Method for controlling a compressed air supply device

Abstract

The invention relates to a method for controlling a compressed-air supply device of a utility vehicle, wherein the compressed-air supply device has a compressor and a compressed-air processing unit, wherein current operating data (v, ML, BS, KS, S1, S2a, S2b, S2c) of the vehicle (1) is read in, wherein control signals (S3, S4, S5) are formed from input signals, which contain at least the current operating data, by means of an AI model (50), which control signals contain at least one compressor control signal (S4) for controlling a compressor, preferably also an RV control signal (S3) for controlling a regeneration valve device, wherein the control signals are used to switch between at least two states of the compressed air supply device, in particular between a delivery mode and a regeneration mode, and possibly also other states, for example in the case of multi-chamber air dryers. The AI model can contain, in particular, a neural network (52) having a plurality of layers (L1, L2, L3, L4, L5) and can be pre-trained or self-learning.

Description

[0001] Method for controlling a compressed air supply device

[0002] Background of the invention

[0003] The invention relates to a method for controlling a compressed air supply device, a control unit for carrying out the method and a compressed air treatment unit with such a control unit.

[0004] A compressed air supply system in a commercial vehicle generally comprises a compressor and a downstream compressed air treatment unit (APU). The compressed air treatment unit (APU) in turn comprises an air dryer unit, a regeneration valve device, a control unit, and preferably a multi-circuit protection valve (MCPV) to which consumer circuits of the commercial vehicle are connected. The compressor draws in air and compresses it, so that it is subsequently passed through the air dryer unit, where it is cleaned and dried. The dried compressed air is then fed to an outlet area, to which the consumer circuits, in which the delivered compressed air is stored in compressed air reservoirs, are connected via the multi-circuit protection valve.During conveying phases, the compressed air is conveyed through the air dryer unit to the outlet area; during regeneration phases, the previously conveyed and dried compressed air from the outlet area is returned via the regeneration valve device, preferably partially expanded, and passed through the air dryer unit to remove moisture. Thus, the condition of the air dryer unit and the degree of dryness of the compressed air are adjusted by setting the phases, in particular the time and duration of the regeneration phases. Furthermore, systems with multiple dryer units, e.g. tandem dryers, are known in which it is possible to switch between multiple dryer cartridges so that, for example, one cartridge is in conveying mode while the other is regenerated.

[0005] In commercial vehicles, the compressor can be driven directly by the combustion engine, in particular, and can be mounted directly on the engine shaft of the combustion engine. Furthermore, vehicles with a connected, e.g., electrically driven compressor are known, e.g., commercial vehicles with hybrid drive. Energy is thus generated during the delivery phases, in particular by connecting the compressor to the drive train of the combustion engine. During the regeneration phases, energy is lost due to the return of the previously delivered, dried compressed air.

[0006] Air management control systems generally operate based on predefined parameters and are installed as programs in the control unit of the compressed air treatment unit. The control unit controls the compressor and valves of the multi-circuit protection valve. The control unit or program can read various sensors of the compressed air system and the vehicle, e.g., current pressure readings from the compressed air reservoirs, and also receive data via the vehicle's internal data bus. Delivery phases can be defined depending on the pressure values ​​of the compressed air reservoirs, so that, for example, a delivery phase is initiated when the pressure falls below a lower limit.

[0007] To control complex processes, artificial intelligence (AI) models have recently been used. These models involve self-learning processes, whereby input signals are then used to optimize output signals based on mathematical models. An example of such AI models are neural networks, which have multiple layers, each with multiple nodes. The nodes in one layer receive input signals and use the input signals to create a mathematical function or mathematical operation, the output value of which they output to the nodes in the subsequent layer. After one or more middle or hidden layers, the acquired data is passed to the output nodes of an output layer. The output nodes represent, for example, probabilities for detecting a state or a probability for an advantageous state.

[0008] Description of the invention

[0009] The invention is based on the object of creating a method for controlling a compressed air supply system that enables flexible adaptation to driving situations, particularly with low energy consumption. This object is achieved by a method according to claim 1. The subclaims describe preferred developments. Furthermore, a control unit for implementing the method, a compressed air treatment unit with the control unit, a commercial vehicle with the compressed air treatment unit, and a fleet of commercial vehicles are provided.

[0010] Thus, according to the invention, input signals are received which comprise current operating data or driving dynamics data of the vehicle, and control signals for controlling the compressed air supply device are formed from the input signals via a Kl model, wherein the control signals comprise at least compressor control signals for controlling the compressor.

[0011] This alone achieves a number of advantages. For example, control can be carried out flexibly depending on the dynamic operating data of the vehicle. Because the input signals are received and processed by a Kl model, several different signals can be processed flexibly and efficiently. In this way, a Kl model can learn or develop control depending on relationships that cannot initially be incorporated into a fixed program by a user. For example, a Kl model can determine from the operating data that a phase with higher fuel consumption and lower compressed air consumption by the brakes is currently taking place, without a user explicitly defining this as uphill driving; in this case, the Kl model can recognize orlearn that such a phase is highly likely to be followed by a downhill section within a relevant driving distance, enabling cost-effective compressed air extraction by creating delivery phases during the engine's overrun phases. Thus, the Kl model can automatically determine that during an uphill phase, shorter periods of time are used for delivery phases and that the compressed air stored in the compressed air reservoirs can also be reduced.

[0012] Thus, according to one embodiment, the Kl model can advantageously determine that a lower limit of the compressed air reservoir pressure can be reduced when driving uphill, wherein the self-learning model can dynamically determine an adjustment depending on the properties of the respective gradient, ie length, angle of inclination.

[0013] According to a preferred embodiment, the Kl model outputs output values ​​in its output layer that correspond to the states of the compressed air supply system to be set. Thus, the output values ​​can be directly converted into suitable control signals, in particular for a switching state of the compressor.

