Apparatus, method and machine learning system for determining a transmission state of a vehicle

Abstract

Device, method and machine learning system for determining a transmission state of a vehicle. The method comprises: providing an input of a first model, in particular a first generative model, depending on route information, a vehicle speed, a probabilistic variable, in particular a noise, and providing an output of a second model, in particular a second physical model; determining an output of the first model in response to the input of the first model, wherein the output of the first model characterizes a state, wherein the first model comprises a first layer trained to map the input of the first model determined depending on the route information, the vehicle speed and the probabilistic variable to an intermediate state, wherein the first model comprises a second layer trained to map the intermediate state to the state depending on the output of the second model; providing an input of the second physical model depending on at least one vehicle state and / or the route information; determining an output of the second model in response to the input of the second model, wherein the output of the second model characterizes at least one limitation for the intermediate state.

Description

Background Technology

[0001] To enable the prediction of the gear to be engaged in a future engine-driven vehicle, wherein the vehicle has a transmission including at least two selectable gears for transmission and / or reducing engine torque, DE 10 2006007717 A1 proposes: detecting and evaluating the current driving situation, and, depending on the evaluation of the current driving situation, determining probability values ​​for at least two gears, and predicting the gear for which the highest probability value has been determined.

[0002] US 5557521A discloses a control system for changing automatic gears, which operates based on a fuzzy logic approach. All fuzzy generation rules are divided into at least three parts: a set of basic rules for determining gear change points in a consumption-oriented driving style; a set of adaptation rules for modifying the set of basic rules depending on the current driving style; and a set of identification rules for classifying the driver / her according to their driving style.

[0003] WO 2019 / 127231 A1 discloses a method for generating training data for machine learning, the method comprising: generating simulated training data for a target neural network, and transforming the simulated training data using a training data transformer to form transformed training data, the training data transformer being trained to increase the consistency between the transformed training data and the simulated training data; and training the target neural network using the transformed training data.

[0004] Physical models, or probabilistic models, can be used to determine various aspects of a vehicle's transmission system. While physical models are deterministic, their accuracy depends on the observability of the system being modeled. Probabilistic models can model unobservable behavior. However, it is difficult to prove that the transmission states predicted by a probabilistic model are physically feasible.

[0005] Reliable and accurate determination of the vehicle's transmission status is desirable. Summary of the Invention

[0006] This is achieved through the methods, apparatus, and machine learning systems claimed in the independent claims.

[0007] A method for determining the transmission state of a vehicle includes: providing inputs to a first model, particularly a first generative model, depending on route information, vehicle speed, and probability variables (especially noise), and providing outputs to a second model, particularly a second physical model; determining the output of the first model in response to the inputs to the first model, wherein the output of the first model represents a state, wherein the first model includes a first layer trained to map the inputs of the first model, determined depending on the route information, vehicle speed, and probability variables, to intermediate states, wherein the first model includes a second layer trained to map intermediate states to states depending on the outputs of the second model; providing inputs to the second physical model depending on at least one vehicle state and / or route information; and determining the output of the second model in response to the inputs to the second model, wherein the output of the second model represents at least one constraint for intermediate states. The first model is a generative model. The output of the second layer depends on the output of the second model. The second layer may apply physical constraints to the intermediate outputs of the first layer. The second layer may also include, for example, a conversion from acceleration to velocity. The second model provides at least one constraint for the output of the generative model. This constraint depends on physical constraints. Therefore, the accuracy and reliability of the transmission state, such as a gear position, are significantly improved. Another advantage of building upon a purely data-based model is that the relationships modeled in the physical model help extrapolate the model to characteristics of the test data that are not present in the training data of the data-based model. For example, if only 6-speed vehicles are available in the training data and at test time, then gears for 5-speed vehicles can also be generated. The physical model helps to make useful transformations for this new, unseen test case.

[0008] Preferably, at least one vehicle state is defined by at least one geographical variable (in particular at least one gradient trajectory value) or at least one vehicle variable (in particular at least one speed trajectory value). Route information (such as road gradient or vehicle speed) provides physical information that can be used to determine limitations for transmission gears.

[0009] Preferably, at least one vehicle state is defined by at least one vehicle parameter, which in particular is at least one of the following: transmission gear ratio, differential gear ratio, tire circumference, vehicle engine idle speed, safety margin, gradient, speed, engine speed, power required by auxiliary equipment, vehicle drag, and vehicle engine full-load power-speed curve. These vehicle parameters provide physical information that the physical model can use to determine limitations for transmission gears.

[0010] Preferably, the route information includes route features, wherein the method includes extracting route features from map information. These route features can improve the mapping of the generative model.

