Controller for providing a classification of sensor data for a motor vehicle by means of an optical neural network, and method for operating the controller
Patent Information
- Application Number
- EP2022834537
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-01-14
- Filing Date
- 2022-12-07
- Publication Date
- 2024-11-06
AI Technical Summary
Conventional neural networks are insufficient for real-time classification of high-resolution camera images in motor vehicles due to high computational demands, limiting the frame rate and resolution for reliable automatic or semi-automatic driving.
An optical neural network utilizing atomic ensembles as neurons with electromagnetically induced transparency properties, enabling fast and precise processing of sensor data by transmitting and evaluating optically encoded information at the speed of light.
The optical neural network facilitates rapid classification of sensor data, overcoming the limitations of conventional NNs by providing real-time, high-resolution object classification essential for automatic driving maneuvers.
Smart Images

Figure 1.1
Abstract
Description
[0001] Description
[0002] Control device for providing a classification of sensor data for a motor vehicle by means of an optical neural network and method for operating the control device
[0003] The invention relates to a control device for providing a classification of sensor data for a motor vehicle using an optical neural network. The invention also relates to a motor vehicle having such a control device and a method for operating such a control device.
[0004] A motor vehicle can be configured for automatic or at least semi-automatic driving. For this purpose, the motor vehicle has, for example, a computing device configured to control the longitudinal and / or lateral guidance of the motor vehicle, for example by controlling a drive device, a braking system, and / or a steering system of the motor vehicle. For automatic or at least semi-automatic driving, reliable detection and evaluation of the surroundings of the motor vehicle is essential; this means that the motor vehicle should be provided with reliable environmental perception. For this purpose, the surroundings of the motor vehicle, which can alternatively be referred to as the surroundings of the motor vehicle, are detected by means of at least one sensor device, for example by means of a radar device, lidar device, and / or a camera, and at least one object detected in this way in the surroundings is classified.In other words, it determines which class of objects the detected object can be assigned to. During classification, it is checked, for example, whether the detected object is another road user, such as another motor vehicle or a bicycle, a person, a road surface, and / or an infrastructure element, such as a building. Such a classification can be performed, for example, using a neural network (NN) or artificial neural network (ANN) that has been previously trained to recognize objects from different classes.
[0005] US 2020 / 0327403 A1 shows an optical neural network that uses light beams and optical components to provide layers of a neural network, having an input layer, zero or more intermediate layers, and an output layer. An optical nonlinear operation is implemented using a nonlinear optical medium that exhibits electromagnetically induced transparency properties. An optical neural network is also described in the scientific publication "All-optical neural network with nonlinear activation functions" by Ying Zuo, Bohan Li, Yujun Zhao, Yue Jiang, You-Chiuan Chen, Peng Chen, Gyu-Boong Jo, Junwei Liu, and Shengwang Du (Optica, Vol. 6, No.9, September 2019, pages 1132-1137), according to which laser-cooled rubidium-85 isotopes are excited in a two-dimensional magneto-optical trap (MOT) using a sample laser and a coupling laser to provide a nonlinear optical activation function with electromagnetically induced transparency properties.
[0006] The object of the invention is to provide a solution by means of which sensor data for a motor vehicle can be classified particularly quickly.
[0007] The problem is solved by the subject matter of the independent patent claims and the subordinate patent claim.
[0008] A first aspect of the invention relates to a control device for providing a classification of sensor data for a motor vehicle using an optical neural network (NN). The control device comprises the optical neural network (NN). The NN can alternatively be referred to as an artificial neural network (ANN).
[0009] The invention is based on the finding that for automatic or at least partially automatic driving of a motor vehicle, holistic detection of the vehicle's surroundings over 360 degrees and in three dimensions (360° 3D environment detection) is essential so that all static and / or dynamic objects in the environment can be detected and subsequently classified. The camera, in particular, plays a key role in redundant, robust detection of the environment, as this type of sensor can precisely measure angles in the environment and be used to classify the environment. However, the processing and classification of camera images captured by the camera is computationally intensive and, in particular, architecturally complex.360° 3D environment detection, in particular, can be problematic, as numerous individual images must be classified and processed, resulting in a typically high computational effort compared to evaluating other sensor data. A high-performance NN or ANN offers the possibility of classifying camera images and / or data from other sensor devices in the vehicle, such as a radar device and / or lidar device, with frame rates of several hertz, which can, for example, be between greater than 0 hertz and less than 10 hertz. However, this frame rate is insufficient for reliable environment detection with high resolution (i.e., within a desired resolution range) in real time, as modern camera systems operate at a frame rate of 30 hertz. Furthermore, the data load increases with the increasing resolution of the camera images.An automotive-grade camera, for example, may have a resolution of 8 megapixels or more. Classifying the camera images from this camera in a vehicle is not possible in real time using conventional NN if numerous, distinct classes must be distinguished. The limiting factor is the speed of the processor or graphics processor used, which, even with high-performance computers, has typically not been sufficient to fully classify the high-resolution and numerous images per second in real time. Classification is particularly necessary for understanding the surroundings and to be able to determine and execute an automatic or at least semi-automatic driving maneuver for the vehicle based on the environment.An incomplete and / or incorrect classification is undesirable for the use of the automatic or at least partially automatic driving function and / or an automatic or at least partially automatic driver assistance system.
[0010] A NN is used in the context of an automatic or at least semi-automatic driving function, primarily for environmental classification. Sensor data is passed electronically to several neurons of the NN via weights. Within each neuron, a computational operation is performed using a linear or nonlinear response function. An output signal from the neurons of one layer of the NN, which is passed to the neurons of a subsequent layer of the NN, can be represented, for example, by a sigmoid function of the sum of weighted response functions, where the sigmoid function is an example of a nonlinear response function:
[0011] Here, a is a neuron of a layer of several neurons, coj is the weight, a; 0 the neurons and o the sigmoid function. The NN thus forms the following function: Here, n and k are natural numbers, a' are individual neurons (for i = 0 to i = n), and y' are function values (for i = 0 to i = k) that are output as class information. The output signal of the neurons corresponds to the classification. For example, for an image area of a camera image, the sensor data classified by the NN indicates which class of object is represented in the image area.
[0012] When a NN is implemented on a conventional computer architecture, slow data processing is usually observed, meaning that the class information for the high-resolution camera images or other sensor data cannot be provided in real time.
[0013] In order to classify sensor data faster than with a conventional NN, an optical NN should be used. The use of atomic ensembles as neurons of a NN enables highly precise quantum control of electromagnetic radiation. The atomic ensemble, which is assumed to be the neuron of the optical NN, can be designed as a nonlinear optical medium. It is important that the nonlinear optical medium exhibits electromagnetically induced transparency properties, i.e., the effect of electromagnetically induced transparency (EIT). The nonlinear optical medium can, for example, be designed as a quantum structure. By exploiting the EIT effect, at least one electromagnetic atomic state of the nonlinear optical medium can be used as an optical neuron.