[0014] The compressor can be switched directly; furthermore, the compressor control signal can also be output to a clutch for engaging and disengaging the compressor, or to an unloader valve device for pneumatically controlling the compressor.

[0015] According to a preferred embodiment, the determined control signals not only control the compressor, but also at least one regeneration valve device. Thus, the compressed air supply device can be switched between at least

[0016] - a delivery mode as a first state in which the compressor is switched on and the regeneration valve device is in a first state, in particular basic state, so that compressed air is delivered and dried, and

[0017] - a regeneration mode as a second state in which the compressor is switched off and the regeneration valve device is in a second state, in particular an actively switched state, in order to regenerate the air dryer.

[0018] - preferably furthermore an idle mode, wherein the air dryer can also have, for example, several chambers which are operated in the same or different ways, so that the regeneration valve device can also be switched, for example, such that one chamber is in conveying mode and the other in regeneration mode, and / or one chamber is in idle mode.

[0019] This enables control of all relevant components of the compressed air supply system, whereby, in particular, the output values ​​of the Kl model can be directly assigned to the two or three modes. In addition, a multi-circuit protection valve or output valves of the multi-circuit protection valve can be switched, so that compressed air backflow from the compressed air reservoirs of the consumer circuits, especially between the consumer circuits, as well as, for example, the successive connection and filling of the consumer circuits, can be controlled by the Kl model.

[0020] One or more of the following data can be recorded and used as operating data or driving dynamics data: driving speed, engine load, engine speed, brake signal (i.e. brake on, off), clutch signal (i.e. clutch engaged, disengaged), measured or indirectly determined pressure signals, in particular pressure signal in the compressed air treatment unit, and / or pressure signals of the consumer circuits, an inclination signal of an inclination sensor, if present.

[0021] For example, uphill driving can be determined from a current driving situation, e.g., with high engine load, low engine speed, and low compressed air consumption due to braking, as well as, if applicable, a transmission signal or a gear selection signal, even if no incline signal is present. The Kl model preferably learns the advantageous gearshift indirectly from this data without explicitly recording this condition as uphill driving. The Kl model can therefore, for example, learn that in such situations a small supply of compressed air needs to be placed in the compressed air reservoirs, since a downhill drive with overrun phases will soon occur. Accordingly, a downhill drive with overrun phase, for example, can be detected even without an incline signal.

[0022] In particular, the engine load and engine speed determine whether a drive phase, overrun phase or sailing mode (disengaged state) is present,

[0023] Thus, for example, a coasting phase can be detected when there is a low engine load and a high engine speed, and the clutch signal indicates that the clutch is engaged.

[0024] Furthermore, the current and / or previous states of the compressed air treatment unit and / or the compressor can also be recorded as input signals. According to a preferred embodiment, one or more external signals or data are also recorded as input signals, in particular from the group that includes:

[0025] Outside temperature, outside humidity, outside air pressure, position data, e.g. GPS data in a Global Positioning System and / or in a map, route data, traffic data.

[0026] The external data can be recorded, for example, via a telecommunications interface or determined by measurements in the commercial vehicle itself, such as the outside temperature and the position of the vehicle.

[0027] The external data helps to better adapt the control system to the ambient conditions. The invention recognizes that this integration is possible with little effort and is very effective, since outside temperature, outside air pressure, and outside humidity, in particular, have a direct influence on the moisture absorbed. Data on the position and route, in particular, improves the adjustment of consumption and the division of phases along the respective route.

[0028] The data about one or more routes can be stored or transmitted via remote data transmission.

[0029] According to a preferred embodiment, at least some of the input signals are recorded as time-dependent values. The input signals can be recorded with their respective input signal histories, in particular as functions over time. For this purpose, time can be divided into successive time steps, and at least some of the input signals can be recorded with the values ​​they have in the successive time steps.

[0030] This allows for better consideration of the temporal behavior of the input signals. This allows for the evaluation of how long certain states or input variables existed and thus contributed to the current state. For example, the duration of a specific driving state can be evaluated. The input signals can be recorded, for example, as a function over the time steps or as a tuples of numbers over the time steps. Furthermore, or alternatively, only a value in a previous time step can be recorded.

[0031] For example, the previous compressor state or the state of the compressed air treatment unit or the time duration, ie number of time steps, of the compressor state or the state of the compressed air treatment unit can be included in the evaluation by the Kl model.

[0032] Two different models can be used as Kl models:

[0033] A pre-trained model, where a machine learning model is developed, e.g., for a specific bus route and / or route, which learns and stores the typical vehicle behavior on this specific route, with the corresponding specific air consumption. This model can thus, for example, minimize fuel consumption and / or wear, e.g., of the valves. Such a pre-trained model can, for example, be learned in advance depending on the input signals, so that the algorithm thus determined can subsequently be directly applied to identical or similar vehicles on the same route. Possible use cases for a pre-trained model are = a specific route and / or = a specific bus route, and / or

[0034] = motorway traffic in a defined area, e.g. in Europe.

[0035] An active self-learning model, also called an active machine learning model, can preferably actively adjust the control strategy from the input signals. Such a self-learning model can determine the driver's specific driving style and / or typical uses of the compressed air braking system, e.g., depending on road characteristics, driving speed, or gradient, and adjust the control strategy, in particular to achieve desired requirements. A desired requirement can, for example, be a reduction in energy consumption, in particular fuel consumption, i.e., an ecological objective, and / or an economic objective such as considering and reducing wear, e.g., a valve service life that is limited by switching processes.The self-learning system can be implemented in the control unit of the compressed air supply system itself; alternatively, part or all of the model can also be implemented outside the compressed air supply system, e.g., in a central control unit of the vehicle, and in particular also in a computer system or control system outside the vehicle, so that input signals are transmitted to the external central unit by remote data transmission via a telecommunications interface, processed there with the self-learning model, and subsequently output signals or control signals are transmitted back to the vehicle.

