Method and system for control based on MEMS resonators and a neural network

Abstract

The invention relates to the combination of micro-electromechanical systems (MEMS) with a neural network system. In many processes, sound is a strong indicator of efficiency and / or wear. The type and speed of open-loop and / or closed-loop control has previously been problematic for use. Neural networks afford significant advantages in this respect. Specifically the combination of the processing of generated noise by means of a neural network affords significant advantages in a large number of applications. In this respect, it is particularly advantageous to use MEMS resonators to feed the neural network, because, as a result, given an appropriate design of the MEMS resonators, very precise information about the frequency spectrum can be obtained, at least in a frequency range that is covered by the MEMS resonators, without transformation and thus particularly quickly. Thus, not only can the use of ultrasound to increase the efficiency in electrolysis be considerably improved with such a procedure, but so can the control of other processes, for example grinding, cutting and / or mixing processes.

Description

[0001] Method and system for control based on MEMS resonators and neural network

[0002] The invention relates to the combination of micro-electromechanical systems (microelectromechanical system - MEMS) with a neural network system (artificial neural network - ANN).

[0003] In principle, sound during a process is a strong indicator of efficiency and / or wear. This can therefore be used for control purposes. However, the type and speed of control have so far been problematic. This is where neural networks offer significant advantages.

[0004] The effect of waves on electrolysis is also known from "Acoustically-Induced Water Frustration for Enhanced Hydrogen Evolution Reaction in Neutral Electrolytes", Yemima Ehrnst, Peter C. Sherrell, Amgad R. Rezk, Leslie Y. Yeo, 4 December 2022, https: / / doi.org / 10.1002 / a-enm.202203164.

[0005] WO 2013 / 023068 A1 discloses controlling an ultrasonic generator to compensate for environmental influences on sound generation. A MEMS can be used as the ultrasonic generator, and control is performed by a neural network whose input nodes are fed with environmental sensors, such as temperature sensors.

[0006] MEMS are known, for example, from EP 2 713 509 A1, which can also be used for the present invention. However, the combination of processing generated sound using a neural network offers significant advantages in a variety of applications, as the inventor has recognized. It is particularly advantageous to use MEMS resonators to feed the neural network because this allows very precise information regarding the frequency spectrum, at least in a frequency range covered by the MEMS resonators, to be obtained without transformation and thus particularly quickly. With appropriate design of the MEMS resonators, this information can be obtained.

[0007] Not only can the use of ultrasound to increase the efficiency of electrolysis be significantly improved with such an approach, but also the control of other processes, such as grinding, machining and / or mixing processes.

[0008] It is also a great advantage to operate an actuator with a constant change and to control the type of change, particularly its center position, width, and / or speed, via the neural network. This allows the neural network to provide particularly good feedback and can even partially avoid local extremes in favor of global extremes. For example, the ultrasound generation in an electrolyzer can be optimally controlled so that the efficiency can be increased by a factor of more than five.

[0009] The object is achieved in particular by a system comprising at least one component acoustically coupled to a processing space and / or processing medium and comprising a neural network and comprising a means, in particular comprising an actuator, in particular a means containing a sound generator, a shaping, cutting and / or grinding element and / or a dosing device, configured to act on a processing medium arranged in the processing space by generating sound, wherein the system is configured to control the means by means of the neural network, characterized in thatthat the at least one component has a plurality of micro-electro-mechanical resonators with a plurality of different resonance frequencies and is configured to convert sound from the processing space and / or processing medium into electrical signals at the resonance frequencies of the resonators and to transmit information about the detected sound to the neural network via at least one, in particular a plurality of, information lines, in particular one information line per resonator and / or per resonance frequency, by means of the electrical signals, wherein the neural network is coupled to the at least one means and the system is designed such that it influences the behavior of the at least one means, in particular controls the at least one means, on the basis of the information detected by the component.

[0010] The problem is also solved by a method for controlling at least one process medium arranged in a processing room, generating sound.

