Method for optimizing the thermal load of a light source and light system

Abstract

The invention relates to a method for optimizing the thermal load of a light source (1), wherein the light source (1) comprises at least four light-emitting elements (1aa-1nm) arranged in an n x m matrix. The invention further relates to a light system (LS), in particular headlamp, for radiating light (L).

Description

[0001] The present invention relates to a method for optimizing the thermal load of a light source, wherein the light source comprises at least four light-emitting elements arranged in an n x m matrix, wherein n, m ≥ 2.

[0002] Thermal management is critical for light sources, particularly those employing light-emitting diodes (LEDs), as their performance, lifespan, and safety are highly dependent on maintaining junction temperatures below a specified maximum threshold. Exceeding the maximum junction temperature can result in device damage, reduced efficiency, and shorter operational life. Consequently, various methods have been developed in the state of art to optimize the thermal load of light sources and prevent temperature overshoots.

[0003] Known techniques for managing the thermal load of light sources usually limit the power provided to the light sources. A common method for reducing the average power output is by manipulating the time duration of an "on"-phase when applying pulse width modulation (PWM). These approaches are widely applied in systems where temperature control is crucial, such as in automotive headlamps, which frequently utilize pixelated light sources like a multichip array or a matrix LED comprising multiple LEDs arranged in a specific pattern, e.g. multiple LEDs arranged in an array. Current temperature monitoring systems for LED arrays often rely on a limited number of temperature sensors distributed across the respective LED array. However, this coarse monitoring can lead to an inaccurate assumption of the temperature in between the temperature sensors and therefore to inadequate derating measures for individual LEDs.

[0004] It is an object of the present invention to optimize the thermal load of light-emitting elements of a light source via avoiding or at least reducing temperature overshoots.

[0005] The present invention relates to a method comprising the following steps: a) providing a first data stream of light images, b) operating the light source in a predetermined environment and training a first neural network, said step b) comprising the following sub-steps: b1) controlling, preferably by a control unit, the light source and its light-emitting elements to radiate the light images, wherein each light image is radiated by the light source for a predetermined period of time, b2) while performing sub-step b1), measuring temperature values of at least one of the light-emitting elements over the predetermined period of time of sub-step b1), b3) before, while and / or after the latest sub-steps b1) and b2), feeding the first data stream to the first neural network to calculate first estimated temperature values of the at least one light-emitting element based on the first data stream, b4) comparing the first estimated temperature values of sub-step b3) with the measured temperature values of sub-step b2), and in case that a difference between at least one of the first estimated temperature values and at least one of the measured temperature values meets, for the same light-emitting element, a predetermined success condition, proceeding with step c), otherwise repeating sub-steps b1) to b4) to reduce the difference until the success condition is met, and hereby training the first neural network via supervised learning, c) operating the light source in a non-predetermined environment while using a second neural network, said step c) comprising the following sub-steps: c1) providing a second data stream of non-verified light images, c2) feeding the second data stream of non-verified light images to the first neural network to calculate second estimated temperature values of each of the light-emitting elements based on the second data stream of non-verified light images, and comparing the second estimated temperature values with a predetermined temperature threshold to obtain a verified data stream of verified light images, c21) wherein at least in case that at least one of the second estimated temperature values exceeds the predetermined temperature threshold, the second data stream of non-verified light images and the corresponding second estimated temperature values are fed to the second neural network, and the second neural network manipulates at least one of the non-verified light images of the second data stream with the goal of avoiding exceeding the predetermined temperature threshold in any of the light-emitting elements while minimizing the extent of manipulation of the at least one of the non-verified light images, hereby obtaining the verified data stream of verified light images, and c22) wherein in case that none of the second estimated temperature values exceeds the predetermined temperature threshold, either still a manipulation according to sub-step c21) takes place, or the second data stream of non-verified light images is not manipulated and set as the verified data stream of verified light images, c3) controlling, preferably by the control unit, the light source and its light-emitting elements to radiate the verified light images of the verified data stream.

[0006] The method according to the present invention allows to train the first neural network in a predetermined (i.e. controlled) environment via supervised learning, such to render it capable of calculating first estimated temperature values of at least one of the light-emitting elements of the light source, preferably of all light-emitting elements. When subsequently using the trained, first neural network in a non-predetermined environment, e.g. in an automotive headlamp during operation of the vehicle, it is possible to calculate second estimated temperature values of the light-emitting elements based on the second data stream of non-verified light images without the use of any temperature sensors. The obtained second estimated temperature values are then compared with the predetermined temperature threshold. If a temperature overshoot is predicted, the non-verified light images are manipulated by the second neural network with the goal of avoiding exceeding the predetermined temperature threshold while minimizing the extent of manipulation. As such, temperature overshoots of the light-emitting elements can be considerably reduced or even avoided. The present invention thus enhances the performance and lifespan of the light source, hereby enormously contributing to sustainability and safety.

[0007] Preferably, the manipulation in step c21) is executed in a manner that a luminous flux radiated by the light source remains unchanged. Preferably, the luminous flux remains unchanged for each verified light image or over a sequence of two or more verified light images. An unchanged luminous flux ensures that the illumination required for the specific driving situation is still met, even if the light image is slightly changed. Preferably, the resolution of all the light images of the first data stream, and / or all the non-verified light images of the second data stream matches (i.e. equals) the resolution determined by the number of light-emitting elements. That is, each light image preferably comprises information for each light-emitting element, e.g. about whether it is in a switched-on- or switched-off-state, and / or about its intensity, which is preferably modulated via PWM parameters. As such, every light image preferably represents the light distribution of all light-emitting elements.

[0008] It is preferred that the light source is a pixelated light source, preferably a multichip array, particularly preferably a light-emitting diode array. Since headlamps, particularly automotive headlamps, generally comprise a multichip array, particularly an LED array, the present invention is suitable for use in such headlamps.

[0009] The light source preferably comprises 24 to 100 000 light-emitting elements, more preferably 30 to 80 000, yet more preferably 60 to 70 000, particularly preferably 100 to 65 000. As such, the light source is suitable for use in the automotive industry. For example, in the high-end sector, about 60 000 light-emitting elements are currently used.

