Vehicle torque distribution method, electronic equipment and vehicle

Abstract

The invention provides a vehicle torque distribution method, electronic equipment and a vehicle, which are applied to the technical field of vehicle power distribution, and the method comprises the following steps: acquiring a topographic image and environmental data outside the current vehicle; feature extraction is carried out on the topographic image through a topographic feature extraction model obtained through pre-training, and an image embedding vector is obtained; and converting the environment data into an environment semantic embedding vector through a large language model. And fusing the image embedding vector and the environment semantic embedding vector to obtain a fusion vector. Based on the fusion vector, matching with a predetermined power distribution strategy is carried out, and a torque distribution result of the current vehicle is obtained; wherein the power distribution strategy is obtained through reinforcement learning training. The understanding ability of the terrain and the environment can be improved through the fusion vector, the torque distribution method under different driving scenes is optimized, the vehicle energy consumption can be remarkably reduced, and the vehicle slipping phenomenon is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle power distribution, and in particular to a vehicle torque distribution method, an electronic device and a vehicle. BACKGROUND

[0002] The existing off-road vehicle power distribution strategy mostly adopts fixed rules or simple judgments based on a preset map, and cannot accurately perceive real-time terrain changes and environmental complexity, resulting in unreasonable torque output in sand, steep slope, wet and slippery scenes, and further causing problems such as high vehicle energy consumption and frequent skidding. SUMMARY

[0003] Therefore, the present application aims to provide a vehicle torque distribution method, an electronic device and a vehicle to solve the problem of unreasonable vehicle torque output in off-road scenes.

[0004] To achieve the above purpose, the first aspect of the present application provides a vehicle torque distribution method, comprising: obtaining a terrain image and environmental data outside the current vehicle; extracting features of the terrain image through a pre-trained terrain feature extraction model to obtain an image embedding vector; converting the environmental data into an environmental semantic embedding vector through a large language model; fusing the image embedding vector and the environmental semantic embedding vector to obtain a fusion vector; based on the fusion vector, matching with a pre-determined power distribution strategy to obtain a torque distribution result of the current vehicle.

[0005] Optionally, the training method of the terrain feature extraction model comprises: constructing an initial feature extraction model, wherein the initial feature extraction model comprises a backbone network and a classification head; training the initial feature extraction model through a constructed training data set to obtain a trained initial feature extraction model; removing the classification head of the trained initial feature extraction model to obtain the terrain feature extraction model.

[0006] Optionally, the environmental data comprises structured environmental data collected at the vehicle end and multi-modal environmental data obtained from the outside; the conversion of the environmental data into an environmental semantic embedding vector through a large language model comprises: converting the structured environmental data into natural language environmental data; constructing a prompt statement based on the natural language environmental data and the multi-modal environmental data; inputting the prompt sentence into the large language model, and outputting the environment semantic embedding vector by the large language model.

[0007] Optionally, the fusion of the image embedding vector and the environment semantic embedding vector to obtain a fusion vector comprises: The attention weight of the image embedding vector is adjusted according to the semantic information in the environment semantic embedding vector through a cross-modal attention mechanism, and the fusion vector is obtained.

[0008] Optionally, the method for determining the power distribution strategy comprises: Collecting terrain image samples and environment data samples under different driving scenarios; Constructing a state space based on the terrain image samples and the environment data samples; Defining an action space and a reward function; Performing reinforcement learning training based on the state space, the action space, and the reward function, and updating an initial power distribution strategy through interaction with a training environment; In response to the convergence of the initial power distribution strategy, the power distribution strategy is obtained.

[0009] Optionally, the vehicle driving data comprises energy consumption related data, dynamics data, and motor operation data; The definition of the action space and the reward function comprises: Constructing an action space based on a plurality of pre-set vehicle torque distribution results; Constructing an energy saving index according to the energy consumption related data, determining a stability index according to the dynamics data, and determining an efficiency index according to the motor operation data; Constructing a reward function based on the energy saving index, the stability index, and the efficiency index.

[0010] Optionally, the torque distribution result of the current vehicle is obtained by matching the fusion vector with a pre-determined power distribution strategy, comprising: Based on the fusion vector, the torque distribution result is determined by querying a corresponding relationship pre-constructed according to the power distribution strategy; or, Based on the fusion vector, the torque distribution result is determined by a distribution model trained according to the power distribution strategy.

[0011] Optionally, it further comprises: In response to the current vehicle not being a historical vehicle, before converting the environment data into an environment semantic embedding vector by a large language model, it further comprises: Constructing a prior label field according to the vehicle structure parameters of the current vehicle; adding the prior label field to the environment data.

[0012] Based on the same inventive concept, a second aspect of the present application also provides a vehicle torque distribution device, comprising: an acquisition module configured to acquire a terrain image and environment data of a current vehicle exterior; a first embedding module configured to perform feature extraction on the terrain image by a pre-trained terrain feature extraction model to obtain an image embedding vector; a second embedding module configured to convert the environment data into an environment semantic embedding vector by a large language model; a fusion module configured to fuse the image embedding vector and the environment semantic embedding vector to obtain a fusion vector; a matching module configured to obtain a torque distribution result of the current vehicle by matching the fusion vector with a pre-determined power distribution strategy, wherein the power distribution strategy is obtained by reinforcement learning.

[0013] Based on the same inventive concept, a third aspect of the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0014] Based on the same inventive concept, a fourth aspect of the present application also provides a non-transitory computer readable storage medium, which stores computer instructions for causing a computer to execute the method described above.

[0015] Based on the same inventive concept, a fifth aspect of the present application also provides a computer program product, comprising computer program instructions, which, when executed on a computer, cause the computer to execute the method described above.

[0016] Based on the same inventive concept, a sixth aspect of the present application also provides a vehicle comprising the electronic device described above.

[0017] As described above, the vehicle torque distribution method, electronic device, and vehicle provided in this application include: acquiring terrain images and environmental data of the current vehicle exterior; extracting features from the terrain image using a pre-trained terrain feature extraction model to obtain an image embedding vector, which can express rich surface structure features; converting the environmental data into an environmental semantic embedding vector using a large language model, introducing multi-source heterogeneous environmental information through the environmental semantic embedding vector, improving environmental perception capabilities, and facilitating accurate determination of the torque distribution result; fusing the image embedding vector and the environmental semantic embedding vector to obtain a fused vector; and matching the fused vector with a pre-determined power distribution strategy to obtain the torque distribution result of the current vehicle, wherein the power distribution strategy is obtained through reinforcement learning training. The fused vector enhances the understanding of terrain and environment, and the power distribution strategy obtained through reinforcement learning training enables adaptive torque distribution, thereby achieving accurate dynamic torque distribution based on rich state perception, optimizing the torque distribution method under different driving scenarios, and significantly reducing vehicle energy consumption and vehicle slippage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of the vehicle torque distribution method according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the vehicle torque distribution device according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0020] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the embodiments of the present application shall have the common meaning understood by one of ordinary skill in the art to which the embodiments of the present application belong. The terms "first", "second", and similar terms used in the embodiments of the present application do not represent any order, number, or importance, but are only used to distinguish different components. The terms "include", "contain", and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, without excluding other elements or objects. The terms "connect" or "connected" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right", and the like are only used to represent relative positional relationships, and when the absolute positions of the described objects change, the relative positional relationships can also change accordingly.