[0014] The control device is designed to provide optically coded sensor data. The sensor data can be acquired by a sensor device of the motor vehicle and provided to the control device, for example by means of a control device of the sensor device, in particular via an internal vehicle communication connection. Alternatively or additionally, the sensor data can be provided by an external device, for example, by another motor vehicle, a server device, and / or an infrastructure device. The sensor data can then be transmitted wirelessly to the motor vehicle, for example, via a vehicle-to-vehicle and / or a vehicle-to-infrastructure connection.The sensor data are, for example, camera data from a camera, for example a vehicle-mounted front camera, rear camera and / or side camera on the one hand and / or a stationary traffic observation camera on the other hand, radar data from a radar device of the motor vehicle and / or lidar data from a lidar device of the motor vehicle (lidar for “light detection and ranging”).
[0015] The control device is also designed to evaluate the provided optically coded sensor data using the optical neural network. The optical neural network has a plurality of optical neurons with electromagnetically induced transparency properties, which thus behave according to the EIT effect. The control device is also designed to provide optically coded classification information as a result of the evaluation by an evaluation device of the optical neural network. The optically coded classification information describes the classification of the sensor data. The optically coded classification information therefore includes, for example, data from which the classification of the sensor data can be derived.The evaluation device can be a central device of the control apparatus, which is designed to evaluate the data provided by all optical neurons of the optical NN and thereby determine the optically coded classification information for the sensor data. The optically coded classification information describes, for example, for each image pixel of a camera image as sensor data, the class of the object to which the image pixel is assigned. The class of the object differentiates, for example, between the objects roadway, road marking, vehicle, pedestrian, vegetation, and / or infrastructure element.
[0016] The control device has a conversion device. By means of this, the control device is designed to convert the optically encoded classification information into electronically encoded classification information. An electronic signal or electronic information is thus generated from the optically encoded classification information previously provided as an optical signal or optical information. The content of the optically encoded classification information is not changed in the process. This means that the electronically encoded classification information also describes the classification of the sensor data determined using the optical NN. The electronically encoded classification information includes data that can be evaluated by an electronic device, such as a computing device of the motor vehicle.The computing device can, for example, be designed to control the longitudinal and / or lateral guidance of the motor vehicle, for example by controlling a drive device, a braking system, and / or a steering system of the motor vehicle. The electronically coded classification information is thus suitable for being provided to an automatic or at least partially automatic driving function and / or an automatic or at least partially automatic driver assistance system. The control device is therefore designed to provide the electronically coded classification information for the motor vehicle. The control device is preferably a component of the motor vehicle, so that it makes the electronically coded classification information available to itself and / or to the computing device of the motor vehicle, which is arranged, for example, separately from the control device.The provision may comprise transmitting the electronically coded classification information from the control device to another component of the motor vehicle, for example to the computing device for the automatic or at least partially automatic driving function or the automatic or at least partially automatic driver assistance system.
[0017] The optical NN is first trained before being used to classify the sensor data. For this purpose, for example, camera data of a possible surroundings of a motor vehicle, preferably multiple camera images, are provided to the optical NN, which learns the various classes of objects described by the sensor data. Alternatively or additionally, radar data and / or lidar data can be provided to the optical NN for training. This training data was preferably acquired using a sensor device of the motor vehicle, i.e., it preferably corresponds to actual sensor data from the surroundings of the motor vehicle. The training process can include a manual review of the optically encoded classification information obtained using the optical NN.The optical NN is thus trained to assign at least one object described by the sensor data to a class of objects, i.e., to classify the sensor data. Through training, a connection is created for the optical NN between typical sensor data and the typical classes of objects in the environment of the motor vehicle. This connection is made available for future applications of the optical NN, for example, by storing the connection in the evaluation unit of the control device.
[0018] By utilizing the optical NN, data can be transmitted and evaluated between individual neurons at the speed of light, since optically encoded data can be transmitted and processed so quickly. The control device thus enables particularly fast classification of the sensor data for the motor vehicle, since it is optically encoded data that is processed, not electronically encoded data as in a conventional ANN or NN. One embodiment provides that the at least one optical neuron is formed from a nonlinear optical medium, in particular as a quantum dot, quantum wire, quantum well, and / or vapor cell. The vapor cell has at least one atom and / or molecule with a predetermined energy level, in particular a Rydberg atom and / or a Bose-Einstein condensate. The vapor cell can therefore have highly excited atoms. The vapor cell can be filled with a gas or a solution.
[0019] A quantum dot is a nanoscopic material structure, made, for example, from a semiconductor material, in which charge carriers (electrons, holes) in the quantum dot are so restricted in their mobility in all three spatial directions that their energy can only assume discrete values. The quantum dot therefore behaves similarly to an atom, but the shape, size, and / or number of electrons in the quantum dot can be influenced, so that the electronic and / or optical properties of the quantum dot can be adapted to a given requirement. A quantum wire is a spatial potential structure in which the freedom of movement of charge carriers is restricted to one dimension. A quantum well is a potential curve that limits the freedom of movement of a particle in one spatial dimension, so that only a planar region can be occupied.The width of the quantum well largely determines the quantum mechanical states the particle can assume. This leads, in particular, to the particle being able to assume only discrete energy states. The nonlinear optical medium thus represents a system subject to quantum confinement, as is the case for quantum dots, quantum wires, and / or quantum wells.
[0020] Thus, coupled quantum mechanical states are used for the optical NN, which are present, for example, in at least one atom, highly excited atom, Rydberg atom, molecule, a solution, and / or the system subject to quantum confinement, in order to utilize at least one electromagnetic atomic state as an optical neuron using EIT. Thus, versatile design options for the nonlinear optical medium are possible, all of which can exhibit the EIT effect. The optical neurons are thus particularly flexible in their design.
[0021] An additional embodiment provides that the optical NN has at least one layer containing multiple optical neurons. In particular, the optical NN has an input layer, no intermediate layer or at least one intermediate layer (“hidden layer”), and an output layer. Each of these layers can be understood as a column of multiple neurons. At least one first optical waveguide leads to each optical neuron. The optical waveguide is designed to transmit light. The optical waveguide guides light, in particular, into fibers made of quartz glass or into a polymer optical fiber made of plastic, or into a waveguide integrated into a semiconductor structure, or into a waveguide structure integrated into a solid. A sample laser beam from a sample laser and a coupling laser beam from a coupling laser can be coupled into the optical neuron via the first optical waveguide.The probe laser and the coupling laser are included in the optical NN.