[0010] The light-emitting elements are preferably light-emitting diodes (LEDs). Preferably, a maximum expansion (e.g. a maximum diameter) of a light-emitting element, particularly LED, in its direction transversal to a direction of radiation of light images ranges from 5 to 100 µm, preferably from 10 to 50 µm. Pixel pitches of the light-emitting elements, particularly LEDs, are preferably below 80 µm, more preferably below 60 µm, particularly preferably of from 40 to 50 µm. These dimensions allow for a change in the light distribution caused by manipulation with the inventive method not to, at least not significantly, impact the perception of the light distribution by human beings.

[0011] The first data stream of light images and / or the second data stream of non-verified light images provided in step a) and / or in sub-step c1) of the method, respectively, preferably comprise digital representations of the light images. It is preferred that the light images of the first and / or second data stream comprise pulse width modulation (PWM) parameters, wherein the PWM parameters preferably comprise current, voltage, actual video data and / or actual image data.

[0012] In step b), the light source is operated in a predetermined environment. The predetermined (i.e. controlled) environment preferably comprises a predetermined temperature, a predetermined thermal gradient ranging from -40 °C to 160 °C and / or a predetermined heat path from the light source to a corresponding colling element, for example a heat sink, of the light module. This allows for good control of the predetermined environment. Preferably, the predetermined environment is comprised in a closed area. This ensures stable training conditions for the first neural network.

[0013] In sub-step b1), the light source and its light-emitting elements are controlled, preferably by a control unit, to radiate the light images of the first data stream. Preferably, said light images are radiated in a random sequence from a predetermined set of image data. This enhances the adaptability of the first neural network to various conditions and reduces the likelihood of the first neural network becoming too specialized on a particular order of inputs. Alternatively, the light images of the first data stream are preferably radiated in a predetermined sequence from the predetermined set of image data. This allows for gradual learning, systematic testing and debugging, and further facilitates performance comparison of different models or training configurations (e.g. differences in the predetermined environment, such as a different temperature; differences in the configuration of the first neural network). Alternatively, it is preferred that one or more boundary conditions for the light source to match real life conditions are given (e.g. legal provisions, customer specifications). The light images are then created randomly, such to create as many different images as possible.

[0014] Further in sub-step b1), each light image of the first data stream is radiated by the light source for a predetermined period of time. The predetermined period of time preferably is 100 ms or below, more preferably 50 ms or below, yet more preferably 10 ms or below. It is preferred that the predetermined period of time ranges from 1 to 100 ms, preferably 30 to 50 ms. This allows to closely align the training conditions with conditions experienced in the non-predetermined environment during step c).

[0015] In sub-step b2), temperature values of at least one of the light-emitting elements are measured over the predetermined period of time of sub-step b1). Preferably, temperature values of at least 60% of all light-emitting elements are measured, more preferably temperature values of at least 80%, particularly preferably temperature values of at least 95%. Additionally, or alternatively, it is preferred that two or more temperature values of at least one of the light-emitting elements are measured, more preferably five or more, yet more preferably ten or more. That is, preferably, a sequence of temperature values of at least one of the light-emitting elements is measured. The more light-emitting elements and / or the more temperature values are used for temperature measurement, the better and faster can the first neural network be trained. Particularly preferably, in sub-step b2), temperature values of every light-emitting element are measured over the predetermined period of time of sub-step b1). This considerably improves the training efficiency of the first neural network. Alternatively, it is preferred that in sub-step b2), a temperature value of at least one cluster of light-emitting elements is measured. Preferably, the cluster comprises eight light-emitting elements arranged in a 4x4 matrix, more preferably four light-emitting elements arranged in a 2x2 matrix. When using light-emitting elements, particularly LEDs, with a small size, temperature measurements of a single LED may lead to inaccuracy due to small pixel pitches. This issue can be overcome by grouping a certain number of light-emitting elements into a cluster and measuring temperature values of said cluster. Preferably, at least 60% of all light-emitting elements are grouped into clusters, more preferably at least 80%, yet more preferably at least 90%, particularly preferably at least 99%. The temperature values are preferably measured with a temperature sensor, preferably with an infrared sensor, particularly preferably in a wavelength range of from 0.75 to 1.5 µm.

[0016] In sub-step b3), the first data stream is fed to the first neural network to calculate first estimated temperature values of the at least one light-emitting element based on the first data stream. Sub-step b3) is performed before, while and / or after the latest sub-steps b1) and b2). Which sequence of the sub-steps to choose depends on various factors, inter alia on the predetermined period of time during which each light image is radiated in sub-step b1), on the number of light images radiated in sub-step b1), on the measurement speed in sub-step b2), and on the calculation speed of the first neural network in step b3). As sub-steps b1) to b4) can be iteratively performed, the latest sub-steps b1) and b2) intend to refer to the sub-steps occurring in the same iteration as current sub-step b3).

[0017] In sub-step b4), the first estimated temperature values of sub-step b3) are compared with the measured temperature values of sub-step b2). In case that the difference between at least one of the first estimated temperature values and at least one of the measured temperature values meets, for the same light-emitting element, and preferably at the same point in time, a predetermined success condition, it is proceeded with step c). The predetermined success condition is preferably a temperature difference of 10% or below, more preferably 5% or below, yet more preferably 3% or below, particularly preferably 1% or below. As such, it is possible to obtain a well-trained first neural network before proceeding with step c).

[0018] If the predetermined success condition is not met in sub-step b4), sub-steps b1) to b4) are repeated to reduce the difference until the predetermined success condition is met. The first neural network is hereby trained via supervised learning. To train the first neural network via supervised learning, inputs and corresponding desired outputs are used. The training process is based on a function which is able to map new data to a desired output. Over time, the supervised guidance helps the first neural network to improve its accuracy in making predictions, hereby optimizing the function. The first neural network's algorithm preferably uses the measured temperature values (i.e. current temperature distribution) as input and the calculated first estimated temperature values (i.e. future temperature distribution) as output, and preferably finds a function that assigns an input to an output with as few errors as possible, preferably whilst performing neither an underfitting (i.e. no representation of the function) nor an overfitting (i.e. mapping of each individual data point). The speed of supervised learning of the first neural network depends on various parameters, such as the number of light-emitting elements, the step size and the predetermined success condition.