[0021] The power distribution strategy of existing off-road vehicles is not suitable for complex and variable real-time environments, and its inherent limitations lead to various problems in performance, energy consumption, and safety. Specifically, the following aspects are included: First, the torque output is unreasonable. For example, when the vehicle enters sandy land, a soft and linear torque output is needed to prevent the wheel end torque from being too large to dig a pit. However, the fixed strategy may still use the response mode for paved roads, such as too fast torque release when accelerating, which causes the tires to spin momentarily and the vehicle to sink, which is not conducive to the vehicle getting out of trouble.

[0022] Second, the energy consumption is too high. In long-distance desert or snowfield driving, the continuous low adhesion environment will bring higher risk of slipping. The fixed strategy cannot intelligently identify this continuous state and may always use high torque distribution or frequently intervene ESP (Electronic Stability Program) / TCS (Traction Control System), which will cut off the power or brake when working. Its essence is to maintain stability through energy consumption. The motor or engine runs continuously under high load, and the battery power or fuel is consumed at a very high speed, which seriously shortens the vehicle's range and poses a serious threat to driving safety in extreme environments such as uninhabited areas.

[0023] Third, slipping occurs frequently. When the vehicle is climbing a wet and slippery steep slope, the fixed strategy may not be able to predict the change in road adhesion at the top of the slope (such as from mud to rock). If the torque distribution is improper when the current wheel reaches the top, it may spin, causing the vehicle to stall or even skid. The vehicle's operating stability is reduced, not only unable to pass through obstacles, but also likely to cause accidents and endanger personal safety.

[0024] The above problems mainly occur due to unreasonable vehicle power distribution, lack of joint modeling capability of multi-modal environmental information, and lack of adaptive learning mechanism of power control logic. In view of the lack of joint modeling capability of multi-modal environmental information, specifically, the sensor data (such as wheel speed, acceleration, and yaw rate) relied on by the traditional system is the reaction of the vehicle itself, which is usually lagging and one-sided. Through the sensor data, it can only be known that the tire is slipping, but it is not known why it is slipping and what is in front of it that may cause the tire to slip. The traditional system has the problems of visual information loss and semantic information loss. The visual information refers to the terrain image information captured by the vehicle-mounted camera, and through the terrain image information, it can be determined that the front road surface is sand, water, or delay, and the change of adhesion can be judged. The semantic information belongs to prior information, including semantic information (such as entering the swamp mode, or the front is a long and steep slope, etc.) provided by the driving log or map. The traditional system cannot understand and fuse such abstract semantic information, and cannot realize the coordination of human and vehicle intentions.

[0025] In view of the lack of adaptive learning mechanism of power control logic, specifically, the traditional torque distribution strategy is usually a large number of "if-else" conditional statements (such as if the wheel speed difference > X, then reduce the torque) preset by the user. This way cannot enumerate all possible off-road scene combinations, and there may be conflicts between different conditional statements, and it is extremely difficult to optimize the parameters, mainly relying on expert experience. The fixed rules cannot adapt to the driving styles of different drivers, and cannot realize personalization. Usually, the torque distribution strategy is fixed after the vehicle is offline, and cannot be updated and optimized in real time.

[0026] Therefore, the present application proposes a vehicle torque distribution method, which solves the problem of insufficient perception capability of the traditional system by fusing visual information and semantic information, realizes high-dimensional environmental understanding, so that the system has environmental perception capability, which is beneficial to subsequent vehicle torque distribution. The present application also introduces a reinforcement learning mechanism, which enables the control strategy to learn and optimize itself through interaction with the environment, to autonomously learn what kind of power distribution action should be taken in what kind of complex state for the purpose of reducing energy consumption and improving stability. Through multi-source environmental perception, the reinforcement learning provides rich, accurate, and forward-looking input states, and the reinforcement learning learns a dynamically adjustable and refined vehicle torque distribution strategy using these high-quality input states, and finally realizes the purpose of energy saving and stability in a complex driving environment.

[0027] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0028] The present application proposes a vehicle torque distribution method, referring to Figure 1 , comprising the following steps: Step 102, acquiring a terrain image and environmental data outside the current vehicle.

[0029] Specifically, the terrain image can be obtained according to a camera mounted outside the vehicle. For example, the terrain image is obtained by a front camera and a bottom camera. The front camera can identify the terrain change, obstacle type, etc. in front of the vehicle, and the bottom camera can identify the road surface condition directly below the vehicle. The terrain image needs to have a unified resolution to facilitate subsequent terrain feature extraction.

[0030] The environmental data includes data sensed by sensors mounted on the vehicle, and also includes data provided by an interface such as a high-precision map or a weather system. Some of the data sensed by the sensors can be collected by a T-BOX (Telematics BOX), such as the ambient temperature, humidity, etc. The inertial measurement unit can collect the vehicle's own attitude and dynamic data, such as the slope angle, altitude, yaw angular velocity, acceleration, and ground adhesion coefficient. The high-precision map can provide road condition information and environmental information of the future road section, and the weather system can provide future weather information to provide a prediction time for the vehicle. The structured semantic input is generated by combining the data provided by the high-precision map and the weather system, for example, "currently located in a gravel road section with a medium slope, and short-term rainfall is expected".

[0031] The sampling frequency of the terrain image and the environmental data can be set to 1-10 Hz, which can ensure the continuity of the information while avoiding data explosion, achieving a balance between efficiency and accuracy. Further, in order to ensure consistency after subsequent feature extraction, all collected data needs to be preprocessed to achieve consistent representation of the information. The preprocessing includes: aligning the collected data based on a unified timestamp; performing brightness normalization, view standardization, etc. on the terrain image; mapping the environmental data (such as temperature, slope, etc. data with different dimensions and values) to the [0, 1] interval according to the field normalization; and the structured semantic description is in the form of a structured key-value pair, which is convenient for subsequent processing by the model.

[0032] Step 104, performing feature extraction on the terrain image by using a terrain feature extraction model pre-trained to obtain an image embedding vector.

[0033] Specifically, the terrain feature extraction model in the embodiment can be a convolutional neural network model (CNN). The terrain feature extraction model is used to automatically extract geometric features such as slope, undulation, and surface obstacles in the terrain image, and output a high-dimensional feature vector, that is, an image embedding vector. For example, the basic network structure of the terrain feature extraction model can use ResNet50, and the size of the input terrain image is 224x224x3. The front layers of the network are used to extract edge texture and structural gradient information, and the middle and rear layers are used to extract deeper slope undulation patterns and surface morphology expressions through residual connection. The terrain feature extraction model outputs a 2048-dimensional image embedding vector representing the terrain structure features of the current driving section.

[0034] Step 106: converting the environment data into an environment semantic embedding vector by using a large language model.