[0036] In the embodiment with at least partially external detection, the central unit can, in particular, record and process data from multiple vehicles, enabling self-learning in a variety of different situations with different input signals. The results can, in turn, subsequently be offered to all vehicles.

[0037] The Kl model can, in particular, comprise a neural network having multiple layers. The neural network can, in particular, comprise an input layer with multiple nodes (neurons) that receive the input signals. Furthermore, an output layer is provided, the output nodes of which each correspond to a control signal or an output state. For example, two output nodes can be provided that indicate different switching states, e.g., compressor on, compressor off, so that the output values ​​can be used directly as control signals to control the compressor. In an embodiment with the control of both the compressor and the regeneration valve device, the output signals can also indicate states or phases, e.g.,Delivery mode, regeneration mode, so that, depending on the output values, both the compressor and the regeneration valve device, and if necessary, the valves of a multi-circuit protection valve, are controlled. More than two output values ​​can also be set here, e.g., an additional idle mode with the compressor switched off and the regeneration valve device in the default state, i.e., without regeneration. Furthermore, the output values ​​can also indicate changes or dynamic processes, e.g., "maintain switching state of the compressed air supply device, change switching state."

[0038] Thus, a control unit for carrying out the method, furthermore a compressed air treatment unit with the control unit and a corresponding commercial vehicle and in particular also a fleet of several commercial vehicles can be provided, which can be controlled centrally with common data processing.

[0039] The invention is explained in more detail below with reference to some embodiments in the accompanying drawings. They show:

[0040] Fig. 1 shows a compressed air system of a commercial vehicle according to an embodiment of the invention;

[0041] Fig. 2 a block diagram of relevant system elements of the commercial vehicle,

[0042] Fig. 3 is a circuit diagram of a multi-circuit protection valve device according to the embodiment;

[0043] Fig. 4 the commercial vehicle while driving on a route;

[0044] Fig. 5 shows a neural network according to an embodiment, the neural network being suitable for use in the method according to the invention;

[0045] Fig. 6 shows a node of a middle level of the neural network of Figure 5 and a mathematical operation applied in the node,

[0046] Fig. 7 shows a simplified representation of the layers and nodes of the neural network compared to Fig. 5; and

[0047] Fig. 8 is a flow chart of a method according to the invention.

[0048] According to Figure 1, a commercial vehicle 1 has a compressed air system 2 with a compressed air supply device 15 (shown here in dash-dotted lines) and a consumer stage 4 with multiple consumer circuits 14a to 14d. Furthermore, according to Figure 2, the commercial vehicle 1 has, among other things, a drive system 5 with an engine control unit 27 and a transmission control 28, as well as at least one internal vehicle data bus 6. The compressed air supply device 15 in turn has a compressor 7 and a compressed air processing unit (APU) 3 (shown in dashed lines) connected to the compressor. The compressed air supply device 3 in turn has an air dryer unit 8, a regeneration valve device 9, and an electronic control unit (APU-ECU) 10 connected to the vehicle bus 6. The air dryer unit 8 is connected with its dryer inlet 8a to the outlet of the compressor 7, and with its dryer outlet 8b via e.g.a check valve 13 is connected to an output area 11 of the compressed air treatment unit 3, to which in turn a multi-circuit protection valve 12 is connected, to which the consumer stage 4 with several consumer circuits 14a, 14b, 14c, 14d is connected.

[0049] As consumer circuits 14a to 14d, for example, a first service brake circuit 14a, a second service brake circuit 14b, a parking brake circuit 14c, e.g. combined with a trailer brake circuit, as well as a further consumer circuit 14d for auxiliary consumers or transmission divider can be connected. Compressed air reservoirs 18a, 18b are connected to some or all of the consumer circuits 14a, 14b, e.g. a first compressed air reservoir 18a to the first service brake circuit 14a and a second compressed air reservoir 18b to the second service brake circuit 14b. An internal pressure sensor 22a is connected to the dryer outlet 8b, and / or an internal pressure sensor 22b is connected to the output area 11. The internal pressure sensor 22a and / or 22b outputs an internal pressure signal S1 to the electronic control unit 10. Furthermore, for example, B. pressure sensors are connected to the service brake circuits 14a, 14b or their compressed air reservoirs 18a, 18b, thus e.g.a first consumer circuit pressure sensor 23a and a second consumer circuit pressure sensor 23b, which also output their pressure measurement signals S2a and S2b to the electronic control unit 10, or an external control unit.

[0050] Additionally, a humidity sensor 24 can be provided in a consumer circuit, e.g., the first service brake circuit, which outputs a humidity measurement signal S6 to the electronic control unit 10; Accordingly, the humidity sensor 24 can also be provided in the compressed air supply device, e.g., at the dryer outlet 8b, or in a return path 31 from the regeneration valve device 9 to the dryer outlet 8b.

[0051] In the feedback path 31, for example, a throttle 25 and a second check valve 26 can be provided in a manner known per se, so that the pressure sensor 22a is connected, for example, in the feedback path 31 between these elements.

[0052] The electronic control unit 10 controls the compressor 7 with a compressor control signal S4 and the regeneration valve device 9 with an RV control signal S3. In a delivery phase, the compressor 7 delivers compressed air 30 through the dryer inlet 8a, the air dryer unit 8 and its dryer outlet 8b, as well as the check valve 13, to the outlet area 11 of the compressed air treatment unit 3, so that it can be made available to the multiple consumer circuits 14a, 14b, 14c, 14d via the multi-circuit protection valve 12 of the consumer stage 4 and stored in their compressed air reservoirs 18a, 18b. In a fundamentally known manner, the multi-circuit protection valve 12 can open the consumer circuits 14a to 14d successively or with different inlet pressures, for example, to preferentially supply air to the service brake circuits.

[0053] In this case, for example, according to Fig. 3, MCPV control signals S5 can be output from the electronic control unit 10 to valves V1, V2, V3, V4 of the multi-circuit protection valve 12, which are assigned to the consumer circuits 14a to 14d.

