Torque adjustment method and vehicle

Abstract

This application discloses a torque adjustment method and a vehicle, belonging to the field of intelligent driving technology. It includes: determining a target deviation score based on driving operation information, vehicle state information, and vehicle environmental perception information; the target deviation score representing the degree of deviation between the driver's operational intention and the operational requirements in the current environment; when the target deviation score exceeds a preset threshold, determining the vehicle's desired output torque in the current environment based on the driving operation information, environmental perception information, and the target deviation score; and controlling the vehicle to adjust its torque based on the desired output torque. Thus, when there is a mismatch between driving intention and operational requirements under environmental constraints, by correcting the original output torque to an output torque that matches the environmental requirements, effective intervention in the vehicle's power output can be achieved, thereby improving vehicle driving safety.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and in particular to a torque adjustment method and a vehicle. BACKGROUND

[0002] Driving safety has always been one of the important research directions in the field of vehicle technology. Usually, manufacturers will set various safety auxiliary functions in the vehicle development process, such as lane keeping assistance, automatic emergency braking, etc. However, safety auxiliary functions may not completely cover all safety scenarios, which may lead to driving safety problems in some scenarios.

[0003] For example, in some complex road conditions or congestion scenarios, the driver may misoperate the accelerator due to nervousness, misjudgment or operation habits, such as mistakenly deep pressing the accelerator that should be lightly pressed, or deep pressing the accelerator when it should be released, so that the power output that should be reduced suddenly increases, which may cause energy waste and even safety hazards. Therefore, there is an urgent need for a solution that can reasonably correct the misoperation of the accelerator. SUMMARY

[0004] The present application provides a torque adjustment method, a vehicle and a storage medium, which can determine a desired output torque in the current environment based on the deviation degree and vehicle data and environmental data in each dimension when the deviation between the driver's operation intention and the operation demand in the current environment is large. Finally, the driving intervention is combined with the desired output torque to adjust the output of the vehicle torque, so that when the driving operation and the environmental demand do not match, the original output torque is corrected to an output torque that matches the environmental demand, so that effective intervention on the vehicle power output can be realized, thereby ensuring the driving safety of the vehicle. The technical solution includes the following contents.

[0005] In a first aspect, a torque adjustment method is provided, the method comprising: determining a target deviation score based on driving operation information, vehicle state information and environment perception information of an environment in which the vehicle is currently located, the target deviation score being used to represent a deviation degree between an operation intention of a driver and an operation demand in the current environment; in a case where the target deviation score is greater than a preset score threshold, determining a desired output torque of the vehicle in the current environment based on the driving operation information, the environment perception information and the target deviation score; controlling the vehicle to perform torque adjustment based on the desired output torque.

[0006] In the present application, by first determining a target deviation score based on driving operation information, vehicle state information and environment perception information of the vehicle, the deviation degree between the operation intention of the driver and the operation demand in the current environment can be known, and then it is judged whether the operation intention of the driver matches the operation demand in the current environment according to the deviation degree. When the deviation degree is large, it means that the operation intention of the driver does not match the operation demand in the current environment, and then the power output of the vehicle can be intervened. Specifically, the expected output torque of the vehicle in the current environment can be determined based on the driving operation information, the environment perception information and the target deviation score, and then the vehicle is controlled to adjust the torque based on the expected output torque. The target deviation score represents the risk degree of the current operation of the driver. The risk degree of the driving operation is different, and the driving scene represented by the risk degree is also different. Therefore, the degree of driving intervention is different, and the output torque is also different. By combining the target deviation score in the process of determining the expected output torque, the target deviation score can affect the degree of intervention. Therefore, the expected output torque that matches the current environment can be accurately determined based on the driving operation information, the environment perception information and the target deviation score, that is, the torque that should be output by the vehicle in the current environment is accurately determined. Finally, the original output torque of the vehicle is adjusted based on the torque that should be output by the vehicle in the current environment, which can realize effective intervention on the driving operation. Compared with the scheme of directly cutting off the power response for driving intervention in the prior art, the occurrence of unexpected accidents can be avoided, and the driving intervention effect is improved. For example, in a scene where the driver should slow down, the driver mistakenly operates the accelerator pedal. Directly cutting off the power response in the prior art will cause the vehicle to lose power quickly. If there is a vehicle following behind, a rear-end accident may occur. However, in the present scheme, the original output torque is adjusted to the expected output torque that matches the scene, so the power response is not directly switched, and the rear-end accident in the prior art does not occur. Therefore, the driving safety can be improved.

[0007] In a second aspect, a torque adjustment device is provided, which comprises: a deviation calculation module configured to determine a target deviation score based on driving operation information, vehicle state information and environment perception information of the a torque determination module configured to determine an expected output torque of the vehicle in the current environment based on the driving operation information, the environment perception information and the target deviation score when the target deviation score is greater than a preset score threshold. a torque adjustment module configured to control the vehicle to adjust the torque based on the expected output torque.

[0008] In a third aspect, a vehicle is provided, and the vehicle comprises: a memory for storing executable program code; a processor for invoking and running the executable program code from the memory, so that the vehicle performs the torque adjustment method described above.

[0009] In a fourth aspect, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the torque adjustment method described above.

[0010] In a fifth aspect, a computer program product containing instructions is provided, and when the computer program product is run on a computer, the computer is caused to perform the steps of the torque adjustment method described above.

[0011] It can be understood that the beneficial effects of the second aspect, the third aspect, the fourth aspect and the fifth aspect described above can be referred to the related description in the first aspect described above, and will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0013] Figure 1 is a scene diagram of a torque adjustment method provided by an embodiment of the present application; Figure 2 is a flowchart of a torque adjustment method provided by an embodiment of the present application; Figure 3 is a model structure diagram of a deviation detection model provided by an embodiment of the present application; Figure 4 is a model structure diagram of an expected torque matching model provided by an embodiment of the present application; Figure 5 is a structure diagram of a double-channel neural perturbation regulator provided by an embodiment of the present application; Figure 6 is a flowchart of another torque adjustment method provided by an embodiment of the present application; Figure 7 is a structure diagram of a torque adjustment device provided by an embodiment of the present application; Figure 8 is a structure diagram of a vehicle provided by an embodiment of the present application. DETAILED DESCRIPTION

[0014] In order to make the purposes, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0015] It should be understood that the "multiple" mentioned in the present application refers to two or more. In the description of the present application, unless otherwise specified, " / " represents the meaning of or, for example, A / B can represent A or B; "and / or" herein only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, in order to clearly describe the technical solutions of the present application, the same items or similar items with basically the same functions and roles are distinguished by using "first", "second", etc. The skilled in the art can understand that "first", "second", etc. do not limit the quantity and execution order, and "first", "second", etc. also do not necessarily mean different.

[0016] First, the terms related to the embodiments of the present application are explained.

[0017] 1. Graph Neural Network (GNN): Graph Neural Network is a kind of deep learning model specially used for processing graph structure data, which can effectively capture the complex relationship and topology structure between nodes in the graph.

[0018] The core idea is: for each node in the graph structure, each node updates its own feature representation by aggregating the features of other neighbor nodes. After multiple iterations, the representation of each node will contain the feature information of all nodes in its local neighborhood, so as to enrich its own feature representation. Finally, based on the node representation, classification, prediction and other tasks can be realized.

[0019] 2. Graph Attention Network (GAT): GAT is a graph neural network architecture that integrates attention mechanism. Its core is that each node can learn to assign different importance weights to different neighbors when aggregating neighbor information through attention mechanism, so that the graph neural network can learn to selectively pay attention and can improve the processing ability of the model for complex and heterogeneous graph data.

[0020] 3. Membership degree: is a mathematical tool used to represent fuzzy sets, which can describe the membership relationship of an element to a fuzzy set. Generally, the membership degree of an element can be [0, 1]. Different values can represent the true degree of an element belonging to a fuzzy set.

[0021] For example, fat is a fuzzy set, and the membership degree of a person weighing 80 kg belonging to fat can be 0.9, and the membership degree of a person weighing 70 kg belonging to fat can be 0.8.

[0022] 4. Differentiable neural dictionary network (DND): The differentiable neural dictionary network refers to a class of models that deeply integrate dictionary learning mechanisms and differentiable neural network architectures. The core feature is to construct a structured and learnable key-value memory structure (i.e., a dictionary) inside the network. Both the dictionary itself and the encoding process are designed to be differentiable, so they can be jointly optimized with the task objective through the backpropagation algorithm.

[0023] At present, the handling of the driver's mispressing the accelerator in the vehicle mainly focuses on the following two ways.

[0024] One way in the related art is to collect the driving speed, the accelerator pedal depression degree and the accelerator pedal opening degree of the vehicle in real time during vehicle driving. Whether the accelerator mispressing behavior occurs is determined by judging the driving speed, the accelerator pedal depression degree and the accelerator pedal opening degree, such as determining that the accelerator mispressing occurs when the driving speed is small, but the accelerator pedal degree is large and the accelerator pedal opening degree is large. In this case, driving intervention is performed by stopping fuel supply / power supply to suppress the vehicle power response, thereby ensuring driving safety.

[0025] Another way in the related art is to continuously monitor the accelerator pedal state, the driver state and identify whether the vehicle has a collision risk during vehicle driving. When it is monitored that the vehicle has a collision risk and the driver is not focused, if the accelerator pedal is depressed too deeply, driving intervention is performed by suppressing the vehicle power response, thereby ensuring driving safety.

[0026] However, in the above two ways, whether the accelerator mispressing behavior occurs is determined by judging whether the vehicle state and the driving state meet the corresponding conditions, and the subsequent driving intervention operation is also a simple power response suppression. This simple power response suppression scheme may cause unintended accidents (such as rear-end accidents) to occur, thereby making the intervention operation ineffective. In some cases, these unintended phenomena may even cause psychological panic in the driver, thereby adversely affecting driving safety.