[0035] Specifically, a prompt sentence for inputting the large language model is constructed based on the environment data, and a 1024-dimensional environment semantic embedding vector is output by the large language model. The environment semantic embedding vector is a kind of “semantic fingerprint” or “thought vector” of the current environment state, and has the ability to compress and express complex environment semantics. All input information and its complex correlation are compressed and encoded in the environment semantic embedding vector, for example: high humidity + low temperature + slope, forming a freezing risk; gravel + uphill + high altitude, high power demand and easy to slip; expected rainfall + muddy, adhesion will decrease soon and need to be predicted in time.

[0036] The large language model in the embodiment can use an open-source large language model and be fine-tuned on driving scene data. During the fine-tuning of the large language model, LoRA (Low-Rank Adaptation) technology can be used to fine-tune part of the intermediate layers and position encoding in the large language model, retain the backbone knowledge ability, and enhance the modeling ability of the large language model for professional terrain terms and off-road terms. LoRA technology is a parameter-efficient fine-tuning technology. LoRA technology freezes the original parameters of the pre-trained model and only trains a small number of low-rank matrices to quickly adapt the model to new tasks. This method uses matrix decomposition technology to replace the update of high-dimensional weight matrices with low-rank matrices, effectively reducing the computing and storage requirements.

[0037] Step 108: fusing the image embedding vector and the environment semantic embedding vector to obtain a fusion vector.

[0038] Specifically, the image embedding vector represents the visual information of the current road surface, such as slope, undulation and obstacles, and is a concrete and spatial information. The environment semantic embedding vector represents the non-visual information of the current climate, ground state, adhesion condition, slope level and the like, and is an abstract and semantic information. The fusion of the image embedding vector and the environment semantic embedding vector is equivalent to the fusion and correlation of the visual and semantic information, and a fusion vector is extracted for reinforcement learning decision-making. Since the image embedding vector and the environment semantic embedding vector come from different models, the distribution and dimension of the values may be completely different, and cannot be directly fused. Before fusion, a feature normalization operation is required. For example, the image embedding vector and the environment semantic embedding vector are standardized respectively, and the image embedding vector and the environment semantic embedding vector are converted into the same range, so that both kinds of information can contribute effectively to the fusion vector.

[0039] After fusion, a fixed-length fusion vector is outputted. Exemplarily, the dimension of the fusion vector can be 256, which is beneficial to subsequent reinforcement learning training and final vehicle deployment. The fusion vector contains visual details and environment semantics, and is the most comprehensive and in-depth digital expression of the current driving scene.

[0040] Step 110, based on the fusion vector, a torque distribution result of the current vehicle is obtained by matching with a predetermined power distribution strategy; wherein the power distribution strategy is obtained by reinforcement learning training.

[0041] Specifically, the predetermined power distribution strategy gives the corresponding relationship between state and action, and the fusion vector is used as the state, so that the corresponding action can be queried in the power distribution strategy, that is, the corresponding torque distribution result is queried. The torque distribution result includes the front and rear axle torque distribution ratio of the vehicle or the wheel-end torque distribution ratio of each wheel. After the torque distribution result is determined, the vehicle controller parses the torque distribution result into a target motor torque instruction, and sends it to the motor controller through the vehicle controller local area network for execution. The embodiment realizes the optimization learning of the adaptive power distribution strategy by constructing a reinforcement learning training process with Q-learning as the core and using the fusion vector and the defined action space.

[0042] Based on the steps 102 to 110, the vehicle torque distribution method provided by the embodiment includes: obtaining a terrain image and environment data of a current vehicle outside. The terrain feature extraction model obtained by pre-training is used to extract features of the terrain image to obtain an image embedding vector, and the image embedding vector can express rich surface structure features. The environment data is converted into an environment semantic embedding vector by a large language model, the environment semantic embedding vector introduces multi-source heterogeneous environment information, improves the environment perception ability, and is conducive to accurately determining the torque distribution result. The image embedding vector and the environment semantic embedding vector are fused to obtain a fusion vector. Based on the fusion vector, the torque distribution result of the current vehicle is obtained by matching with a pre-determined power distribution strategy; wherein the power distribution strategy is obtained by reinforcement learning. The fusion vector can improve the understanding of the terrain and the environment, and the power distribution strategy obtained by reinforcement learning can realize adaptive torque distribution, thereby realizing accurate dynamic torque distribution based on rich state perception, optimizing the torque distribution method in different driving scenarios, and significantly reducing the energy consumption and the vehicle slip phenomenon of the vehicle.

[0043] In some embodiments, the training method of the terrain feature extraction model includes: An initial feature extraction model is constructed, and the initial feature extraction model includes a backbone network and a classification head. The initial feature extraction model is trained by using a constructed training data set to obtain a trained initial feature extraction model. The classification head of the trained initial feature extraction model is removed to obtain the terrain feature extraction model.

[0044] Specifically, an initial feature extraction model is constructed, and the initial feature extraction model includes a backbone network and a classification head. The terrain features of the terrain image are extracted by the backbone network, and the terrain type corresponding to the terrain image can be output by the classification head. The terrain images in the training data set are derived from real data in different driving scenarios, covering most typical driving scenarios (such as sand, gravel, wet and muddy, shallow potholes, and rocky slopes). The terrain images in the training data set are labeled by laser radar scanning and manual labeling to form labels. The label categories include obstacle level, slope level, and ground adhesion estimation. The laser radar can provide high-precision and objective geometric true values (such as accurate slope values and obstacle heights). Manual labeling can provide semantic and classified true values (such as adhesion coefficient estimation “high / medium / low” and obstacle type).

[0045] The initial feature extraction model is trained through a training data set and a label, and the network discrimination ability is optimized by combining a cross-entropy loss and a center loss. The cross-entropy loss is a standard loss for a classification task, which ensures that the predicted category is as accurate as possible. The center loss not only requires the category to be classified correctly, but also requires the feature vectors of all samples in the same category to be gathered together in space, and the feature vectors of different categories to be far away from each other. This makes the finally extracted feature vectors have small intra-class differences and large inter-class differences, which is beneficial to subsequent fusion and differentiation, and can greatly improve the discrimination ability of the model in unknown scenarios.

[0046] In addition, the generalization ability of the initial feature extraction model can also be improved through an image enhancement method. The training data set is expanded through brightness disturbance, rotation and cropping, etc. The generalization ability of the model is greatly improved, and overfitting is avoided. After training, the backbone network is retained and the classification head is removed, which is equivalent to retaining only the encoding part that can accurately extract image features, and a terrain feature extraction model is obtained. Then, a high-dimensional feature vector rich in terrain information is extracted through the terrain feature extraction model. The subsequent feature fusion provides rich visual context, so that the fusion process can weigh different information sources more finely.

[0047] In addition, during model training, the model inference delay is controlled within 30 ms to meet the performance requirements of vehicle deployment.

[0048] The method of the embodiment gives the training method of the terrain feature extraction model, so that the terrain feature extraction model can perform deep hierarchical feature extraction on the terrain image, obtain a more rich, accurate and discriminative high-dimensional feature representation, and provide reliable and comprehensive visual information basis for subsequent decision-making.