[0054] The electronic control unit 10 switches from such a conveying phase to a regeneration phase by outputting an RV control signal S3 to the regeneration valve device 9 and a compressor control signal S4 directly or indirectly to the compressor 6 in order to switch off the compressor 6; thus, the previously stored compressed air 30 from the outlet area 11 and / or the compressed air reservoirs 18a, 18b can reach the dryer outlet 8b via the regeneration path 31, i.e. via the switched regeneration valve device 9, the throttle 25, the second check valve 26, and also through the air dryer unit 8 and the air dryer inlet 8a to a compressed air outlet 32 ​​in order to regenerate the air dryer unit 8.

[0055] In a commercial vehicle 1 with an internal combustion engine as the drive system 5, the compressor 6 can be provided directly on the engine shaft. The compressor 6 can be switched over, for example, by switching to idle mode or switching off via a clutch, for example; furthermore, indirect control can be achieved by outputting the compressor control signal S4 not directly to the compressor, but rather to a pneumatic valve of an unloader unit, whose pneumatic output signal controls a pneumatic control input of the compressor 7.

[0056] The electronic control unit 10 receives input signals, which include current operating signals, in particular the pressure signals S1 of the compressed air treatment unit 3 and the pressure signals S2a, S2b of the consumer stage 4, as well as, if applicable, a humidity measurement signal S6. Furthermore, the electronic control unit 10 receives external data or external signals via the vehicle's internal data bus 6.

[0057] In particular, the electronic control unit 10 can record current operating data from the group that contains:

[0058] Driving speed v, engine load ML, engine speed, brake signal BS, clutch signal KS, pressure signals, in particular the pressure signals S1 in the compressed air treatment unit, and / or the pressure signals S2a, S2b of the consumer circuits, inclination signal of an inclination sensor, the humidity measurement signal S6.

[0059] Accordingly, the electronic control unit 10 can receive external signals from the group that includes:

[0060] Outside temperature (Ta), outside air humidity, outside air pressure, position data, e.g. GPS data in a Global Positioning System and / or in a map, route data, traffic data.

[0061] Furthermore, according to Fig. 2, a data memory 100 for, for example, external data or vehicle-specific data is connected directly or indirectly to the electronic control unit 10, wherein the data memory 100 can also be connected, for example, via the vehicle-internal data bus 6.

[0062] According to Fig. 4, the commercial vehicle 1 travels on a roadway 33 along a route 34 or route, wherein the route 34 may, for example, have uphill gradients 35 and downhill gradients 36. The commercial vehicle 1 can, via its transmission control 28 and the engine control device 27, for example, set a driving mode with fuel consumption, and also set an idle and a coasting mode with the engine engaged, wherein the compressor 7 is passively driven in the coasting mode; furthermore, the compressor 7 can also be passively driven, for example, during braking on a straight stretch of road by driving the compressor 7 instead of controlling the service brake circuits 14a, 14b. The route 34 can be stored in the commercial vehicle 1 or in a memory of the commercial vehicle 1, or it can be transmitted to the vehicle 1 as route data by remote data transmission using the telecommunications device 40. The vehicle 1 initially travels on a level stretch 37 and, for example,subsequently on a gradient 35. The gradient 35 can be determined, for example, via an inclination sensor 42 in the vehicle 1, or from the engine load ML, which is transmitted from the engine control unit 27 of the drive system 5 via the vehicle's own data bus 6.

[0063] The data processing can take place in the commercial vehicle 1 itself, ie in particular in the electronic control unit 10, or in the external location 120, in that the electronic control unit 10 transmits input signals received via the telecommunication interface 40 and receives calculated data, in particular the control signals S1 to S5, again from the external location 120.

[0064] The evaluation of the input signals, ie in particular the operating signal S1, the pressure signals S2a, S2b, S3, S6, the driving speed v, engine load ML, engine speed, brake signal BS, clutch signal KS, is carried out via a Kl model 50, which in particular has a neural network 52, which determines the control signals S3, S4, S5 from the input signals and subsequently outputs them.

[0065] The Kl-Model 50 can be pre-trained, ie with programmed presetting, whereby it was thus trained in a learning phase, which can also have taken place in another, similar vehicle, for example.

[0066] Furthermore, the Kl model can be actively self-learning and thus continuously develop itself.

[0067] In the example of driving on route 34, the control model 50 has learned, for example, that after the gradient 35, particularly depending on the gradient angle α and the length of the gradient 35, within, for example, a distance X35, X36, there will be another gradient 36. This has the following effect on the control of the compressed air treatment unit 3:

[0068] - When driving on a gradient of 35°, compressed air consumption is considered to be low, since the driving speed v is lower and a deceleration desired by the driver on a gradient of 35° is often possible without using the brakes, and / or

[0069] - During travel on the downhill gradient 36, which occurs with a certain probability W35, W36 subsequently, for example, within a distance X35, overrun operation is possible, so that in the vehicle 1, in which the compressor 7 is arranged on the motor shaft of the drive system 5, a delivery phase of the compressor 6 can be set solely by the downhill gradient 36, without additional fuel consumption; the operated compressor 6 supports the effect of the engine brake.

[0070] In particular, these relationships between the incline 35 and the decline 36, and the other parameters of the angle, distance X35, etc., are not explicitly stored in the electronic control unit 10; however, the at least statistical relationship is stored by the previous learning phase (pre-training) or as a learned relationship (active learning), so that the electronic control unit 10, for example, does not unnecessarily fill the compressed air reservoirs 18a, 18b, since the overrun mode can presumably be used subsequently to cost-effectively fill the compressed air reservoirs 18a, 18b.Furthermore, the electronic control unit 10 can switch to a regeneration phase by outputting a compressor control signal S4 and a regeneration valve control signal S3 in order to regenerate the air dryer unit 8 with the currently stored compressed air of the compressed air reservoirs 18a, 18b, so that the compressed air reservoirs 18a, 18b can then subsequently be filled cost-effectively in the gradient 36.