[0019] Preferably, the supervised learning of the first neural network is performed by using a loss function (also referred to as cost function) with the goal to minimize said loss function. A loss function is understood to be a mathematical function that measures how well or poorly the first neural network's predictions match the actual training data, it provides feedback on the first neural network's performance and guides the learning process by minimizing errors during training. Preferably the loss function is a classification function or a regression function, particularly preferably a regression function. Preferably, the loss function is selected from Mean Square Error (MSE; corresponds to Quadratic Loss function or L2 Loss), Mean Absolute Error (MAE; corresponds to L1 Loss), and Smooth Mean Absolute Error (combination of MSE and MAE; corresponds to Huber Loss). MSE is preferably used, inter alia because of its mathematical simplicity and its convex nature, which simplifies the training process and makes optimization efficient even for large datasets. However, it is preferred to not use MSE in case the training data stream comprises many outliers. MAE is preferably used if the training data stream comprises many outliers. As will be readily evident by a person skilled in the art, the loss function will need to be selected in accordance with the specific application, including the specific training data stream provided. Also, other loss functions as those listed above may be suitable, such as Log-Cosh-Loss or Quantile Loss.

[0020] In step c), the light source is operated in a non-predetermined environment while using the second neural network. The non-predetermined environment is e.g. the environment of a car while driving on a country road or in a city.

[0021] Preferably, in sub-step c1), a continuous stream of non-verified light images is provided, and in sub-step c3), a continuous stream of verified light images is radiated, wherein each non-verified light image is preferably processed individually in sub-step c2), wherein sub-step c3) is provided with a consecutive continuous stream of verified light images, wherein in sub-step c3), each verified light image is preferably radiated until it is replaced by a consecutive verified light-image. That is, preferably, every single light image is used individually (i.e. separately) by the first neural network for calculation of second estimated temperature values, which predict the future temperatures of the input temperatures given. That each verified light image is preferably radiated until it is replaced by a consecutive verified light image means that a first verified light image is preferably continued to be radiated in case the time required to obtain and radiate a consecutive, second verified light image exceeds a predetermined period of time (e.g. due to prolonged required manipulation time), preferably a time of 20 ms or below, more preferably 15 ms or below, yet more preferably 10 ms or below. This ensures continuous radiation of verified light images.

[0022] Preferably, in sub-step c1), a continuous stream of non-verified light images is provided, and in sub-step c3), a continuous stream of verified light-images is radiated, wherein in sub-step c2), the calculation of second estimated temperature values and manipulation according to sub-step c21), or potentially sub-step c22), is preferably performed for sets of non-verified light images, wherein each set preferably comprises two or more non-verified light images and the extent of manipulation is preferably distributed, particularly preferably evenly distributed, between the non-verified light images of each set, wherein in sub-step c3), each verified light-image is preferably radiated until it is replaced by a consecutive verified light-image. This can be useful if an interruption routine is implemented according to which only light images of the same second data stream provided in step c2) (i.e. of the same light function) can be used for averaging (instead of second data streams provided in two or more consecutive steps c2)). In that case, a change in the light function sets time boundaries to light images usable for averaging. A set of non-verified light images preferably comprises a non-verified light image wherein a temperature overshoot is noticed, and at least one other non-verified light image.

[0023] This allows to spread the manipulation on the entire set of non-verified light images, hereby reducing the extent of manipulation of each individual non-verified light image.

[0024] Preferably, the size of the set of non-verified light images and the speed of the execution of sub-step c2) are designed in a manner that a time delay between providing a non-verified light image in sub-step c1) and radiating a corresponding verified light image in sub-step c3) is 20 ms or below, more preferably 15 ms or below, yet more preferably 10 ms or below. The toleration of a certain time delay (i.e. time buffer) allows for a more efficient procedure.

[0025] Preferably, the predetermined temperature threshold is below a value of 80% to 99% of a maximum junction temperature, wherein the maximum junction temperature preferably ranges from 150 °C to 160 °C. This provides enhanced safety in terms of avoiding temperature overshoots in the light-emitting elements. Preferably, the predetermined temperature threshold is below a value of 85% to 99%, more preferably 90% to 99%, particularly preferably 95% to 99%. The maximum junction temperature preferably ranges from 110 °C to 200 °C, more preferably from 130 °C to 180 °C, yet more preferably from 150 °C to 160 °C. The selected predetermined temperature threshold depends inter alia on the non-predetermined environment in which step c) is performed. For example, in cities where rather slow driving speeds are required, changes of light images of automative headlamps while driving are comparatively slow, resulting in a continuous increase of the temperature of the light-emitting elements, which allows for the predetermined temperature threshold to be set in close proximity to the maximum junction temperature, e.g. to a value below 95% to 99% of the maximum junction temperature. In contrast, when driving on a country road, the high beam needs to be switched on and off, depending on whether there is oncoming traffic. This results in a rapid rise of the temperature of the light-emitting elements such that an earlier intervention is required to avoid reaching the maximum junction temperature. As such, the predetermined temperature threshold is then preferably below a value of 80% to 90% of the maximum junction temperature, more preferably 80% to 85 %. Preferably, the predetermined temperature is determined by a lookup table (LUT). The LUT preferably correlates the rate of temperature increase (i.e. the steepness of rise) and the respective predetermined temperature threshold. This allows to quickly react to sudden temperature increases.

[0026] Sub-step c21) is performed at least in case that at least one of the second estimated temperature values exceeds the predetermined temperature threshold (i.e., if a temperature overshoot is detected). In sub-step c21), the second data stream of non-verified light images and the corresponding second estimated temperature values are fed to the second neural network. The second neural network then manipulates at least one of the non-verified light images of the second data stream with the goal of avoiding exceeding of the predetermined temperature threshold in any of the light-emitting elements while minimizing the extent of manipulation of the at least one of the non-verified light images. The verified data stream of verified light images is obtained hereby. By minimizing the extent of manipulation, the second neural network takes care to alter the light distribution as little as necessary. For example, instead of continuing to use a light-emitting element with a temperature sufficiently close to the predetermined threshold at 100% intensity, four adjacent light-emitting elements could each be used at intensities lower than 100%, which allows for the light-emitting element with a temperature sufficiently close to the predetermined threshold to cool off and avoid a temperature overshoot whilst also changing the light distribution only in an insignificant manner which can not be recognized by the human eye.