[0022] The function of the two lasers results from the EIT effect. This effect is based on the fact that the probe laser beam is irradiated onto the nonlinear optical medium by the probe laser and the coupling laser beam is irradiated onto the nonlinear optical medium by the coupling laser. The nonlinear optical medium can have energy states E0, E1, and E2. By applying the probe laser beam to the nonlinear optical medium, at least one electron of an atom of the nonlinear optical medium, which has the energy state E0, can be raised to the energy state E1. By applying the coupling laser beam, it can also be raised from the energy state E1 to the energy state E2. In other words, the coupling laser couples the energy states E1 and E2, which can be referred to as "coupled states" or "dressed states."The atom's electron can thus transition from the energy state E0 to the energy state E2 by absorbing one photon each from the probe laser and the coupling laser. The energy state E2 is preferably long-lived, meaning, for example, a metastable state. In this case, the nonlinear optical medium becomes transparent to the probe laser when there are no more electrons in the nonlinear optical medium that can absorb a photon from the probe laser beam, since they have not done so before, and the transition from the energy state E2 to the energy state E0 is forbidden. This forbidden transition can be referred to as a "dark state." Therefore, repopulation of the energy state E0 is not possible.By means of the two lasers, the optical neuron, which has the nonlinear optical medium, can be excited in such a way that, due to the electromagnetically induced transparency properties of the optical neuron, the optical NN can serve as such, i.e., functions as an optical NN.
[0023] It is preferably provided that the optical NN has two first optical waveguides for each neuron, wherein a first of the two first optical waveguides is provided for the sample laser beam and a second of the two first optical waveguides is provided for the coupling laser beam. Thus, the two laser beams can be radiated onto the respective neuron from, for example, two different directions. At least one second optical waveguide leads away from each optical neuron. The second optical waveguide is designed analogously to the first optical waveguide. The at least one second optical waveguide leads to another optical neuron or directly to the evaluation device. The sample laser beam can be coupled into the at least one second optical waveguide after it has passed the optical neuron. Thus, the sample laser beam is coupled into each of the second optical waveguides after it has passed the optical neuron.The test laser beam entering the optical neuron via the at least one first optical fiber can interact or not interact with at least one electron of the optical neuron, i.e., the nonlinear optical medium. After the interaction or non-interaction, the test laser beam is forwarded via the second optical fiber, into which the test laser beam that has passed through the optical neuron is coupled. It is provided that for each neuron of a first layer, a respective second optical fiber is available to each individual neuron of the adjacent second layer. For each neuron of the second layer, this respective optical fiber is the first optical fiber. In the case of the last layer, i.e., the output layer, each neuron is connected to the output device via the second optical fiber.It is therefore intended that the individual neurons of the preferably multiple layers of the optical NN are interconnected. This ensures the typical structure of a conventional NN or ANN, which is not necessarily an optical NN, where the neurons of the different layers can influence each other.
[0024] It is possible for the optical NN to be configured to adjust a frequency and / or wavelength of the probe laser and / or coupling laser depending on the sensor data. By selecting a parameter setting of the probe laser and / or coupling laser for the input layer of optical neurons, for example, image pixel information from the sensor data can be transferred to the optical NN. The sensor data here represents, in particular, input information that describes, for example, a pixel value of an image pixel of a camera image from the camera, a local value of the radar data from the radar device, and / or a local value of the lidar data from the lidar device.This can, for example, ensure that each neuron in the input layer assumes a sensor-data-specific, elevated energy state that may differ from other neurons in the input layer. This sensor-data-specific information is communicated to the neurons in the subsequent layers, since the test laser beam coupled into the second optical fiber still contains this information, as this information can be extracted, for example, from the transmission spectrum and / or the absorption spectrum of the coupled test laser beam. Ultimately, a reliably functioning optical NN is thus provided.
[0025] Furthermore, one embodiment provides that the optical NN is configured to evaluate a transmission spectrum and / or an absorption spectrum of the sample laser beam coupled into the at least one second optical fiber to provide the optically encoded classification information. For this purpose, in particular, an extreme value of a frequency and / or an amplitude of the transmission spectrum or the absorption spectrum is evaluated. The extreme value is, for example, a maximum of the frequency and / or the amplitude. Alternatively or additionally, a phase of the transmission spectrum or the absorption spectrum can be evaluated. The evaluation can be carried out by means of the evaluation device.If the optical neuron, i.e., in particular the nonlinear optical medium, is irradiated by the probe laser and the coupling laser, a minimum in the absorption spectrum of the probe laser beam forms as soon as a time and / or intensity of laser irradiation greater than a minimum time or minimum intensity specified for the optical neuron used has been exceeded. In this case, photons from the probe laser can pass through the optical neuron without interaction, i.e., propagate through it without interacting with, for example, the nonlinear optical medium. This state of the optical neuron can, in principle, be detected by recording the absorption spectrum and / or the transmission spectrum of the probe laser. The transmission spectra and / or absorption spectra of the output layer are evaluated, and the optically encoded classification information is obtained as a result of this evaluation.
[0026] According to one embodiment, the optical NN comprises at least one optical modulator. Preferably, a separate optical modulator is provided for each optical neuron. The optical modulator is designed to influence at least one property of the sample laser beam. In particular, it is designed to influence an amplitude, polarization, frequency, wavelength, and / or phase of the sample laser beam. The optical modulator can be designed to provide the optical weighting of the optical NN using the amplitude, polarization, frequency, wavelength, and / or phase of the sample laser beam, i.e., to weight the amplitude, polarization, frequency, wavelength, and / or phase of the sample laser beam. Using the optical modulator, the sample laser beam received by a respective neuron can be adjusted based on the sensor data, for example, given constant properties of the sample laser and / or coupling laser.The optical modulator can thus be designed to influence a row of optical neurons in several columns, i.e., several layers of optical neurons, in such a way that, for example, image pixel information from the sensor data is specified using the sample laser beam received by the respective neuron. Furthermore, the optical modulator can be used to transmit data to the evaluation device. The optical modulator can also be used to optically weight the optical neural network.
[0027] In addition, one embodiment provides that the optical NN has a radiation source designed to irradiate the at least one optical neuron with electromagnetic radiation. An external electromagnetic field is therefore applied to the respective optical neuron. This causes the transmission spectrum of the test laser beam coupled into the at least one second optical fiber to split, utilizing the dynamic Stark effect, which can alternatively be referred to as the AC Stark effect (AC for "alternating current"). The electron of the optical neuron, raised to a higher energy state by the test laser beam, is excited by the irradiation with the electromagnetic radiation from the energy state E2 to an energy state E3, wherein the energy states E2 and E3 differ from one another. This causes the transmission spectrum to split.The optical NN is designed to evaluate the split transmission spectrum to provide the optically encoded classification information. In particular, an extreme value and / or a distance between two extreme values of a frequency and / or an amplitude of the transmission spectrum is evaluated. The distance can be determined for two maxima as extreme values. Furthermore, a minimum lying between the two maxima can be evaluated, for example, with regard to its frequency and / or amplitude.