[0049] In some embodiments, the environment data includes structured environment data collected by the vehicle and multi-modal environment data obtained from the outside; The converting the environment data into an environment semantic embedding vector by the large language model comprises: Converting the structured environment data into natural language environment data; Based on the natural language environment data and the multi-modal environment data, a prompt sentence is constructed; The prompt sentence is input into the large language model, and the environment semantic embedding vector is output by the large language model.

[0050] Specifically, the environment data includes structured environment data collected by the vehicle and multi-modal environment data obtained from the outside. The structured environment data includes sensor collected data such as temperature, humidity, slope angle, friction coefficient, and altitude, etc. The multi-modal environment data comes from maps, weather systems, and driving logs, and is described in natural language, such as “current gravel uphill, continuous curve within 500 meters ahead, and expected rain”. Inputting the multi-modal environment data into the large language model is equivalent to injecting prior knowledge into the large language model, so that the large language model can automatically search for data similar in semantics to the multi-modal environment data.

[0051] In order to facilitate the understanding of the large model, it is necessary to convert the structured environment data into natural language environment data. Exemplarily, the structured environment data “temperature = 30, humidity = 85, slope = 12, surface adhesion = 0.35” is converted into natural language “temperature 30 degrees Celsius, humidity 85%, slope 12 degrees, and surface muddy”. In this way, the large model can process both structured environment data and multi-modal environment data.

[0052] Based on the natural language environment data and the multi-modal environment data, a prompt statement is constructed. The prompt statement can be generated through a constructed prompt template. Exemplarily, the prompt template is “The current vehicle environment sensor readings are as follows: the slope is {slope} degrees, the altitude is {altitude} meters, the lateral acceleration is {lateral acceleration} g, the outdoor temperature is {temperature} degrees Celsius, and the current road surface estimated adhesion coefficient is {adhesion coefficient}. In addition, navigation and weather information prompt: {natural language description}”. Correspondingly, the prompt statement filled according to the prompt template can be “The current vehicle environment sensor readings are as follows: the slope is 5.2 degrees, the altitude is 1850 meters, the lateral acceleration is 0.15 g, the outdoor temperature is 12 degrees Celsius, and the current road surface estimated adhesion coefficient is 0.45. In addition, navigation and weather information prompt: 500 meters ahead into continuous sharp curve sandstone road, current wind speed increases”.

[0053] The prompt sentence is input to the large language model, and after the large language model receives the prompt sentence, the prompt sentence is converted into word pieces. The large language model processes these word pieces through its powerful self-attention mechanism. In this process, the large language model establishes complex associations between word pieces. For example, it associates the concepts of "sharp turn", "gravel road", "lateral acceleration", and "coefficient of adhesion 0.45" and understands that they all point to "the risk of skidding". The large language model finally generates and outputs a high-dimensional feature representation, i.e., an environmental semantic embedding vector. The environmental semantic embedding vector integrates the comprehensive understanding of the current driving environment by the large language model, including the status, risks, trends, and associations. Through the method of this embodiment, the powerful abstraction and understanding ability of the large language model is used to integrate scattered, multi-modal environmental information into a high-dimensional vector containing rich semantics, providing rich semantic context for the subsequent vector fusion process and improving the system's environmental perception ability.

[0054] In some embodiments, the fusion of the image embedding vector and the environmental semantic embedding vector to obtain a fusion vector includes: Through the cross-modal attention mechanism, the attention weight of the image embedding vector is adjusted according to the semantic information in the environmental semantic embedding vector to obtain the fusion vector.

[0055] Specifically, in order to further improve the perception sensitivity to high-risk scenarios, a cross-modal attention mechanism is introduced in this embodiment, which allows visual features to automatically and selectively focus on key elements in the environment according to semantic content. That is, semantic information tells visual features where to look and what to focus on. By introducing the cross-modal attention mechanism, dynamic and selective information fusion is achieved instead of static concatenation.

[0056] The cross-modal attention mechanism can be implemented through a cross-attention mechanism. The key (Key) of the environmental semantic embedding vector and the image embedding vector are calculated for similarity to generate an attention weight map. Through the attention weight map, it can be represented which parts or features in the image are important for the current semantic context and the corresponding importance. The value (Value) of the image feature is weighted and summed using the attention weight map, thereby obtaining a new visual feature that is modulated and enhanced by semantic information.

[0057] In a specific implementation, since the dimensions and sources of the semantic vector and the image feature are different, they need to be projected into the same public subspace first to make the environmental semantic embedding vector and the image embedding vector comparable. The environmental semantic embedding vector S is projected into a new dimension through a trainable fully connected layer Wq to obtain Q = S · Wq. Two different trainable fully connected layers Wk and Wv are used to process the image embedding vector I respectively to obtain K = I · Wk and V = I · Wv. The transpose of the Q and K matrices is calculated to obtain Attention Weights = Q · K T . The Attention Weights are scaled and normalized to obtain the attention weight map Alpha = Softmax( (Q · K T ) / sqrt(d) ), where d represents the dimension. Each value in the attention weight map represents an importance score of each spatial position in the image feature. The higher the importance score, the more important the image region corresponding to the spatial position is to the current semantic context. Then, the attention weight Alpha is multiplied with the V matrix to obtain the new visual feature New_Visual_Feature = Alpha · V.

[0058] For example, when the environmental semantic embedding vector contains semantics such as "wet and slippery" and "low adhesion", the cross-attention mechanism can give higher weights to low-contrast regions, reflective regions, and shadow regions in the image embedding vector. When the environmental semantic embedding vector contains semantics such as "steep slope" and "climb", the cross-attention mechanism can give higher weights to the position of the horizon, the ratio of the sky to the ground, and other geometric features that can reflect the slope in the image embedding vector.

[0059] The fusion feature obtained by the fusion method of the embodiment contains visual details and environmental semantics. Through the cross-modal attention mechanism, the perception ability is intelligentized, so that the system can actively focus on potential risk areas, greatly improving the perception sensitivity and response speed to high-risk scenes.

[0060] In some embodiments, the method for determining the power distribution strategy comprises: collecting terrain image samples and environmental data samples under different driving scenarios; constructing a state space based on the terrain image samples and the environmental data samples; defining an action space and a reward function; performing reinforcement learning training based on the state space, the action space, and the reward function, and updating the initial power distribution strategy through interaction with the training environment; In response to the initial power distribution strategy converging, the power distribution strategy is obtained.

[0061] Specifically, the power distribution strategy is obtained through reinforcement learning. First, data samples in different driving scenarios are collected, including terrain image samples and environment data samples. The terrain image samples are collected by the front and bottom cameras mounted on the vehicle. The environment data samples are collected by sensors mounted on the vehicle, and also include structured semantic descriptions extracted from high-precision maps and weather systems. The environment data samples specifically include, for example, ambient temperature, humidity, slope angle, altitude, yaw angular velocity, acceleration, ground adhesion coefficient, etc. At the same time, vehicle controller local area network signals (such as vehicle speed, motor torque, wheel speed, etc.) are synchronously collected.