[0027] Therefore, the embodiments of the present application provide a torque adjustment method, which can be applied to the scene of vehicle driving. This scheme can be applied throughout the entire driving process of the vehicle, especially in the scene where driving intervention is needed when the driving intention and the environmental demand do not match. For example, Figure 1is a scene schematic diagram of a torque adjustment method provided by an embodiment of the present application, which includes a first vehicle 101 and a second vehicle 102. The first vehicle 101 travels on the road at a travel speed of 100 km / h (kilometers per hour), and in the process of traveling of the first vehicle 101, the second vehicle 102 suddenly merges into the front of the first vehicle 101. In this scenario, the torque adjustment method can be applied to first judge whether the driving intention of the first vehicle 101 matches the environmental demand, and when the two do not match, the driving intervention can be further performed on the first vehicle 101.

[0028] Specifically, by first determining a target deviation score based on driving operation information, vehicle state information, and environment perception information of the vehicle, the deviation degree between the driving intention of the driver and the operation demand in the current environment is identified. When the deviation between the driving intention of the driver and the operation demand in the current environment is large, it means that the current driving operation of the driver is risky, and then driving intervention needs to be performed in time. When driving intervention is performed, an expected output torque is determined based on the driving operation information, the environment perception information, and the deviation degree. Since the target deviation score represents the risk degree of the current operation of the driver, the risk degree of the driving operation is different, and the corresponding driving scene also has differences, so the degree of driving intervention is also different, and the output torque also has differences. By combining the target deviation score in the process of determining the expected output torque, the target deviation score can affect the intervention degree, so that an expected output torque that matches the current environment can be accurately determined based on the driving operation information, the environment perception information, and the target deviation score, that is, the torque that the vehicle should output in the current environment is accurately determined. Finally, the original output torque of the vehicle is adjusted in combination with the torque that the vehicle should output in the current environment, which can realize precise intervention on the driving operation. Compared with the scheme of directly cutting off the power response for driving intervention in the prior art, the occurrence of unintended accidents can be avoided, and the driving intervention effect is improved.

[0029] The torque adjustment method provided by the embodiment of the present application is explained and described in detail below.

[0030] Figure 2 is a flowchart of a torque adjustment method provided by an embodiment of the present application. The method can be applied to a domain controller of a vehicle. Referring to Figure 2 , the method includes the following steps 201-203.

[0031] Step 201, determining a target deviation score based on driving operation information, vehicle state information, and environment perception information, the target deviation score being used to represent the deviation degree between the driving intention of the driver and the operation demand in the current environment.

[0032] The driving operation information is used to represent the driving operation of the driver on the vehicle, mainly embodied in the control on the power output. In some embodiments, the driving operation information can include the pedal opening degree of the accelerator pedal and the pedal opening degree change rate.

[0033] The vehicle state information is used to represent the driving state of the vehicle, which can include the current output torque, the current vehicle speed, the gear, the engine speed / drive motor speed, etc.

[0034] The environment perception information is used to represent the road conditions of the road where the vehicle is currently driving, the traffic rules of the road section, etc. For example, the environment perception information can include the relative distance and relative speed of the vehicle and the preceding vehicle, the obstacle distribution on the current road, the slope of the current road, the speed limit information and traffic signs, etc.

[0035] It should be understood that the above driving operation information, vehicle state information and environment perception information are dimensional parameter information at the current time. In the embodiments of the present application, during the driving of the vehicle, the driving operation information, vehicle state information and environment perception information can be collected in real time from the vehicle controller and various sensors through the CAN (Controller Area Network) bus, high-speed Ethernet, TSN (Time-Sensitive Networking) interface, etc. After collecting the above data, the collected data can be preprocessed to obtain the driving operation information, vehicle state information and environment perception information at the current time. The preprocessing includes but is not limited to operations such as outlier rejection, timestamp alignment and value normalization, etc.

[0036] In the above manner, by obtaining the driving operation information, vehicle state information and environment perception information of the environment where the vehicle is currently located, it can be known what kind of vehicle state change will be caused by the current driving operation of the driver in the current environment, and it can be known what the operation demand of the vehicle in the current environment should be. For example, if the current vehicle speed is fast and there is an obstacle in front, the driver should perform a deceleration operation, i.e., the vehicle should avoid at low speed. By the driving operation information, it can be known what state the vehicle will present under the current driving operation of the driver, so that the deviation between the current operation of the driver and the operation demand in the current environment can be known according to the above parameters, and then it can be known whether the current operation of the driver matches the operation demand in the current environment, which provides an effective basis for subsequent driving intervention.

[0037] In a possible manner, the operation of step 201 can be: performing feature encoding on the driving operation information, the vehicle state information, and the environment perception information to obtain a plurality of variable features; taking the plurality of variable features as node features of a plurality of graph nodes respectively, and connecting the plurality of graph nodes based on a target connection relationship to obtain a target topological graph; and inputting the target topological graph into a deviation detection model, and outputting a target deviation score by the deviation detection model.

[0038] The plurality of variable features include features of each parameter in the driving operation information, features of each parameter in the vehicle state information, and features of each parameter in the environment perception information, so that after the plurality of variable features are taken as node features of the plurality of graph nodes, one graph node corresponds to one parameter, and the plurality of graph nodes can include three types of graph nodes: driving operation nodes, vehicle state nodes, and environment nodes.

[0039] The target connection relationship is used to represent an association relationship between the driving operation information, the vehicle state information, and the environment perception information. In the embodiments of the present application, the association relationship can be determined based on prior constraint knowledge, which refers to physical laws, traffic regulations, and safe driving common sense in the driving field, and which specifies the relationship direction and influence strength between specific nodes. For example, the prior constraint knowledge can include driving common sense such as "the front vehicle distance should be reduced when the front vehicle distance is too close" and "acceleration should be prohibited under the red light state", and the association relationship between "the front vehicle distance" and "the vehicle speed" and the association relationship between "the traffic light state" and "the vehicle speed" are covered.

[0040] The target topological graph can include a plurality of graph nodes and a plurality of edges. The plurality of graph nodes correspond to the plurality of variable features one by one, and the node feature of any one of the plurality of graph nodes is the corresponding variable feature, so that the plurality of graph nodes refer to the driving operation information, the vehicle state information, and the environment perception information. The plurality of edges are edges generated after the plurality of graph nodes are connected based on the target connection relationship. Each edge is used to represent an association relationship between two adjacent graph nodes.

[0041] In the construction process of the target topological graph, parameters with an association relationship in the prior constraint knowledge can be connected. For example, there is an association relationship between "the front vehicle distance" and "the vehicle speed", so there is a connection relationship between "the front vehicle distance" and "the vehicle speed" in the target connection relationship, and the graph nodes corresponding to "the front vehicle distance" and "the vehicle speed" can be connected. This operation is performed on all two parameters with an association relationship in the driving operation information, the vehicle state information, and the environment perception information, and the target topological graph can be obtained.

[0042] In the above manner, by converting the relationship among the driving operation information, the vehicle state information, and the environment perception information into a graph structure, and writing the driving operation information, the vehicle state information, and the environment perception information as node features of graph nodes into the graph structure, a target topology graph is obtained, so that the driving operation information, the vehicle state information, and the environment perception information are converted into a graph structure with prior constraints. Subsequently, the target deviation score is output based on the target topology graph, so that the conflict between the driving intention and the environmental constraint is converted from the traditional numerical threshold judgment to a structured relationship reasoning, so that the risk consequences of the current driving operation and the operation demand under the environmental constraint can be accurately reasoned. Subsequently, it is equivalent to comparing the vector distance between the actual consequences and the expected consequences. In this way, the abstract driving intention-environment demand deviation is converted into a calculable geometric deviation, which is a reasoning process consistent with physical intuition. Therefore, the target deviation score can be accurately reasoned by the deviation detection model.

[0043] It is worth noting that the above deviation detection model can be a neural network model capable of processing graph structures, such as a graph neural network model, specifically a graph attention neural network model. In this case, when the deviation detection model aggregates the graph node features in the target topology graph, the deviation detection model can quantitatively calculate the importance of each edge in the target topology graph, and further reason to obtain a more accurate target deviation score.

[0044] For example, Figure 3 is a model structure diagram of a deviation detection model provided by an embodiment of the present application.

[0045] As Figure 3 shown, in some embodiments, the deviation detection model 300 can include an input layer 301, an attention module 302, a path aggregation module 303, a deviation calculation module 304, and an output layer 305. The input layer 301 is configured to receive input data (a target topology graph). The attention module 302 is configured to quantitatively calculate the importance of each edge in the target topology graph. The path aggregation module 303 is configured to obtain a response chain corresponding to a driving operation intention and a response chain under a current environmental constraint from the target topology graph. The deviation calculation module 304 is configured to calculate a target deviation score.

[0046] In this case, after the target topology graph is input into the deviation detection model, the following steps (1)-(5) are performed by the deviation detection model to output the target deviation score.

[0047] (1) For any one of the multiple edges of the target topology graph, the attention module determines the edge weight between the two adjacent graph nodes on this edge based on the node features and the prior constraint features of the two adjacent graph nodes on this edge.

[0048] The prior constraint feature is a feature vector obtained by encoding prior constraint knowledge, which can be encoded into the target topology graph in an explicit logical relationship to affect the calculation of the edge weight between adjacent graph nodes, thereby guiding the target deviation score calculation of the deviation detection model.

[0049] The edge weight is used to represent the influence strength between two adjacent graph nodes. The greater the edge weight, the greater the influence strength of one graph node on another graph node. The smaller the edge weight, the smaller the influence strength of one graph node on another graph node. For example, when the distance to the front vehicle is close, the weight of the edge “distance to front vehicle → vehicle speed” will be relatively high, which means that the distance to the front vehicle has a greater influence on the vehicle speed. When the distance to the front vehicle is far, the weight of the edge “distance to front vehicle → vehicle speed” will be relatively low, which means that the distance to the front vehicle has a smaller influence on the vehicle speed.

