Method and apparatus for providing a trained machine learning model to identify a hands-on / hands-off state at a steering wheel of a vehicle

Abstract

The present invention relates to a method for providing a trained machine learning model to identify the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50), comprising: training at least one teacher model (20) with a training dataset (10); training a smaller student model (21) in a teacher-student training manner with the at least one teacher model (20), wherein a loss function is used for this purpose, the loss function consisting of a weighted sum of a student loss and a distillation loss, wherein a weight factor (λ) for each training data or group of training data in the training dataset (10) used in the teacher-student training is determined taking into account at least one preset criterion (11); and providing the trained student model (21). Furthermore, the present invention relates to an apparatus (1) for providing a trained machine learning model to identify the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50). Furthermore, the present invention also relates to a method and apparatus for identifying the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50).

Description

Technical Field

[0001] This invention relates to a method and apparatus for providing a trained machine learning model to identify the Hands-On / Hands-Off state of a vehicle's steering wheel. Furthermore, this invention relates to a method and apparatus for identifying the Hands-On / Hands-Off state of a vehicle's steering wheel. Background Technology

[0002] In vehicles, sensors, such as capacitive steering wheels, are used to monitor driver activity. These steering wheels use capacitive sensors to identify whether the driver touches (“Hands-On”) or doesn't touch (“Hands-Off”) the steering wheel. The results are transmitted to user functions, such as longitudinal and / or lateral guidance assist systems. Driver activity and attention can be inferred from hand touches on the steering wheel. For example, it can be programmed to prompt the driver to return their hands to the steering wheel if it is detected that the hands are not on the steering wheel for a preset time during lateral guidance.

[0003] To save on the additional cost of capacitive sensors in the steering wheel, it is known to monitor driver activity by means of machine learning models, particularly artificial neural networks, based on the torque (hand torque) detected at the steering wheel. This method is known, for example, from DE 102019211 016 A1.

[0004] A significant challenge in identifying steering torque-based torque lies in distinguishing driver-induced torque from the measured (noisy) torque (steering torque). Many factors can contribute to noisy torque, particularly the location of the sensor (which is typically part of the steering drivetrain or steering assist system, thereby generating torsional vibrations via the elasticity of the steering column and its connection to the steering wheel, the dynamics of which make accurate measurement of driver-induced torque difficult); the intensity of friction in the steering system; counter-excitation from road roughness; the weight of the steering wheel / steering system; and vibrations of the steering wheel caused by assist functions (such as tactile feedback when leaving the lane).

[0005] Furthermore, the characteristics of the measured torque (which is used for Hands-On / Hands-Off identification) may change due to external influences, such as temperature, vehicle loading status, presence of trailers, tire type and / or tire condition, steering system changes over its service life, lane gradient / inclination / tilt position, etc.

[0006] Due to the sheer number of influencing factors, it is necessary to equip machine learning models, especially neural networks, with sufficient degrees of freedom, i.e., weights, so that the model can map all influences. However, this results in large models, and therefore greater demands on storage and computational resources. In particular, the ratio between the number of weights and the desired mapping quality of the objective function is not linear, but rather grows approximately exponentially.

[0007] The use of teacher-student training is known, for example, by DE 11 2020 001 663 T5 or DE 102021 200643 B3. Furthermore, the use of teacher-student training is known by G. Hinton et al., “Distilling the Knowledge in a Neural Network”, arXiv: 1503.02531 v1 [stat.ML], March 9, 2015. Summary of the Invention

[0008] The objective of this invention is to improve a method and apparatus for providing a trained machine learning model to identify the Hands-On / Hands-Off state of a vehicle's steering wheel. Furthermore, the objective of this invention is to improve a method and apparatus for identifying the Hands-On / Hands-Off state of a vehicle's steering wheel.

[0009] This task, according to the invention, is solved by a method having the features of claim 1 and an apparatus having the features of claim 11. The latter task is solved by a method having the features of claim 10 and an apparatus having the features of claim 12. Advantageous embodiments of the invention are derived from the dependent claims.