[0071] For example, a lower pressure limit in one or more of the pressure signals S2a to S2d, at which a filling process is initiated, can be adjusted depending on the determination, so that, for example, during a long uphill journey, the lower pressure limit for the service brake circuits 14a, 14b is lowered. According to a preferred embodiment, the free determination of the pressure limit values ​​can only be freely selected within a non-safety-relevant range. If the pressure exceeds such pressure limit values, which can be specified, for example, by legal frameworks, a normal algorithm can preferably take over control according to fixed rules, e.g., in an emergency.

[0072] Furthermore, the valves V1, V2, V3, V4 shown in Fig. 3 can be controlled in the multi-circuit protection valve 12, for example, to control a filling process in the delivery phase, so that, for example, the service brake circuits 14a, 14b are filled with priority or subordination, depending on whether a high braking requirement is to be expected

[0073] Thus, such data processing can be performed by the Kl model 50 based on the current input signals. Furthermore, or alternatively, the Kl model 50 can also be controlled based on previously known route data, i.e., the input data for route 34 with gradients 35 and 36 are used to output the control signals S3, S4, and S5.

[0074] A further embodiment relates to the travel of a vehicle 1 as a bus, e.g., a scheduled bus, along a fixed route 34. Here, too, the Kl model 50 learns from the driving data, e.g., driving speed v, engine load ML, ..., a specific behavior of the vehicle 1, which can be generally assumed, so that, for example, the imminent approach to a bus stop can be recognized from the driving data, at which, for example, a rolling brake or automatic holding of the service brake occurs. Furthermore, however, a fixed route, e.g., a fixed bus line, can also be learned. In such an embodiment, a pre-trained system can therefore also be used, in particular.

[0075] The structure of the Kl model 50 is shown by way of example in Figure 5: The neural network 52 has an input layer L1 with first nodes N 1 -i, with i=1 to i=8, i.e. N1-1 to N1-8, which each receive an input signal, such that the input layer L1 can receive eight input signals, e.g. the above-mentioned input signals v, ML, BS; KS, S1, S2a, S2b, S2c, S2d. The input signals are recorded in particular as time-dependent values. In this case, the input signals with their respective input signal history can be recorded in particular as functions over time. For this purpose, time can be divided into successive time steps and at least some of the input signals can be recorded with the values ​​they have in the successive time steps.

[0076] This allows for better consideration of the temporal behavior of the input signals. This allows for the evaluation of how long certain states or input variables existed and thus contributed to the current state. For example, the duration of a specific driving state can be evaluated. The input signals can be recorded, for example, as a function over the time steps or as a tuples of numbers over the time steps. Furthermore, or alternatively, only a value can be recorded in a previous log.

[0077] For example, the previous compressor state or the state of the compressed air treatment unit or the time duration, ie number of time steps, of the compressor state or the state of the compressed air treatment unit can be included in the evaluation by the Kl model.

[0078] Furthermore, a second layer L2, a third layer L3, a fourth layer L4 and an output layer L5 are provided, so that the layers L2, L3, L4 form intermediate layers or hidden layers.

[0079] The second layer L2 has second nodes N2-j, e.g., with j=1 to 22, which are each connected to the first nodes N1-i, i = 1 to 8. The neural network 52 of this embodiment is fully connected, i.e., all first nodes N1-i are connected to all second nodes N2-j. The second nodes N2-j each perform a mathematical operation, which is described further below. The third layer L3 has third nodes N3-k, e.g., with k=1 to 22, which are each connected to the second nodes N2-j. The output layer L5 then has two fifth nodes or output nodes N5-m, i.e., with m=1, 2, which are each connected to all fourth nodes N4-L, which correspond to a state or state to be set of the compressed air treatment unit 3. In the embodiment with control of only compressor 7, these are, for example, the states with N5-1 as “compressor on” and N3-2 as “compressor off”, so that the value orthe result of the output layer L5 is output as compressor control signal S4 to compressor 7.

[0080] In Fig. 5, the computation in layers L2, L3, L4, and L5 is indicated by lines between the consecutive nodes and by indicating activation functions AF2, AF3, and AF5 in layers L2, L3, and L5. These inter-layer computations are shown in more detail in Fig. 6 and Fig. 7 for simplified layers LL1, LL2, and LL3 with fewer nodes NN1, NN2, and NN3 and partially simplified indexing:

[0081] According to Figs. 6 and 7, input signals xi are fed to the first nodes NN1-i, where i = 1, 2, of the first layer LL1. In Fig. 6, n first nodes NN1-i are provided, ie, i = 1 ... n; in Fig. 7, two first nodes NN1-i are provided, ie, i = 1 or i = 2.

[0082] These input signals xi are subsequently fed from the first nodes NN 1 -i to the individual second nodes NN2-j. For the sake of simplicity, Fig. 6 shows the connection of the plurality of first nodes NN 1 -i to a single node NN2. Between the first nodes NN 1 -i of the first layer LL1 and the second nodes NN2-j of the second layer LL2, first connections C1-i are provided which, as described below, serve to pass on the values ​​and signals between the layers LL1 and LL2. According to the illustration in Fig. 6, the data xi, i.e. x1 to xn, are each multiplied by a specific weighting or weight wi in the second node NN2 and then added, i.e. the sum Sigma is formed, preferably with a constant additive bias b, which here has the bias value WO.

[0083] Thus, the sum Sigma is formed as sum E = WO + W1 * x1 + W2 * x2 + ... + W1* xn. or with double indexing for the first nodes of the first layer and the index i, the sum Sigma is formed as

[0084] Sum S = WO + W1,1 * x1 + W1,2 * x2 + ... + Wn,1* xn.