[0027] Preferably, in sub-step c21), the second data stream of light images and the corresponding second estimated temperature values are fed to the second neural network regardless of whether or not at least one of the second estimated temperature values exceeds the predetermined temperature threshold, and in case that none of the second estimated temperature values exceeds the predetermined temperature threshold, it is proceeded with sub-step c22) without or with reduced manipulation. Thus, sub-step c21) is preferably performed not only in case the predetermined temperature threshold is exceeded, which allows for a manipulation, at least a reduced manipulation, to still be performed. This is preferred if any temperature values of the light-emitting elements are already in close proximity to the predetermined temperature threshold, e.g. of from 5% to 10% below the predetermined temperature threshold. A reduced manipulation can then be performed as a precaution. Alternatively, optionally, no manipulation is performed in case that none of the second estimated temperature values exceeds the predetermined temperature threshold, and the second data stream of non-verified light images is set as the verified data stream of verified light images.

[0028] In step c3), the light source and its light-emitting elements are controlled, preferably by a control unit, to radiate the verified light images of the verified data stream. Preferably, the control unit is the same control unit used in step b1).

[0029] Preferably, sub-steps c1) to c3) are iteratively performed while receiving a continuous stream of non-verified light images to obtain and radiate verified light images, wherein in sub-step c2), also previous verified light images are provided to the first neural network, and wherein the impact of the manipulation performed by the second neural network is fed back to the second neural network to enable its optimization. This enables reinforcement learning of the second neural network. The impact of the manipulation is understood to refer to a difference between the temperature associated with a non-verified light image and a corresponding verified light image.

[0030] It is preferred that in sub-step c2), also previously verified light images are provided to the first neural network, and the impact of the manipulation performed by the second neural network is fed back to the second neural network to enable its optimization, and in case that the goal of avoiding exceeding the predetermined temperature threshold is not met in an iteration of sub-step c21), another iteration of sub-steps c2) and c21) is initiated until the goal of avoiding exceeding the predetermined temperature threshold is met. As such, reinforcement learning of the second neural network occurs, and a temperature overshoot can be completely avoided.

[0031] Preferably, after sub-steps c2) or c3), the verified data stream is fed to the first neural network and sub-steps c1) and c2) are repeated, wherein in sub-step c2), second estimated temperature values of each of the light-emitting elements are calculated based on the second data stream of latest sub-step c2), and verified estimated temperature values of each of the light-emitting elements are calculated based on the verified data stream; wherein the second estimated temperature values calculated based on the second data stream are compared with the respective verified estimated temperature values calculated based on the verified data stream; wherein the second neural network is rewarded in case the number of verified estimated temperature values calculated based on the verified data stream that exceed the predetermined temperature threshold is lower than the number of the respective second estimated temperature values calculated based on the second data stream; and wherein the second neural network is punished in case the number of verified estimated temperature values calculated based on the verified data stream that exceed the predetermined temperature threshold is equal to or higher than the number of the respective second estimated temperature values calculated based on the second data stream. The second neural network hereby learns via reinforcement learning.

[0032] The second neural network is preferably optimized (i.e. continuously trained) via reinforcement learning, as described in the above-outlined preferred features. Reinforcement learning is an iterative process that allows the second neural network to learn a strategy for decision-making in a given environment. Preferably, the algorithm is selected from Deep Q-Network, Proximal Policy Optimization (PPO), Q-Learning, Asynchronous Advantage Actor-Critic (A3C), State-Action-Reward-State-Action (SARSA), REINFORCE, Twin Delayed DDPG (TD3), Dyna-Q, Monte Carlo Control, a Monte Carlo variant of a policy gradient algorithm, and Deep Deterministic Policy Gradient. Particularly preferably, the algorithm is a Monte Carlo variant of a policy gradient algorithm. This algorithm works particularly well in continuous and / or complex environments. Policy gradient algorithms are intended to describe a class of algorithms that optimize the policy (i.e., the strategy that the second neural networks follows) directly, rather than learning a value function. These algorithms focus on updating the parameters of the second neural network based on the rewards received and aim to increase the probability of actions that lead to higher cumulative rewards. Examples for policy gradient algorithms are REINFORCE (Monte Carlo Policy Gradient), Proximal Policy Optimization (PPO), Trust Region Policy Optimization (TRPO), Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3). Via reinforcement learning, the second neural network can be trained over its entire lifetime. That is, step c) is performed iteratively and without interruption, hereby continuously improving the performance of the second neural network.

[0033] Moreover, the invention relates to a light system, in particular to a headlamp, for radiating light while performing at least sub-steps c2) to c3) of the inventive method, said light system comprising the light source comprising the at least four light-emitting elements arranged in a n × m matrix, wherein n, m ≥ 2, a control unit for controlling the light source and its light-emitting elements, a computing device comprising * a memory wherein the first and the second neural network are stored, and * an interface for receiving the second data stream of non-verified light images according to sub-step c1) of the inventive method, wherein the computing device is configured to execute the calculations according to sub-step c2) and to provide the control unit with a verified data stream of verified light images to enable the control unit to execute sub-step c3) of the inventive method.

[0034] The control unit is preferably an electronic control unit (ECU). Since modern vehicles are generally equipped with an ECU, this improves the integration of the inventive light system, particularly a headlamp, into a vehicle.

[0035] Preferably, the light system, particularly the light source, preferably the pixelated light source, particularly preferably the multichip array, is free of temperature sensors. Since the light system comprises the first neural network trained via supervised learning according to step b) of the method, the first neural network is configured to calculate first estimated temperature values for each of the light-emitting elements. As such, no temperature sensors are required which brings not only immense advantages in terms of costs and construction efforts, but also allows for a space-saving design and arrangement of the light system.

[0036] The light system according to the present invention is preferably used in a vehicle, particularly a motorized vehicle (e.g. car, truck, bus, motorbike, plane).

[0037] The invention is not limited to the embodiments shown but is defined by the entire scope of protection of the claims. Individual aspects of the invention or of the embodiments can also be taken up and combined with one another. Any reference signs in the claims are exemplary and serve only the purpose of allowing easier review without restricting the claims.

[0038] In the following, in order to further demonstrate the present invention, illustrative and non-restrictive embodiments are discussed, as displayed in the figures, which show: Fig. 1aa schematic view of a vehicle equipped with a light system according to the invention, Fig. 1ban exemplary view of the components of the light system of Fig. 1a, Fig. 1can exemplary view of the components of a predetermined environment to train a first neural network, Fig. 2a schematic block diagram showing the training of the first neural network, Fig. 3another schematic block diagram showing the training of the first neural network, Fig. 4a schematic block diagram showing the use of a second neural network, Fig. 5illustrating how the second neural network is trained via reinforcement learning, Fig. 6a schematic block diagram to illustrate the operation of the second neural network, and Fig. 7another schematic block diagram to illustrate the operation of the second neural network.