[0010] In particular, in the first aspect, a method is provided for providing a trained machine learning model to identify the Hands-On / Hands-Off state at the steering wheel of a vehicle, comprising: training at least one teacher model with a training dataset; training a smaller student model in a teacher-student training manner with the at least one teacher model, wherein a loss function is used for this purpose, the loss function consisting of a weighted sum of a student loss and a distillation loss, wherein the weight factor for each training data or group of training data in the training dataset used for teacher-student training is determined taking into account at least one preset criterion; and providing the trained student model.

[0011] Furthermore, particularly in the second aspect, an apparatus is created for providing a trained machine learning model to identify the Hands-On / Hands-Off state at the steering wheel of a vehicle, comprising a data processing device, wherein the data processing device is configured to train at least one teacher model with a training dataset, and to train a smaller student model in a teacher-student training manner with the at least one teacher model, wherein a loss function is used for this purpose, which consists of a weighted sum of student loss and distillation loss, and the weight factors for each training data or group of training data in the training dataset used during teacher-student training are determined taking into account at least one preset criterion.

[0012] Furthermore, particularly in a third aspect, a method is provided for identifying the Hands-On / Hands-Off state at the steering wheel of a vehicle, comprising: detecting at least one steering parameter at the steering wheel of the vehicle; using a student model trained according to an embodiment of the first or second aspect to identify the Hands-On / Hands-Off state, wherein the detected at least one steering parameter is fed into the input of the trained student model for this purpose; and outputting the Hands-On / Hands-Off state estimated by the trained student model.

[0013] Finally, in the fourth aspect, in particular, an apparatus for identifying the Hands-On / Hands-Off state at the steering wheel of a vehicle is created, comprising: at least one sensor for detecting at least one steering parameter at the steering wheel; and a controller, wherein the controller is configured to acquire the detected at least one steering parameter, provide and apply a student model trained according to an embodiment of the first or second aspect to identify the Hands-On / Hands-Off state, and for this purpose, input the detected at least one steering parameter to the input of the trained student model, and output the Hands-On / Hands-Off state estimated by the trained student model.

[0014] The methods and apparatus according to the first and second aspects allow for the provision of a machine learning model with improved storage and computational resource requirements. To this end, a student model is configured, which is structurally simpler or less complex than at least one teacher model. For example, the student model may have fewer neurons or fewer internal layers than the teacher model, and therefore fewer parameters (especially weights). To prevent a reduction in functional quality, the student model is configured during training to use a loss function consisting of a weighted sum of student loss and distillation loss, wherein weight factors for each training data point or group of training data in the training dataset used during teacher-student training are determined considering at least one pre-defined criterion. The pre-defined at least one criterion here specifically defines those factors considered in determining the weight factors or those factors from which the weight factors are determined. By enabling the determination of weight factors individually for each training data point or group of training data, the strength at which the student model learns from at least one teacher model within the framework of teacher-student training can be influenced for each training data point in the training dataset. Therefore, it is particularly possible to make emphasis settings during teacher-student training, through which, for example, higher quality can be achieved when recognizing the Hands-Off state than when recognizing the Hands-On state. The student model trained in this way is provided, and in particular, output. The trained student model is used, especially in vehicles, to recognize the Hands-On / Hands-Off states. Based on this, for example, an assistance system can be run.

[0015] One advantage of this method and apparatus, in particular, is that it can provide a student model with improved functional quality. Thus, it can provide a machine learning model with reduced storage and computational resource requirements while maintaining improved functional quality.

[0016] The model (at least one teacher model and one student model) is specifically designed to be trained or be trained to estimate the Hands-On / Hands-Off states starting from at least one steering parameter.

[0017] At least one teacher model and student model are designed, in particular, as artificial neural networks, especially as deep neural networks, which include multiple inner layers. The models are, in particular, artificial recurrent neural networks that process input data X at each time point t. t And output at least one Hands-Off probability y in [0,1]. t y t = p(X t |X 0:t-1Preferably, the neural network outputs probabilities for the Hands-On and Hands-Off states at each time point t. Here, the neural network specifically has a so-called memory h, in which information from previous time steps is stored and can be used for the output at the current time step. This output is further processed, for example, by filtering, before being processed by a consumer-side function (e.g., a lateral guidance assist system). Specifically, it can be configured to provide a binary Hands-Off signal (with two states: "Hands-Off identified" and "Hands-Off not identified") based on a comparison of the Hands-Off probability with a preset threshold. Accordingly, this can be done for the Hands-On state, or a common signal with two signal states, "Hands-On" or "Hands-Off", can be provided.