[0062] Based on the terrain image samples and the environment data samples, a state space is constructed, which includes state vectors (i.e., fusion vectors). The terrain image samples and the environment data samples are aligned and preprocessed: all data samples are aligned based on a unified timestamp. The terrain image samples are subjected to brightness normalization and perspective standardization. The sensor-perceived data in the environment data samples are normalized to the [0, 1] interval by field, and the structured semantic description is converted into a structured key-value pair format. The preprocessed terrain image samples are subjected to feature extraction by a terrain feature extraction model to obtain image embedding vector samples. The environment data samples are converted into environment semantic embedding vector samples by a large language model. After the image embedding vector samples and the environment semantic embedding vector samples are fused, a fusion vector is obtained. The acquisition of the image embedding vector samples and the environment semantic embedding vector samples and the generation of the fusion vector are the same as in the foregoing embodiments, and will not be described again here. Based on the obtained fusion vector (i.e., state vector), a state space is generated.

[0063] Further, the vehicle driving data includes energy consumption related data, dynamics data, and motor operation data; The action space and the reward function are defined, including: The action space is constructed based on a plurality of vehicle torque distribution results pre-set; The energy saving index is constructed according to the energy consumption related data, the stability index is determined according to the dynamics data, and the efficiency index is determined according to the motor operation data; The reward function is constructed based on the energy saving index, the stability index, and the efficiency index.

[0064] Specifically, a discrete action space is constructed based on a plurality of pre-set vehicle torque distribution results. For example, eight typical front axle / rear axle torque distribution ratio combinations [50:50, 40:60, 30:70, 20:80, 60:40, 70:30, 80:20, 0:100] are pre-defined to constitute the action space. When selecting a specific action in the action space according to the state vector, the policy network only needs to select an action from a limited and explicit option, which reduces the exploration dimension and accelerates the convergence speed. On the basis of the discrete action, the embodiment also adds a plurality of continuous action intervals. For example, the policy can not only select the discrete point of 70:30, but also select a specific value in the continuous interval of 65:35 to 75:25. The continuous action change can bring a smoother torque switching process and reduce the impact of the vehicle. Generally, each continuous interval is associated with a discrete action as a fine-tuning interval of the discrete action. For example, the continuous interval of 65:35 to 75:25 is the fine-tuning interval of the discrete action of 70:30.

[0065] The action (whether discrete or continuous) selected by the reinforcement learning policy is first received by the vehicle controller. The vehicle controller calculates the target torque values of the front axle and the rear axle according to the torque distribution ratio, the current driver demand torque (from the accelerator pedal), and the maximum / minimum torque output capability of the vehicle. For distributed drive, the vehicle controller further distributes the axle torque to the left and right wheel ends according to the differential, anti-skid and other requirements. The final target torque instructions calculated are transmitted to the motor controllers through the Controller Area Network (CAN) bus. After receiving the instructions, the motor controllers control the motor to output accurate torque.

[0066] It should be noted that the action instruction cannot exceed the maximum torque that the motor can provide (limited by the motor itself and the battery discharge capability) and the minimum torque (usually the regenerative braking torque), so as to avoid torque instructions that the output motor cannot realize and protect the vehicle hardware. The action change frequency (i.e., the control frequency of the reinforcement learning) must be slower than the response time of the motor system, otherwise the instructions will be stacked, causing control imbalance. The torque distributed to a front axle, a rear axle or a wheel end cannot exceed the limit that can be borne by the current adhesion, otherwise it will immediately cause skidding.

[0067] In addition, for complex off-road conditions, an action space interface can be reserved to support expansion, such as adding a differential lock control state, a torque limiting factor, an instantaneous escape mode, etc., to improve the adjustment ability and safety margin of the policy in extreme environments.

[0068] By constructing a reward function, the immediate effect of executing a specific action a (torque allocation strategy) in a specific state s (fusion vector) can be accurately and quantitatively evaluated, and this can guide the agent to learn to balance the three indicators of energy saving, stability and efficiency. Ensure that the torque allocation strategy formulated after learning can improve the vehicle's endurance and optimize the driving quality.

[0069] An energy-saving indicator is constructed according to energy consumption-related data, which can include the voltage, current of the vehicle controller or battery management system, and the driving distance of the vehicle. The energy-saving indicator is to minimize the energy consumed per unit distance, which can be quantified by the unit mileage power consumption or oil consumption. The energy-saving indicator can be represented by the following formula: R energy = -k energy × (ΔE / Δd) Where ΔE represents the energy consumed in a control period Δt, ΔE = P × Δt, the instantaneous power P = I × V, I represents the instantaneous current of the vehicle controller or battery management system, and V represents the voltage of the vehicle controller or battery management system. Δd represents the distance traveled in Δt obtained by a wheel speed sensor or the like. k energy represents the weight coefficient of the energy-saving indicator.

[0070] A stability indicator is determined according to dynamics data, which includes wheel speed difference, lateral acceleration fluctuation, wheel speed, tire radius, vehicle speed, etc. The stability indicator is to ensure the passability, controllability and safety of the vehicle on various terrains, and to prevent instability. The stability indicator is a composite indicator, which usually includes wheel speed difference, lateral acceleration fluctuation and slip judgment. The wheel speed difference represents the speed difference between the left and right wheels of the same vehicle. A larger wheel speed difference usually represents that one of the wheels is slipping, and the vehicle loses power or stability. The sharp change in lateral acceleration indicates that the vehicle may be sliding or spinning. Slip judgment is determined by slip rate, slip rate λ = (ω × r - v) / max(ω × r, v), where ω is the wheel speed, r is the tire radius, and v is the actual speed of the vehicle. (ω × r - v) represents the slip speed, i.e. the speed of the tire tread relative to the ground. If the tire is pure rolling, the contact point between the tire tread and the ground is instantaneous and stationary, and the slip speed is 0. If the slip speed is positive, it means that the tire is rotating faster than the vehicle speed (ω × r > v), and the tire is slipping. λ = 0, the tire is pure rolling, ω × r = v, no slip. λ > 0, slip starts. The larger the λ value, the more serious the slip. λ approaches 1, serious slip.

[0071] According to the motor operation data, the efficiency index is determined, and the motor operation data includes the rotating speed and the output torque of the motor. Through the efficiency index, the driving motor can work in the high-efficiency interval as much as possible, and the purpose of energy saving is indirectly realized. The efficiency index can be represented by the deviation degree of the current operating point of the motor and the highest efficiency interval. According to the actual rotating speed and the output torque of the current motor, the current actual efficiency is obtained by table lookup. For the same rotating speed, the maximum possible efficiency at the rotating speed is found by table lookup. The higher the ratio of the current actual efficiency to the maximum efficiency, the higher the motor operating efficiency, and the higher the reward.