[0011] By means, in particular comprising an actuator, in particular containing a sound generator, wherein at least one component acoustically coupled to the processing space and / or processing medium and a neural network are used to control the means, characterized in that the at least one component has a plurality of micro-electro-mechanical resonators with a plurality of different resonance frequencies and

[0012] Sound from the processing space and / or processing medium is converted into electrical signals at the resonance frequencies of the resonators and, by means of the electrical signals, information about the detected sound is transmitted to the neural network via at least one, in particular a plurality of, information line(s), in particular one information line per resonator and / or per resonance frequency, wherein the neural network is coupled to the at least one means and, on the basis of the information detected by the component, influences the behavior of the at least one means, in particular controls the at least one means.

[0013] The acting means can be, for example, an ultrasonic generator, a control actuator in a milling machine or for pressing an abrasive, a valve during a dosing process, a forming, cutting and / or grinding means, a laser, an electrical and / or electromagnetic transmitter, or even a wave or sound generator itself. The important thing is that a medium is acted upon and sound is generated. For example, the sound generator can act directly on a medium such as air or water and generate sound and transmit it to the medium, or a dosing device can control a volume flow that flows into another volume, whereby sound is generated during mixing. A control actuator can also act on a milling head, whereby the milling head mills, generating sound.For example, the influencing means could be an extrusion device, where the temperature and / or heating current and / or rotation speed of an extrusion screw can be controlled. A casting device, for example, for investment casting, can also be used as an influencing means, where the volume flow and / or the casting speed can be controlled. The welding current and / or voltage of a welding gun can also be controlled by the neural network. A mixer could also be used as an influencing means, where the rotation speed, for example, is controlled. An ultrasound generator of an ultrasonic flow meter can also represent the influencing means.

[0014] The effect of the agent can also be supported by another agent, or only part of it can be controlled. The desired change may also not be possible without another agent. For example, electrolysis can be enhanced by the action of an ultrasound generator, even if electrolysis would not be possible with ultrasound alone.

[0015] The processing medium can, for example, be water in a container into which a different current is metered or into which ultrasound is coupled, or even the piece of metal that is milled. But also air in which a sound is generated. In particular, however, it is a medium that is materially and / or permanently changed. It can therefore be a solid, liquid, or gaseous medium. The material and / or permanent change does not have to be caused by the influencing medium alone; it is sufficient if the influencing medium plays a promoting role, as is the case, for example, with an ultrasound generator as a means of acting on the electrolytic splitting of water into hydrogen and oxygen. In particular, the material and / or permanent change takes place in the processing space.

[0016] Particularly advantageously, the system is configured and / or the method is controlled such that the neural network is used to influence sound generation, particularly in a predetermined frequency range, particularly across the predetermined frequency range, and / or averaged over time, to be minimized, maximized, or brought closer to a predetermined value. This allows particularly good results to be achieved in a simple manner. The amplitude of the oscillations of the MEMS resonators or the voltages they output, averaged and / or summed at a given point in time (across the resonators, but in particular not over time), can be used as a measure of sound generation.

[0017] The predetermined value can be fixed or dynamic, and can also be one- or multi-dimensional. For example, it could be an audio file played back by a sound generator, such as a loudspeaker. The neural network can then be configured to detect the difference between the audio file and influence the sound generator to minimize or at least reduce the deviation. This not only optimizes sound generation but also achieves active noise suppression.

[0018] Particularly advantageously, the system is configured and / or the method is conducted such that the at least one component comprises micro-electro-mechanical resonators with different spatial orientations and / or comprises a plurality of components with micro-electro-mechanical resonators that are spatially spaced and / or arranged with different spatial orientations, wherein the components in particular have and / or comprise micro-electro-mechanical resonators for the same resonant frequencies. This allows for a particularly reliable and interference-resistant implementation, especially under inhomogeneous conditions in the processing space.

[0019] Particularly advantageously, the system is configured and / or the method is conducted such that each component has at least 100, in particular at least 10,000, resonators and / or at least 10, in particular at least 100 resonators, in particular per spatial orientation, of which at least three are present, of the resonators, with adjacent resonant frequencies, wherein the frequency spacing of the respective adjacent resonant frequencies is no more than 5 Hz, in particular no more than 1 Hz. This allows for a particularly precise conversion to be achieved in a very simple manner, since a large number of resonators can be arranged in a single chip.