[0028] Alternatively or additionally, a phase of the transmission spectrum can be evaluated. The transmission spectrum is preferably viewed in frequency domain, i.e. as a function of frequency. One advantage of the applied external electromagnetic field is that the amplitude of the transmitted radiation is reduced by spectral splitting in the transmission spectrum, and a phase of the incident test laser radiation can be reconstructed by a time-resolved measurement of the transmission spectrum. The minimum as an extreme value in the transmission spectrum provides information about the frequency of the external field and thus about the radiation source due to the previously defined atomic transition to the energy state E3. There are therefore numerous other characteristic properties of the transmission spectrum that can be used to determine the classification result of the optical NN.Alternatively, or in addition to the transmission spectrum, the absorption spectrum can be considered, including its splitting due to the dynamic Stark effect. The procedure is then analogous to that for the transmission spectrum.
[0029] A preferred embodiment comprises the optical NN being configured to adjust a frequency, a phase characteristic, and / or an amplitude of the electromagnetic radiation depending on the sensor data. The sensor data here represent, in particular, the input information, which describes, for example, the pixel value of the image pixel of the camera image, the local value of the radar data of the radar device, and / or the local value of the lidar data of the lidar device. By appropriately selecting the applied external electromagnetic field, the connection to the sensor data can be provided, for example, for the optical neurons of the input layer, so that, for example, each optical neuron is specified in a sensor-data-specific manner using the respective applied external electromagnetic field.
[0030] It can also be assumed that, if the applied external electromagnetic field is detected sufficiently quickly, a phase response of the electromagnetic field can be measured over time. This makes it possible to measure the following parameters: the frequency f of the electromagnetic field when compared with a known spectral line (e.g., an integrated semiconductor-based optical frequency comb), a phase response of the electromagnetic field, and an amplitude response of the electromagnetic field. When the radiation source is activated, the generated electromagnetic field can be used as additional optical weight for the optical NN.
[0031] At least one of the following options for optical weighting can be selected: modulating the amplitude, phase, and / or frequency using a corresponding modulator to control the probe laser and / or a pump laser for the probe laser, on the one hand, and / or selecting the externally irradiated electromagnetic radiation, for example, in the gigahertz range, on the other. By selecting the externally irradiated electromagnetic radiation in the gigahertz range, the absorption spectrum can be modulated in such a way that the transmitted portion of the probe laser is modulated, which acts as a weighting.
[0032] Furthermore, according to one exemplary embodiment, the control device comprises a computing unit, in particular a computer and / or a graphics processor, and a further conversion device. The computing unit is designed to provide electronically coded sensor data. The computing unit can receive this data, for example, from a control device of a sensor device of the motor vehicle. The computing unit comprises, for example, at least one microprocessor and / or at least one microcontroller and / or at least one FPGA (Field Programmable Gate Array) and / or at least one DSP (Digital Signal Processor). The further conversion device is designed to convert the provided electronically coded sensor data into the optically coded sensor data and to provide this data.The control device is therefore preferably provided with electronically coded sensor data, which can, however, be converted into optically coded sensor data by the control device so that they can be fed to the optical NN for further evaluation. Thus, optically coded sensor data does not have to be available from the outset for the sensor data to be evaluated by the optical NN.
[0033] Furthermore, according to one exemplary embodiment, it is provided that the conversion device of the control apparatus is designed to interfere with local oscillator information provided by a local oscillator with the provided optically coded classification information in order to provide the electronically coded classification information. The local oscillator can alternatively be referred to as a local oscillator. A reliable method can therefore be used to provide the optically coded classification information as electronically coded classification information. This can then be provided, for example, to the computing device for the automatic or at least partially automatic driving function or the automatic or at least partially automatic driver assistance system for further evaluation or consideration.
[0034] In one embodiment, it can be provided that the optical NN is arranged on at least one semiconductor die. The at least one semiconductor die can be designed as at least one of the following components: as an electronically-photonically co-integrated semiconductor chip, which can alternatively be referred to as an electronically-photonically integrated semiconductor chip; a photonic integrated circuit (IC); a multi-chip module; and / or a chip mounted using flip-chip assembly. Various basic modules or basic elements are thus possible on the basis of which the optical NN is formed. For example, commercially available basic modules or basic elements can be used, whereby the optical NN can be manufactured easily and cost-effectively in this respect.According to one exemplary embodiment, the control device has a communication interface to a control unit of a sensor device of the motor vehicle. The control device is designed to receive sensor data acquired by the sensor device via the communication interface. The sensor device is in particular the camera, the radar device and / or the lidar device. The control device is at least designed to control the sensor device, for example to activate or deactivate it and / or to specify an acquisition parameter for the sensor data acquisition by means of the sensor device. The control device therefore has at least one microprocessor and / or microcontroller. The sensor data is therefore preferably acquired in the motor vehicle itself and evaluated in the motor vehicle directly by means of the control device with the optical NN.The evaluated data, i.e., the electronically encoded classification information, is then preferably used in the vehicle, for example, by the computing device. This allows for significant benefits from the rapid classification of sensor data using the optical NN in the vehicle, since no connection to an external device is required to perform the classification.
[0035] A further aspect of the invention relates to a motor vehicle with a control device as described above. The exemplary embodiments described in connection with the control device, individually and in combination with one another, apply accordingly, where applicable, to the motor vehicle according to the invention. The motor vehicle is, for example, a passenger car, a truck, a bus, a motorcycle, and / or a moped.
[0036] An exemplary embodiment of the motor vehicle according to the invention provides that the motor vehicle has at least one sensor device with a control device. The control device is connected to the communication interface of the control device via at least one optical fiber connection. The control device of the sensor device is designed to transmit the sensor data acquired by the sensor device to the control device via the at least one optical fiber connection. In other words, the control device is designed to receive the sensor data acquired by the sensor device from the control device via the at least one optical fiber connection. The optical fiber connection can be formed, for example, from polymeric optical fibers (POF for “polymeric optical fiber” or “plastic optical fiber”).In other words, a third optical fiber can be arranged between the control device and the control apparatus. This is preferably made of plastic and is used to transmit the sensor data to the control apparatus. This enables reliable and fast internal communication in the motor vehicle between the sensor device and the control apparatus using the optical NN.
[0037] A further aspect of the invention relates to a method for operating a control device for providing a classification of sensor data for a motor vehicle by means of an optical neural network. The method comprises the following steps: providing optically encoded sensor data; evaluating the provided optically encoded sensor data by means of an optical neural network having a plurality of optical neurons with electromagnetically induced transparency properties, wherein, as a result of the evaluation by means of an evaluation device of the optical neural network, optically encoded classification information is provided which describes the classification of the sensor data; converting the optically encoded classification information into electronically encoded classification information by means of a conversion device; and providing the electronically encoded information for the motor vehicle.The exemplary embodiments described in connection with the control device, individually and in combination with one another, apply accordingly, to the method according to the invention, as far as applicable. The method can be considered, at least in part, a computer-implemented method.