[0050] In the above manner, the edge weight is dynamically calculated through the attention mechanism, so that the deviation detection model can adaptively adjust the relationship strength between variables according to the real-time scene, so that the same pair of graph nodes can have different semantic connections when the node features are different (different driving situations), and thus a more accurate target deviation score can be inferred.

[0051] In one possible manner, the first edge weight of the edge is determined based on the node features of the two adjacent graph nodes on the edge through the attention module; the second edge weight of the edge is determined based on the first edge weight and the prior constraint feature on the edge through the attention module; and the first edge weight and the second edge weight are fused to obtain the edge weight of the edge.

[0052] The first edge weight is used to represent the actual influence strength between two adjacent graph nodes, and the second edge weight is used to represent the expected influence strength between two adjacent graph nodes in accordance with the prior constraint knowledge. For example, for an edge with a physical causal relationship in the target topology graph (the edge between the throttle and the torque), the expected weight can be close to 1, and for an edge with a logical constraint (such as the distance to the front vehicle → slow down), the edge weight can be the first edge weight. For a forbidden edge (such as red light → accelerate), the edge weight can be set to 0.

[0053] In the above manner, the first edge weight and the second edge weight are calculated, and the first edge weight and the second edge weight are fused to calculate the edge weight of the edge. In this process, the prior constraint feature is introduced to calculate the edge weight, so that the degree to which the current operation of the driver deviates from the reasonable rules can be quantified, thereby the influence relationship between the graph nodes in accordance with the physical rules and knowledge can be calculated, and then the accurate edge weight can be calculated.

[0054] In some embodiments, the attention module can be implemented by using a multi-head attention mechanism, in which case the attention module can include a plurality of attention units. In this case, the operation of determining the first edge weight of the edge based on the node features of the two adjacent graph nodes on the edge can be as follows: for any one of the attention units in the multi-head attention mechanism, the node features of the two adjacent graph nodes on the edge are multiplied by a transformation matrix through the attention unit to obtain two transformed features; the two transformed features are spliced to obtain a spliced feature, and the spliced feature is dot-multiplied with an attention vector, and the dot product result is sent to an activation layer to obtain an initial attention coefficient; the initial attention coefficient is normalized to obtain an attention weight calculated by the attention unit; and the attention weights calculated by the plurality of attention units are aggregated to obtain the first edge weight.

[0055] The transformation matrix and the attention vector can be set in advance. In the above operation, the activation layer can be implemented by using a LeakyReLU activation function, and other activation functions can also be used, which are not limited in the embodiments of the present application.

[0056] The operation of aggregating the attention weights calculated by the plurality of attention units to obtain the first edge weight includes, but is not limited to, calculating the average value, weighted sum, etc. of the attention weights calculated by the plurality of attention units.

[0057] In the above manner, the multi-head attention mechanism is used to calculate a plurality of attention weights, and then the plurality of attention weights are aggregated to calculate the first edge weight, so that the correlation between different features in different channels can be learned, and a more accurate first edge weight can be determined.

[0058] The operation of fusing the first edge weight and the second edge weight to obtain the edge weight of the edge can be implemented by using the following formula (1).

[0059] (1) wherein, is the edge weight between the two adjacent graph nodes (graph node i and graph node j) of the edge finally determined, is the first edge weight between the two adjacent graph nodes (graph node i and graph node j) of the edge, is the second edge weight between the two adjacent graph nodes (graph node i and graph node j) of the edge, is the influence weight of the second edge weight, which can be set in advance, is a constant, which can be set in advance.

[0060] (2) The path aggregation module takes the graph node corresponding to the driving operation information in the multiple graph nodes as a first starting node, and obtains graph nodes traversed to reach a target node from the first starting node, to construct a driving intention path.

[0061] The driving intention path is used to represent an actual response result caused by the current driving operation under the current environment, for example, the vehicle speed is too fast, an obstacle appears, and the current driving operation is deep pressing of the accelerator. The driving intention path can indicate the actual response state of the vehicle after deep pressing of the accelerator.

[0062] The first starting node is used to represent a starting point of the driving intention path. The target node is used to represent an expected response of the environment to the driving operation. In the embodiment of the present application, the target node can be set in advance, for example, the target node can be set as the vehicle speed, that is, whether the current driving operation is consistent with the expected response under the environment constraint is measured by the vehicle speed.

[0063] Since the actual response result caused by the current driving operation is inferred from the target topology graph in this step, in the above step (2), the graph node corresponding to the driving operation information (that is, the driving operation node) can be taken as the starting point of the driving intention path, and all graph nodes and edges traversed from the first starting node to the target node are obtained by continuously traversing according to the connection relationship between the graph nodes, which can constitute the driving intention path.

[0064] (3) The path aggregation module takes the graph node corresponding to the environment perception information in the multiple graph nodes as a second starting node, and obtains graph nodes traversed to reach a target node from the second starting node, to construct an environment constraint path.

[0065] The environment constraint path is used to represent a theoretical response result that the vehicle should present under the current environment, in other words, the environment constraint path is used to indicate what response the driver is expected to make or what state (such as acceleration or deceleration) the vehicle is expected to present under the current environment after logical constraint.

[0066] The second starting node is used to represent a starting point of the environment constraint path.

[0067] Since what the vehicle should respond to under the current environment is inferred from the target topology in this step, in the above step (3), the graph node corresponding to the environment perception information (that is, the environment perception information node) can be taken as the starting point of the environment constraint path, and all graph nodes and edges traversed from the second starting node to the target node are obtained by continuously traversing according to the connection relationship between the nodes, which can constitute the environment constraint path.

[0068] (4) The bias calculation module determines the feature of the driving intention path based on the node features of the multiple graph nodes traversed by the driving intention path and the edge weights between adjacent graph nodes, and determines the feature of the environmental constraint path based on the node features of the multiple graph nodes traversed by the environmental constraint path and the edge weights between adjacent graph nodes.

[0069] In one possible manner, the edge weights of the edges included in the driving intention path are multiplied to obtain the path weight of the driving intention path, and the edge weights of the edges included in the environmental constraint path are multiplied to obtain the path weight of the environmental constraint path; the node features of the multiple graph nodes traversed by the driving intention path are aggregated to obtain first fusion features, and the first fusion features are multiplied by the path weight of the driving intention path to obtain the feature of the driving intention path; and the node features of the multiple graph nodes traversed by the environmental constraint path are aggregated to obtain second fusion features, and the second fusion features are multiplied by the path weight of the environmental constraint path to obtain the feature of the environmental constraint path.

[0070] (5) The bias calculation module determines the target bias score based on the difference between the feature of the driving intention path and the feature of the environmental constraint path.

[0071] For example, the least squares method can be used to calculate the difference between the feature of the driving intention path and the feature of the environmental constraint path, and then the difference is normalized to obtain the target bias score.

[0072] In another example, the Euclidean distance can be used to calculate the difference between the feature of the driving intention path and the feature of the environmental constraint path, and then the difference is normalized to obtain the target bias score.

[0073] The normalization of the difference includes but is not limited to inputting the difference into a Sigmoid function for normalization to obtain the target bias score.

[0074] In the above steps (1)-(5), the driving intention path and the environmental constraint path are inferred from the target topology graph, which is equivalent to inferring the actual response result caused by the current driving operation, such as the current driving operation causing the vehicle speed to rise, and the theoretical response result that the vehicle should present under the current environmental constraint, such as reducing the vehicle speed when the distance to the front vehicle is close. This makes it possible to express the driving intention path and the environmental constraint path in the graph structure, and then by combining the node features and edge weights of the graph nodes traversed by each path, the importance between adjacent graph nodes can be combined to evaluate the difference between the two, and a more accurate target bias score can be output.

[0075] The above step 201 can identify and quantify the deviation between the driving intention under the current operation and the operation demand under the current environment, and output the deviation degree between the driving intention and the operation demand under the current environment. After obtaining the deviation degree (target deviation score) between the driving intention and the operation demand under the current environment, it can also be determined whether the driving intention and the operation demand under the current environment match according to the target deviation score.

[0076] One possible way is that, in the case that the target deviation score is greater than a preset score threshold, it is determined that the driving intention and the operation demand under the current environment do not match; in the case that the target deviation score is less than or equal to the preset score threshold, it is determined that the driving intention and the operation demand under the current environment match.

[0077] The preset score threshold can be set according to actual needs. For example, when the intervention degree of the vehicle needs to be larger, the preset score threshold can be set lower, for example, the preset score threshold can be set to 0.2, so that the vehicle can actively intervene when the deviation between the driving intention and the operation demand under the current environment is slight. For example, in the case that the vehicle does not need to frequently intervene in the driving operation, the preset score threshold can be set higher, for example, the preset score threshold can be set to 0.5, so that the vehicle actively intervenes only when the deviation between the driving intention and the operation demand under the current environment is large enough to cause safety problems, thereby ensuring the safety of vehicle driving.

[0078] It should be understood that, when it is determined that the driving intention and the operation demand under the current environment match, subsequent driving intervention operations can not be performed, and when it is determined that the driving intention and the operation demand under the current environment do not match, subsequent steps 202-203 can be executed to perform driving intervention operations.

[0079] Step 202, in the case that the target deviation score is greater than the preset score threshold, determining an expected output torque of the vehicle under the current environment based on the driving operation information, the environment perception information and the target deviation score.

[0080] The expected output torque refers to a torque value that matches the current driving scene and meets the constraints of the current environment.

[0081] Since the target deviation score represents the risk degree of the current operation of the driver, the risk degree of the driving operation is different, and the driving scene represented by the target deviation score also differs, and then the degree of driving intervention also differs, so that the output torque also differs. In the above step 202, the target deviation score is combined in the process of determining the expected output torque, so that the target deviation score can affect the intervention degree, and thus an expected output torque that matches the current environment can be accurately determined based on the driving operation information, the environment perception information and the target deviation score, that is, the torque that the vehicle should output under the current environment can be accurately determined.