[0018] During training (i.e., during the corresponding training phase), the model is trained using training data, which includes pairs in which data on steering parameters, particularly torque data, are paired with a corresponding Hands-On / Hands-Off state as basic facts. Data on at least one steering parameter, particularly torque data, is included, especially time series of at least one steering parameter detected at the steering wheel, particularly time series of torque detected at the steering wheel. The training data is obtained, in particular, through test driving and / or in a simulator. In principle, the provision of training data can be, in particular, according to the method described in DE102019211 016 A1. Training is conducted in a manner known per se, particularly in a supervised learning manner. The detection and / or collection of training data for the training dataset can be part of both the first and second aspects. The Hands-On / Hands-Off state is detected, in particular, with the aid of an additional sensor system (e.g., with the aid of a capacitive steering wheel).

[0019] The error for each weight is computed and optimized during training within the framework of backpropagation via a loss function. Binary cross-entropy is particularly useful in identifying Hands-On / Hands-Off states. L = -y*log(Y) + (1-y)*log(1-Y), Here, Y is the ground truth, and y is the output of the model.

[0020] According to the present invention, for student training within a teacher-student training framework, an adapted loss function is used: L=λ*student-loss+(1-λ)*distillation-loss, Wherein, student-loss is the binary cross-entropy between the student model's output and the ground truth, and distillation-loss is the binary cross-entropy between the student model's output and the teacher model's output. The weighting factor λ is determined according to the present invention for each training data point or group of training data in the training dataset used during teacher-student training, taking into account at least one preset criterion. Therefore, the weighting factor λ is not a constant, but is determined considering at least one preset criterion.

[0021] Within the framework of teacher-student learning, the student model is trained, in particular, with the aid of at least one teacher model. Here, the same training data is fed as input to both the at least one teacher model and the student model. In addition to basic facts, the output of the at least one teacher model is used to adapt the parameters (especially the weights) of the student model, and the student model is trained in this manner. The consideration of these two parameters is made via the adaptation loss function listed above.

[0022] Methods for providing trained machine learning models to identify the Hands-On / Hands-Off states of a vehicle's steering wheel are implemented, in particular, as computer-based methods.

[0023] Steering parameters are, in particular, parameters that represent and / or describe the current state of the steering wheel. Steering parameters are especially torque, which is detected primarily by means of a torque sensor at the steering wheel. In principle, the steering parameter can also be another parameter that is detected directly or indirectly at the steering wheel. For example, it can be configured to detect the current at the motor at the steering wheel and use it as a steering parameter. The identification of the Hands-On / Hands-Off state can be based solely on the steering parameters detected at the steering wheel, especially the detected torque. However, it is also possible, in particular, to use other (steering) parameters detected at the steering wheel (e.g., steering wheel angle and / or steering wheel angular velocity, etc.), and for the model to identify the Hands-On / Hands-Off state while also considering these other parameters(s). Furthermore, quantities not detected at the steering wheel can also be considered, such as vehicle speed, lateral acceleration, yaw rate, wheel pulsation, damper information, and / or other driving dynamics parameters, etc. However, capacitive sensors at the steering wheel are not typically used.

[0024] The Hands-Off state is particularly characterized by a state in which no steering wheel is touched by the driver. Specifically, none of the driver's fingers are in contact with the steering wheel. Recognition of the Hands-Off state may, in particular, include providing a Hands-Off state signal. This signal may include, for example, a Hands-Off probability, or coded signals for the states "Hands-Off recognized" and "Hands-Off not recognized". The Hands-On state is particularly characterized by a state in which the steering wheel is touched by the driver. Specifically, a Hands-On / Hands-Off state is provided, for example, as a Hands-On / Hands-Off state signal having, in particular, two signal states (e.g., "Hands-On recognized" or "Hands-Off recognized").

[0025] Parts of the device, particularly data processing equipment, can be constructed individually or in combination as a combination of hardware and software, for example as program code implemented on a microcontroller or microprocessor. Alternatively, these parts can be configured individually or in combination as application-specific integrated circuits (ASICs) and / or field-programmable gate arrays (FPGAs).