[0038] The invention also includes combinations of the features of the described embodiments.
[0039] Exemplary embodiments of the invention are described below. Shown are:
[0040] Fig. 1 is a schematic side view of a motor vehicle with a control device and several sensor devices;
[0041] Fig. 2 is a schematic front view of a motor vehicle with a control device and several sensor devices;
[0042] Fig. 3 is a schematic rear view of a motor vehicle with a control device and several sensor devices;
[0043] Fig. 4 is a schematic representation of energy states in electromagnetically induced transparency; Fig. 5 is a schematic representation of the energy states according to Fig. 4 in an external electromagnetic field;
[0044] Fig. 6 is a schematic representation of an absorption spectrum for electromagnetically induced transparency;
[0045] Fig. 7 is a schematic representation of an absorption spectrum according to Fig. 4 in an external electromagnetic field;
[0046] Fig. 8 is a schematic representation of an EIT signal depending on the intensity of a coupling laser;
[0047] Fig. 9 is a schematic representation of an optical neuron according to a first embodiment;
[0048] Fig. 10 is a schematic representation of an optical neuron according to a second embodiment;
[0049] Fig. 11 is a schematic representation of an optical neuron according to a third embodiment;
[0050] Fig. 12 is a schematic diagram of an optical neuron according to a fourth embodiment;
[0051] Fig. 13 is a schematic representation of an optical neural network; and
[0052] Fig. 14 is a schematic representation of a control device with an optical neural network.
[0053] The exemplary embodiments explained below are preferred exemplary embodiments of the invention. In the exemplary embodiments, the described components each represent individual features of the invention that can be considered independently of one another. These features also further develop the invention independently of one another and are therefore to be considered as components of the invention, either individually or in a combination other than that shown. Furthermore, the described exemplary embodiments can also be supplemented by further features of the invention already described. In the figures, functionally identical elements are each provided with the same reference numerals.
[0054] Fig. 1 shows a schematic side view of a motor vehicle 1. The motor vehicle 1 has a control device 2 and a plurality of sensor devices 3. Each of the sensor devices 3 has a control device 4. This is designed, for example, to control the respective sensor device 3, i.e., to activate or deactivate it. Furthermore, by means of the respective control device 4, sensor data of the sensor device 3 can be stored and / or at least one setting parameter for capturing the sensor data by means of the sensor device 3 can be specified. The setting parameter describes, for example, a resolution of the sensor data captured by the sensor device 3. The sensor device 3 is either a camera 5, here a side camera, a radar device 6, or a lidar device 7. In a vertical direction of the motor vehicle 1, the lidar device 7 is arranged at the top and the radar devices 6 at the bottom.
[0055] The motor vehicle 1 has a plurality of optical fiber connections 8, via which the respective control device 4 is connected to the control device 2. It can be provided that the optical fiber connection 8 connects the control device 4, for example, a communication interface 38 (see reference numeral 38 in Fig. 14) of the control device 4, to the communication interface 38 of the control device 2. The respective control device 4 can be configured to transmit the sensor data acquired by the sensor device 3 to the control device 2 via the optical fiber connection 8.
[0056] Fig. 2 shows a schematic front view of the motor vehicle 1 from Fig. 1. Here, the lidar device 7 is arranged at the top center in the vertical direction of the motor vehicle 1, and the radar devices 6 are arranged at the top and side of the lidar device 7, with center and side referring to a transverse direction of the motor vehicle 1. The camera 5 is arranged at the bottom in the vertical direction.
[0057] Fig. 3 shows a schematic rear view of the motor vehicle 1 from Figs. 1 and 2. Here, the lidar device 7 is arranged at the top center in the vertical and transverse directions of the motor vehicle 1. Three radar devices 6 are arranged one above the other in the vertical direction, but only on one side in the transverse direction, namely the left side here. There is also a radar device 6 arranged at the bottom in the vertical direction and on the right in the transverse direction. The positions of the sensor devices 3 shown in Figs. 1 to 3 are purely exemplary. Alternative positions of the sensor devices 3 in the motor vehicle 1 are possible and are not sketched here.
[0058] Fig. 4 schematically illustrates the effect of electromagnetically induced transparency (EIT). This effect is based on a probe laser 29 (see reference numeral 29 in Fig. 13) irradiating a probe laser beam 9 and a coupling laser beam 10 (see reference numeral 30 in Fig. 13) irradiating a nonlinear optical medium 17 (see reference numeral 17 in Fig. 9), which can alternatively be referred to as an atomic system. The nonlinear optical medium 17 has the energy states E0, E1, and E2. The energy state E1, E1, E2 can alternatively be referred to as an energy level. By applying the probe laser beam 9 to the nonlinear optical medium 17, at least one electron of an atom of the nonlinear optical medium 17 having the energy state E0 can be raised to the energy state E1.By applying the coupling laser beam 10, it can also be raised from the energy state E1 to the energy state E2. In other words, the coupling laser 30 couples the energy states E1 and E2, which can be referred to as "coupled states" or "dressed states." The electron of the atom can thus make the transition from the energy state E0 to the energy state E2 by absorbing one photon each from the probe laser 29 and the coupling laser 30. The E2 state is preferably long-lived, i.e., for example, a metastable state. In this case, the nonlinear optical medium 17 becomes transparent to the probe laser when there is no longer any electron in the nonlinear optical medium 17 that can absorb a photon from the probe laser beam 9, since it has not done so yet, and the transition from the energy state E2 to the energy state E0 is forbidden. The forbidden transition can be referred to as the "dark state" and is shown in Fig.4 with a crossed arrow. Therefore, repopulation of the energy state E0 is not possible.
[0059] Fig. 5 shows the energy states E0, E1, E2 of the nonlinear optical medium 17 from Fig. 4, whereby an additional external electromagnetic field is now radiated. Thus, the nonlinear optical medium 17 is irradiated with electromagnetic radiation 27 (see reference numeral 27 in Fig. 12) by means of a radiation source 26 (see reference numeral 26 in Fig. 12). The electromagnetic radiation 27 is typically in the millimeter range and can, for example, have a frequency of 70 gigahertz. The electromagnetic radiation 27 shifts the energy state E2 to an energy state E3. The reason for this is the dynamic Stark effect, also called the AC Stark effect, or the Autler-Townes effect.
[0060] If the nonlinear optical medium 17 is now irradiated by means of the probe laser 29 and the coupling laser 30, a minimum in the absorption spectrum of the probe laser beam 9 forms as soon as a time and / or intensity of laser irradiation greater than a minimum time or minimum intensity specified for the nonlinear optical medium 17 used has been exceeded. In this case, photons of the probe laser 29 can pass through the nonlinear optical medium 17 without interaction, i.e., propagate through it without interacting with the nonlinear optical medium 17. This state of the nonlinear optical medium 17 can therefore be detected by recording the absorption spectrum of the probe laser 29. Alternatively or additionally, a transmitted power of the probe laser beam 9 can be measured directly.