[0082] In one possible implementation, the operation of step 202 can be: inputting the driving operation information, the environment perception information, and the target deviation score into a desired torque matching model, and determining the desired output torque based on historical behavior data through the desired torque matching model.

[0083] The historical behavior data includes torque values that the vehicle should output in different driving scenarios in the history. This includes torque output values when the vehicle is successfully intervened in different driving scenarios, and torque output values based on theoretical calculations in different driving scenarios. In some embodiments, different driving scenarios can be understood as driving scenarios referred to in different environment perception information, different driving operations, and different deviation scores. Therefore, the historical behavior data includes torque values that the vehicle should output in different environment perception information, different driving operations, and different deviation scores.

[0084] To improve the accuracy of the driving scenario description, vehicle state information can also be added to the description of the driving scenario, that is, the corresponding driving scenario is described by the environment perception information, the driving operation information, the deviation score, and the vehicle state information. In this case, the historical behavior data includes torque values that the vehicle should output in different environment perception information, different driving operations, different deviation scores, and different vehicle states.

[0085] In the embodiments of the present application, the historical behavior data can be stored in the form of a dictionary, in other words, the historical behavior data can be composed of a plurality of key-value pairs, the plurality of key-value pairs include a plurality of key vectors and a plurality of value vectors, the plurality of key vectors and the plurality of value vectors correspond one by one, and one key vector and one value vector constitute one key-value pair. The plurality of key vectors are used to represent a plurality of driving scenarios, which can be the vehicle state, the driving operation, the environment perception information, and the comprehensive indication of the deviation between the driving intention and the environmental constraint when the vehicle is successfully intervened in the history. Each key vector can be an embedding vector obtained by embedding the vehicle state, the driving operation, the environment perception information, and the deviation between the driving intention and the environmental constraint when the vehicle is successfully intervened in the history. The embedding vector represents a driving scenario. The plurality of value vectors are used to represent the torque output value when the vehicle is successfully intervened in each driving scenario, in other words, a value vector refers to a torque value matching the corresponding driving scenario.

[0086] In the above manner, by encoding a historical behavior data in the desired torque matching model, the prediction task of the desired torque matching model for the desired output torque is converted into a retrieval task. By combining the historical behavior data, more reasonable and accurate desired output torque can be determined by combining the intervention data in the historical driving scenario.

[0087] In some embodiments, the desired torque matching model can be a differentiable neural dictionary network (DNN). An exemplary, Figure 4 is a model structure diagram of a desired torque matching model provided by an embodiment of the present application, as Figure 4 shown, the desired torque matching model 400 can include an input layer 401, an encoder 402, a differentiable memory dictionary 403, a query module 404, and an output layer 405, wherein the input layer 401 is used to receive input data (driving operation information, vehicle state information, environment perception information, target deviation score), the encoder 402 is used to encode the features of the input data. The differentiable memory dictionary 403 maintains a historical behavior data composed of a plurality of key-value pairs. The query module 404 is used to infer the expected output torque from the historical behavior data. In some embodiments, the desired torque matching model 400 can also include a write control module 406, through which the differentiable memory dictionary can maintain the torque output value of the successful intervention in the new driving scene to keep the historical behavior data updated continuously.

[0088] In this case, the driving operation information, the environment perception information and the target deviation score are input into the desired torque matching model, and the operation of determining the expected output torque based on the historical behavior data through the desired torque matching model can be: the driving operation information, the environment perception information and the target deviation score are encoded by the encoder to obtain the key vector of the current driving scene; the similarity between the key vector of the current driving scene and the plurality of key vectors is calculated by the query module, and based on the similarity, the weights of the plurality of value vectors are determined; the plurality of value vectors are weighted and summed by the query module based on the weights of the plurality of value vectors to obtain the expected output torque.

[0089] Since the driving scene is almost never exactly the same as a certain driving scene in history, there will always be some subtle differences, for example, the environment in which the vehicle is located is rainy, and the following scene, but the following distance in history is 13 meters, and the following distance in the current driving scene is 15 meters, the deviation score corresponding to the driving scene in history is 0.72, and the target deviation score corresponding to the current driving scene is 0.68. If the torque output value corresponding to a driving scene in the historical behavior data is accurately matched, the key vector of the current driving scene needs to be exactly the same as a certain key vector in the historical behavior data, but in actual situation, it is almost impossible to match the exact same key.

[0090] In the above manner, the similarity between the current driving scene and the plurality of key vectors in the historical behavior data is calculated first, and then the similarity is used to assign a corresponding weight to each value vector. The value vector corresponding to the key vector with high similarity can obtain a higher weight, and the value vector corresponding to the key vector with low similarity can obtain a lower weight. Finally, the plurality of value vectors in the historical behavior data are weighted and summed to obtain the expected output torque by comprehensively considering all the value vectors, so that a more reasonable expected output torque can be output.

[0091] The encoder of the expected torque matching model can be implemented by a multi-layer perceptron (MLP). Since the core idea of the MLP is to use multiple layers of simple nonlinear transformations to extract features layer by layer to fit the complex relationship between data. Therefore, by inputting the driving operation information, the environment perception information, and the target deviation score into the multi-layer perceptron, the multi-layer perceptron can fully capture the complex relationship between the driving operation information, the environment perception information, and the target deviation score to obtain a feature vector of the driving operation information, the environment perception information, and the target deviation score as the key vector of the current driving scene.

[0092] Since the driving scene in the historical behavior data can be described by the environment perception information, the driving operation information, the deviation score, and the vehicle state information, in some embodiments, the encoder can also be used to encode the driving operation information, the vehicle state information, the environment perception information, and the target deviation score to obtain the key vector of the current driving scene. In this way, the key vector of the current driving scene is calculated by the driving operation information, the vehicle state information, the environment perception information, and the target deviation score, which can more accurately represent the current driving scene, so that more accurate similarity calculation of the key vector can be achieved in the historical behavior data, and a more accurate expected output torque can be obtained.

[0093] In one possible manner, based on the similarity, the operation of determining the weights of the plurality of value vectors can be: normalizing the similarity between the key vector of the current driving scene and a key vector to obtain the weight of the value vector corresponding to the key vector. In this way, the determination of the weights of the plurality of value vectors is more convenient.

[0094] In some embodiments, the expected torque matching model can output the confidence of the expected output torque while outputting the expected output torque. The confidence is used to represent the credibility of the expected output torque.

[0095] In one possible manner, the expected torque matching model can obtain the maximum weight from the weights of the plurality of value vectors; and multiply the maximum weight by a consistency factor to obtain the confidence of the expected output torque.

[0096] It should be understood that the maximum weight is used to reflect the matching degree of the current driving scene and the most similar driving scene in the historical behavior data, and the greater the maximum weight, the more similar and matched the current driving scene is to a certain driving scene in the historical behavior data. The consistency factor is used to represent the concentration of the weight distribution of the plurality of value vectors, and it should be understood that the more concentrated the weight distribution, the higher the consistency factor, and the more dispersed the weight distribution, the lower the consistency factor.

[0097] In the above steps, when the maximum weight is close to 1 and the weight distribution is relatively concentrated, the confidence of the expected output torque is higher, which indicates that there is more reliable historical experience in the historical behavior data, and the expected output torque is reliable. On the contrary, when the maximum weight is low (such as close to 0) or the plurality of weight distributions are relatively dispersed, the confidence of the expected output torque is lower, which can indicate that the historical behavior data lacks similar historical experience to the current driving scene.

[0098] In the above manner, by multiplying the maximum weight by the consistency factor, it is equivalent to comprehensively determining the confidence degree of the expected output torque by the matching degree of the current driving scene and the most similar driving scene in the historical behavior data and the weight distribution of the plurality of value vectors. Thus, a more accurate confidence of the expected output torque can be determined.

[0099] It should be noted that the expected output torque is a torque output value matched with the current driving scene, that is, under the current driving scene, driving intervention by the vehicle outputting the expected output torque can avoid unsafe events caused by driver operation errors. However, if the vehicle is directly controlled to output the expected output torque, there may be a sudden torque drop or a sudden torque rise that causes a driving discomfort. Therefore, in the embodiments of the present application, the target deviation score and the expected output torque can be combined to determine an output torque that matches the current driving operation and can avoid sudden torque drop or sudden torque rise.

[0100] In step 203, the vehicle is controlled to perform torque adjustment based on the expected output torque.

[0101] Since the expected output torque is the torque that should be output under the current environmental constraints, adjusting the original output torque of the vehicle in combination with the expected output torque can achieve effective intervention on the driving operation. Compared with the prior art scheme of directly cutting off the power response for driving intervention, the occurrence of unexpected accidents can be avoided, and the driving intervention effect is improved.

[0102] One possible way is to first obtain the confidence of the expected output torque before controlling the vehicle to perform torque adjustment based on the expected output torque. Then, in the case that the confidence of the expected output torque is greater than or equal to a preset confidence threshold, the vehicle is controlled to perform torque adjustment based on the expected output torque.

[0103] The confidence of the expected output torque is used to represent the degree of reliability of the expected output torque. The greater the confidence, the more reliable the expected output torque is. The smaller the confidence, the less reliable the expected output torque is.

[0104] Thus, in the above manner, when the confidence of the expected output torque is greater than or equal to the preset confidence threshold, it indicates that the degree of reliability of the expected output torque is high, that is, the current driving scene is more similar to a driving scene in the historical behavior data, and then the obtained expected output torque matches the current driving scene, so that the vehicle is controlled for torque adjustment based on the expected output torque, and the original output torque of the vehicle can also be corrected.

