value x₁To x_nBy means of a correspondingly assigned weight w₁To w_nAre weighted and summed, thereby obtaining a weighted sum x_w. This weighted sum is provided to the activation function F of the neuron 33. The activation function F also depends on the threshold S. In particular, the value of the activation function F is output in the form of a neuron output value y, which depends on the weighted sum x_wSubtract the difference of threshold S: y = xw-S. The triggering of the neuron 33 is predefined by means of a threshold value S.

Instead of considering the threshold S as an input parameter in the activation function F, it is also possible to use so-called ordered neurons (on-neuron), in which the weighted sum x is used_wDuring the calculation of (2) with the neuron input value x₀Takes into account the threshold S.

In fig. 5, one possible activation function F is illustrated by way of example only. According to this example, a rectifier (ReLU) is selected as the activation function F. Other known activation functions for the neuron 33, preferably non-linear, may also be used. In this embodiment, the neuron output value y is limited to the maximum value F of the activation function F_maxFor example 6. Due to the limitation of the maximum value, the rarely occurring features can be better taken into account during the learning of the artificial neural network 19.

The artificial neural network 19 is trained based on known parameters and data and may be used for open or closed loop control after training. During operation, the artificial neural network 19 may be updated. No continuous learning is provided in the preferred embodiment of the fan system 10, as the sensors 21, 22 are preferably not provided.

At least one determined neuron output value y may be fed back from a subsequent layer to a previous layer. This feedback may be implemented within the artificial neural network 19, or may be implemented external to the artificial neural network 19. For example, the artificial neural network 19 may be constructed as a recurrent neural network. According to this example, at least one actual flow value is determined in the hidden layer 31 or alternatively in the output layer 32, which actual flow value is fed back in one of the previous layers, in particular in the input layer 30. In the embodiment shown in fig. 2 and 3, the actual value of the output pressure pa is determined in at least one hidden layer 31 and fed back as an input variable to the input layer 30.

Since the output pressure pa is fed back into the input layer 30, a very good stability of the closed-loop control can be achieved if the other input variables provided are constant. In addition, during the recalculation of the output pressure, an average calculation may be made by using the determined actual output pressure.

If the flow state of the air flow G changes due to an external influence, the fan system reacts by changing the fan speed n, which can be recognized in the change of the motor current I (actual system variable). The trained artificial neural network 19 determines an assigned actual volumetric flow rate value for the volumetric flow rate Q. The regulator 20 then adjusts the motor voltage U (the correcting variable) so as to minimize the control deviation D, which is in turn fed back to the artificial neural network 19.

The regulator 20 may be implemented as a software module and/or a hardware module. The determination of the control deviation and the correction variable by the regulator 20 takes place, according to the example, outside the artificial neural network 19 and, as an alternative, also inside the artificial neural network 19.

During training, different output pressures pa may be adjusted, for example by using venturi nozzles for differential pressure determination, and the artificial neural network 19 may learn based on the value of the output pressure pa without measuring the actual value of the volumetric flow rate Q. As a parameter for training or learning that characterizes the flow state, the output pressure pa is used instead of the volume flow rate Q, so that the volume flow rate sensor can be omitted. The use of a volumetric flow rate sensor during training may occur, if necessary, in order to achieve more accurate training results. Based on this training, the artificial neural network 19 can be adapted with sufficient accuracy so that good control results for controlling the volume flow rate Q are obtained during start-up.

The artificial neural network 19 also reacts to external changes during operation of the wind turbine system 10, which have an effect on operation and effect, e.g., changes in the wind turbine speed n. In the control device 11, a comparison model for the state of the artificial neural network 19 can be stored, which is assigned to known disturbance variables or known environmental parameters. Therefore, by comparing the actual state of the artificial neural network 19 with the comparison model, the control device 11 can also determine whether and what changes have occurred in the environment or system. For example, an increase in the load on the filter or a filter blockage may be determined. Such changes are not abrupt with respect to time-dependent behavior, but rather are slower compared to other external influences. Since the time-dependent change or the chronological order of the values is taken into account, conclusions can be drawn about the kind of external influence, at least with regard to the actual system variables and, where appropriate, one or more additional input variables in the artificial neural network 19. For example, due to a sudden change in the operating state of the fan system 10 (e.g., fan speed n), the opening or closing of a door or damper in the intake volume and/or the output flow volume may be determined.

Since the time-dependent behavior, in particular the motor current I and/or the fan speed n and/or the motor voltage U, is taken into account, the operating point can be unambiguously determined and, for example, the actual value of the volume flow rate Q can be unambiguously assigned to the motor current I. For the illustration in fig. 9, this means that the characteristic line may be selected from the characteristic line group based on the actual value of the fan speed n (measured or calculated from the motor current I or the motor voltage U). Based on the change in the motor current I, it may be determined whether the motor current I is increased or decreased. Accordingly, the actual operating point is in the increasing part of the characteristic line up to the maximum value of the motor current I or in the decreasing part of the characteristic line at a volume flow rate value which is greater than the volume flow rate value characterized by the maximum value of the motor current I.

The present invention relates to a method for operating a fan system 10 and such a fan system 10. The fan system 10 has a control device 11 with an artificial neural network 19. The control device 11 controls an electric motor 13 of the backward-curved centrifugal fan 12. The centrifugal fan 12 generates an air flow G, which is characterized by an actual flow value, in particular an actual value of the volume flow rate Q. This actual flow value is not detected by the sensor device but is determined from the input variables by means of the artificial neural network 19 and, on the basis of this, the electric motor 13 is controlled open-loop or closed-loop by means of the control device 11. The motor current I and the motor voltage U and their time-dependent behavior, which may be a time differential (e.g. a first order gradient) or may be at least one previous value at a previous point in time, are provided to the input layer 30 of the artificial neural network 19. It is particularly advantageous if the artificial neural network 19 determines the actual value of the output pressure pa fed back internally or externally, thus forming an input variable for the input layer 30. Alternatively, additional input variables can be additionally taken into account.

List of reference numerals:

10 blower system

11 control device

12 centrifugal fan

13 Motor

14 rotor

15 Fan blade

19 Artificial neural network

20 regulator

21 rotation speed sensor

22 sensor

23 differentiator

30 input layer

31 coating layer

32 output layer

33 neurons

B expected flow value

D control deviation

dI motor current variation

dU motor voltage variation

dpa output pressure variation

F activation function

F_maxMaximum value of activation function

G gas flow

h air humidity

I motor current

kT actual time point (time dispersion)

n fan speed

pa output pressure

pe input pressure

Q volume flow rate

S threshold

t_aktActual time points (time continuous)

Voltage of U motor

w_iWeight i (i =1-n)

x_i Neuron input value i (i =1-n)

x_wWeighted sum

The y neuron outputs a value.