[0061] Such an absorption spectrum 11, i.e., the absorbed power of the test laser beam 9 as a function of the frequency f, is sketched in Fig. 6. The absorption spectrum 11 was recorded without an activated radiation source 26. The absorption spectrum 11 exhibits a singlet structure as a function of the frequency f around the central frequency f = 0. The absorption spectrum 11 is therefore maximum around the central frequency f = 0 and thus has an extreme value 12 there. The absorbed radiation of the test laser beam 9 is therefore maximum at the central frequency f = 0.
[0062] Fig. 7 shows the absorption spectrum 11 for the case where the radiation source 26 is activated. This causes the absorption spectrum 11 to be split in the frequency domain, i.e., as a function of the frequency f, and a minimum to form around the sample laser wavelength. This results in a so-called "spectral splitting" in the absorption spectrum 11. In the energy state E3 excited by the external electromagnetic radiation 27, the resonance frequency shifts, resulting in the spectral splitting around the central frequency f = 0. The amplitude of the transmitted radiation decreases. A time-resolved measurement of the absorption spectrum 11 allows a phase of the incident sample laser radiation to be reconstructed. The minimum provides information about the frequency f of the external field, i.e., the radiation source 26, through the previously defined atomic transition to the energy state E3.
[0063] The width of the spectral splitting is shown here as the frequency separation Af between the two maximum extreme values 12 in the absorption spectrum 11. The frequency separation Af can be calculated using the following formula:
[0064] Here, E is the electric field strength of the external electromagnetic field achieved by the electromagnetic radiation 27, D is an atomic dipole moment for the transition of the energy state E2 to the energy state E3 induced by the electromagnetic radiation 27, A1 is a wavelength of the probe laser 29, and A2 is the wavelength of the coupling laser 30. Furthermore, the formula includes n for the number Pi and h for the reduced Planck constant.
[0065] If the electromagnetic field is detected sufficiently quickly, a phase profile of the electromagnetic field can be measured over time by the electromagnetic radiation 27. This provides the possibility of measuring the following additional parameters: the frequency f of the electromagnetic field when compared with a known spectral line (e.g., an integrated semiconductor-based optical frequency comb), a phase profile of the electromagnetic field, and an amplitude profile of the electromagnetic field. When the radiation source 26 is activated, the electromagnetic field generated thereby can be used as additional optical weight for an optical neural network (NN) 14 (see reference numeral 14 in Fig. 9).
[0066] Fig. 8 shows an EIT signal 13 as a function of the intensity I of the coupling laser 30. This function resembles a sigmoid function, i.e., the function frequently used as a nonlinear function of an optical neuron 16 (see reference numeral 16 in Fig. 9) within the NN 14. Fig. 8 thus illustrates that the transmission induced by the coupling laser 30 can be recorded as a nonlinear function similar to the sigmoid function, while the coupling laser 30 acts as an optical weight. In an off state, the coupling laser 30 can maximize its absorption by the nonlinear optical medium 17, which corresponds to the weighting. It is also possible to induce a rabioscillation by the probe laser using a pump laser, which can also act as a weighting. The intensity I or a frequency detuning of the coupling laser 30 determines a strength of the EIT signal 13.
[0067] Furthermore, optical weighting can be achieved by modulating the amplitude, phase, and / or frequency f using a corresponding modulator for controlling the probe laser 29 and / or a pump laser for the probe laser 29. Fig. 9 shows a first exemplary embodiment of an optical neuron 16 for the optical NN 14, wherein only one optical neuron 16 is sketched here as an example. It schematically shows an optical neuron 16 based on an EIT cell on a semiconductor die 15, which is embodied here as an electronic-photon cointegrated semiconductor chip (EPIC). The semiconductor die 15 can alternatively be embodied as a photonic integrated circuit, as a multi-chip module, and / or as a chip mounted using flip-chip assembly. The optical neuron 16 here has the nonlinear optical medium 17, which is exemplary formed as a quantum dot (for example made of InAs or GaAs), quantum wire or quantum well.
[0068] The quantum dot, quantum wire or quantum well as a nonlinear optical medium 17 has an optional reflective coating 18 on its two outer sides.
[0069] At least one first optical waveguide 19 leads to the optical neuron 16, via which the probe laser beam 9 of the probe laser 29 and the coupling laser beam 10 of the coupling laser 30 can be coupled into the optical neuron 16. At least one second optical waveguide 20 also leads away from the optical neuron 16. An optocoupler 21 is also shown here, via which an optical waveguide 22 can be coupled into the first optical waveguide 19. The first optical waveguide 19 and the second optical waveguide 20 are arranged on the semiconductor die 15 at least from and up to the optocoupler 21.
[0070] An optional optical modulator 23 can also be arranged on the semiconductor die 15. Furthermore, an electronic circuit, designed, for example, for diagnostic purposes, can be provided on the semiconductor die 15. The electronic circuit is an evaluation device 24 of the optical neuron 14. The second optical fiber 20 leads directly to the evaluation device 24. The test laser beam 9 can be coupled into the second optical fiber 20 after it has passed through the optical neuron 16.
[0071] In other words, Fig. 9 shows that the first optical waveguide 19 couples radiation from the probe laser 29 and coupling laser 30 with the data to be processed, which may be the sensor data from the sensor device 3, into a photonic semiconductor, sketched here by the semiconductor die 15. These are guided to the nonlinear optical medium 17 through integrated photonic structures. The laser radiation, i.e., the probe laser beam 9 and the coupling laser beam 10, interact with the nonlinear optical medium 17, creating the EIT described above. The frequency and / or amplitude information of the transmission spectrum and / or the absorption spectrum 11 is passed on via the second optical waveguide 20 to the next layer of optical neurons 16 (not sketched here) before being transmitted to the evaluation device 24. The evaluation device 24 is used for diagnosis and control.Optionally, the optical modulator 23 can also be integrated to influence the amplitude, polarization, frequency, or phase, which can also be used for data transmission to a central unit, i.e., the evaluation device 24. Several of the EIT cells shown in Fig. 9 can be combined on an electronic-photonically cointegrated chip to form an optical NN 14.