[0072] After the energy-saving index, the stability index and the efficiency index are defined, the reward function is constructed, and the reward function can be represented by the following formula: R total = w energy × R energy + w stab × R stability + w eff × R efficiency Among them, R energy represents the energy-saving index, w energy represents the weight of the energy-saving index, R stability represents the stability index, w stab represents the weight of the stability index, R efficiency represents the efficiency index, and w eff represents the weight of the efficiency index. Each weight is not fixed and can be dynamically adjusted according to the current state. Through the dynamic weighting mechanism, the system can dynamically adjust the weights of each index, guide the agent to make reasonable trade-offs in different scenarios, and achieve global optimization.

[0073] For example, in the driving scene of a slippery downhill, the stability of the vehicle is the most important, and energy saving is secondary. The system will automatically increase w stab and reduce w energy . The strategy will tend to choose a torque distribution that can better maintain the grip. In the driving scene of a long slope climbing, the motor is in long-term high-load operation, and it is particularly important to prevent the motor from overheating and maintain high-efficiency operation. The system will increase w eff to encourage the motor to work in the high-efficiency interval, while also indirectly saving energy.

[0074] The reward signal is generated for each state-action pair as the basis for updating the Q value in the Q-learning algorithm. Q-learning is one of the most classic and important algorithms in reinforcement learning, which belongs to the value function learning method. Its core idea is to learn a Q function. Q function Q(s, a): represents the cumulative expected value of all future rewards that can be obtained after choosing to perform action a in state s. The Q value is a score of the long-term return of a "state-action pair". The higher the Q(s, a) value, the more profitable it is to choose action a in state s in the long run. The ultimate goal is to learn an optimal strategy π, which can choose the action a that maximizes the Q(s, a) value in any state s, π(s) = argmax a Q(s, a).

[0075] To improve the stability of the learning process, a penalty term can also be added to the reward function R total A penalty term is introduced to suppress undesirable behaviors of the vehicle, such as long-term wheel slip, dramatic output torque fluctuations, and repeated switching of drive modes. After the agent performs action a in each control period, the environment (vehicle simulation model) enters a new state s' and calculates an instant reward value r, which is fed back to the agent.

[0076] In the reinforcement learning process, the strategy usually adopts random actions before training (such as using a 50:50 torque distribution ratio). The agent obtains the current state vector (i.e., the fusion vector) from the simulation environment. The simulation environment calculates all the current vehicle parameters and external information based on a high-precision vehicle model and terrain model. The agent selects an action from the action space based on the state vector according to the ε-greedy strategy (for example, selects a torque distribution mode of "front axle 40%, rear axle 60%"). The environment executes this action, and the vehicle dynamics model calculates the result after execution and enters the next state. At the same time, the environment calculates a reward value according to the reward function. The agent stores the experience of this interaction in the experience replay pool. Then, it randomly selects a batch of past experiences from the pool for updating its strategy. For example, using the Q-learning algorithm, the Bellman equation is used to update the "value judgment". For example, in the "wet uphill curve" (state), choosing "heavy acceleration" (action) brings poor long-term returns (low Q value). The next time it encounters a similar state, it will be less willing to choose this action. By adjusting the parameters of the neural network through gradient descent, the prediction of good actions becomes more and more accurate, and the prediction of bad actions becomes worse and worse. After multiple interactions, the optimal power distribution strategy is obtained through continuous updating of the initial power distribution strategy.

[0077] It should be noted that the ε-greedy strategy is adopted in the training process to balance exploration and utilization, a higher random exploration ratio is set in the initial stage to fully cover diversified off-road scene samples, and the value of ε is gradually reduced in the later stage to converge to a stable strategy. In order to improve the training efficiency and generalization ability, the environment interaction adopts a "virtual off-road terrain-vehicle model" platform based on simulation, and the scene atlas is generated from the test field data collected in reality, which is used for high-frequency motion evaluation and strategy optimization, and reduces the cost of real vehicle training.

[0078] Through the method of the embodiment, an automatic process of continuously trying and correcting the learning torque distribution strategy is given. By designing a reward function including three indexes of energy saving, stability and efficiency, an intelligent power distribution strategy that can cope with complex off-road environments and comprehensively consider energy saving and stability is obtained after reinforcement learning.

[0079] In some embodiments, the torque distribution result of the current vehicle is obtained by matching the fusion vector with a predetermined power distribution strategy, including: Based on the fusion vector, the torque distribution result is determined by querying the corresponding relationship constructed in advance according to the power distribution strategy; or, Based on the fusion vector, the torque distribution result is determined by a distribution model trained according to the power distribution strategy.

[0080] Specifically, after completing the reinforcement learning through the foregoing embodiments, a converged and high-performance power distribution strategy is obtained. There are two ways to deploy the power distribution strategy on the vehicle end: One is to use the offline Q table (corresponding relationship). The offline Q table is a pre-computed table that contains the corresponding relationship between states and actions. Each value in the table is the Q value of the state-action pair. By giving the state vector (fusion vector), the corresponding action can be queried. The Q table can be directly burned into the memory of the vehicle end controller. After the vehicle end real-time perception of the current state, all Q values corresponding to the current state are found in the Q table, and the action corresponding to the maximum Q value is returned. The offline Q table is easy to query and has small calculation overhead.

[0081] Secondly, an approximate function is used. A small neural network (such as a multi-layer perceptron) is used to simulate or approximate the trained Q function to learn the internal mapping relationship between the state features and the optimal action. The decision data set is formed by collecting the decision results of the Q function. The small neural network is trained through the decision data set, so that the output of the small neural network is as close to the Q function as possible. The trained small neural network is deployed to the vehicle-side controller. After the vehicle-side real-time perception of the current state, the current state is input into the trained small neural network, and the optimal action is output through the trained small neural network. The calculation speed of the small neural network is fast, and the required floating point operation amount and memory occupation are small, which can meet the real-time requirements (such as millisecond-level response) of the vehicle-level chip. Compared with the offline Q table, the small neural network has a certain reasoning ability, and can reason the corresponding optimal action for a new state.

[0082] The method of the embodiment gives a deployment method of the power distribution strategy on the vehicle side. Users can choose a suitable method for deployment according to actual needs, and provides a possibility for real-time power distribution on the vehicle side.

[0083] In some embodiments, further comprising: In response to the current vehicle not being the historical vehicle, before converting the environment data into an environment semantic embedding vector through the large language model, further comprising: constructing a prior label field according to vehicle structure parameters of the current vehicle; adding the prior label field to the environment data.

[0084] Specifically, due to the differences in vehicle types and terrains, the torque distribution strategy trained on a specific vehicle type and specific test scene data may have a significant performance decline or even failure when directly applied to other vehicle types or other driving scenes. In order to enhance the adaptability of the method of the application and avoid strategy migration failure, the embodiment proposes an adaptation method for different vehicle types. In specific implementation, a prior label field is added to the prompt sentence input into the large language model. The prior label field can include vehicle structure parameters such as driving form, wheelbase, minimum ground clearance, and electric drive peak power. By introducing the prior label field, the attempt and reasoning ability of the large model can be stimulated to generate an environment semantic embedding vector related to the vehicle type. In this way, even if there is no data for a new vehicle type, as long as the key parameters of the new vehicle type are provided, the system can immediately generate a relatively reasonable state estimation without starting from complete randomness for retraining.