[0105] In another possible implementation, when the confidence of the expected output torque is less than the preset confidence threshold, the target deviation score and the driving operation information are obtained. Thus, the operation of step 203 can be: controlling the vehicle for torque adjustment based on the target deviation score, the driving operation information, and the expected output torque.

[0106] When the confidence of the expected output torque is less than the preset confidence threshold, it indicates that the degree of reliability of the expected output torque is low, that is, there is no driving scene in the historical behavior data that is similar to the current driving scene, and then the obtained expected output torque does not necessarily match the current driving scene. In this case, the vehicle cannot be completely controlled for torque adjustment based on the expected output torque.

[0107] In the above manner, since the target deviation score represents the degree of risk of the current driving operation of the driver, the driving operation information can indicate the original output torque. By obtaining the target deviation score and the driving operation information, and based on the target deviation score, the driving operation information, and the expected output torque, an output torque that is more matched to the current driving scene can be determined, so that precise intervention on the driving operation can be realized.

[0108] In one possible manner, the operation of controlling the vehicle for torque adjustment based on the target deviation score, the driving operation information, and the expected output torque can be implemented through the following steps (1) to (4).

[0109] (1) Fuzzing the target deviation score to determine a plurality of membership degrees.

[0110] The plurality of membership degrees are used to represent the likelihood of the target deviation score belonging to a plurality of deviation levels. In the embodiments of the present application, a plurality of deviation levels (such as "slight", "moderate", and "severe") are provided, which are used to indicate the deviation level between the driving intention of the driver and the environmental demand. It should be understood that the deviation level of "slight" means that the deviation between the driving intention of the driver and the environmental demand is small, the deviation level of "moderate" means that the deviation between the driving intention of the driver and the environmental demand is neither small nor large, and the deviation level of "severe" means that the deviation between the driving intention of the driver and the environmental demand is large. The plurality of deviation levels correspond to a plurality of fuzzy intervals, for example, the fuzzy interval corresponding to the "slight" deviation level is [0.2, 0.3), the fuzzy interval corresponding to the "moderate" deviation level is [0.3, 0.6), and the fuzzy interval corresponding to the "severe" deviation level is [0.6, 1].

[0111] Specifically, the operation of fuzzifying the target deviation score to determine a plurality of membership degrees can be: for any one of the plurality of deviation levels, the membership degree corresponding to the deviation level is determined based on the target deviation score and the fuzzy interval corresponding to the deviation level by the following formula (2).

[0112] (2) wherein, is the membership degree corresponding to a deviation level, represents the likelihood of the target deviation score belonging to a deviation level, x is the target deviation score, a is the minimum deviation score of the fuzzy interval corresponding to a deviation level, b is the median of the fuzzy interval corresponding to a deviation level, which is also the point at which the membership degree reaches the maximum value 1, and c is the maximum deviation score of the fuzzy interval corresponding to a deviation level.

[0113] It should be understood that the calculation of the above formula (2) is performed for each deviation level, and then the likelihood of the target deviation score belonging to the plurality of deviation levels can be obtained.

[0114] (2) input the driving operation information into each of the plurality of torque determination models, and output the candidate torque when driving intervention is performed through the plurality of torque determination models.

[0115] In some embodiments, a plurality of torque determination models can also be provided, which correspond one-to-one to a plurality of deviation levels, and different torque determination models have the ability of nonlinear mapping between driving operation input and output torque under different deviation levels, for example, for the "moderate" deviation level, the torque determination model corresponding to the "moderate" deviation level has the ability to map the driving operation input to a reasonable output torque when the deviation between the driving intention and the operation demand under the current environment is small.

[0116] Each candidate torque is substantially a torque that the vehicle should output under a different deviation level, i.e., a torque that can eliminate the driving deviation under the different deviation level.

[0117] In some embodiments, one torque determination model can be implemented by a Gaussian process model. Since the Gaussian process is a non-parametric Bayesian regression method that can predict an output value while giving a prediction reliability, it fits a non-linear relationship by a kernel function. Thus, each torque determination model can be set with a kernel function to fit a non-linear mapping between the driving operation input and the output torque under different deviation levels. After the driving operation information is input into the plurality of torque determination models, the plurality of torque determination models can output the torque for driving intervention under different deviation levels. In addition, since the Gaussian process can give a prediction uncertainty of the predicted output value, each torque determination model can output the candidate torque and the prediction uncertainty of the candidate torque.

[0118] (3) Based on the plurality of membership degrees, the candidate torques output by the plurality of torque determination models are weighted and summed to obtain a reference torque.

[0119] The reference torque refers to a torque that the vehicle should output under the current driving operation based on the target deviation score, i.e., an output torque that enables the vehicle to meet the environmental constraints under the current deviation level.

[0120] In step (3), the membership degree corresponding to one deviation level is taken as the weight of the torque determination model corresponding to the deviation level, and then the candidate torques output by the plurality of torque determination models are weighted and summed to obtain the reference torque.

[0121] In the above manner, the higher the membership degree corresponding to one deviation level, the greater the possibility that the target deviation model belongs to the deviation level, and thus the greater the weight of the torque determination model corresponding to the deviation level. Therefore, the adoption proportion of the candidate torque output by the torque determination model in the determination of the output torque can be greater, and thus a more reasonable and accurate reference torque can be determined subsequently.

[0122] In some embodiments, the confidence of the reference torque can be determined at the same time as the reference torque, so that the reliability of the reference torque can be known.

[0123] One possible manner is to obtain the prediction uncertainty of the plurality of candidate torques; for any one of the plurality of candidate torques, a preset value is subtracted from the prediction uncertainty of the candidate torque to obtain the confidence of the candidate torque; and the confidence of the plurality of candidate torques is weighted and summed based on the plurality of membership degrees to obtain the confidence of the reference torque.

[0124] The preset value can be set in advance, for example, the preset value can be set as 1. In this case, the prediction uncertainty of the candidate torque is subtracted from the preset value, and the prediction certainty of the candidate torque, that is, the confidence level of the candidate torque, can be obtained. Then, the confidence levels of different candidate torques are weighted according to the adoption ratios of the candidate torques. Compared with directly calculating the average of the confidence levels of multiple candidate torques, the confidence level of the final determined reference torque is more reasonable, and thus a more accurate confidence level of the reference torque can be obtained.

[0125] In some embodiments, in a case where the confidence level of the reference torque is greater than the first confidence threshold and the target deviation score is greater than the first score threshold, the vehicle can be controlled to output the reference torque.

[0126] Since each candidate torque is essentially an output torque that can eliminate driving deviation at different deviation levels, when the confidence level of the reference torque is high, it indicates that the output torque determined based on each candidate torque has high reliability. When the target deviation score is large, it indicates that the deviation between the driving intention and the environmental demand is very large, which can cause unsafe phenomena. In this case, the vehicle can be controlled to output the reference torque to quickly avoid the occurrence of unsafe phenomena.

[0127] In some other embodiments, in a case where the confidence level of the reference torque is less than or equal to the first confidence threshold, the reference torque and the expected output torque can be fused to obtain a final output torque.

[0128] (4) Controlling the vehicle to perform torque adjustment based on the reference torque and the expected output torque.

[0129] In the above steps (1)-(4), the target deviation score is first fuzzified to determine multiple membership degrees. Then, the output torque when driving intervention is performed at different deviation levels is predicted by using multiple preset torque determination models. Then, the output torque when driving intervention is performed at the current driving operation is determined based on the candidate torques output by the torque determination models. Finally, the torque and the expected output torque are fused to obtain a more reasonable and accurate output torque, so that driving intervention can be performed more smoothly and stably.

[0130] One possible way is to take the average of the reference torque and the expected output torque as a target torque, and control the vehicle to perform torque adjustment based on the target torque.

[0131] In another possible manner, the target deviation score and the confidence of the expected output torque can be obtained first; then, based on the target deviation score and the confidence of the expected output torque, a fusion ratio of the reference torque and the expected output torque is determined; the reference torque and the expected output torque are fused according to the fusion ratio to obtain a target torque, and the vehicle is controlled based on the target torque for torque adjustment.

[0132] It should be understood that, since the expected output torque is a torque output value of a similar driving scene determined from historical behavior data, the torque output value of the historical driving scene should have a relatively high adoption ratio. However, if the confidence of the expected output torque is not high, it indicates that the reliability of the expected output torque is not high, and there is no particularly similar driving scene in the history, so the value of the expected output torque is not matched with the current driving scene. Therefore, in this case, the expected output torque cannot be completely relied on.

[0133] In the above manner, by determining the fusion ratio of the reference torque and the expected output torque based on the target deviation score and the confidence of the expected output torque, a more reasonable fusion ratio of the reference torque and the expected output torque can be determined. Subsequently, the reference torque and the expected output torque are fused according to the fusion ratio to obtain a more reasonable target torque, so that subsequent torque adjustment of the vehicle based on the target torque can be more smooth and stable.

[0134] The operation of determining the fusion ratio of the reference torque and the expected output torque based on the target deviation score, and the confidence of the expected output torque can be: in a case where the target deviation score is greater than a first score threshold and the confidence of the expected output torque is greater than a second confidence threshold, determining the fusion ratio of the reference torque and the expected output torque as a first ratio; in a case where the target deviation score is greater than the first score threshold and the confidence of the expected output torque is less than or equal to a third confidence threshold, determining the fusion ratio of the reference torque and the expected output torque as a second ratio; in a case where the target deviation score is less than or equal to the first score threshold and greater than a second score threshold, obtaining the fusion ratio by dividing a value obtained by subtracting the confidence of the expected output torque from a preset value by the confidence of the expected output torque.

[0135] The first score threshold and the second score threshold can be set in advance, and the first score threshold is greater than the second score threshold. For example, the first score threshold can be set to 0.7, and the second score threshold can be set to 0.4.