[0072] Fig. 10 shows a second embodiment in which the EIT cell is integrated, for example, on a purely photonic chip, i.e., without the use of electronic components. The nonlinear optical medium 17 is embodied here as a vapor cell. This comprises, for example, at least one atom and / or molecule with a predetermined energy level, in particular a Rydberg atom and / or a Bose-Einstein condensate. Fig. 10 thus shows a simplified optical neuron 16 based on the photonic chip. Instead of the optocoupler 21, a photonic coupler 25 is arranged at the input of the first optical waveguide 19 and at the output of the second optical waveguide 20. The first optical waveguide 19 again couples the probe laser beam 9 and the coupling laser beam 10 into the photonic chip, which is then guided to the integrated EIT cell, i.e., the nonlinear optical medium 17.Within the vapor cell, which can alternatively be referred to as a gas cell, the coupling laser 30 and the probe laser 29 interact with the nonlinear optical medium 17, creating the above-described EIT. This occurs particularly when an external electromagnetic field (not shown here) interacts with the nonlinear optical medium 17. The frequency and / or amplitude information is fed back via the second optical fiber 20 to a central data processing module, for example, the evaluation device 24, and optically detected.
[0073] Fig. 11 shows a third embodiment, which can be referred to as a highly integrated design with an EIT cell with a vapor cell / Bose-Einstein condensate as the nonlinear optical medium 17. In other words, Fig. 11 shows a schematic representation of the optical neuron 16 based on an EIT cell on an electronic-photonic cointegrated semiconductor chip (EPIC). The first optical waveguide 19 couples coherent radiation into a photonic semiconductor and leads to the integrated vapor cell / Bose-Einstein condensate. Within the EIT cell, the coupling laser 30 and the probe laser 29 interact with the nonlinear optical medium 17, creating the EIT described above. Optionally, additional optical modulators 23 can be integrated to influence the amplitude, polarization, frequency, or phase; these can also be used to transmit data to the evaluation device 24.
[0074] The optical weighting of the optical NN 14 can be achieved by selecting the externally radiated electromagnetic radiation 27, for example, in the gigahertz range. By selecting the externally radiated electromagnetic radiation 27 in the gigahertz range, the absorption spectrum 11 can be modulated such that the transmitted portion of the probe laser 29 is modulated, which acts as a weighting.
[0075] Fig. 12 shows a fourth embodiment in which the nonlinear optical medium 17 can again be designed as a quantum dot, quantum wire, and / or quantum well, for example, as an InAs / GaAs quantum dot. This embodiment is characterized by the use of the external radiation sources 26 as an optical weight, wherein the electromagnetic radiation 27 is radiated onto the optical neuron 16 by means of the external radiation source 26. The semiconductor die 15 can be configured analogously to Fig. 9. The optical NN 14 is configured to adjust a frequency, a phase response, and / or an amplitude of the electromagnetic radiation 27 depending on the sensor data.
[0076] The sketched first optical waveguide 19 and the sketched second optical waveguide 20 can alternatively be interchanged in Figs. 9 to 12.
[0077] Fig. 13 shows an example of the optical NN 14 with multiple optical neurons 16, specifically the integration on a photonic-electronic cointegrated chip. Here, the probe laser 29 is sketched, which couples the probe laser beam 9 into a plurality of first optical waveguides 19 via the optocoupler 21. Also sketched is the coupling laser 30, which couples the coupling laser beam 10 into three parallel layers of optical neurons 16 via a plurality of optocouplers 21. The optocoupler 21 can be configured to perform a wavelength division multiplexing method and / or a time division multiplexing method in order to encode the image information, i.e., the optically encoded sensor data, in the wavelength, i.e., to be able to provide the weighting by the coupling laser 30. Alternatively or additionally, multiple channels can be provided, and the weighting can thus be achieved by means of the coupling laser 30. The optic neurons are as shown in Fig.9 Here each exemplified as a quantum dot, quantum wire and / or quantum well.
[0078] For the sketched optical NN 14, a plurality of optical weights 31 are also shown, each of which is assigned to one of the optical neurons 16. Furthermore, the optical NN 14 can comprise a conversion device 32 in addition to the evaluation device 24. The combination of evaluation device 24 and conversion device 32 can be referred to as an electro-optical interface or electronic data processing device and can have precisely these functions. The evaluation device 24 together with the conversion device 32 provide electronically coded classification information 33 that describes the classification of the sensor data as an electronic signal or electronic data. Furthermore, Fig. 13 shows that the second optical waveguide 20 leads from each optical neuron 16 of one layer to each optical neuron 16 of the adjacent layer.Alternatively, only selected and thus individual optical neurons 16 of a layer can be connected to selected optical neurons 16 of the adjacent layer via the respective second optical waveguide 20.
[0079] Fig. 14 shows the control device 2 in detail. The control device 2 has, for example, a computing unit 34, in particular a computer and / or a graphics processor. The computing unit 34 is designed to provide electronically coded sensor data. Furthermore, a further conversion device 32' of the control device 2 can be designed to convert the provided electronically coded sensor data into optically coded sensor data 36 and to provide them. This means that the control device 2 is designed to provide optically coded sensor data 36 for the optical NN 14.
[0080] The control device 2 is designed to evaluate the provided optically coded sensor data 36 by means of the optical NN 14, wherein the optical NN 14 has the plurality of optical neurons 16 with electromagnetically induced transparency properties, and as a result of the evaluation by means of the evaluation device 24 of the optical NN 14, to provide optically coded classification information 37 which describes the classification of the sensor data, specifically as an optical signal or optical data.
[0081] The conversion device 32 of the control device 2 is designed to interfere with local oscillator information provided by a local oscillator 35 of the control device 2 with the provided optically coded classification information 37 in order to provide the electronically coded classification information 33. The control device 2 is thus designed to convert the optically coded classification information 37 into electronically coded classification information 33 by means of the conversion device 32 and ultimately to provide the electronically coded classification information 33 for the motor vehicle 1. Instead of or in addition to the local oscillator 35, a phase lock can be provided.
[0082] The control device 2 can also have the communication interface 38, via which the sensor data can be received from the control devices 4 of the sensor devices 3.
[0083] It can further be provided that the control device 2 has a diagnostic and / or control interface 39, a digital interface 40, a processing unit 41, which can be provided optionally and can, for example, perform a fast Fourier transformation for processing the electronically encoded classification information 33, and an optical switch 42. The conversion device 32 can alternatively or additionally be designed for optical detection, homodyne or heterodyne detection, and / or for frequency downconversion. Alternatively or additionally, the probe laser 29 and the coupling laser 30 can be arranged at the location of the further conversion device 32'. The further conversion device 32' can have an optical frequency comb and / or a microring resonator.
[0084] In Fig. 14, transmission paths for electronically encoded data are outlined with a solid line, and transmission paths for optically encoded data are outlined with a dashed line. Furthermore, an electronic transmission path to the neurons 16 can be provided, which is sketched here as an arrow without an endpoint emanating from the further conversion device 32'.
[0085] The control device 2 shown in Fig. 14 enables: combined processing of the data in a spatially separated central unit; coherent processing; integration of optical frequency combs for broadband sampling of the frequencies of external electromagnetic fields; and integration of the local oscillator 35 or a phase-locking unit to electronically or optically phase-lock neurons 16 and use them for frequency mixing in the detector. Furthermore, the classification of data can be performed in the 5G frequency band or beyond for: data processing for vehicle-to-infrastructure applications (software updates, map updates, infrastructure signals), gesture recognition, and / or user interface applications.