[0085] Subsequently, a non-linear activation function AF2 is applied to this sum E, which represents a linear combination of the input signals, e.g., with AF2 as tanh, ie the hyperbolic tangent, or the ReLU function, which assigns zero as the output value to all negative input values ​​up to zero and a linear function, e.g., the identity function y=x, to all positive values.

[0086] Thus, the respective second node NN2-j receives an output value formed in this way, which it outputs to the subsequent third layer LL3. Accordingly, the other nodes N2-j of the second layer LL2 receive different weights and also a different bias b. A different activation function AF2,j can be provided for each second node NN2-j, or the activation functions AF2 for the second layer LL2 can also be the same.

[0087] Examples of non-linear activation functions are:

[0088] Hyperbolic tangent. ReLU function, ELU function, Swish function, Sigmoid function, and as activation functions for the output layer, especially the softmax function or the argmax function.

[0089] This is also shown accordingly in Fig. 7 for a neural network 52 with three layers LL1, LL2, LL3, where the multiple connections and weightings of the first node NN1 are simplified as W1 and the multiple connections and weightings of the second node NN2 are simplified as W2, correspondingly with a bias b1 and b2. In the second layer LL2, as also indicated in Fig. 5, the sum is first formed and then the activation function AF is applied, for example, as a ReLU function, which is indicated as a double symbol for the nodes NN2 and NN3. In the output layer LL3 with the two third nodes NN3, probability values ​​are initially obtained, to which the softmax function is applied as the activation function AF, which has the value 0 or 1 as the output, so that a compressor control signal S4 is output as S4=0 or S4=1. In the mathematical description, the input values ​​therefore represent vectors orone-dimensional matrices that are mapped to vectors, i.e. the first mathematical sub-operation of the linear mapping represents a matrix.

[0090] In Fig. 7, a two-vector x is mapped to a three-vector, ie, with a 3x2 matrix.

[0091] Thus, the following example calculation results in the neural network 52 of Fig. 7. Here, too, a simplified representation is shown, since time information is preferably added, e.g., over successive time steps TS. The units are adjusted accordingly. For example, the weights Wi are generally not assigned a unit. For example, the unit for the vehicle speed v is km / h, and the unit for the pressure is bar. y - softmax (b2 + ReLU (61 + xx MZ4) xy / 2) x = (30, 10)

[0092] Z?1 = (3, -4, 7)

[0093] Ö2 = (1, -5) The result y of the example calculation for Fig. 7 is a probability distribution that aims to turn on compressor 7 with approximately 100% probability (the left vector element represents ON; the right one represents OFF). The selection can be done via a sampling operation, where a random action is drawn with the specified probabilities of the distribution. Alternatively, an argmax operation can be used, which returns the action with the highest value, in contrast to the max operator, which only returns the highest value, to select the action of the node Ni with the highest probability value.

[0094] As an alternative to the embodiment shown, a not-fully connected neural network or a neural network with several hidden layers can also be used.

[0095] Alternatively, in the embodiment shown with two output values, instead of the states "compressor on" and "compressor off", the states "change compressor state" and "do not change compressor state" can also be output; according to a further embodiment, three states "compressor on", "compressor off", and "compressor idle" can also be controlled. In air dryers with two or more chambers, one of several possible states can also be output accordingly.

[0096] In the version with control

[0097] - both the compressor 7 via the compressor control signal S4 a

[0098] - as well as the regeneration valve device 12 via the RV control signal RV, the RV control signals S3 are also output accordingly.

[0099] Thus, the neural network 52 serves for the entire control of the compressed air treatment unit 3, ie different phases or modes of the compressed air treatment unit 3 are set, in particular:

[0100] - a delivery phase or delivery mode with the compressor 7 switched on and the regeneration valve device 9 in the basic position, - a regeneration phase or regeneration mode with the compressor 7 switched off or the compressor 7 in the idle phase and the regeneration valve device 9 switched over, in which compressed air from a compressed air reservoir, e.g. from the compressed air reservoirs 18a, b of the consumer circuits 14a, 14b, is thus returned through the air dryer unit 8 via the regeneration line and the regeneration valve device 9, the throttle 25 and the second check valve 26 and is output via the compressed air outlet 32,

[0101] - in the case of multi-chamber air dryers, also combined conveying phases and regeneration phases.

[0102] In the method according to the invention, according to the flow chart in Fig. 8, after the start in step ST0, input signals with current operating data P1, P2, n, ML of the commercial vehicle 1 are subsequently read in in the recording step ST1, then in a calculation step ST2, control signals (S3, S4, S5) are formed from input signals by means of the Kl model 50 with the neural network 52, which control signals contain at least compressor control signals S4, and in the output step ST3 the control signals, e.g. the compressor control signals S4 and RV signals S3, and optionally MCPV signals S5, are output.

[0103] In an active learning Kl model 50 or neural network 52, parameters of the neural network 52 are then changed in a step ST4, or even after several runs, by, for example, changing the weights Wi.

[0104] Reference symbol (part of the description)