[0020] Particularly advantageously, the system is set up and / or the method is carried out such that the neural network is a convolutional neural network (CNN), in particular with max-pooling layers, and / or has at least one intermediate layer and / or the input layer has a plurality of nodes, wherein each node is coupled to resonators of a resonant frequency, in particular only to resonators of a resonant frequency, in particular only to resonators of a resonant frequency and a spatial orientation and / or wherein a three-dimensional matrix is ​​generated from the outputs of the resonators, wherein the amplitudes are plotted on a two-dimensional plane and each point in the plane is assigned to one, in particular exactly one, resonator, wherein the CNN is in particular a U-Net or DEEPLabvß. This means that good results can also be achieved with known and / or simple networks.It is particularly advantageous if the system is configured and / or the process is conducted in such a way that the neural network has backpropagation and / or a target value. This type of configuration allows for a simple, good result to be achieved without training the network and even under changing environmental conditions and / or operating states.

[0021] Particularly advantageously, the system is set up and / or the method is carried out in such a way that the neural network has a plurality of layers, each of which at least partially has a plurality of nodes, and wherein the nodes of at least one layer with a plurality of nodes are processed in parallel.

[0022] Particularly advantageously, the system is configured and / or the method is conducted such that the means is operated under a, in particular continuous and / or in particular cyclical, change of at least one parameter, in particular at least one frequency, amplitude, speed, rotational speed, casting speed, feed rate, temperature, contact pressure, position, current, voltage and / or opening width and / or rate of change and / or range of change of one or more of the aforementioned. This change can in particular be modulated onto the usual control or, in the case of a control that is already constantly changing, such as in the playback of an audio file of a piece of music, can already be provided by this control.

[0023] By doing this, the neural network can be constantly supplied with training data, local extremes can be avoided in favor of global ones, and a high level of control stability with minimal fluctuations around an optimal operating point can be achieved.

[0024] Particularly advantageously, the system is configured and / or the method is conducted in such a way that the influence on the behavior is and / or includes the change of a center position, a change range, and / or a change rate, in particular of the at least one, in particular at least two, particularly preferably at least three or exactly three, parameters. In this way, a relatively simple control can be achieved, which can be implemented with a simple neural network with only three output variables.

[0025] With particular advantage, the system is designed and / or the method is carried out in such a way that the means comprises or is a material removal means, a dosing means or a wave generator, in particular a sound wave generator, a light generator and / or an electro- and / or magnetic transmitter.

[0026] It is particularly advantageous if the system is set up and / or the process is carried out in such a way that the means is a machine.

[0027] Particularly advantageously, the system is designed and / or the method is carried out in such a way that the means comprises a material flow, in particular a bulk material flow, a liquid flow and / or a gas flow, and / or a valve.

[0028] Particularly advantageously, the system is designed and / or the method is carried out in such a way that the means comprises one, in particular a plurality of, ultrasound generators, in particular arranged in / on the water-conducting chamber and / or the membrane of an electrolyzer.

[0029] With particular advantage, the system is set up and / or the process is carried out in such a way that the process comprises electrolysis, production of a product by mixing, shaping and / or processing.

[0030] In such applications, the invention can achieve particularly good results.

[0031] Particularly advantageously, the system is configured and / or the method is conducted such that the electromechanical resonators are piezoelectric resonators. These are particularly easy to manufacture and sufficient for the purposes of the invention; their instability is largely compensated for by the adaptation of the neural network.

[0032] It is particularly advantageous if the system is configured and / or the method is carried out in such a way that the MEMS resonators are designed as comb and / or ring structures. These structures have proven to be particularly efficient and are entirely sufficient for the implementation. It is particularly advantageous if the system is configured and / or the method is carried out in such a way that the resonators, in particular based on an output voltage generated by the piezoelectric properties of the respective piezoelectric resonator, continuously transmit information, in particular regarding the amplitude of the oscillation of the respective resonator, to the neural network, in particular over time.

[0033] It is particularly advantageous if the system is configured and / or the method is conducted such that the resonance frequencies of the plurality of resonators are at least partly or exclusively in the range from 1 kHz to 1.0 GHz. These frequencies have proven particularly favorable for control by the neural network.