[0011] Preferably, the route information includes at least one of the following: geographical characteristics, particularly absolute elevation or road gradient characteristics; traffic flow characteristics, particularly time-related average traffic speed; road characteristics, particularly the number of lanes, road type, and / or road curvature; traffic control characteristics, particularly speed limits, the number of traffic lights, the number of specific types of traffic signs, the number of stop signs, the number of yield signs, and / or the number of pedestrian crossing signs; and weather characteristics, particularly rainfall, wind speed, and / or the presence of fog at a predetermined time. This information can be used to predict the vehicle's transmission gears using a generative model and to train the generative model.

[0012] Preferably, the method includes: providing inputs to a third model depending on route information, vehicle speed, and state; determining the output of the third model in response to the inputs of the third model, wherein the output of the third model represents a score indicating the realism estimate of the state, wherein the third model is trained to map the inputs of the third model determined depending on the route information, vehicle speed, and state to the output of the third model, the output of the third model representing a score indicating the realism estimate of the state. The score can output true or false information regarding whether the gear time series represented by the output determined for a time period matches a real-world gear time series of a transmission for the same speed limit and gradient within that time period.

[0013] Preferably, route information and vehicle speed are defined by a continuous or discrete value sequence over a time period, wherein the first model includes a first recurrent neural network adapted to process the values ​​in the value sequence over the time period, and the third model includes a second recurrent neural network adapted to process the values ​​in the value sequence over the time period, wherein the method includes processing the same values ​​of route information and vehicle speed in the value sequence over the time period by both the first model and the third model.

[0014] Preferably, the probability variable and the state are defined by a continuous or discrete sequence of values ​​over time within that time period.

[0015] Preferably, the method includes determining the gear characteristics over time based on multiple inputs to a second model and multiple inputs to a first model.

[0016] Preferably, the method includes estimating the vehicle's exhaust characteristics based on the gear characteristics and / or score over time and based on the vehicle engine model.

[0017] A corresponding device includes at least one processor and a storage device adapted to perform the steps of the method.

[0018] A machine learning system includes a first model, a second model, and a third model, and is adapted to perform the steps of the method.

[0019] A method for training a machine learning system includes: determining a state in response to training data containing input data defining a first model and a second model, depending on the outputs of the first and second models; determining the output of a third model representing a score indicating the degree of realism of an estimate of the state; and determining at least one parameter of the first, second, and / or third models based on the score; wherein the input data for a training period includes the state, route information, and at least one vehicle state. The physical model represented by the second model may have free parameters that need to be adjusted. In this respect, these can be trained using exactly the same training algorithm as when training the first model. Attached Figure Description

[0020] Further advantages can be derived from the following description and figures. In the figures: Figure 1 The diagram schematically depicts a device used to determine the state of a vehicle's transmission. Figure 2 The aspects of the model are schematically depicted. Figure 3 The steps in the method for determining state variables are schematically depicted. Figure 4 A machine learning system is schematically depicted. Figure 5 It schematically depicts another aspect of the physical model.

[0021] Figure 1 Device 100 is schematically depicted. Device 100 may include at least one processor and a storage device, which can be adapted to perform the model and method steps described below.

[0022] The device 100 includes a first model, particularly a first generative model 102, and a second model, particularly a second physical model 104.

[0023] The first model 102 is a data-based model. The first model 102 includes a first layer 102a, which is adapted to map inputs representing route information 106a, vehicle speed 106b, and probability variables 108 (particularly noise) to outputs representing intermediate states 110 of the vehicle's transmission. The inputs representing route information 106a and vehicle speed 106b are sequences of state variables. In this example, intermediate state 110 is a sequence of state variables indicating the generated unconstrained gear.

[0024] Route information 106a can be determined, for example, based on a sequence of longitude and latitude coordinates. Route information 106a in this example includes map features. Map features may include speed limits, topological information (such as slope), and weather conditions. Map features may include at least one of the following: geographical characteristics, particularly absolute altitude or road slope characteristics; traffic flow characteristics, particularly time-related average traffic speed; road characteristics, particularly the number of lanes, road type, and / or road curvature; traffic control characteristics, particularly speed limit characteristics, the number of traffic lights, the number of specific types of traffic signs, the number of stop signs, the number of yield signs, and / or the number of pedestrian crossing signs; and weather characteristics, particularly rainfall at a predetermined time, wind speed, and / or the presence of fog.

[0025] The second model 104 is adapted to determine the output representing at least one constraint 112 for intermediate state 110, depending on at least one vehicle state 114 and / or route information 106a, particularly the gradient.