[0026] In one embodiment, the trained student model is loaded into the memory of at least one controller of the vehicle within a provided framework after training. Thus, the trained student model can be subsequently applied in the vehicle to provide Hands-On / Hands-Off recognition. Loading into memory specifically includes loading a structural description and parameters (weights, hyperparameters, etc.) of the student model into memory.

[0027] In one implementation, the values ​​of the weighting factors are determined based on a standard derived from the corresponding training data. This allows the content of the training data to be considered. In particular, the weighting factors can be determined, for example, by considering the semantic content mapped in the training data.

[0028] In one implementation, the at least one criterion relates to the values ​​of fundamental facts of the corresponding training data, wherein the fundamental facts include at least the values ​​"Hands-On" or "Hands-Off". Thus, weighting factors can be determined based on the fundamental facts of the corresponding training data. For example, the weighting factors can be configured to have different values ​​depending on whether the fundamental facts contain the state "Hands-On" or "Hands-Off".

[0029] In one implementation, the weighting factors are chosen such that the student model learns more strongly from at least one teacher model for the state "Hands-Off". This allows the focus to be placed on the knowledge already distilled in at least one teacher model for the state "Hands-Off". Specifically, when the basic fact is "Hands-Off", a value of λ < 0.5 is chosen for the weighting factors according to the equation given above.

[0030] In one implementation, at least one teacher model is configured to be more specialized and / or robust for Hands-Off state recognition within the training framework of at least one teacher model. This allows the focus to be placed on achieving high functional quality in Hands-Off state recognition during teacher model training. Since Hands-Off state recognition is more critical than Hands-On state recognition in application scenarios, this further improves robustness in assistive systems.

[0031] In one implementation, the weighting factors are determined such that the student model learns from the training data more effectively for the state "Hands-On" than from at least one teacher model. This mitigates the specialization effect of the teacher model for the state "Hands-Off". For this purpose, for this state of basic fact, according to the equations given above, a value λ > 0.5 is chosen for the weighting factors.

[0032] In one implementation, the at least one criterion relates to a performance parameter of at least one teacher model. This allows the functional quality of the teacher model to be considered directly when selecting weighting factors. In particular, it allows for flexibility in responding to the specialization of the teacher model; that is, for example, greater weight can always be given to the teacher model where it is specialized and where it provides better results in terms of functional quality. The performance parameter could, for example, be the binary cross-entropy between the teacher model's output and the underlying facts.

[0033] In one implementation, at least one teacher model comprises multiple teacher models, each trained for a specific context. This allows the student model to be trained context-dependently using the teacher models. For example, specialized teacher models can be configured for one or more of the following contexts: vehicles with trailers, vehicles with snow chains, rough roads (cobblestones, potholes, speed bumps, etc.), maximum assisted intervention (e.g., steering wheel vibration as haptic feedback), extreme external temperatures (extreme heat or cold), etc. Specifically, within the framework of teacher model training, context-specific training datasets are generated and used separately. For this purpose, the original training dataset is enriched and / or sparsified with respect to the training data it contains, such that the corresponding desired context is more strongly mapped by the training data contained in the training dataset. The specialized teacher models are initialized, in particular, using the parameters (weights, etc.) of a fully trained (non-specialized) (standard) teacher model. Thus, foundational knowledge is already available at the start of training the corresponding specialized teacher model, which has not yet been specialized for a specific context. Therefore, the training of specialized teacher models can be accelerated. Within the teacher-student training framework, the teacher model is either a standard model, one of specialized teacher models, or a mixture of these models. Then, based on the corresponding data points (or domains) in the training dataset used, a specialized teacher model is selected (e.g., for data points corresponding to the context of a attached trailer, a teacher model specialized for the trailer context is selected, and so on). If there is overlap in these contexts (or domains), all teacher models utilizing context specialization are selected. Multiple specialized teacher models can then be considered via weights in the loss function.