Claims

Claims 1 . Method for controlling at least one processing medium arranged in a processing room, generating sound Means, in particular a means containing a and / or sound generator, wherein at least one component acoustically coupled to the processing space and / or processing medium and a neural network are used to control the means, characterized in that the at least one component has a plurality of micro-electro-mechanical resonators with a plurality of different resonance frequencies and Sound from the processing space and / or processing medium is converted into electrical signals at the resonance frequencies of the resonators and, by means of the electrical signals, information about the detected sound is transmitted to the neural network via at least one, in particular a plurality of, information line(s), in particular one information line per resonator and / or per resonance frequency, wherein the neural network is coupled to the at least one means and, on the basis of the information detected by the component, influences the behavior of the at least one means, in particular controls the at least one means.

2. Method according to claim 1, wherein the neural network is used to influence the sound generation, in particular in a predetermined frequency range, in particular over the predetermined frequency range, and / or on average over time, to be minimized, maximized or approximated to a predetermined value.

3. Method according to one of the preceding claims, wherein the at least one component comprises micro-electro-mechanical resonators with different spatial orientation and / or a plurality of Components with micro-electro-mechanical resonators that are spatially spaced and / or arranged with different spatial orientations, wherein the components in particular have and / or comprise micro-electro-mechanical resonators for the same resonance frequencies.

4. Method according to one of the preceding claims, wherein each component has at least 100, in particular at least 10,000, resonators and / or at least 10, in particular at least 100 resonators, in particular per spatial orientation, of which in particular at least three are present, of the resonators, with adjacent resonant frequencies, wherein the frequency spacing of the respectively adjacent resonant frequencies is not more than 5 Hz, in particular not more than 1 Hz.

5. The method according to any one of the preceding claims, wherein the neural network is a convolutional neural network (CNN), in particular with max-pooling layers, and / or has at least one intermediate layer and / or the input layer has a plurality of nodes, wherein each node is coupled to resonators of a resonant frequency, in particular only to resonators of a resonant frequency, in particular only to resonators of a resonant frequency and a spatial orientation and / or wherein a three-dimensional matrix is ​​generated from the outputs of the resonators, wherein the amplitudes are plotted on a two-dimensional plane and each point in the plane is assigned to one, in particular exactly one, resonator, wherein the CNN is in particular a U-Net or DEEPLabv3.

6. Method according to one of the preceding claims, wherein the neural network has a backpropagation and / or a target value 7. Method according to one of the preceding claims, wherein the neural network has a plurality of layers, each of which at least partially comprises a plurality nodes and wherein the nodes of at least one layer with a plurality of nodes are processed in parallel.

8. Method according to one of the preceding claims, wherein the means is operated under a, in particular continuous and / or in particular cyclical, change of at least one parameter, in particular frequency, speed and / or opening width.

9. Method according to one of the preceding claims, wherein the influence on the behavior comprises the change of a central position, a change range and / or a change rate, in particular of the at least one parameter 10. Method according to one of the preceding claims, wherein the means comprises or is a material removal means, a dosing means or a wave generator, in particular a sound wave generator, a light generator and / or an electro- and / or magnetic transmitter. 1 1. Method according to one of the preceding claims, wherein the means is a machine.

12. Method according to one of the preceding claims, wherein the means comprises a material flow, in particular a bulk material flow, a liquid flow and / or a gas flow, and / or a valve.

13. Method according to one of the preceding claims, wherein the means comprises one, in particular a plurality of, ultrasound generator(s), in particular arranged in / on the water-conducting chamber and / or the membrane of an electrolyzer.

14. A process according to any one of the preceding claims, wherein the process comprises electrolysis, production of a product by mixing, shaping and / or processing.

15. Method according to one of the preceding claims, wherein the electro-mechanical resonators are piezoelectric resonators.

16. Method according to the preceding claim, wherein, the resonators, in particular based on an output voltage resulting from the piezoelectric properties of the respective piezoelectric resonator, each transmit information, in particular about the amplitude of the oscillation of the respective resonator, to the neural network, in particular continuously over time.

17. Method according to one of the preceding claims, wherein the resonance frequencies of the plurality of resonators are at least also or exclusively in the range from 1 kHz to 1.0 GHz.

18. Method according to one of the preceding claims, wherein the MEMS resonators are formed as comb and / or ring structures.