[0026] Device 100 is adapted to determine the state 116 of a transmission, specifically the gear position of a vehicle. The transmission may include a clutch or torque converter, at least one drive gear, and several pinions. In this example, the transmission gear refers to a fixed gear ratio resulting from a selected combination of drive gears and pinions. In this example, state 116 is a sequence of state variables that, for each time step, indicate one of six forward driving gears via an integer value of 1, 2, 3, 4, 5, or 6. For example, specifically, gear values ​​for each time step are generated with the same discretization as the input data (e.g., route information 106a). Reverse driving gear, neutral, or parking gear can similarly be indicated by other integer values. For continuously variable transmission systems, the gear can be a continuous floating-point variable, for example, indicating the actual gear ratio instead.

[0027] The first model 102 is a generative model for predicting the output representing state 116. The second model 104 includes a physical model for physically modeling reasonable gear positions and for determining constraints based on them. The first model 102 includes a second layer 102b, which is adapted to apply constraints to the intermediate state 110 generated by the first layer 102a of the first model 102.

[0028] In this context, the first layer 102a and the second layer 102b are model layers, i.e., components of the first model 102. In this context, the term "layer" does not correspond to a single layer in a neural network, but rather a model layer can consist of multiple neural network layers.

[0029] In this respect, device 100 is adapted to constrain the output of the generative first model 102 by limitations depending on physical constraints. A second model 104 enriches the purely data-based generator output of the first model 102 with prior knowledge from physics. This provides a "hybrid" model because it combines a data-based model with a physics-based model.

[0030] Device 100 may include a third model 118. The third model 118 may be a data-based model. The third model 118 is adapted to map inputs representing route information 106a, vehicle speed 106b, and state 116 to an output representing a score 120 indicating the realism estimate of state 116. This score may output true or false information regarding whether the gear represented by the output matches a real-world gear for the same speed limit and gradient. The score is not necessarily binary. The score may be a regression value, e.g., positive for true and negative for false.

[0031] Device 100 may include a training device 122 adapted to determine at least one parameter of a first model 102 and / or a second model 104 based on a score 120. In this example, to train the first model 102, a gradient descent method (e.g., ADAM) may be iteratively repeated based on training data used to train the first model 102 to map inputs representing route information 106a, vehicle speed 106b, and probability variable 108 to outputs representing intermediate states 110, which represent the predicted gear or gear ratio of the transmission. A third model 118 may be trained. Input data for training may include state 116, route information 106a, vehicle speed 106b, and / or at least one vehicle state 114. In this context, "training" refers to determining parameters for the first model 102 or the third model 118, or alternately, parameters for both, based on the training data. Input data for a training cycle may include state 116, route information 106a, vehicle speed 106b, and at least one vehicle state 114.

[0032] In one example, device 100 includes a generative adversarial network (GAN), where a first model 102 and a second model 104 are configured as generators and trained, and a third model 118 is configured and trained to score the authenticity or falsity of a given state 116. Jenson-Shannon divergence, particularly regularized Jenson-Shannon divergence, can be used as an objective function for training the GAN. In one example, the GAN is a Wasserstein GAN. In this case, the third model 118 is a critic or discriminator that scores the authenticity or falsity of state 116 based on the Wasserstein distance. Regularized Wasserstein distance can be used.

[0033] Generative adversarial networks are configured as follows: Given some real-world data based on gear position x and environmental condition c In this context, generative adversarial networks learn from distributed systems. Samples are extracted from the generator network. Trained to use another distribution From the distribution Sampling is performed in this other distribution. Approximate real data distribution More specifically, minimizing the difference through adversarial training. The difference in this example is the Wasserstein distance or Jensen-Shannon divergence, specifically the regularized Jensen-Shannon divergence. In a given generator network... g And critique f In this case, it corresponds to the following minimax objective: in This means estimating the divergence, and the generator attempts to minimize that divergence. The Wasserstein distance is, for example, defined as... in F Composed of all 1-Lipschitz functions, i.e. .

[0034] Therefore, the minimum-maximum objective is refer to Figure 2 Further details are provided on other aspects of the model.

[0035] At least one vehicle state 114 can be defined by at least one vehicle parameter 202, which in particular is at least one of the following: transmission gear ratio, differential gear ratio, tire circumference, vehicle engine idle speed, safety margin, gradient, speed, engine speed, power required by one or more auxiliary devices, vehicle drag, and vehicle engine full-load power-speed curve. These vehicle parameters provide physical information that a physical model can use to determine limitations for transmission gears.

[0036] At least one vehicle state 114 can be defined by at least one geographical variable 204 (in particular at least one slope trajectory value) or at least one vehicle variable 206 (in particular at least one speed trajectory value).