[0136] The second confidence threshold and the third confidence threshold can be set in advance, and the second confidence threshold is greater than the third confidence threshold. For example, the second confidence threshold can be set to 0.7, and the third confidence threshold can be set to 0.3.

[0137] The first ratio and the second ratio can be set in advance, the first ratio can be set as 0:1, and the second ratio can be set as 1:0.

[0138] In the case that the target deviation score is greater than the first score threshold and the confidence of the expected output torque is greater than the second confidence, it is indicated that the deviation between the driving intention and the environmental demand is large, and driving intervention needs to be performed as soon as possible, and the confidence of the expected output torque is high, which indicates that there is a high similarity between the historical driving scene and the current driving scene, and since the expected output torque is the torque output value when the intervention is successful in the similar driving scene, the expected output torque can be completely adopted, and in this case, the fusion ratio is set as 0:1, that is, the expected output torque is directly adopted as the output torque.

[0139] In the case that the confidence of the expected output torque is less than or equal to the third confidence threshold, it is indicated that there is no intervention data similar to the current driving scene in the historical driving scene, and in this case, the expected output torque cannot be trusted, and since the deviation between the driving intention and the environmental demand is large, driving intervention needs to be performed as soon as possible, and therefore in this case, the fusion ratio can be set as 1:0, that is, the reference torque is completely adopted to avoid the occurrence of unsafe accidents as soon as possible.

[0140] In the case that the target deviation score is less than or equal to the first score threshold and greater than the second score threshold, it is indicated that the deviation between the driving intention and the environmental demand is moderate, and in this case, the torque output can be as smooth as possible to avoid abrupt driving, and therefore in this case, the fusion ratio can be determined according to the confidence of the expected output torque. For example, the confidence of the expected output torque is determined as the adoption ratio of the expected output torque, and the value obtained by subtracting the confidence of the expected output torque from 1 is determined as the adoption ratio of the reference torque, and therefore the ratio between the adoption ratio of the reference torque and the adoption ratio of the expected output torque is the fusion ratio.

[0141] The operation of controlling the vehicle based on the target torque can include: determining a transition torque based on the original output torque and the target torque; and controlling the vehicle to output the transition torque.

[0142] The transition torque is between the original output torque and the target torque.

[0143] In the above manner, in order to avoid abrupt driving, an output torque between the original output torque and the target torque is determined based on the original output torque and the target torque, so that the vehicle does not output an excessively large torque or an excessively small torque at once, but gradually transitions from the original output torque to the target torque, and smooth transition of the output torque of the vehicle can be achieved, thereby avoiding abrupt driving caused by sharp switching.

[0144] In a possible manner, the operation of determining the transition torque based on the original output torque and the target torque can be: inputting the original output torque and the target torque into a double-channel neural perturbation regulator, and determining the transition torque through the double-channel neural perturbation regulator.

[0145] For example, Figure 5 is a structural schematic diagram of a double-channel neural perturbation regulator provided by an embodiment of the present application, as Figure 5 shown, the double-channel neural perturbation regulator 500 is composed of two parallel processing channels (a master channel 501 and a reference channel 502) and a residual coupling layer 505. The master channel 501 is used to receive the target torque, for providing the expected post-intervention torque value, and the reference channel 502 receives the original output torque, for retaining the original operation intention of the driver. The residual coupling layer 503 receives the difference between the outputs of the two channels, and performs weighted calculation on the difference to obtain a fused difference value, and then superimposes the fused difference value on the original output torque, so that a smooth transition value between the original output torque and the target torque can be calculated accordingly.

[0146] For ease of understanding, the overall flow of the torque adjustment method provided by the present application will now be described. Figure 6 For ease of understanding, the overall flow of the torque adjustment method provided by the present application will now be described.

[0147] As Figure 6 shown, the driving operation information, vehicle state information, and environmental perception information are first converted into a target topology graph, and then the target topology graph is input into a deviation detection model. The deviation detection model combines prior constraint features to reason in the target topology graph, aggregates the actual response of the vehicle caused by the current driving operation, and the response state that the vehicle should reach under the current environmental constraints, and calculates the deviation degree between the current driving intention of the driver and the operation demand under the current environment based on the actual response of the vehicle caused by the current driving operation and the response state that the vehicle should reach under the current environmental constraints, to obtain a target deviation score.

[0148] However, the target deviation score, driving operation information, vehicle state information, and environmental perception information are output together into an expected torque matching model, and the expected output torque matching the current driving scene is output through the expected torque matching model.

[0149] Then, the target deviation score is fuzzified to calculate the possibility of the target deviation score belonging to multiple deviation levels, to obtain multiple membership degrees. Since multiple torque determination models are set corresponding to the multiple deviation levels, the driving operation information is input into each of the multiple torque determination models, and each torque determination model outputs a candidate torque. The multiple torque determination models output multiple candidate torques, and then the multiple candidate torques are weighted and fused based on the multiple membership degrees to obtain a reference torque.

[0150] Finally, the reference torque and the expected output torque are fused to obtain a target torque. In order to further ensure smooth transition of the vehicle output torque, the target torque and the original output torque can be input into a double-channel neural disturbance regulator, and a transition torque between the original output torque and the target torque is output by the double-channel neural disturbance regulator. Finally, the vehicle outputs the transition torque, so that the vehicle is driven to intervene, thereby avoiding unsafe phenomena caused by the mismatch between the driving intention and the environmental demand, and improving the safety of driving.

[0151] It should be noted that the deviation detection model, the expected torque matching model, and the plurality of torque determination models introduced in the embodiments of the present application can be trained before actual inference. The training methods of the above-mentioned models are introduced as follows.

[0152] (1) Training the deviation detection model.

[0153] The original training data is obtained, and a first training data set is constructed based on the original training data. The first training data set includes a plurality of first training samples.

[0154] The original training data includes driver operation samples, vehicle state samples, and environment samples. The original training data is the driving data such as driver operation information, vehicle state information, environment perception information, and output torque continuously obtained in the driving process of the vehicle in the history.

[0155] It should be understood that during the driving of the vehicle, the driving operation information, the vehicle state information, and the environment perception information can be collected in real time from the vehicle controller and various sensors through a CAN (Controller Area Network) bus, a high-speed Ethernet, a TSN (Time-Sensitive Networking) interface, etc. The collection frequency can be set according to the data type, for example, for throttle and torque data, the data can be collected at a frequency of 100 Hz (Hertz), and for image and point cloud data, the data can be collected at a frequency of 30 Hz. After the above-mentioned data is collected, the above-mentioned data can be sent to the cloud, and the driver operation information, the vehicle state information, and the environment perception information of each driving process of each vehicle are stored through the cloud. After accumulating to a certain extent, a sliding window multi-channel alignment algorithm can be used to pre-process the collected data to obtain driving operation samples, vehicle state samples, and environment samples of each time frame, forming multi-source time series data. The pre-processing includes but is not limited to operations such as outlier rejection, timestamp alignment, and value normalization. In some embodiments, in order to prevent high-frequency noise errors caused by throttle operation jitter, low-pass filtering and dynamic weighted median filtering can be used to reject outliers of the above-mentioned data.

[0156] After obtaining the original training samples, first, samples labeled as false acceleration behaviors are filtered out from the original training samples to obtain candidate training samples. Then, for the driver operation sample, the vehicle state sample and the environment sample of each time frame in the candidate training samples, the driver operation sample, the vehicle state sample and the environment sample are converted into a sample topology graph of the time frame to constitute an input data of a first training sample, and then the true deviation score of the time frame is manually labeled to constitute a sample label of the first training sample. By performing the above sample construction operation on each time frame in the candidate training samples, a plurality of first training samples can be obtained.

[0157] Then, the specific operation of training the deviation detection model can be: for any one of the plurality of first training samples, the input data of the first training sample can be input into the original graph neural network to obtain the predicted deviation score of the first training sample; based on the difference between the true deviation score and the predicted deviation score of the first training sample, the original graph neural network is trained to obtain the deviation detection model.

[0158] It should be understood that in the above training process, multiple rounds of iterative training are usually set. In the process of iterative training, the input data of a first training sample is input into the original graph neural network, the predicted deviation score of the first training sample is output through the original graph neural network, and then the model parameters of the original graph neural network are updated based on the difference between the true deviation score and the preset deviation score of the first training sample to obtain the model obtained by the first round of training. Then in the next round of training, the input data of the first training sample is continuously input into the model, and the predicted deviation score is obtained, and then the model parameters of the first round of model are updated based on the difference between the true deviation score and the preset deviation score of the first training sample in this round. According to the difference between the true deviation score and the corresponding predicted deviation score of the plurality of first training samples, the original graph neural network is trained in multiple rounds of iteration, and the model obtained after the multiple rounds of iterative training is the deviation detection model.

[0159] In some embodiments, in order to ensure that the graph structure learned by the original graph neural network is smooth and continuous, a series of graph snapshots can be sampled from continuous multiple time frames during the model training process, so that the model not only learns the edge weights of the sample topological graph corresponding to the current time frame, but also refers to the edge weights of adjacent time frames, and requires that the edge weights of the same edge corresponding to adjacent time frames cannot change too much. In addition, in order to avoid overfitting of the model, a data enhancement strategy of disturbance robustness learning can be used to add slight noise to the input data or the graph structure, so that the deviation detection model learns to tolerate slight errors of sensors or non-ideal states of the environment, so that the model can still make correct reasoning in the disturbed case, thereby improving the robustness of the deviation detection model.

[0160] (2) Training the expected torque matching model.

[0161] A second training data set is obtained, and the second training data set includes a plurality of second training samples. The input data in the plurality of second training samples includes a driver operation sample, a vehicle state sample and an environment sample of a time frame, and a true deviation score of the time frame. The second training sample further includes a sample label, and the sample label can obtain a corrected output torque collected at the time frame when the vehicle is successfully intervened from the original training data.