[0034] Other features used in designing this device are derived from the description of the design scheme of this method. The advantages of this device are the same as those described in the design scheme of this method. Attached Figure Description

[0035] The present invention will now be described in more detail with reference to the accompanying drawings and preferred embodiments. Herein: Figure 1 A schematic diagram is shown illustrating an embodiment of a device for providing a trained machine learning model to identify the Hands-On / Hands-Off state at the steering wheel of a vehicle. Figure 2 A schematic flowchart is shown to illustrate an implementation of a method for providing a trained machine learning model to identify the Hands-On / Hands-Off state at the steering wheel of a vehicle. Detailed Implementation

[0036] Figure 1A schematic diagram is shown illustrating an embodiment of a device 1 for providing a trained machine learning model to identify the Hands-On / Hands-Off state at the steering wheel of a vehicle 50. Device 1 is configured to implement the methods described in this disclosure.

[0037] The apparatus 1 includes a data processing device 2. The data processing device 2 includes at least one computing device 2-1 and at least one memory 2-2. The computing device 2-1 can perform computational operations on data stored in the memory 2-2 to execute the method.

[0038] Data processing device 2 is configured to train at least one teacher model 20 using training dataset 10, and to train a smaller student model 21 using at least one teacher model 20 in a teacher-student training manner. For this purpose, a loss function is used, which consists of a weighted sum of student loss and distillation loss. The weight factor λ for each training data or group of training data in the training dataset 10 used during teacher-student training is determined here, taking into account at least one preset criterion 11.

[0039] Figure 2 A schematic flowchart illustrating one implementation of a method for providing a trained machine learning model to identify the Hands-On / Hands-Off state at the steering wheel of a vehicle is shown. The method is implemented, for example, on device 1, as it is in… Figure 1 As shown in the image.

[0040] In measure 100, at least one teacher model is trained using a training dataset. This is done in a manner known per se, such as through supervised learning.

[0041] In step 101, a smaller student model is trained using at least one teacher model in a teacher-student training manner. Step 101 includes steps 101a to 101d for this purpose. In step 101a, training data or training data sets are selected from the training dataset for one pass. In step 101b, weight factors of the loss function used during training are determined for the training data or training data sets, taking into account at least one preset criterion. In step 101c, the student model is trained in a teacher-student training manner, where a loss function with the determined weight factors is used. Subsequently, in step 101d, it is checked whether the preset function (target) quality of the student model has been achieved. If not, the process continues with the next training data or the next set of training data, as described in step 101a. If so, the trained student network is provided, and in particular, output, in step 102.

[0042] It can be set up so that the trained student model 21 ( Figure 1After training, it is loaded into the memory 53 of at least one controller 52 of the vehicle 50 within the provided framework. Figure 2 (Measure 103 in the text). Controller 52 is part of device 51 for identifying the Hands-On / Hands-Off state 40 at the steering wheel of vehicle 50. Device 51 further includes at least one sensor 54 for detecting at least one steering parameter 55 at the steering wheel. The at least one steering parameter 55 includes, in particular, hand torque. Controller 52 is configured to obtain the detected at least one steering parameter 55, provide and apply a trained student model 21 to identify the Hands-On / Hands-Off state 40, and for this purpose, input the detected at least one steering parameter 55 to the input of the trained student model 21, and output the Hands-On / Hands-Off state 40 estimated by the trained student model 21. The Hands-On / Hands-Off state 40 can then be fed to another controller 56 of vehicle 50 for further processing. In particular, it can be configured to operate the vehicle 50's assistance systems starting from the Hands-On / Hands-Off state 40. This could be, for example, a lateral guidance assist device.

[0043] It can be set such that the value of the weighting factor λ depends on being determined from the corresponding training data in a standard manner.

[0044] It can be set such that at least one criterion 11 relates to the value of a fundamental fact of the corresponding training data, wherein the fundamental fact includes at least the value "Hands-On" or "Hands-Off". Then, in particular, two possible values ​​can be set for the weight factor λ, wherein either one value is selected or the other value is selected depending on the value of the fundamental fact.

[0045] The weighting factor λ can be chosen such that the student model 21 learns more strongly from at least one teacher model 20 for the state “Hands-Off”.

[0046] It can be configured such that at least one teacher model 20 is more specialized and / or robust for the identification of Hands-Off states within the training framework of at least one teacher model 20.

[0047] The weight factor λ can be set such that the student model 21 learns more strongly from the training data 10 for the state “Hands-On” than from at least one teacher model 20.