[0037] The second model 104 includes a physical modeling device 208 adapted to model at least one constraint 112. In this example, the at least one constraint 112 is a hard physical constraint on continuous or discrete gear output. In this example, a minimum gear 210 and a maximum gear 212 for that time step are determined. The at least one constraint 112 may be a minimum gear only or a maximum gear only for one or all time steps.

[0038] The second layer 102b may include a limiting device 214. In this example, the limiting device 214 is defined by a limiting function that depends on the minimum gear 210 and the maximum gear 212: in It is x t Compressed to θ Parameterized functions in the [0,1] range, for example in The minimum gear is 210. The maximum gear is 212. x t Characterizes intermediate transmission state 110, and x' t Characterizes transmission state 116. In this example, the gear output is discrete. In this example, by, for example, x' t Round to the nearest integer. x' t Discretization. Alternatively, the discrete values ​​can be determined using the softmax function. This requires masking with respect to the minimum / maximum range. In this context, masking refers to filtering out values ​​outside the limits during training. θThe parameters in this formula are characterized and may be adjusted during the training of the generative adversarial network. Alternatively, they can be set to constants, specifically 1, and not adjusted during training. In this example, in one aspect, intermediate transmission states 110 and 116 are comprised of a time series of values. In this example, the limiting device 214 is adapted to output transmission state 116 as determined by the limiting function.

[0039] Route information 106a may include route features. In this respect, device 100 can be adapted to extract route features from map information. These route features can improve the mapping of generative models.

[0040] Route information 106a may include at least one of the following: geographical characteristics, particularly absolute altitude or road gradient characteristics; traffic flow characteristics, particularly time-related average traffic speed; road characteristics, particularly the number of lanes, road type, and / or road curvature; traffic control characteristics, particularly speed limits, the number of traffic lights, the number of specific types of traffic signs, the number of stop signs, the number of yield signs, and / or the number of pedestrian crossing signs; and weather characteristics, particularly rainfall, wind speed, and / or the presence of fog at a predetermined time. This information can be used to predict vehicle transmission gears using a generative model and to train the generative model.

[0041] Route information 106a and vehicle speed 106b can be defined by a continuous or discrete sequence of values ​​over a time period. In this example, the time period can be divided into time steps that include values ​​used to determine transmission state 116. In this example, each time step determines one transmission state 116.

[0042] Route information (such as road gradient or vehicle speed) provides physical information that can be used to determine limits for transmission gears.

[0043] In this example, probability variable 108 and state 116 are defined by a continuous or discrete sequence of values ​​over time within that time period.

[0044] The following is for reference. Figure 3 The described method assumes that a trained first model 102 is implemented. In one aspect, a third model 118 may exist specifically during training. However, after training, the first model 102 and the second model 104 may be used independently of the third model 118.

[0045] In this example, the first model 102 and the third model 118 include recurrent neural networks. The recurrent neural network can be implemented as a long short-term memory network, a gated recurrent unit, a transformer architecture, or a vanilla recurrent neural network. The first model 102 includes a first recurrent neural network adapted to process values ​​in a sequence of values ​​for that time period.

[0046] The third model 118 includes a second recurrent neural network adapted to process values ​​in the value sequence for that time period. Device 100 is adapted to process the same values ​​of route information 106a and vehicle speed 106b in both the first model 104 and the third model 118. In this example, the first model 102 and the third model 118 are adapted to process the same values ​​of route information 106a and vehicle speed 106b in the value sequence for that time period within the same period. The second model 104 can be adapted to provide its output to the first model 102 based on the route information 106a within the same period. Device 100 is adapted to process the third model 118 using the output of the first model 102 within the same period as input to the third model 118 within the same period.

[0047] The method for determining the transmission state 116 of a vehicle includes step 302: providing inputs to a first model 102 based on route information 106a, vehicle speed 106b, and probability variables 108. In this example, the probability variables are noise, such as white noise.

[0048] Subsequently, in step 304, in response to the input of the first model 102, an intermediate state 110 of the transmission is determined based on the output of the first model 102. The output of the first layer 102a of the first model 102 characterizes the intermediate state 110. In this example, the input of the first layer 102a of the first model 102, determined from the route information 106a, vehicle speed 106b, and probability variable 108, is mapped to the output of the first layer 102a characterizing the intermediate state 110. A state 116 is determined based on the intermediate state 110 and at least one constraint 112. In this example, the value of state 116 is determined by the second layer 102b of the first model 102 that implements the constraint function.

[0049] The method further includes step 306: providing input to the second model 104 depending on at least one vehicle state 114.

[0050] Subsequently, in step 308, in response to the input of the second model 104, at least one constraint 112 is determined for the intermediate state 110, depending on the output of the second model 104. The output of the second model 104 characterizes at least one constraint 112.