[0039] In the following figures, identical reference signs refer to identical features unless expressly depicted otherwise. The reference signs are only for informational purpose and do not delimit the scope of protection.

[0040] Fig. 1a shows a schematic view of a vehicle 2 equipped with a light system LS according to the invention and configured to radiate light L. The light system LS can be realized as a headlamp or any other light emitting device.

[0041] Fig. 1b shows an exemplary view of components of the light system LS, in particular headlamp, for radiating light L. The light system LS is operated in a non-predetermined environment non-prEn and comprises a light source 1. The light source 1 can be a pixelated light source, e.g. a multichip array, preferably a light-emitting diode (LED) array. The light source 1 comprises at least four light-emitting elements 1aa-1nm arranged in a n × m matrix (with n, m ≥ 2). In practice, the number of light-emitting elements 1aa-1nm can be 30 to 80 000, for example about 60 000. A maximum expansion (e.g. a maximum diameter) of the light-emitting elements 1aa-1nm in a direction transversal to a direction of radiation of light images can range from 10 µm to 50 µm. Pixel pitches of the light-emitting elements 1aa-1nm can range of from 40 µm to 50 µm. These dimensions allow for a change in the light distribution caused by manipulation with the inventive method not to, at least not significantly, impact the perception of the light distribution by human beings. As further shown in Fig. 1b, the light system LS comprises a control unit CU for controlling the light source 1 and its light-emitting elements 1aa-1nm, and a computing device CD. The control unit CU is preferably an electronic control unit (ECU). The computing device CD comprises a memory M wherein a first neural network NN1 and a second neural network NN2 are stored, and an interface Int for receiving a second data stream DS2-Im of non-verified light images 2Im1-2Im5. The first neural network NN1 was trained according to the inventive method (see Fig. 2 and Fig. 3) in a predetermined environment prEn (see Fig. 1c). The computing device CD provides the control unit CU with a verified data stream VDS of verified light images VIm1-VIm5. As can be seen from Fig. 1b, the light system LS is free of temperature sensors, which brings not only immense advantages in terms of costs and construction efforts, but also allows for a space-saving design and arrangement of the light system LS.

[0042] Fig. 1c shows an exemplary view of the components of a predetermined environment prEn to train a first neural network NN1 for subsequent use in the light system LS of Fig. 1b. A light source 1 comprising light-emitting elements 1aa-1nm (arranged in a n × m matrix, wherein n, m ≥ 2) is operated in the predetermined environment prEn. The light source 1 is identical to or of the same type as the light source 1 used in the light system LS of Fig. 1b (at least regarding the number, arrangement and type of light-emitting elements 1aa-1nm). As such, the first neural network NN1 can be trained to calculate first estimated temperature values 1ETaa(t)-1ETnm(t) (see Fig. 2) and second estimated temperature values 2ETaa(t)-2ETnm(t) (see Fig. 4) of this specific type of light source 1. A first data stream DS1-Im of light images 1Im1-1Im5 is provided to a computing device CD via an interface Int. The light source 1 and its light-emitting elements 1aa-1nm are controlled by a control unit CU to radiate the light images 1Im1-1Im5. Each light image 1Im1-1Im5 is radiated by the light source 1 for a predetermined period of time, for example for 10 ms. While light images 1Im1-1Im5 are radiated by the light source 1, temperature values MTaa(t)-MTnm(t) of the light-emitting elements 1aa-1nm are measured over the predetermined period of time via an infrared measuring unit IR comprising temperature sensors TSaa-TSnm. Each temperature sensor TSaa-TSnm is configured to measure the temperature of one specific, corresponding light-emitting element 1aa-1nm. The measured temperature values MTaa(t)-MTnm(t) are provided to the computing device CD to train the first neural network NN1 stored on a memory M of the control unit CU, which method is described in detail in Fig. 2 and Fig. 3. The computing device CD (comprising the memory M and the interface Int) and the control unit CU can be identical to or of the same type as the respective computing device CD and control unit CU employed in the light system LS of Fig. 1b.

[0043] In Fig. 2, a schematic block diagram showing the training of the first neural network NN1 is displayed. The light source 1 is operated in the predetermined environment prEn (see Fig. 1c), and the light source 1 and its light-emitting elements 1aa-1nm (see Fig. 1c) are controlled by the control unit CU to radiate the light images 1Im1-1Im5. Each light image 1Im1-1Im5 is radiated for a predetermined period of time, for example 10 ms. Temperature values MTaa(t)-MTnm(t) of the light-emitting elements 1aa-1nm are measured with the infrared measuring unit IR comprising temperature sensors TSaa-TSnm over said predetermined period of time. Preferably, a sequence of temperature values MTaa(t)-MTnm(t) of the light-emitting elements 1aa-1nm is measured which improves the quality and speed of the training of the first neural network NN1. The first data stream DS1-Im is also fed to the first neural network NN1, and the first neural network NN1 calculates first estimated temperature values 1ETaa(t)-1ETnm(t) of the light-emitting elements 1aa-1nm based on said first data stream DS1-Im. Subsequently, the first estimated temperature values 1ETaa(t)-1ETnm(t) and the measured temperature values MTaa(t)-MTnm(t) are compared by a computing device CD, preferably by the first neural network NN1 comprised therein (see Fig. 1c). If a difference between at least one of the first estimated temperature values 1ETaa(t)-1ETnm(t) and at least one of the measured temperature values MTaa(t)-MTnm(t) meets, for the same light-emitting element 1aa-1nm, a predetermined success condition SC (see Fig. 3) in a reproducible manner, the training of the first neural network NN1 is completed. Otherwise, the method is repeated to reduce the difference until the predetermined success condition SC is reproducibly met. The predetermined success condition SC can e.g. be a temperature difference of 5% or below. The first neural network NN1 is hereby trained via supervised learning by minimizing an employed loss function.