[0162] In some embodiments, a second training data set corresponding to different driving styles can also be constructed, and the expected torque matching model corresponding to different driving styles can be trained subsequently. That is, for the same second training data set, it includes a plurality of second training samples of the same driving style. Then in the actual reasoning process of the expected torque matching model, the corresponding expected torque matching model can also be obtained combined with the driver label (such as the driving style of the driver), and then the actual reasoning is performed according to the obtained expected torque matching model.

[0163] In this case, the operation of training the expected torque matching model can be: for any one of the plurality of second training samples, the input data of the second training sample can be input into the original differentiable neural dictionary network to obtain the predicted output torque of the second training sample; based on the difference between the corrected output torque and the predicted output torque of the second training sample, the original differentiable neural dictionary network is trained to obtain the expected torque matching model.

[0164] In some embodiments, during the process of training the expected torque matching model, a differentiable hashing mechanism can be employed for backpropagation. The differentiable hashing mechanism maps a key vector to a vector of ownership probabilities over a plurality of hash buckets through a learnable hashing network, represents the retrieval result as a probability-weighted sum of the values stored in each bucket, and enables gradients to be backpropagated from the output loss to the hashing network parameters and the encoder, thereby enabling joint optimization of the memory storage and retrieval strategy and enabling end-to-end training.

[0165] (3) Training a plurality of torque determination models.

[0166] A plurality of third training data sets are obtained, the plurality of third training data sets corresponding one-to-one to a plurality of bias levels, that is, a third training data set includes a plurality of third training samples belonging to the same bias level. The input data of a third training sample is a driver operation sample of a time frame, and the sample label of the third training sample can be set as the real output torque collected in the time frame after the vehicle is successfully intervened.

[0167] It should be understood that the training process of each of the plurality of torque determination models is similar, and the training process of each of the plurality of torque determination models will be described below by taking the training process of one torque determination model as an example.

[0168] Specifically, for any third training sample in the plurality of third training samples, the input data of the third training sample can be input into the original Gaussian process model to obtain the predicted output torque of the third training sample. Based on the difference between the real output torque and the predicted output torque of the third training sample, the original Gaussian process model is trained to obtain the torque determination model.

[0169] In addition, in the embodiments of the present application, the joint optimization process of the local / cloud can be triggered periodically during the vehicle driving gap or the cloud synchronization stage. On the one hand, a lightweight incremental learning strategy is adopted in the local optimization process to slightly adjust the model parameters of the bias detection model, the expected torque matching model and the kernel function parameters of the torque determination model. The update frequency of the model joint optimization can be adaptively adjusted according to the vehicle usage intensity. On the other hand, after uploading the accumulated large-scale driving data to the cloud, the cloud can periodically retrain each model and push it back to the local deployment after meeting the verification conditions, thereby realizing version update and long-term optimization.

[0170] In the embodiment of the present application, the domain controller can know the deviation degree between the operation intention of the driver and the operation demand in the current environment by first determining a target deviation score based on the driving operation information, the vehicle state information and the environment perception information of the vehicle, and then judging whether the operation intention of the driver matches the operation demand in the current environment according to the deviation degree. When the deviation degree is large, it indicates that the operation intention of the driver does not match the operation demand in the current environment, and then the vehicle power output can be intervened. Specifically, the expected output torque of the vehicle in the current environment can be determined based on the driving operation information, the environment perception information and the target deviation score, and then the vehicle is controlled to adjust the torque based on the expected output torque. The target deviation score represents the risk degree of the current operation of the driver. The risk degree of the driving operation is different, and the driving scene represented by the risk degree is also different. Therefore, the degree of driving intervention is also different, and the output torque is also different. By combining the target deviation score in the process of determining the expected output torque, the target deviation score can affect the intervention degree. Therefore, the expected output torque that matches the current environment can be accurately determined based on the driving operation information, the environment perception information and the target deviation score, that is, the torque that should be output by the vehicle in the current environment is accurately determined. Finally, the original output torque of the vehicle is adjusted based on the torque that should be output by the vehicle in the current environment, which can realize effective intervention on the driving operation. Compared with the scheme of directly cutting off the power response for driving intervention in the prior art, the occurrence of unexpected accidents can be avoided, and the driving intervention effect is improved.

[0171] Figure 7 FIG. 1 is a structural schematic diagram of a torque adjustment device provided by an embodiment of the present application. The torque adjustment device can be realized as part or all of a vehicle by software, hardware or a combination of both. The vehicle can be a vehicle shown in the following Figure 8 FIG. 2. Figure 7 The device includes a deviation calculation module 701, a torque determination module 702 and a torque adjustment module 703.

[0172] The deviation calculation module 701 is configured to determine a target deviation score based on driving operation information, vehicle state information and environment perception information of the vehicle, the target deviation score being used to represent a deviation degree between an operation intention of a driver and an operation demand in a current environment; The torque determination module 702 is configured to determine an expected output torque of the vehicle in the current environment based on the driving operation information, the environment perception information and the target deviation score when the target deviation score is greater than a preset score threshold. The torque adjustment module 703 is configured to control the vehicle to adjust the torque based on the expected output torque.

[0173] Optionally, the deviation calculation module 701 is specifically configured to: feature encode the driving operation information, the environment perception information, and the vehicle state information to obtain a plurality of variable features; take the plurality of variable features as node features of a plurality of graph nodes respectively, and connect the plurality of graph nodes based on a target connection relationship to obtain a target topological graph, the target connection relationship being used to represent an association relationship between the driving operation information, the environment perception information, and the vehicle state information; input the target topological graph into a deviation detection model, and output a target deviation score through the deviation detection model.

[0174] Optionally, the deviation detection model includes an attention module, a path aggregation module, and a deviation calculation module, and the deviation calculation module 701 is specifically configured to: for any one edge of a plurality of edges of the target topological graph, determine an edge weight between two adjacent graph nodes on the edge based on node features of the two adjacent graph nodes and prior constraint features through the attention module, the edge weight being used to represent an influence strength between the two adjacent graph nodes, and the prior constraint features being used to indicate constraint information between the environment perception information, the vehicle state information, and the driving operation; take a graph node corresponding to the driving operation information in the plurality of graph nodes as a first starting node, and obtain graph nodes traversed to reach a target node starting from the first starting node through the path aggregation module to construct a driving intention path, the driving intention path being used to indicate an actual response result caused by a current driving operation; take a graph node corresponding to the environment perception information in the plurality of graph nodes as a second starting node, and obtain graph nodes traversed to reach a target node starting from the second starting node through the path aggregation module to construct an environmental constraint path, the environmental constraint path being used to indicate a theoretical response result that a vehicle should present in a current environment; determine a feature of the driving intention path based on node features of a plurality of graph nodes traversed by the driving intention path and edge weights between adjacent graph nodes, and determine a feature of the environmental constraint path based on node features of a plurality of graph nodes traversed by the environmental constraint path and edge weights between adjacent graph nodes through the deviation calculation module; determine the target deviation score based on a difference between the feature of the driving intention path and the feature of the environmental constraint path through the deviation calculation module.

[0175] Optionally, the torque determination module 702 is specifically configured to: input the driving operation information, the environment perception information, and the target deviation score into an expected torque matching model, and determine an expected output torque based on historical behavior data through the expected torque matching model, the historical behavior data including torque output values when a vehicle is successfully intervened in different driving scenarios.

[0176] Optionally, the expected torque matching model comprises an encoder and a query module, the historical behavior data is composed of a plurality of key-value pairs, the plurality of key-value pairs comprise a plurality of key vectors and a plurality of value vectors, the plurality of key vectors are used to represent a plurality of driving scenes, and the plurality of value vectors are used to represent torque output values of successful interventions of the vehicle in respective driving scenes; the torque determination module 702 is specifically configured to: the encoder is configured to encode the driving operation information, the environment perception information, and the target deviation score to obtain a key vector of the current driving scene; the query module is configured to calculate a similarity between the key vector of the current driving scene and the plurality of key vectors, and determine weights of the plurality of value vectors based on the similarity; the query module is configured to perform weighted summation on the plurality of value vectors based on the weights of the plurality of value vectors to obtain the expected output torque.

[0177] Optionally, the apparatus further comprises: the acquisition module is configured to acquire a confidence degree of the expected output torque, the confidence degree being used to represent a degree of credibility of the expected output torque; and in a case where the confidence degree of the expected output torque is less than a preset confidence threshold, the target deviation score and the driving operation information are acquired; and the torque adjustment module 703 is specifically configured to: based on the target deviation score, the driving operation information, and the expected output torque, control the vehicle to perform torque adjustment.

[0178] Optionally, the torque adjustment module 703 is specifically configured to: fuzzify the target deviation score to determine a plurality of membership degrees, the plurality of membership degrees being used to represent likelihoods of the target deviation score belonging to a plurality of deviation levels; input the driving operation information into each of a plurality of torque determination models, and output candidate torques when performing driving intervention by the plurality of torque determination models respectively, the plurality of torque determination models corresponding to the plurality of deviation levels one by one; based on the plurality of membership degrees, perform weighted summation on the candidate torques output by the plurality of torque determination models respectively to obtain a reference torque; based on the reference torque and the expected output torque, control the vehicle to perform torque adjustment.

[0179] Optionally, the torque adjustment module 703 is further configured to: acquire the confidence degree of the expected output torque and the target deviation score, the confidence degree being used to represent a degree of credibility of the expected output torque, and the confidence degree being output by the expected torque matching model; based on the target deviation score and the confidence degree, determine a fusion ratio of the reference torque and the expected output torque; The reference torque and the expected output torque are fused according to a fusion ratio to obtain a target torque, and the vehicle is controlled based on the target torque to adjust the torque.

[0180] Optionally, the torque adjustment module 703 is further configured to: obtain an original output torque; determine a transition torque based on the original output torque and the target torque, the transition torque being between the original output torque and the target torque; control the vehicle to output the transition torque.