[0086] Overall, the examples demonstrate the use of EIT as an optical neuron 16 in artificial optical NNs 14.
[0087] Reference symbols
[0088] Motor vehicle
[0089] Control device
[0090] Sensor device
[0091] Control device
[0092] camera
[0093] radar device
[0094] Lidar device optical fiber connection
[0095] Test laser beam
[0096] Coupling laser beam
[0097] Absorption spectrum
[0098] Extreme value
[0099] EIT signal optical neural network
[0100] semiconductor chip optical neuron nonlinear optical medium
[0101] Coating of first optical fiber, second optical fiber
[0102] Optocoupler
[0103] Fiber optic cable optical modulator
[0104] Evaluation device for photonic couplers
[0105] Radiation source electromagnetic radiation
[0106] Test laser
[0107] Coupling laser optical weighting
[0108] Conversion device ' further conversion device electronically coded classification information
[0109] Computing unit
[0110] Local oscillator 36 optically coded sensor data
[0111] 37 optically coded classification information
[0112] 38 Communication interface
[0113] 39 Diagnostic and / or control interface
[0114] 40 digital interface
[0115] 41 processing unit
[0116] 42 optical switch E0-E3 energy state f frequency
[0117] Af frequency spacing
[0118] I Intensity
Claims
Claims: Control device (2) for providing a classification of sensor data for a motor vehicle (1) by means of an optical neural network (14), wherein the control device (2) comprises the optical neural network (14) and is configured to: provide optically encoded sensor data; evaluate the provided optically encoded sensor data (36) by means of the optical neural network (14), wherein the optical neural network (14) comprises several optical neurons (16) with electromagnetically induced transparency properties, and provide optically encoded classification information (37) as a result of the evaluation by means of an evaluation device (24) of the optical neural network (14).which describes the classification of the sensor data; to convert the optically encoded classification information (37) into electronically encoded classification information (33) by means of a conversion device (32) of the control device (2); and to provide the electronically encoded classification information (37) to the motor vehicle (1). Control device (2) according to claim 1, wherein the at least one optical neuron (16) is formed from a nonlinear optical medium (17), in particular as a quantum dot, quantum wire, quantum well and / or vapor cell, which has at least one atom and / or molecule with a predetermined energy level, in particular a Rydberg atom and / or a Bose-Einstein condensate. Control device (2) according to any one of the preceding claims, wherein the optical neural network (14) has at least one layer having several optical neurons (16),wherein at least one first optical waveguide (19) leads to each optical neuron (16), via which a probe laser beam (9) of a probe laser (29) and a coupling laser beam (10) of a coupling laser (30) can be coupled into the optical neuron (16), and at least one second optical waveguide (20) leads away from each optical neuron (16), which leads to another optical neuron (16) or directly to the evaluation unit (24) and into which the probe laser beam (9) can be coupled after passing through the optical neuron (16). Control device (2) according to claim 3, wherein the optical neural network (14) is configured to evaluate an absorption spectrum (11) and / or transmission spectrum of the probe laser beam (9) coupled into the at least one second optical fiber (20) to provide the optically encoded classification information (37), in particular an extreme value (12) of a frequency and / or an amplitude on the one hand and / or a phase on the other hand of the absorption spectrum (11) and / or transmission spectrum. Control device (2) according to claim 3 or 4, wherein the optical neural network (14) comprises at least one optical modulator (23) configured to influence at least one property of the probe laser beam (9), in particular an amplitude, polarization, frequency, wavelength and / or phase of the probe laser beam (9). Control device (2) according to any one of claims 3 to 5, wherein the optical neural network (14) comprises a radiation source (26),which is configured to irradiate the at least one optical neuron (16) with electromagnetic radiation (27) such that, by utilizing the dynamic Stark effect, the absorption spectrum (11) and / or the transmission spectrum of the probe laser beam (9) coupled into the at least one second optical fiber (20) is split, wherein the optical neural network (14) is configured to evaluate the split absorption spectrum (11) and / or transmission spectrum to provide the optically encoded classification information (37), in particular an extreme value (12) and / or a distance between two extreme values (12) of a frequency and / or an amplitude of the absorption spectrum (11) and / or transmission spectrum on the one hand, and / or a phase of the absorption spectrum (11) and / or transmission spectrum on the other hand. Control device (2) according to claim 6, wherein the optical neural network (14) is configured toto adjust a frequency, phase profile and / or amplitude of the electromagnetic radiation (27) depending on the sensor data. Control device (2) according to one of the preceding claims, wherein the control device (2) comprises a computing unit (34), in particular a computer and / or a graphics processor, and a further conversion device (32'), wherein the computing unit (34) is configured to provide electronically encoded sensor data, and the further conversion device (32') is configured to is to convert the provided electronically encoded sensor data into optically encoded sensor data and to provide this data. Control device (2) according to one of the preceding claims, wherein the conversion device (32) is configured to interfere with a local oscillator information provided by a local oscillator (35) with the provided optically encoded classification information (37) in order to provide the electronically encoded classification information (33). Control device (2) according to one of the preceding claims, wherein the optical neural network (14) is arranged on at least one semiconductor chip (15) configured as at least one of the following components: electronically-photonically cointegrated semiconductor chip; photonic integrated circuit; Multi-chip module; and / or chip mounted by means of flip-chip assembly. Control device (2) according to one of the preceding claims, wherein the control device (2) has a communication interface (38) to a control unit (4) of a sensor device (3) of the motor vehicle (1) and is configured to receive sensor data acquired by means of the sensor device (3) via the communication interface (38). Motor vehicle (1) with a control device (2) according to one of the preceding claims.Motor vehicle (1) according to claim 12 in its references to claim 11, wherein the motor vehicle (1) has at least one sensor device (3) with a control device (4), the control device (4) being connected to the communication interface (38) of the control device (2) via at least one optical fiber connection (8), and the control device (4) being configured to transmit the sensor data acquired by means of the sensor device (3) to the control device (2) via the at least one optical fiber connection (8). Method for operating a control device (2) for providing a classification of sensor data for a motor vehicle (1) by means of an optical neural network (14), comprising the following steps: Providing optically encoded sensor data; - Evaluating the provided optically encoded sensor data using an optical neural network (14) which has several optical neurons (16) with electromagnetically induced transparency properties, wherein as a result of the evaluation using an evaluation device (24) of the optical neural network (14) an optically encoded classification information (37) is provided which describes the classification of the sensor data; Converting the optically encoded classification information (37) into electronically encoded classification information (33) using a conversion device (23); Providing the electronically coded classification information (33) for the motor vehicle (1).