[0044] Fig. 3 shows another schematic block diagram illustrating the training of the first neural network NN1 in a predetermined environment prEn. As can be seen, the first data stream DS1-Im, which comprises the light images 1Im1-1Im5 (see Fig. 2), is provided. The light source 1, which comprises the light-emitting elements 1aa-1nm (see Fig. 2), is operated, hereby radiating the light images 1Im1-1Im5. The temperature values MTaa(t)-MTnm(t) (see Fig. 2) of each light-emitting element 1aa-1nm are measured. Also, the first neural network NN1 calculates first estimated temperature values 1ETaa(t)-1ETnm(t) (see Fig. 2) based on the first data stream DS1-Im. The measured temperature values MTaa(t)-MTnm(t) and the first estimated temperature values 1ETaa(t)-1ETnm(t) are then compared. If a difference between at least one of the first estimated temperature values 1ETaa(t)-1ETnm(t) and at least one of the measured temperature values MTaa(t)-MTnm(t) is detected such that a predetermined success condition SC is not met, the aforementioned steps are repeated such to train the first neural network NN1 by minimizing the loss function. When the predetermined success condition SC is met for all of the light-emitting elements 1aa-1nm in a reproducible manner, the training of the first neural network NN1 is completed.

[0045] In Fig. 4, a schematic block diagram showing the use of the second neural network NN2 is displayed. A control unit CU feeds a second data stream DS2-Im of non-verified light images 2Im1-2Im5 to the first neural network NN1 trained according to the inventive method (see Fig. 2 and 3). The first neural network NN1 calculates second estimated temperature values 2ETaa(t)-2Etnm(t) for each of the light-emitting elements 1aa-1nm (see Fig. 1b) of the light source 1 based on the second data stream DS2-Im. The second estimated temperature values 2Etaa(t)-2Etnm(t) are then compared with a predetermined temperature threshold TT by a computing device CD, preferably by the first neural network NN1 or by the second neural network NN2, comprised therein (see Fig. 1b). If at least one of the second estimated temperature values 2Etaa(t)-2Etnm(t) exceeds the predetermined temperature threshold TT, sub-step c21) of the inventive method is performed. Namely, the second data stream DS2-Im of non-verified light images 2Im1-2Im5 and the corresponding second estimated temperature values 2Etaa(t)-2Etnm(t) are fed to the second neural network NN2, and the second neural network NN2 manipulates at least one of the non-verified light images 2Im1-2Im5 of the second data stream DS2-Im with the goal of avoiding exceeding the predetermined temperature threshold TT in any of the light-emitting elements 1aa-1nm while minimizing the extent of manipulation of the at least one of the non-verified light images 2Im1-2Im5, hereby obtaining a verified data stream VDS of verified light images Vim1-Vim5. If, instead, none of the second estimated temperature values 2Etaa(t)-2Etnm(t) exceeds the predetermined temperature threshold TT, the second data stream DS2-Im of non-verified light images 2Im1-2Im5 is either not manipulated and set as the verified data stream VDS of verified light images Vim1-Vim5 or alternatively, a manipulation according to sub-step c21) can nevertheless take place. For example, a reduced manipulation can be carried out as a precaution if at least one of the second estimated temperature values 2ETaa(t)-2ETnm(t) is already in close proximity to the predetermined temperature threshold TT, e.g. about 10% below the predetermined temperature threshold TT. The verified data stream VDS is then fed to the light source 1 controlled by the control unit CU to radiate the verified light images VIm1-VIm5. The predetermined temperature threshold TT is preferably below a value of 80 to 99% of a maximum junction temperature, with the maximum junction temperature being e.g. in the range of from 130 °C to 180 °C. The predetermined temperature threshold TT is preferably determined with a lookup table which allows to quickly react to a sudden temperature rise.

[0046] Fig. 5 shows a schematic block diagram illustrating how the second neural network NN2 of the present invention can be trained via reinforcement learning. As described in connection with Fig. 4, second estimated temperature values 2ETaa(t)-2ETnm(t) of each of light-emitting elements 1aa-1nm are calculated based on the second data stream DS2-Im by the first neural network NN1. Furthermore, verified estimated temperature values VETaa(t)-VETnm(t) of each of the light-emitting elements 1aa-1nm are calculated based on the verified data stream VDS obtained as described in connection with Fig. 4. The respective calculations are preferably performed by the second neural network NN2. Alternatively, another component of the computing device CD (see Fig. 1b) can perform the calculations. If the number of verified estimated temperature values VETaa(t)-VETnm(t), calculated based on the verified data stream VDS, that exceed the predetermined temperature threshold TT (see Fig. 4) is lower than the number of the respective second estimated temperature values 2ETaa(t)-2ETnm(t), calculated based on the second data stream DS2-Im, the second neural network NN2 is rewarded. If, instead, the number of verified estimated temperature values VETaa(t)-VETnm(t) calculated based on the verified data stream VDS that exceed the predetermined temperature threshold TT is equal to or higher than the number of the respective second estimated temperature values 2ETaa(t)-2ETnm(t) calculated based on the second data stream DS2-Im, the second neural network NN2 is punished. As such, the second neural network NN2 is trained via an iterative process, which training can continue over its whole lifetime such to continuously improve its performance.

[0047] Fig. 6 shows a schematic block diagram to further illustrate the operation of the second neural network NN2. The second data stream DS2-Im of non-verified light images 2Im1-2Im5 is provided and fed to the first neural network NN1 to predict the temperature of the light-emitting elements 1aa-1nm of the light source 1, hereby predicting second estimated temperature values 2ETaa(t)-2ETnm(t) of each of the light-emitting elements 1aa-1nm by calculation. The light images radiated by the light source 1 are then optimized by manipulating at least one of the non-verified light-images 2Im1-2Im5 to obtain a verified data stream VDS of verified light images VIm1-VIm5, as described above in connection with Fig. 4. If none of the second estimated temperature values 2ETaa(t)-2ETnm(t) exceeds the predetermined temperature threshold TT (see Fig. 4), a reduced manipulation can be performed. The optimized, i.e. verified light images VIm1-VIm5 are then sent to the control unit CU and radiated by the light source 1. Furthermore, the optimized, i.e. verified light images VIm1-VIm5 are evaluated, and feedback is provided to the second neural network NN2 to train it via reinforcement learning, as described above in connection with Fig. 5.