[0085] If the new vehicle model has similarities with the historical vehicle model, strategy initialization and migration can also be performed. In specific implementation, the key structural parameters (drive form, wheelbase, power, etc.) of the vehicle can be coded into a digital vector. Taking this vector as an index, a query is performed in the constructed knowledge base. If a similar vehicle model is queried, a part of the trained Q value can be directly migrated as an initial value, reducing the training time and improving the initial quality of the strategy. For example, the Q value table of the sand strategy trained for a four-wheel drive vehicle with a wheelbase of 2.8 meters and a power of 200 kW has a very high reference value for another four-wheel drive vehicle with a wheelbase of 2.7 meters and a power of 180 kW. Even if there is no similar vehicle model, the system can interpolate between the strategies of similar vehicle models to construct Q values with reference value.

[0086] Correspondingly, in response to the current driving scene not being the historical driving scene, before converting the environment data into an environment semantic embedding vector through the large language model, the method further includes: constructing a prior label field according to terrain information of the current driving scene; adding the prior label field to the environment data.

[0087] The prior label field is added to the prompt sentence input to the large language model, and the prior label field can include topographic information and terrain levels. The topographic information is, for example, sand, gravel, mud, steep slope, etc., and the terrain levels are, for example, low, medium, and high. By introducing the prior label field, the attempt and reasoning ability of the large model can be stimulated to generate an environment semantic embedding vector related to the terrain information. After the prior label field is input, misjudgment can be avoided. For example, if the prior label field is not input, the system will give a general semantic of “steep slope” for a slope of 12°. If the prior label field is input (for example, low steep slope), the system can distinguish between “this is a steep slope that is very dangerous for a small-power rear-drive vehicle” and “this is a steep slope that is very easy for a large-power four-wheel drive vehicle”. In this way, the subsequent environment semantic embedding vector is more targeted, thereby guiding the reinforcement learning algorithm to make more reasonable decisions.

[0088] Through the method of the embodiment, when a new vehicle model or driving scene is encountered, a new torque distribution strategy can be obtained by only a small amount of adaptation and fine-tuning on the existing torque distribution strategy. The method is particularly suitable for unified strategy scheduling of multiple vehicle models and multiple driving scenes, and significantly improves the replicability and scalability of the strategy in engineering deployment.

[0089] It should be noted that the method of the embodiments of the present application can be executed by a single device, for example, a computer or a server, etc. The method of the embodiments can also be applied to a distributed scenario, and be completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only execute one or more steps in the method of the embodiments of the present application, and the multiple devices can interact with each other to complete the method.

[0090] It should be noted that some embodiments of the present application have been described above. In some cases, the actions or steps recorded in the above embodiments can be executed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0091] Based on the same inventive concept, the present application also provides a vehicle torque distribution device corresponding to any of the above-mentioned embodiment methods.

[0092] Reference Figure 2 The vehicle torque distribution device comprises: An acquisition module 202 configured to acquire a terrain image and environment data outside a current vehicle; A first embedding module 204 configured to perform feature extraction on the terrain image through a pre-trained terrain feature extraction model to obtain an image embedding vector; A second embedding module 206 configured to convert the environment data into an environment semantic embedding vector through a large language model; A fusion module 208 configured to fuse the image embedding vector and the environment semantic embedding vector to obtain a fusion vector; A matching module 210 configured to obtain a torque distribution result of the current vehicle by matching the fusion vector with a pre-determined power distribution strategy, wherein the power distribution strategy is obtained through reinforcement learning training.

[0093] In some embodiments, a training module is further included, configured to construct an initial feature extraction model, wherein the initial feature extraction model comprises a backbone network and a classification head; The initial feature extraction model is trained through a constructed training data set to obtain a trained initial feature extraction model; The classification head of the trained initial feature extraction model is removed to obtain the terrain feature extraction model.

[0094] In some embodiments, the environment data includes structured environment data collected by the vehicle and multi-modal environment data obtained from the outside; the second embedding module 206 is configured to convert the environment data into an environment semantic embedding vector by the large language model, including: converting the structured environment data into natural language environment data; constructing a prompt statement based on the natural language environment data and the multi-modal environment data; inputting the prompt statement into the large language model, and outputting the environment semantic embedding vector by the large language model.

[0095] In some embodiments, the fusion module 208 is configured to adjust the attention weight of the image embedding vector according to the semantic information in the environment semantic embedding vector by a cross-modal attention mechanism, to obtain the fusion vector.

[0096] In some embodiments, further comprising a determination module configured to collect terrain image samples and environment data samples under different driving scenarios; constructing a state space based on the terrain image samples and the environment data samples; defining an action space and a reward function; performing reinforcement learning training based on the state space, the action space, and the reward function, and updating an initial power distribution strategy through interaction with a training environment; in response to convergence of the initial power distribution strategy, obtaining the power distribution strategy.

[0097] In some embodiments, the vehicle driving data includes energy consumption related data, dynamics data, and motor operation data; the determination module is further configured to construct an action space based on a plurality of vehicle torque distribution results preset; construct an energy saving index according to the energy consumption related data, determine a stability index according to the dynamics data, and determine an efficiency index according to the motor operation data; and construct a reward function based on the energy saving index, the stability index, and the efficiency index.

[0098] In some embodiments, the matching module 210 is configured to determine the torque distribution result by querying a corresponding relationship pre-constructed according to the power distribution strategy based on the fusion vector; or, determining the torque distribution result by a distribution model trained according to the power distribution strategy based on the fusion vector.

[0099] In some embodiments, in response to the current vehicle not being the historical vehicle, before converting the environment data into an environment semantic embedding vector by the large language model, the method further comprises a migration module configured to construct a priori label field according to vehicle structure parameters of the current vehicle; and adding the priori label field into the environment data.

[0100] For the convenience of description, the above apparatus is described in various modules according to functions. Of course, in the implementation of the present application, the functions of each module can be implemented in one or more software and / or hardware.

[0101] The apparatus of the above embodiments is used to implement the vehicle torque distribution method of any one of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be described here.

[0102] Based on the same inventive concept, the present application also provides an electronic device corresponding to the method of any one of the above embodiments, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the vehicle torque distribution method of any one of the above embodiments.

[0103] Figure 3 A more specific hardware structure schematic diagram of an electronic device provided by the present embodiment is shown, which can include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040 and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030 and the communication interface 1040 are connected to each other through the bus 1050 for communication within the device.

[0104] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application specific integrated circuit (ASIC), or one or more integrated circuits, etc., for executing related programs to implement the technical solutions provided by the present embodiment.

[0105] The memory 1020 can be implemented by a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 1020 can store an operating system and other application programs, and when the technical solutions provided by the present embodiment are implemented by software or firmware, the related program codes are stored in the memory 1020 and executed by the processor 1010.