[0181] In the embodiments of the present application, by first determining a target deviation score based on the driving operation information, the vehicle state information and the environment perception information of the vehicle, the deviation degree between the operation intention of the driver and the operation demand in the current environment can be known, and then whether the operation intention of the driver matches the operation demand in the current environment is determined according to the deviation degree. When the deviation degree is large, it means that the operation intention of the driver does not match the operation demand in the current environment, and then the vehicle power output can be intervened. Specifically, the expected output torque of the vehicle in the current environment can be determined based on the driving operation information, the environment perception information and the target deviation score, and then the vehicle is controlled to adjust the torque based on the expected output torque. The target deviation score represents the risk degree of the current operation of the driver. The risk degree of the driving operation is different, and the driving scene represented by the risk degree is also different. Therefore, the degree of driving intervention is also different, and the output torque is also different. By combining the target deviation score in the process of determining the expected output torque, the target deviation score can affect the intervention degree. Therefore, the expected output torque that matches the current environment can be accurately determined based on the driving operation information, the environment perception information and the target deviation score, that is, the torque that should be output by the vehicle in the current environment is accurately determined. Finally, the original output torque of the vehicle is adjusted based on the torque that should be output by the vehicle in the current environment, which can realize effective intervention on the driving operation. Compared with the scheme of directly cutting off the power response for driving intervention in the prior art, the occurrence of unexpected accidents can be avoided, and the driving intervention effect is improved.

[0182] It should be noted that: the torque adjustment device provided in the above embodiments adjusts the output torque of the vehicle when the driving intervention is performed, and only the division of the above functional modules is exemplified. In actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above.

[0183] The functional units and modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0184] The torque adjustment device and torque adjustment method provided in the above embodiments belong to the same concept. The specific working process and technical effects of the units and modules in the above embodiments can be found in the method embodiments section, and will not be repeated here.

[0185] Figure 8 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application.

[0186] For example, such as Figure 8 As shown, the vehicle 800 includes a memory 81 and a processor 80, wherein the memory 81 stores executable program code 82, and the processor 80 is used to call and execute the executable program code 82 to perform the torque adjustment method described above.

[0187] This embodiment can divide the vehicle into functional modules according to the above method example. For example, each function can be assigned to a separate module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0188] When each functional module is divided according to its corresponding function, the vehicle may include: a deviation calculation module, a torque determination module, and a torque adjustment module. It should be noted that all relevant content regarding the steps involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here.

[0189] The vehicle provided in this embodiment is used to execute the torque adjustment method described above, and therefore can achieve the same effect as the above implementation method.

[0190] When using integrated units, the vehicle may include a processing module and a storage module. The processing module is used to control and manage the vehicle's actions. The storage module is used to support the vehicle in executing corresponding program code and data.

[0191] The processing module can be a processor or a controller, which can implement or execute various exemplary logical blocks, modules, and circuits shown in combination with the disclosure of the present application. The processor can also be a combination of computing functions, such as including one or more microprocessor combinations, digital signal processing (DSP) and microprocessor combinations, etc. The storage module can be a memory.

[0192] The embodiment also provides a computer readable storage medium, which stores computer program codes, and when the computer program codes are run on a computer, the computer executes the related method steps to realize the torque adjustment method in the above embodiment.

[0193] The embodiment also provides a computer program product, which, when run on a computer, causes the computer to execute the related steps to realize the torque adjustment method in the above embodiment.

[0194] The vehicle, the computer readable storage medium, the computer program product or the chip provided by the embodiment are used to execute the method provided above, and therefore, the beneficial effects achieved thereby can refer to the beneficial effects of the method provided above, which will not be described herein again.

[0195] Through the description of the above embodiments, those skilled in the art can understand that, for the convenience and brevity of description, only the above-mentioned division of functional modules is taken as an example for illustration, and in actual application, the above-mentioned functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above.

[0196] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are illustrative, for example, the division of modules or units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another device, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed ones can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0197] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A torque adjustment method characterized by, The method includes: Based on driving operation information, vehicle status information, and vehicle environmental perception information, a target deviation score is determined. The target deviation score is used to represent the degree of deviation between the driver's operating intention and the operating requirements in the current environment. If the target deviation score is greater than a preset score threshold, the desired output torque of the vehicle in the current environment is determined based on the driving operation information, the environmental perception information, and the target deviation score. The desired output torque is used to control the vehicle to adjust its torque.

2. The method of claim 1, wherein, The determination of the target deviation score based on driving operation information, vehicle status information, and vehicle environmental perception information includes: The driving operation information, the environmental perception information, and the vehicle status information are feature-encoded to obtain multiple variable features; The multiple variable features are respectively used as node features of multiple graph nodes, and the multiple graph nodes are connected based on the target connection relationship to obtain the target topology graph. The target connection relationship is used to represent the association between the driving operation information, the environmental perception information, and the vehicle status information. The target topology map is input into the deviation detection model, and the deviation detection model outputs the target deviation score.

3. The method of claim 2, wherein, The deviation detection model includes an attention module, a path aggregation module, and a deviation calculation module. The step of inputting the target topology map into the deviation detection model and outputting the target deviation score through the deviation detection model includes: For any one of the multiple edges of the target topology graph, the attention module determines the edge weight between two adjacent graph nodes on the edge based on the node features and prior constraint features of the two adjacent graph nodes on the edge. The edge weight is used to represent the influence strength between the two adjacent graph nodes, and the prior constraint features are used to indicate the constraint information between environmental perception information, vehicle state information, and driving operation. The path aggregation module takes the graph node corresponding to the driving operation information in the plurality of graph nodes as the first starting node, and starts from the first starting node to obtain the graph nodes traversed to reach the target node, and constructs a driving intention path. The driving intention path is used to indicate the actual response result caused by the current driving operation. The path aggregation module takes the graph node corresponding to the environmental perception information among the multiple graph nodes as the second starting node, and starts from the second starting node to obtain the graph nodes traversed to reach the target node, thus constructing an environmental constraint path. The environmental constraint path is used to indicate the theoretical response result that the vehicle should present in the current environment. The deviation calculation module determines the characteristics of the driving intention path based on the node characteristics of multiple graph nodes traversed by the driving intention path and the edge weights between each adjacent graph node, and determines the characteristics of the environmental constraint path based on the node characteristics of multiple graph nodes traversed by the environmental constraint path and the edge weights between each adjacent graph node. The deviation calculation module determines the target deviation score based on a difference between a feature of the driving intention path and a feature of the environment constraint path.

4. The method of claim 1, wherein, The determination of the expected output torque of the vehicle in the current environment based on the driving operation information, the environment perception information, and the target deviation score includes: inputting the driving operation information, the environment perception information, and the target deviation score into an expected torque matching model, and determining the expected output torque based on historical behavior data by the expected torque matching model, the historical behavior data including torque output values when a vehicle is successfully intervened in different driving scenarios.

5. The method of claim 4, wherein, The expected torque matching model includes an encoder and a query module, the historical behavior data is composed of a plurality of key-value pairs, the plurality of key-value pairs include a plurality of key vectors and a plurality of value vectors, the plurality of key vectors are used to represent a plurality of driving scenarios, and the plurality of value vectors are used to represent torque output values when a vehicle is successfully intervened in each driving scenario; the inputting of the driving operation information, the environment perception information, and the target deviation score into the expected torque matching model and the determination of the expected output torque based on the historical behavior data by the expected torque matching model include: feature encoding of the driving operation information, the environment perception information, and the target deviation score by the encoder to obtain a key vector of a current driving scenario; calculation of a similarity between the key vector of the current driving scenario and the plurality of key vectors by the query module, and determination of weights of the plurality of value vectors based on the similarity; weighted summation of the plurality of value vectors based on the weights of the plurality of value vectors by the query module to obtain the expected output torque.

6. The method according to any one of claims 1 to 5, wherein, The method further includes: obtaining a confidence degree of the expected output torque, the confidence degree being used to represent a credibility of the expected output torque; in a case where the confidence degree of the expected output torque is less than a preset confidence threshold, obtaining the target deviation score and the driving operation information; and the control of the vehicle for torque adjustment based on the expected output torque includes: control of the vehicle for torque adjustment based on the target deviation score, the driving operation information, and the expected output torque.

7. The method of claim 6, wherein, The control of the vehicle for torque adjustment based on the target deviation score, the driving operation information, and the expected output torque includes: fuzzification of the target deviation score to determine a plurality of membership degrees, the plurality of membership degrees being used to represent a possibility that the target deviation score belongs to a plurality of deviation levels; inputting the driving operation information into each of a plurality of torque determination models to respectively output candidate torques when driving intervention is performed by the plurality of torque determination models, the plurality of torque determination models corresponding to the plurality of deviation levels one by one; weighted summation of the candidate torques respectively output by the plurality of torque determination models based on the plurality of membership degrees to obtain a reference torque; control of the vehicle for torque adjustment based on the reference torque and the expected output torque.

8. The method of claim 7, wherein, The method further includes: a confidence level of the expected output torque and the target deviation score; and the controlling the vehicle to make torque adjustment based on the reference torque and the expected output torque comprises: determining a fusion ratio of the reference torque and the expected output torque based on the target deviation score and the confidence level; fusing the reference torque and the expected output torque according to the fusion ratio to obtain a target torque, and controlling the vehicle to make torque adjustment based on the target torque.

9. The method of claim 8, wherein, The method further comprises: obtaining an original output torque; and the controlling the vehicle to make torque adjustment based on the target torque comprises: determining a transition torque based on the original output torque and the target torque, the transition torque being between the original output torque and the target torque; controlling the vehicle to output the transition torque.

10. A vehicle characterized by comprising: The vehicle comprises: a memory for storing executable program code; a processor for invoking and running the executable program code from the memory, so that the vehicle executes the method according to any one of claims 1 to 9.

Images