[0105] 1 commercial vehicle

[0106] 2 compressed air system

[0107] 3 compressed air processing unit, APU, air processing unit

[0108] 4 Consumer level

[0109] 5 Drive system

[0110] 6 vehicle-internal data bus, e.g. N-Bus

[0111] 7 Compressor

[0112] 8 Air dryer unit

[0113] 8a Dryer inlet

[0114] 8b Dryer outlet

[0115] 9 Regeneration valve device

[0116] 10 electronic control unit APU-ECU

[0117] 11 Exit area

[0118] 12 Multi-circuit protection valve, MCPV

[0119] 13 Check valve

[0120] 14a, 14b, 14c,14d Consumer groups

[0121] 14a first service brake circuit

[0122] 14b second service brake circuit

[0123] 14c Parking brake circuit

[0124] 14d Secondary consumer group

[0125] 15 Compressed air supply device

[0126] 18a, 18b Compressed air storage

[0127] 22a, 22b internal pressure sensor in the APU 3

[0128] 23a first consumer circuit - pressure sensor

[0129] 23b second consumer circuit pressure sensor

[0130] 23c, 23d additional consumer circuit pressure sensors

[0131] 24 Humidity sensor

[0132] 25 Throttle

[0133] 26 second check valve

[0134] 27 Engine control unit 28 Transmission control

[0135] 30 compressed air

[0136] 31 Return path

[0137] 32 compressed air outlet

[0138] 33 Roadway

[0139] 34 Route

[0140] 35 gradient

[0141] 36 gradients

[0142] 37 flat route

[0143] 40 Telecommunications interface, telecommunications device

[0144] 42 Tilt sensor

[0145] 50 Kl model

[0146] 52 neural network NN

[0147] 100 data storage

[0148] 120 external central office

[0149] AF2 activation function for the second node K2

[0150] AF3 activation function for a third node K3

[0151] AF5 activation function for a fifth node K5

[0152] AF activation function in Fig. 7 b1, b2 bias

[0153] BS brake signal

[0154] KS coupling signal

[0155] ML engine load

[0156] S1 Pressure signal of the pressure sensor 22a, 22b

[0157] S2a Pressure signal of the first consumer pressure sensor 23a

[0158] S2b Pressure signal of the second consumer pressure sensor 22b

[0159] S2c, S2d pressure signals of the other consumer circuits

[0160] S3 RV control signal to the regeneration valve device 9

[0161] S4 Compressor control signal

[0162] S5 MCPV control signal to the multi-circuit protection valve 12

[0163] S6 Humidity measurement signal of the humidity sensor 24 TS time step

[0164] L1 input layer in Fig. 5

[0165] L2 second layer, hidden layer in Fig. 5

[0166] L3 third layer, hidden layer in Fig. 5

[0167] L4 fourth layer, hidden layer in Fig. 5

[0168] L5 fifth layer, output layer in Fig. 5

[0169] LL1 input layer in Fig. 6, 7

[0170] LL2 second layer, hidden layer in Fig. 7

[0171] LL3 Third layer, output layer in Fig.7

[0172] N 1 -i, with i = 1 - 8 first nodes of the input layer L1 in Fig. 5

[0173] N2-j second nodes of the second layer L2 in Fig. 5

[0174] N3-k third nodes of the third layer L3 in Fig. 5

[0175] N4-L fourth node of the fourth layer L4, in Fig. 5

[0176] N5-m output node of the output layer L5, in Fig. 5

[0177] NN1 first node of the input layer LL1 in Fig. 6, 7

[0178] NN2 second node of the second layer LL2 in Fig. 6, 7

[0179] NN3 output node of the output layer LL3, in Fig. 6, 7

[0180] V1, V2, V3, V4 valves of the multi-circuit protection valve 12

[0181] Wi, W1 weighting or weight

[0182] ST1 - ST4 process steps

Claims

Patent claims 1. Method for controlling a compressed air supply device (15) of a commercial vehicle (1), wherein the compressed air supply device (15) has a compressor (7) and a compressed air processing unit (3), wherein current operating data (v, ML, BS, KS, S1, S2a, S2b, S2c) of the vehicle (1) are read in (ST1), from input signals (xi) containing at least the current operating data ((v, ML, BS, KS, S1, S2a, S2b, S2c)), control signals (S3, S4, S5) are formed by a Kl model (50), which contain at least compressor control signals (S4) (ST2), and the compressor control signals (S4) are output for controlling the compressor (7) (ST3), wherein the control signals (S3, S4, S5) are used to switch between at least two states of the Compressed air supply device (15) is switched on.

2. Method according to claim 1, characterized in that at least two output values ​​are generated from the input signals (xi) by the Kl-Model (50), wherein the output values ​​each correspond to one of the states of the compressed air supply device (15).

3. Method according to claim 1 or 2, characterized in that the compressor control signal (S4) indicates that - the compressor (7) is switched on or off, and / or - a switching state of the compressor (7) is changed.

4. Method according to one of the preceding claims, characterized in that the compressed air treatment unit (3) has an air dryer unit (8) and a regeneration valve device (9), wherein the Kl model (50) further forms an RV control signal (S3), (ST2), which is output to the regeneration valve device (9), (ST3) wherein the control signals (S3, S4, S5) switch the compressed air supply device (15) between at least - a delivery mode as a first state, in which the compressor (7) is switched on and the regeneration valve device (9) is in a first state, in particular ground state, and - a regeneration mode as a second state in which the compressor (7) is switched off and the regeneration valve device (9) is in a second state.

5. Method according to claim 4, characterized in that an idle mode is further provided as the third state, in which the compressor is switched off and the regeneration valve device (9) is in its first state.

6. The method according to claim 4 or 5, characterized in that the air dryer unit (8) has at least two dryer chambers which can be switched into different modes by the regeneration valve device (9), in particular independently of one another into a conveying mode, regeneration mode and / or idle mode, wherein the different modes of the dryer chambers represent different states of the compressed air treatment unit (3), and three or more states can be set by the Kl model, wherein at least in some of the states the dryer chambers are switched differently, e.g. in reverse to one another.

7. Method according to one of the preceding claims, characterized in that the compressed air processing unit (3) further comprises a multi-circuit protection valve (12) which has a plurality of outputs for consumer circuits (14a, 14b, 14c, 14d) and an output valve (V1, V2, V3, V4) at at least one of the outputs, wherein an MCPV control signal (S4) is further formed by the Kl model (50) (ST2) and the at least one output valve (V1, V2, V3, V4) is switched by the MCPV control signal (S4) (ST3).

8. Method according to one of the preceding claims, characterized in that the operating data comprise one or more of the following data: driving speed (v), engine load (ML), engine speed, brake signal (BS), clutch signal (KS), acceleration, Pressure signals, in particular pressure signal (S1) in the compressed air treatment unit (3), and / or pressure signals (S2a, S2b, S2c, S2d) of the consumer circuits, inclination signal of an inclination sensor, a state of the compressed air supply device (15) and / or a switching state of the compressor (7) in a previous time step (TS).