[0106] The input / output interface 1030 is configured to connect an input / output module to realize information input and output. The input / output module can be configured in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. The input device can include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device can include a display, a speaker, a vibrator, an indicator light, etc.

[0107] The communication interface 1040 is configured to connect a communication module (not shown in the figure) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired manner (such as a USB, a network cable, etc.) or a wireless manner (such as a mobile network, WIFI, Bluetooth, etc.).

[0108] The bus 1050 includes a channel to transmit information between various components (such as the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040) of the device.

[0109] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in the specific implementation process, the device can also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device can also only include components necessary for implementing the embodiments of the present specification, and does not necessarily include all the components shown in the figure.

[0110] The electronic device of the above embodiments is used to implement the corresponding vehicle torque distribution method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described here.

[0111] Based on the same inventive concept, corresponding to any of the above embodiment methods, the present application also provides a non-transitory computer readable storage medium storing computer instructions for causing the computer to execute the vehicle torque distribution method according to any of the above embodiments.

[0112] The computer readable medium of the embodiments includes permanent and non-permanent, removable and non-removable media, which can realize information storage by any method or technology. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0113] The storage medium of the above-mentioned embodiments stores computer instructions for causing the computer to execute the vehicle torque distribution method according to any one of the above-mentioned embodiments, and has the beneficial effects of the corresponding method embodiments, which are not repeated here.

[0114] Based on the same concept, the present application also provides a computer program product corresponding to the method of any of the above-mentioned embodiments, comprising computer program instructions, which, when executed on a computer, cause the computer to execute the method according to any one of the above-mentioned embodiments, and have the beneficial effects of the corresponding method embodiments, which are not repeated here.

[0115] It can be understood that before using the technical solutions of the various embodiments in the present disclosure, the user will be informed of the type, scope of use, use scenario, etc. of the personal information involved by appropriate means, and the authorization of the user will be obtained.

[0116] For example, in response to receiving the user's active request, the user is sent prompt information to explicitly prompt the user that the operation requested to be performed will require the acquisition and use of the user's personal information. Thus, the user can voluntarily choose whether to provide personal information to the software or hardware such as electronic devices, application programs, servers or storage media that perform the technical solutions of the present disclosure according to the prompt information.

[0117] As an optional but not limited implementation manner, in response to accepting the user's active request, the way of sending prompt information to the user may, for example, be the pop-up window manner, and the prompt information can be presented in the form of text in the pop-up window. In addition, the pop-up window can also carry a selection control for the user to select "agree" or "disagree" to provide personal information by the electronic device.

[0118] It can be understood that the above notification and obtaining user authorization process is only illustrative, and does not limit the implementation of the present disclosure, and other ways meeting relevant laws and regulations can also be applied to the implementation of the present disclosure.

[0119] Those skilled in the art should understand that the above discussion of any embodiment is only exemplary, and is not intended to limit the scope of the present application to these examples; the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the embodiments of the present application as described above, which are not provided in details for the sake of brevity.

[0120] In addition, in order to simplify the description and discussion, and so as not to make the embodiments of the present application difficult to understand, the well-known power / ground connections of integrated circuit (IC) chips and other components can or can not be shown in the provided drawings. In addition, the apparatus can be shown in the form of a block diagram in order to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram apparatus are highly dependent on the platform to be implemented in the embodiments of the present application (i.e., these details should be fully within the understanding of those skilled in the art). Where specific details (e.g., circuitry) are set forth in order to describe an illustrative embodiment of the present application, it will be apparent to those skilled in the art that the present application can be practiced without these specific details or with an implementation varying from these specific details. Therefore, these descriptions should be considered as illustrative rather than limiting.

[0121] Although the present application has been described in conjunction with the specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art in light of the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) can use the embodiments discussed.

[0122] The embodiments of the present application are intended to cover all such alternatives, modifications and variations as falling within the broad scope of the present application. Accordingly, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A vehicle torque distribution method, characterized by, The method comprises: acquiring a terrain image and environment data of a current vehicle outside; extracting features of the terrain image through a pre-trained terrain feature extraction model to obtain an image embedding vector; converting the environment data into an environment semantic embedding vector through a large language model; fusing the image embedding vector and the environment semantic embedding vector to obtain a fusion vector; based on the fusion vector, matching with a pre-determined power distribution strategy to obtain a torque distribution result of the current vehicle; wherein the power distribution strategy is obtained through reinforcement learning training.

2. The method of claim 1, wherein, The training method of the terrain feature extraction model comprises: constructing an initial feature extraction model, wherein the initial feature extraction model comprises a backbone network and a classification head; training the initial feature extraction model through a constructed training data set to obtain a trained initial feature extraction model; removing the classification head of the trained initial feature extraction model to obtain the terrain feature extraction model.

3. The method of claim 1, wherein, The environment data comprises structured environment data collected by the vehicle and multi-modal environment data obtained from the outside; the conversion of the environment data into an environment semantic embedding vector through a large language model comprises: converting the structured environment data into natural language environment data; constructing a prompt statement based on the natural language environment data and the multi-modal environment data; inputting the prompt statement into the large language model to output the environment semantic embedding vector through the large language model.

4. The method of claim 1, wherein, The fusion of the image embedding vector and the environment semantic embedding vector to obtain a fusion vector comprises: adjusting the attention weight of the image embedding vector according to the semantic information in the environment semantic embedding vector through a cross-modal attention mechanism to obtain the fusion vector.

5. The method of claim 1, wherein, The determination method of the power distribution strategy comprises: collecting terrain image samples and environment data samples under different driving scenarios; constructing a state space based on the terrain image samples and the environment data samples; defining an action space and a reward function; performing reinforcement learning training based on the state space, the action space and the reward function, and updating an initial power distribution strategy through interaction with a training environment; in response to the convergence of the initial power distribution strategy, obtaining the power distribution strategy.

6. The method of claim 5, wherein, The vehicle driving data comprises energy consumption related data, dynamics data and motor operation data; defining an action space and a reward function comprises: constructing an action space based on a plurality of pre-set vehicle torque distribution results; constructing an energy saving index according to the energy consumption related data, determining a stability index according to the dynamics data, and determining an efficiency index according to the motor operation data; constructing a reward function based on the energy saving index, the stability index and the efficiency index.

7. The method of claim 1, wherein, based on the fusion vector, matching with a pre-determined power distribution strategy to obtain a torque distribution result of the current vehicle, comprises: based on the fusion vector, querying and determining the torque distribution result in a corresponding relationship pre-constructed according to the power distribution strategy; or Based on the fusion vector, a distribution model trained according to the power distribution strategy is used to determine the torque distribution result.

8. The method of claim 1, wherein, Also comprising: In response to the current vehicle not being the historical vehicle, before converting the environment data into an environment semantic embedding vector through the large language model, further comprising: Constructing a prior label field according to the vehicle structure parameters of the current vehicle; Adding the prior label field to the environment data.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1 to 8.

10. A vehicle characterized by comprising: The vehicle comprises the electronic device of claim 9. The vehicle comprises the electronic device of claim 9.

Images