[0051] The method may include step 310: providing input to a third model 118 based on route information 106a, vehicle speed 106b, and state 116; and step 312: determining a score 120 indicating the veracity estimate of state 116 based on the output of the third model 118, in response to the input of the third model 118. The output of the third model 118 represents the score 120. In this respect, the third model 118 is trained to map the input of the third model 118 determined based on route information 106a, vehicle speed 106b, and state 116 to the output representing the score 120.

[0052] The method may include step 314: determining at least one parameter of the first model 102 or the third model 118, or both, based on the score 120.

[0053] As an example, at least one constraint 112 is determined based on at least one vehicle state 114 defined by at least one geographic variable 204 or at least one vehicle variable 206. At least one geographic variable 204 may include at least one gradient trajectory value. At least one vehicle variable 206 may include at least one speed trajectory value.

[0054] At least one limitation 112 may be determined by at least one vehicle state 114 defined by at least one vehicle parameter 202, which in particular is at least one of the following: transmission gear ratio, differential gear ratio, tire circumference, vehicle engine idle speed, safety margin, gradient, speed, engine speed, power required by auxiliary equipment, vehicle resistance, and vehicle engine full-load power-speed curve.

[0055] exist Figure 4 The exemplary machine learning system 400 for determining state 116 is schematically depicted in the figure.

[0056] In this example, route information 106a and vehicle speed 106b are defined by a continuous sequence of values ​​over time within time period 402. Instead of a continuous sequence of values ​​over time within time period 402, the sequence of values ​​can be discrete.

[0057] Route information 106a may include route features. At least one of the following may be used: Geographical characteristics, especially absolute altitude or road slope characteristics, Traffic flow characteristics, especially time-related average traffic speed, And road characteristics, especially the number of lanes, road type and / or road curvature; Traffic control characteristics, especially speed limit characteristics, the number of traffic lights, the number of specific types of traffic signs, the number of stop signs, the number of yield signs and / or the number of pedestrian crossing signs; Weather characteristics, particularly rainfall, wind speed, and / or the presence of fog at the scheduled time.

[0058] exist Figure 4 In the diagram, the road gradient characteristic 404 and the speed limit characteristic 406 are schematically depicted as route information 106a.

[0059] In this case, the method may include extracting route features from map information. In this respect, the intermediate state 110 may be determined based on the route information 106a, which includes the route features.

[0060] The first model 102 includes a first layer 102a and a second layer 102b. The machine learning system 400 is adapted to process values ​​in a sequence of values ​​over time period 402. In this example, the values ​​of the probability variable 108 are determined to be from, in particular, a standard normal distribution or a uniform distribution. P z noise in mid-sample z ~ P z Any other distribution can also be used. The input to the first layer 102a is noise. z and inputs for i=0,...,T c i , where T is the number of discrete values ​​in time period 402, determined depending on route information 106a and vehicle speed 106b. In this respect, noise z and input c i These can be concatenated to form the input to the first layer 102a. The output of the first layer 102a in response to this input is an intermediate state 110, which in this example is an unrestricted gear. The second layer 102b is adapted to process the intermediate state 110 based on the output of the second model 104 to determine state 116. In this example, a minimum gear 210 and a maximum gear 212 are provided as the output of the second model 104. In this example, state 116 in time steps has specific values ​​indicating a discrete number of gears. The gear sequence at times 0,...,T is represented by a sequence of different values ​​of state 116.

[0061] The third model 118 is adapted to process the values ​​in the value sequence of that time period as input to the third model 118. The input to the third model 118 is... c i ,enter c i It depends on route information 106a, vehicle speed 106b, and the information being indicated.x' i The state 116 is used to determine this. In this respect, the input... c i and x' i These can be concatenated to form the input to a third model 118. The third model 118, in response to its input, outputs a score 120, which, for example, indicates the authenticity of state 116 with y=1 and the falsity of state 116 with y=0. In the Wasserstein generative adversarial network mentioned above, y is a continuous value.

[0062] In this example, the machine learning system 400 is adapted to process the same values ​​of route information 106a and vehicle speed 106b by both the first model 102 and the third model 118.

[0063] exist Figure 5 The diagram schematically shows a detailed view of the second model 104, which is specifically designed for this purpose.

[0064] In one example, the gears to be generated are only constrained to a physically reasonable range. Given the speed at a given time step, such as at state 116 in a calculation cycle, it is possible to calculate the minimum and maximum gears the vehicle can operate in, since different gears can maintain a certain speed range. These gear ranges can be calculated based on physical quantities such as the power available in each gear and the power required to maintain the target speed. These quantities can be obtained via physical model 500.