[0048] Fig. 7 shows another schematic block diagram to further illustrate the operation of the second neural network NN2. The block diagram of Fig. 7 is similar to that shown in Fig. 6, with the difference that a manipulation does not necessarily occur but rather that it is assessed whether or not an optimization of light images is indeed required. If at least one of the second estimated temperature values 2ETaa(t)-2ETnm(t) exceeds the predetermined temperature threshold TT (see Fig. 4), a manipulation is performed, and feedback is provided to the second neural network NN2, as described in Fig. 4 and 6. If, in turn, none of the second estimated temperature values 2ETaa(t)-2ETnm(t) exceeds the predetermined temperature threshold TT, the second data stream DS2-Im of non-verified light images 2Im1-2Im5 is not manipulated and set as the verified data stream VDS of verified light images VIm1-VIm5. Of course, feedback can also be provided to the second neural network NN2 if no manipulation is performed (not shown in Fig. 7). The verified light images VIm1-VIm5 are subsequently sent to the control unit CU. The light source 1 and its light-emitting elements 1aa-1nm, which are controlled by the control unit CU, then radiate the verified light images VIm1-VIm5.Example

[0049] Tables 1 shows a schematic first and second light distribution of a light source. The light source comprises light-emitting elements arranged in an n x m matrix (n = 5, m = 6). The first light distribution is represented by a non-verified light image, wherein two light-emitting elements, no. (2,5) and (4,2), are switched on at 100% intensity (both marked with an "X" in Table 1). The first neural network calculates first estimated temperature values for all light-emitting elements, which first estimated temperature values are then compared with a predetermined temperature threshold. As it is found that the first estimated temperature values of the light-emitting elements no. (2,5) and (4,2) exceed the predetermined temperature threshold, the non-verified light image is manipulated by the second neural network with the goal of avoiding exceeding the predetermined temperature threshold in the light-emitting elements whilst minimizing the extent of manipulation of the non-verified light image. This is accomplished by replacing, in the radiated light image, light-emitting element no. (2,5) by light-emitting elements no. (3,2), (4,1), (4,3) and (5,2), and by replacing light-emitting element no. (4,2) by light-emitting elements no. (1,5), (2,4), (2,6) and (3,5), and by operating all light-emitting elements used as replacement in the radiated light image at an intensity lower than 100%. For example, each light-emitting element used as replacement can be operated at 25% intensity.

Claims

1. Method for optimizing the thermal load of a light source (1), wherein the light source (1) comprises at least four light-emitting elements (1aa-1nm) arranged in an n x m matrix, wherein n, m ≥ 2, the method comprising the following steps: a) providing a first data stream (DS1-Im) of light images (1Im1-1Im5), b) operating the light source (1) in a predetermined environment (prEn) and training a first neural network (NN1), said step b) comprising the following sub-steps: b1) controlling, preferably by a control unit (CU), the light source (1) and its light-emitting elements (1aa-1nm) to radiate the light images (1Im1-1Im5), wherein each light image (1Im1-1Im5) is radiated by the light source (1) for a predetermined period of time, b2) while performing sub-step b1), measuring temperature values (MTaa(t)-MTnm(t)) of at least one of the light-emitting elements (1aa-1nm) over the predetermined period of time of sub-step b1), b3) before, while and / or after the latest sub-steps b1) and b2), feeding the first data stream (DS1-Im) to the first neural network (NN1) to calculate first estimated temperature values (1ETaa(t)-1ETnm(t)) of the at least one light-emitting element (1aa-1nm) based on the first data stream (DS1-Im), b4) comparing the first estimated temperature values (1ETaa(t)-1ETnm(t)) of sub-step b3) with the measured temperature values (MTaa(t)-MTnm(t)) of sub-step b2), and in case that a difference between at least one of the first estimated temperature values (1ETaa(t)-1ETnm(t)) and at least one of the measured temperature values (MTaa(t)-MTnm(t)) meets, for the same light-emitting element (1aa-1nm), a predetermined success condition (SC), proceeding with step c), otherwise repeating sub-steps b1) to b4) to reduce the difference until the success condition (SC) is met, and hereby training the first neural network (NN1) via supervised learning, c) operating the light source (1) in a non-predetermined environment (non-prEn) while using a second neural network (NN2), said step c) comprising the following sub-steps: c1) providing a second data stream (DS2-Im) of non-verified light images (2Im1-2Im5), c2) feeding the second data stream (DS2-Im) of non-verified light images (2Im1-2Im5) to the first neural network (NN1) to calculate second estimated temperature values (2ETaa(t)-2ETnm(t)) of each of the light-emitting elements (1aa-1nm) based on the second data stream (DS2-Im) of non-verified light images (2Im1-2Im5), and comparing the second estimated temperature values (2ETaa(t)-2ETnm(t)) with a predetermined temperature threshold (TT) to obtain a verified data stream (VDS) of verified light images (VIm1-VIm5), c21) wherein at least in case that at least one of the second estimated temperature values (2ETaa(t)-2ETnm(t)) exceeds the predetermined temperature threshold (TT), the second data stream (DS2-Im) of non-verified light images (2Im1-2Im5) and the corresponding second estimated temperature values (2ETaa(t)-2ETnm(t)) are fed to the second neural network (NN2), and the second neural network (NN2) manipulates at least one of the non-verified light images (2Im1-2Im5) of the second data stream (DS2-Im) with the goal of avoiding exceeding the predetermined temperature threshold (TT) in any of the light-emitting elements (1aa-1nm) while minimizing the extent of manipulation of the at least one of the non-verified light images (2Im1-2Im5), hereby obtaining the verified data stream (VDS) of verified light images (VIm1-VIm5), and c22) wherein in case that none of the second estimated temperature values (2ETaa(t)-2ETnm(t)) exceeds the predetermined temperature threshold (TT), either still a manipulation according to sub-step c21) takes place, or the second data stream (DS2-Im) of non-verified light images (2Im1-2Im5) is not manipulated and set as the verified data stream (VDS) of verified light images (VIm1-VIm5), c3) controlling, preferably by the control unit (CU), the light source (1) and its light-emitting elements (1aa-1nm) to radiate the verified light images (VIm1-VIm5) of the verified data stream (VDS).

2. Method according to claim 1, wherein the manipulation in step c21) is executed in a manner that a luminous flux radiated by the light source (1) remains unchanged.

3. Method according to any of the previous claims, wherein the light source (1) is a pixelated light source, preferably a multichip array, particularly preferably a light-emitting diode (LED) array.