9. Method according to one of the preceding claims, characterized in that external signals are further received as input signals, which comprise one or more of the following data: Outside temperature (Ta), outside air humidity, outside air pressure, position data, e.g. GPS data in a Global Positioning System and / or in a map, route data, traffic data.

10. Method according to one of the preceding claims, characterized in that at least some of the input signals each comprise time-dependent values ​​and the input signals are recorded with their respective input signal history, in particular as functions over time.

11. Method according to claim 10, characterized in that the time is divided into successive time steps (TS) and at least some of the operating data with their values ​​in at least some of the successive time steps (TS) are recorded as input signals.

12. Method according to one of the preceding claims, characterized in that the Kl model (50) comprises a neural network (52), wherein the neural network (52) comprises: - an input layer (L1, LL1) with first nodes (N1-i, NN 1) for receiving one input signal (xi), - an output layer (L5, LL3) with output nodes (N5-m, NN3) representing the control signals (S3, S4, S5) and / or output data for the control signals, and - at least one middle layer (L2, L3, L4; LL2) arranged between the input layer (L1, LL1) and the output layer (L5, LL3) and having a plurality of second nodes (N2-j, N3-k, N4-L; NN2), wherein the nodes of a layer (L1, L2, L3, L4, L5; LL1, LL2, LL3) are connected to at least some of the nodes of the previous layer (L1, L2, L3, L4, L5; LL1, LL2) and / or at least some nodes of the subsequent layer (L2, L3, L4, L5; LL1, LL2, LL3) in connections (Ci), wherein the nodes (N2, N3) of a layer each perform a mathematical operation from the input signals of the connections (Ci) of the previous layer and output the result of the mathematical operation to the respectively connected nodes of the subsequent layer.

13. The method according to claim 12, characterized in that the neural network (52) is designed as a fully connected neural network (52) at least in some layers (L1, L2, L3, L4, L5; LL1, LL2, LL3), in which all nodes of the layer (L1, L2, L3, L4, L5; LL1, LL2, LL3) are connected to all nodes of the previous layer and / or all nodes of the subsequent layer.

14. Method according to claim 12 or 13, characterized in that the mathematical operation comprises at least - multiplying the input signals (xi) of a node (Ni, NNi) by weights (W1, Wi), where each weight (W1, Wi) is determined by the connection (Ci) of the node with the nodes of the previous layer, - a subsequent addition of the input signals (xi) multiplied by the weights to form a sum (Z), in particular by adding a respective bias (WO) which is independent of the input signals, - a subsequent application of a non-linear activation function (AF, AF2, AF3, AF5).

15. The method according to claim 14, characterized in that the activation function (AF, AF2, AF3, AF5) is selected from the group comprising: - a hyperbolic tangent, - a function with partially linear sections with different gradients, - a step function that assigns a fixed value to the input values ​​below a limit and a fixed value above the limit, - a constant function in some areas and linear sections in some areas, e.g. a ReLU function, - a swish function, - a sigmoid function, - a softmax function, - an argmax function.

16. Method according to one of the preceding claims, characterized in that - the Kl-model (50) is pre-learned and trained and defined for a specific task, e.g. = a special route and / or = a special bus route, and / or = motorway traffic in a defined area, e.g. in Europe.

17. Method according to one of claims 1 to 15, characterized in that - the Kl model (50) is an actively learning model which is changed and adapted based on at least one criterion depending on the input signals, - the criterion being formed taking into account one or more of the following properties or characteristics: - a driver's driving style, road conditions, vehicle speed, a reduction in energy consumption, a reduction in wear, e.g. to increase valve lifespan.

18. The method according to claim 17, characterized in that the actively learning model is trained by a plurality of vehicles (1), each of which transmits its data to an external central location (120) in which the input signals of the plurality of vehicles (1) are used to train the actively learning model.

19. Method according to one of the preceding claims, characterized in that the Kl model further sets and / or changes parameters of the compressed air treatment unit (3), in particular parameters from the group containing: a lower pressure limit value for a measured compressed air value, for switching on a conveying phase, an upper pressure limit value for a measured compressed air value, an upper humidity limit value for a measured humidity value.

20. Control unit (10) of a compressed air treatment unit (3) of a commercial vehicle (1), which is designed to carry out a method according to one of the preceding claims, wherein - the control unit (10) carries out the step (ST2) of forming the output signals (S3, S4, S5) from the input signals (xi) by the Kl model (50) itself or - the step (ST2) of forming the output signals (S3, S4) from the input signals (xi) by the mathematical operation is carried out wholly or partly by an external central location (120), wherein the control unit exchanges signals with the external central location (120) by remote data transmission.

21. Compressed air treatment unit (3) for a compressed air supply device (15) of a commercial vehicle (1), comprising: a control unit (10) according to claim 20, a regeneration valve device (9), and a multi-circuit protection valve (12).

22. Commercial vehicle (1) which has: - a compressed air treatment unit (3) according to claim 21, - a compressor (7), and - a plurality of consumer circuits (14a, 14b, 14c, 14d) connected to the compressed air treatment unit (3), wherein the consumer circuits comprise two service brake circuits (14a, 14b), a parking brake circuit (14c), preferably further an air suspension circuit and / or an auxiliary consumer circuit (14d), - a drive system (5), and - an internal vehicle data system (6) for transmitting data and signals between at least the control unit (10) of the compressed air treatment unit (3) and the drive system (5).

23. A fleet comprising a plurality of commercial vehicles (1) according to claim 22 and an external central location (120), wherein the control units (10) of the compressed air supply device (3) transmit the input signals to the external central location (120) via a respective telecommunications interface, and the central location (120) is configured and designed to generate output signals from the input signals of the individual vehicles (1). le of the mathematical operation and to transmit it back to the commercial vehicles (1) via the telecommunications interface.

Images