[0065] Figure 5 The text provides a detailed description of physical quantities and their interactions for a specific example.

[0066] Exemplary inputs used for calculations within the cycle include a cycle slope trajectory value 502 and a cycle velocity trajectory value 504. The cycle velocity trajectory value 504 is referred to as velocity. v The periodic slope trajectory value of 502 is called slope 502.

[0067] Other input parameters are gear ratio 506, differential ratio 508, and tire circumference 510.

[0068] velocity at a given time step v In this case, the second model 104 calculates the physically reasonable minimum reasonable gear 210 and maximum reasonable gear 212 as follows: For each gear i The total gear ratio of 512 is calculated as follows: Using a total gear ratio of 512 and a given speedv Each gear required to drive at that speed i The (theoretical) engine speed of 514 is calculated as follows: Because the clutch is used in, for example, the first two gears, the engine speed 514 calculated in previous steps may not be valid. For instance, the driver might use the clutch when leaving a stationary driving position (acceleration) or coming to a stop (deceleration). In this example, it is assumed that the first two gears are affected by clutch use. The minimum engine speed used for acceleration and deceleration can be given as input by a domain expert, or it can be initialized to 1.15 * engine idle speed 516 based on the input of engine idle speed 516. Then, it is calculated whether the acceleration or deceleration mode is applicable. This can be based on the speed at the current and previous time steps. v Depending on the result of this calculation, replace engine speed 514 with the minimum engine speed for acceleration or the minimum engine speed for deceleration.

[0069] This provides the resulting engine speed of 518.

[0070] In the previous steps, for each gear... i It achieved the speed maintenance v The required engine speed, in RPMs, is 518. This engine speed of 518 is then converted into a gear. i The available power is 520. This is the power that the engine will produce at the obtained engine speed 518. This conversion from engine speed to power can be accomplished via the engine full-load power-speed curve 522, which is typically obtained from vehicle specifications and provided as input.

[0071] Prior to this conversion, the engine speed of 518 could be limited as follows: a. Engine speed 518 can be trimmed to the maximum possible engine speed 524. The maximum possible engine speed 524 can be a manufacturer-specified input.

[0072] b. The vehicle speed v can be adjusted to fall within a certain range of safety margin 526. Safety margin 526 can be set by a domain expert as input. An example of safety margin 526 is: i. Engine speed should not be lower than idle speed.

[0073] ii. Engine speed should not be too close to maximum speed.

[0074] In this example, at the gear position i The available power of 520 is determined individually for each gear.

[0075] Additionally, determine the maintenance speed.v The required power is 528. This power of 528 is calculated as follows: in, These are based on vehicle resistance f 0、 f 1. f 2 and inertial force parameters f 惯性 The calculated driving resistance, pitch angle resistance, and inertial resistance. f eta This is the transmission system efficiency factor. Vehicle resistance. f 0、 f 1. f 2 and inertial force parameters f 惯性 and transmission system efficiency factor f eta It can be input parameter 530, which can be obtained from the vehicle specifications. P aux It is the power used by auxiliary equipment (such as headlights and air conditioning). P aux The input value can be 532. Additionally, a It is acceleration, which can be measured from the velocity at the current and previous time steps. v To calculate. m It refers to the quality of the vehicle.

[0076] Based on each gear i The available power is 520, and the speed is... v Required power 528, select gear 534 to achieve gear i Available power 520 ≥ for speed v The required power is 528. This may result in a physically reasonable gear 534 or a list of physically reasonable gears 534 that meet the condition. From this list, the physically reasonable minimum gear 210 and maximum gear 212 are obtained. If only one gear 534 is available, that gear can be used for both.

[0077] Optionally, neutral or gear 0 can be added to a physically valid gear list. This process can be performed heuristically. For example, neutral can be added to a physically valid gear 534 if: a. The vehicle is stationary and its acceleration is 0.

[0078] b. The vehicle gradually slowed down and came to a stop.

[0079] c. Domain experts can add more conditions.

[0080] Based on the second model 104, at least one vehicle state 114 can be defined by at least one vehicle parameter 202, which in particular is at least one of the following: transmission gear ratio 506, differential gear ratio 508, tire circumference 510, vehicle engine idle speed 516, safety margin 526, gradient 502, speed 504, engine speed 514, power required by one or more auxiliary devices 532, vehicle resistance 530, and vehicle engine full-load power-speed curve 522.