19. System comprising at least one component acoustically coupled to a processing space and / or processing medium and comprising a neural network and comprising a means, in particular an actuator, sound generator, shaping, cutting and / or grinding element and / or dosing device, configured to act on a processing medium arranged in the processing space by generating sound, wherein the system is configured to control the means by means of the neural network, characterized in that the at least one component has a plurality of micro-electro-mechanical resonators with a plurality of different resonance frequencies and is configured to convert sound from the processing space and / or processing medium into electrical signals at the resonance frequencies of the resonators and to transmit information about the detected sound by means of the electrical signals via at least one, in particular a plurality,Information line(s), in particular one information line per resonator and / or per resonance frequency, is transmitted to the neural network, wherein the neural network is coupled to the at least one means and the system is designed such that it can, on the basis of the information from, Component information influences the behavior of the at least one means, in particular controls the at least one means.

20. System according to the preceding claim 19, wherein the neural network is used to influence the sound generation, in particular in a predetermined frequency range, in particular over the predetermined frequency range, and / or on average over time, to be minimized, maximized or approximated to a predetermined value and / or state.

21. System according to one of the preceding claims, wherein the at least one component comprises micro-electro-mechanical resonators with different spatial orientations and / or comprises a plurality of components with micro-electro-mechanical resonators that are spatially spaced and / or arranged with different spatial orientations, wherein the components in particular have and / or comprise micro-electro-mechanical resonators for the same resonance frequencies.

22. System according to one of the preceding claims, wherein each component has at least 100, in particular at least 10,000, resonators and / or at least 10, in particular at least 100 resonators, in particular per spatial orientation, of which in particular at least three are present, of the resonators, with adjacent resonant frequencies, wherein the frequency spacing of the respectively adjacent resonant frequencies is not more than 5 Hz, in particular not more than 1 Hz.

23. System according to one of the preceding claims, wherein the neural network is a Convolutional Neural Networks (CNN), in particular with max-pooling layers, and / or has at least one intermediate layer and / or the input layer has a plurality of nodes, wherein each node is coupled to resonators of a resonance frequency, in particular only to resonators of a resonance frequency, in particular only to resonators of a resonance frequency and a spatial orientation and / or wherein the system is configured to generate a three-dimensional matrix from the outputs of the resonators, wherein the amplitudes are plotted on a two-dimensional plane and each point in the plane is assigned to one, in particular exactly one, resonator, wherein the CNN is in particular a U-Net or DEEPLabvß.

24. System according to one of the preceding claims, wherein the neural network has backpropagation and / or a target value.

25. System according to one of the preceding claims, wherein the neural network has a plurality of layers, each of which at least partially has a plurality of nodes, and wherein the nodes of at least one layer with a plurality of nodes are processed in parallel.

26. System according to one of the preceding claims, wherein the means is operated under a, in particular continuous and / or in particular cyclical, change of at least one parameter, in particular frequency, speed and / or opening width.

27. System according to one of the preceding claims, wherein the system is arranged such that the influence on the behavior comprises the change of a center position, a range of change and / or a rate of change, in particular of the at least one parameter.

28. A system according to any preceding claim, wherein the means is a machine.

29. System according to one of the preceding claims, wherein the means comprises a material flow, in particular a bulk material flow, a liquid flow and / or a gas flow, and / or a valve 30. System according to one of the preceding claims, wherein the means comprises one, in particular a plurality of, ultrasonic generator(s), in particular arranged in / on the water-conducting chamber and / or the membrane of an electrolyzer.

31. System according to any one of the preceding claims, wherein the system comprises a device for electrolysis and / or for producing a product by mixing, shaping and / or processing.

32. System according to one of the preceding claims, wherein the electro-mechanical resonators are piezoelectric resonators.

33. System according to the preceding claim 32, wherein the system is configured such that the resonators, in particular based on an output voltage resulting from the piezoelectric properties of the respective piezoelectric resonator, each transmit information, in particular about the amplitude of the oscillation of the respective resonator, to the neural network, in particular continuously over time.

34. System according to one of the preceding claims, wherein the resonators comprise and / or are resonators having resonance frequencies in the range of 1 kHz to 1.0 GHz.

35. System according to one of the preceding claims, wherein the MEMS resonators are formed as comb and / or ring structures.