[0048] It can be configured such that at least one criterion 11 involves a performance parameter of at least one teacher model 20. In particular, the binary cross-entropy between the output of teacher model 20 and the ground facts can be used as a performance parameter. For example, it can be configured such that the larger the value of this performance parameter, the stronger the weight is placed on teacher model 20. In other words, the better at least one teacher model 20 is able to estimate the Hands-On and / or Hands-Off states, the stronger the student model 21 learns from at least one teacher model 20.

[0049] It can be configured such that at least one teacher model 20 includes multiple teacher models 20, which are trained for specific situations.

[0050] List of reference numerals in the attached diagram: 1 device 2. Data processing equipment 2-1 Computing equipment 2-2 Memory 10 Training Dataset 11 Preset Standards 20 Teacher Models 21 Student Model 40 Hands-On / Hands-Off Status 50 vehicles Device 51 52 controllers 53 Memory 54 sensors 55 Steering Parameters 56. Other controllers 100-103 Measures of this method λ is the weighting factor.

Claims

1. A method for providing a trained machine learning model to identify the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50), comprising: At least one teacher model (20) is trained using the training dataset (10). The smaller student model (21) is trained using the at least one teacher model (20) in a teacher-student training manner. For this purpose, a loss function is used, which consists of a weighted sum of the student loss and the distillation loss. The weight factor (λ) for each training data or training data group in the training dataset (10) used during teacher-student training is determined taking into account at least one preset criterion (11). Provide trained student models (21).

2. The method according to claim 1, characterized in that, The trained student model (21) is loaded into the memory (53) of at least one controller (52) of the vehicle (50) within the provided framework after training.

3. The method according to claim 1 or 2, characterized in that, The value of the weight factor (λ) depends on being determined from the corresponding training data in a standard manner.

4. The method according to any one of the preceding claims, characterized in that, The at least one criterion (11) relates to the value of a basic fact of the corresponding training data, wherein the basic fact includes at least the value "Hands-On" or "Hands-Off".

5. The method according to claim 4, characterized in that, The weight factor (λ) is chosen such that the student model (21) learns more strongly from the at least one teacher model (20) for the state "Hands-Off".

6. The method according to any one of the preceding claims, characterized in that, The at least one teacher model (20) is further specialized and / or robust for the identification of Hands-Off states within the training framework of the at least one teacher model (20).

7. The method according to any one of the preceding claims, characterized in that, The weight factor (λ) is determined such that the student model (21) learns from the training data (10) for the state "Hands-On" more effectively than from the at least one teacher model (20).

8. The method according to any one of the preceding claims, characterized in that, The at least one criterion (11) relates to the performance parameters of the at least one teacher model (20).

9. The method according to any one of the preceding claims, characterized in that, The at least one teacher model (20) includes multiple teacher models (20), each of which is trained for a specific context.

10. A method for identifying the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50), comprising: Detect at least one steering parameter (55) at the steering wheel of the vehicle (50). A student model (21) trained using the method according to any one of claims 1 to 9 is used to identify the Hands-On / Hands-Off states, wherein, for this purpose, the detected at least one steering parameter (55) is fed into the input of the trained student model (21). Output the Hands-On / Hands-Off states (40) estimated by the trained student model (21).

11. An apparatus (1) for providing a trained machine learning model to identify the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50), comprising: Data processing equipment (2), The data processing device (2) is configured to train at least one teacher model (20) using a training dataset (10). The smaller student model (21) is trained using the at least one teacher model (20) in a teacher-student training manner. For this purpose, a loss function is used, which consists of a weighted sum of the student loss and the distillation loss, and The weight factor (λ) for each training data or training data group in the training dataset (10) used during teacher-student training is determined with regard to at least one preset criterion (11).

12. A device (51) for identifying the Hands-On / Hands-Off state (40) at the steering wheel of a vehicle (50), comprising: At least one sensor (54) for detecting at least one steering parameter (55) at the steering wheel, Controller (52), The controller (52) is configured to obtain the detected at least one steering parameter (55), provide and apply a student model (21) trained according to any one of claims 1 to 9 to identify the Hands-On / Hands-Off state (40), and for this purpose, feed the detected at least one steering parameter (55) to the input of the trained student model (21), and output the Hands-On / Hands-Off state (40) estimated by the trained student model (21).

Images