[0081] In one aspect, the time series of the values ​​of slope 106a and the time series of the values ​​of vehicle speed 106b. It serves as the input to both the data-based model (i.e., model 102) and the physical model (i.e., model 104). Time series. They may have the same source. If the vehicle speed 106b indicates a time series of speeds. Then, the velocity trajectory value of 504 for one cycle comes from the time series. A subset of the values. The same applies to slope, i.e., a value in the time series of slope 502 is a subset of the values ​​in the slope time series of 106a.

[0082] The machine learning system 400 calculates the possible physically plausible range of gears at time step t. .

[0083] To ensure that the generator of the machine learning system 400 only generates data within the specified range The gear positions within the range are generated using the previously mentioned equation. x t (For example, intermediate state 110) scaled to x' t (For example, state value 116).

[0084] If the gear (i.e., state 116) generated by the first model 104 is a continuous value, after constraining the value and providing it as state value 116 to the downstream task, the value can be rounded to the nearest integer value.

[0085] In one respect, by utilizing prior knowledge from physics, gear shift trajectories can be generated from the output of a physical model (i.e., the second model 104). For gear shift trajectories generated in this way, it is ensured that the generated trajectories are within physically plausible limits. Therefore, these gear shift trajectories have higher accuracy compared to, for example, trajectories generated solely by the first model 102.

[0086] Compared to training without the second model 102, training of the generative adversarial network including the second model 104 is faster and more stable because the second model 104 does not have to learn physical relationships from the data.

[0087] The aforementioned method and device 100, and especially the machine learning system 400, are better generalized to vehicle or road conditions on which the training data cannot be observed, because some relationships between input and output are described by explicit relationships and therefore not learned from the data. Specifically, for example, gear ratios are highly dependent on the vehicle of interest. By integrating a physical model, the resulting device 100 or machine learning system 400 can, for example, generate gears for gear ratios not observed in the training data, even with varying numbers of gears.

[0088] Compared to using only a physical model, some additional advantages are as follows: Physics-based gear shifting models only output a range of feasible gear values. To generate a realistic gear shifting sequence, a gear value must be selected for each time step from this range. Using a generative adversarial network (GAN)-based model, this selection is learned to be made from the minimum and maximum ranges. For example, using heuristics to select the optimal gear for engine performance, to randomly select from the range, or to select from the top k quantiles of the range will lead to worse accuracy because these heuristics are often violated in the real world due to many potential unobserved factors (e.g., driver behavior, overtaking other vehicles). This difference in unobserved variables is addressed in the probabilistic properties of the GAN-based generator (i.e., the first model 102).

[0089] Alternatively, discrete values ​​can be determined by a trained artificial neural network, which includes a suitable activation function (e.g., softmax), and the physical model is implemented as a second model 104. In this case, the artificial neural network can be trained to constrain the input (e.g., intermediate state 110) to a range. .

[0090] The physical model (i.e., the second model 104) depends on some variables that are not directly input to the model. Furthermore, intermediate features are computed from some input variables. These intermediate features are determined based on expert knowledge. Additionally or alternatively, since it is known from the physical model that not only the feasible gear range but also the actual gears depend on these expert features, these features can be used to train the first model 102 (i.e., the generator), for example, by adding them as input to the first model 102. In this way, the relationship between these features and the actual gears at time step t can be learned from the first model 102.

[0091] The trained model, device 100, and / or machine learning system 400 can also perform the methods described above for predicting gear characteristics over time.

[0092] Gear characteristics can be used to simulate vehicle emissions. This simulation can be used to make probabilistic judgments on compliance with exhaust limits, to parameterize the vehicle's engine, or to optimize vehicle control. The simulation can be used for predictive vehicle control, powertrain management, and / or for torque regeneration in vehicle powertrains that include internal combustion engines, electric engines, or both.

[0093] After training, the second model 104 and / or the third model 118 may not exist. When the third model 118 exists after training, it can be used to determine or distinguish between safe and unsafe gear characteristics over time. This ensures higher accuracy through physical compliance, and this in turn ensures higher accuracy in downstream tasks.

[0094] Therefore, downstream tasks can avoid unstable and / or unsafe mechanisms, because physically unreasonable inputs can make downstream systems unstable and / or unsafe.

[0095] In a preferred embodiment, the gear characteristics over time are determined in response to multiple inputs to the first model 102 and depending on multiple inputs to the second model 104. Additionally, the vehicle's exhaust characteristics may be estimated based on the gear characteristics over time and on a vehicle engine model.