4. Method according to any of the previous claims, wherein the light source (1) comprises 24 to 100 000 light-emitting elements (1aa-1nm), preferably 30 to 80 000, more preferably 60 to 70 000, particularly preferably 100 to 65 000.

5. Method according to any of the previous claims, wherein in sub-step b2), temperature values (MTaa(t)-MTnm(t)) of every light-emitting element (1aa-1nm) are measured over the predetermined period of time of sub-step b1).

6. Method according to any of the previous claims, wherein in sub-step c1), a continuous stream of non-verified light images (2Im1-2Im5) is provided, and wherein in sub-step c3), a continuous stream of verified light-images (VIm1-VIm5) is radiated, wherein each non-verified light image (2Im1-2Im5) is processed individually in sub-step c2), and providing sub-step c3) with a consecutive continuous stream of verified light-images (VIm1-VIm5), wherein in sub-step c3), each verified light-image (VIm1-VIm5) is radiated until it is replaced by a consecutive verified light-image (VIm1-VIm5).

7. Method according to any of the previous claims, wherein in sub-step c1), a continuous stream of non-verified light images (2Im1-2Im5) is provided, and wherein in sub-step c3), a continuous stream of verified light-images (VIm1-VIm5) is radiated, wherein in sub-step c2), the calculation of second estimated temperature values (2ETaa(t)-2ETnm(t)) and manipulation according to sub-step c21), or potentially sub-step c22), is performed for sets of non-verified light images (2Im1-2Im5), wherein each set comprises two or more non-verified light images (2Im1-2Im5) and the extent of manipulation is distributed, in particular evenly distributed, between the non-verified light images (2Im1-2Im5) of each set, wherein in sub-step c3), each verified light-image (VIm1-VIm5) is radiated until it is replaced by a consecutive verified light-image (VIm1-VIm5).

8. Method according to claim 7, wherein the size of the set of non-verified light images (2Im1-2Im5) and the speed of the execution of sub-step c2) are designed in a manner that a time delay between providing a non-verified light image (2Im1-2Im5) in sub-step c1) and radiating a corresponding verified light image (VIm1-VIm5) in sub-step c3) is 20 ms or below.

9. Method according to any of the previous claims, wherein the predetermined temperature threshold (TT) is below a value of 80% to 99% of a maximum junction temperature, wherein the maximum junction temperature preferably ranges from 150 °C to 160 °C.

10. Method according to any of the previous claims, wherein in sub-step c21), the second data stream (DS2-Im) of light images (2Im1-2Im5) and the corresponding second estimated temperature values (2ETaa(t)-2ETnm(t)) are fed to the second neural network (NN2) regardless of whether or not at least one of the second estimated temperature values (2ETaa(t)-2ETnm(t)) exceeds the predetermined temperature threshold (TT), and in case that none of the second estimated temperature values (2ETaa(t)-2ETnm(t)) exceeds the predetermined temperature threshold (TT), it is proceeded with sub-step c22) without or with reduced manipulation.

11. Method according to any of the previous claims, wherein sub-steps c1) to c3) are iteratively performed while receiving a continuous stream of non-verified light images (2Im1-2Im5) to obtain and radiate verified light images (VIm1-VIm5), wherein in sub-step c2), also previous verified light images are provided to the first neural network (NN1), and wherein the impact of the manipulation performed by the second neural network (NN2) is fed back to the second neural network (NN2) to enable its optimization.

12. Method according to any of the preceding claims, wherein in sub-step c2), also previous verified light images are provided to the first neural network (NN1) and wherein the impact of the manipulation performed by the second neural network (NN2) is fed back to the second neural network (NN2) to enable its optimization, and wherein in case that the goal of avoiding exceeding the predetermined temperature threshold (TT) is not met in an iteration of sub-step c21), another iteration of sub-steps c2) and c21) is initiated until the goal of avoiding exceeding the predetermined temperature threshold (TT) is met.

13. Method according to any of the previous claims, wherein after sub-steps c2) or c3), the verified data stream (VDS) is fed to the first neural network (NN1) and sub-steps c1) and c2) are repeated, wherein in sub-step c2), second estimated temperature values (2ETaa(t)-2ETnm(t)) of each of the light-emitting elements (1aa-1nm) are calculated based on the second data stream (DS2-Im) of latest sub-step c2), and verified estimated temperature values (VETaa(t)-VETnm(t)) of each of the light-emitting elements (1aa-1nm) are calculated based on the verified data stream (VDS), wherein the second estimated temperature values (2ETaa(t)-2ETnm(t)) calculated based on the second data stream (DS2-Im) are compared with the respective verified estimated temperature values (VETaa(t)-VETnm(t)) calculated based on the verified data stream (VDS), wherein the second neural network (NN2) is rewarded in case the number of verified estimated temperature values (VETaa(t)-VETnm(t)) calculated based on the verified data stream (VDS) that exceed the predetermined temperature threshold (TT) is lower than the number of the respective second estimated temperature values (2ETaa(t)-2ETnm(t)) calculated based on the second data stream (DS2-Im), and wherein the second neural network (NN2) is punished in case the number of verified estimated temperature values (VETaa(t)-VETnm(t)) calculated based on the verified data stream (VDS) that exceed the predetermined temperature threshold (TT) is equal to or higher than the number of the respective second estimated temperature values (2ETaa(t)-2ETnm(t)) calculated based on the second data stream (DS2-Im).

14. Light system (LS), in particular headlamp, for radiating light (L) while performing at least sub-steps c2 to c3) of the method according to any of the preceding claims, said light system (LS) comprising - the light source (1) comprising the at least four light-emitting elements (1aa-1nm) arranged in a n × m matrix, wherein n, m ≥ 2, - a control unit (CU) for controlling the light source (1) and its light-emitting elements (1aa-1nm), - a computing device (CD) comprising * a memory (M) wherein the first (NN1) and the second neural network (NN2) are stored, and * an interface (Int) for receiving the second data stream (DS2-Im) of non-verified light images (2Im1-2Im5) according to sub-step c1) of the method according to any of the preceding claims, wherein the computing device (CD) is configured to execute the calculations according to sub-step c2) and to provide the control unit (CU) with a verified data stream (VDS) of verified light images (VIm1-VIm5) to enable the control unit (CU) to execute sub-step c3) of the method according to any of the preceding claims.

15. Light system (LS) according to claim 14, wherein the light system (LS) is free of temperature sensors.

Images