Claims

1. A method for determining the state (116) of a vehicle's transmission, characterized in that: The input to (302) a first model (102) is provided based on route information (106a), vehicle speed (106b), and probability variable (108), and the output of a second model (104) is provided, wherein the first model (102) is a generative model and the second model (104) is a physical model; the output of the first model (102) is determined in response to the input of the first model (102), wherein the output of the first model (102) represents the state (116), wherein the first model (102) includes a first layer (102a), the first layer (102a) being trained to determine the first model (102) based on route information (106a), vehicle speed (106b), and probability variable (108). The input is mapped to an intermediate state (110), wherein the first model (102) includes a second layer (102b) which is trained to map the intermediate state (110) to a state (116) depending on the output of the second model (104); the input of the second model (104) is provided depending on at least one vehicle state (114) and / or route information (106a); the output of the second model (104) is determined in response to the input of the second model (104), wherein the output of the second model (104) characterizes at least one constraint (112) for the intermediate state (110), wherein the state of the transmission is the gear of the transmission, and the intermediate state characterizes the predicted gear or gear ratio of the transmission.

2. The method according to claim 1, characterized in that, The at least one vehicle state (114) is defined by at least one geographical variable (204) or at least one vehicle variable (206).

3. The method according to claim 1 or 2, characterized in that, The at least one vehicle state (114) is defined by at least one vehicle parameter (202), which is at least one of the following: transmission gear ratio (506), differential gear ratio (508), tire circumference (510), vehicle engine idle speed (516), safety margin (526), ​​gradient (502), speed (504), engine speed (514), power required for auxiliary equipment (532), vehicle resistance (530), and vehicle engine full-load power-speed curve (522).

4. The method according to claim 1 or 2, characterized in that, Route information (106a) includes route characteristics, The method described therein includes extracting route features from map information.

5. The method according to claim 1 or 2, characterized in that, Route information (106a) includes at least one of the following: geographical characteristics, traffic flow characteristics, road characteristics, traffic control characteristics, and weather characteristics.

6. The method according to claim 1 or 2, characterized in that: The input to the third model (118) is provided based on route information (106a), vehicle speed (106b), and state (116); the output of the third model (118) is determined in response to the input of the third model (118), wherein the third model (118) is trained to map the input of the third model (118) determined based on route information (106a), vehicle speed (106b), and state (116) to the output of the third model (118), the output of the third model (118) representing a score (120) indicating the truthiness estimate of state (116).

7. The method according to claim 6, characterized in that, Route information (106a) and vehicle speed (106b) are defined by a continuous or discrete sequence of values ​​over a time period, wherein a first model (102) includes a first recurrent neural network adapted to process the values ​​in the sequence of values ​​over the time period, and a third model (118) includes a second recurrent neural network adapted to process the values ​​in the sequence of values ​​over the time period, wherein the method includes processing the same values ​​of route information (106a) and vehicle speed (106b) in the sequence of values ​​over the time period by both the first model (102) and the third model (118).

8. The method according to claim 7, characterized in that, The probability variable (108) and the state (116) are defined by a continuous or discrete sequence of values ​​over time within the time period.

9. The method according to claim 6, characterized in that, The gear characteristics over time are determined based on multiple inputs to the second model (104) and multiple inputs to the first model (102).

10. The method according to claim 9, characterized in that, The exhaust characteristics of the vehicle are estimated based on the gear characteristics over time and / or the score (120) and the vehicle engine model.

11. The method according to claim 1, characterized in that, The probability variable (108) is noise.

12. The method according to claim 2, characterized in that, The at least one geographical variable (204) is at least one slope trajectory value, and the at least one vehicle variable (206) is at least one speed trajectory value.

13. The method according to claim 5, characterized in that, The geographical characteristics are absolute altitude or road gradient characteristics; and the traffic flow characteristics are time-related average traffic speeds; the road characteristics are the number of lanes, road type, and / or road curvature; the traffic control characteristics are speed limit characteristics, the number of traffic lights, the number of stop signs, the number of yield signs, or the number of pedestrian crossing signs; and the weather characteristics are rainfall, wind speed, and / or the presence of fog at a predetermined time.

14. A device (100) for determining the state (116) of a vehicle's transmission, characterized in that, The device is adapted to perform the steps of the method according to any one of claims 1 to 13.

15. A machine learning system (400), characterized in that, The system includes a first model (102), a second model (104), and a third model (118), and is adapted to perform the steps of the method according to claim 6 or 7.

16. A method for training a machine learning system (400) according to claim 15, characterized in that, In response to training data that defines the input data of the first model (102) and the second model (104), a state (116) is determined based on the outputs of the first model (102) and the second model (104), wherein the output of the third model (118) that determines the score (120) representing the truthiness estimate of the state (116) is determined, and wherein at least one parameter of the first model (102) and / or the third model (118) is determined based on the score (120), wherein the input data for the training period includes the state (116), route information (106a), and the at least one vehicle state (114).

17. A computer program product, characterized in that, The computer program product includes computer-readable instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 13.

Images