Multi-dimensional fine-grained intelligent driving system evaluation method

Abstract

The invention belongs to the field of data analysis and processing, and discloses a multi-dimensional fine-grained intelligent driving system evaluation method. Comprising the steps that road postures, environment data and vehicle state parameters in the intelligent driving process are collected, three driving modes including high-speed defense, reference cruising and environment penetration speed are dynamically selected based on a cognitive calculation type operation normal form tree, and a gradual change state machine is used for conducting smooth switching of the driving modes; the method comprises the following steps: switching driving modes in a diversified test scene, and based on each driving mode selected by switching, performing judgment by using a multi-layer perception evaluation system to obtain scores of the intelligent driving system in safety, efficiency and comfort and a comprehensive score; wherein the multi-layer perception evaluation system comprises a three-dimensional association network, the three-dimensional association network performs granularity refining on multiple dimensions to form multi-level three-dimensional representation, a complete three-dimensional association network is formed through node simulation and directed edge connection, and multi-dimensional fine granularity evaluation of the intelligent driving system is achieved.

Description

Technical Field

[0001] The present invention relates to the field of data analysis and processing, and more specifically, to a multi-dimensional fine-grained intelligent driving system evaluation method. Background Art

[0002] When evaluating intelligent driving systems, it is generally necessary to evaluate from the perspectives of safety (whether the system can avoid accidents and respond to emergencies or adverse situations, such as sudden rain or snow, driving at night, or pedestrians suddenly crossing the road), efficiency (the system's performance in terms of time, energy consumption, path optimization, etc.), and comfort (user experience, including acceleration and deceleration smoothness, lane changing strategies, etc.). However, there are often trade-offs or conflicts between different evaluation dimensions.

[0003] These conflicts make it a challenge to scientifically balance and comprehensively evaluate intelligent driving when designing the evaluation system. Moreover, in the evaluation process, safety, comfort, and efficiency are relatively general and vague concepts, and how to refine the evaluation dimensions into quantifiable indicators is also a technical challenge. For example, the performance of the perception module can be decomposed into object recognition rate, false detection rate, missed detection rate, etc., but these indicators need to be defined and measured.

[0004] Therefore, design and innovation are needed when conducting multi-dimensional and fine-grained evaluation of intelligent driving systems to meet actual needs. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: a multi-dimensional fine-grained intelligent driving system evaluation method, comprising: collecting road posture, environmental data and vehicle state parameters during intelligent driving, dynamically selecting three driving modes of high-speed defense, baseline cruise and environmental penetration speed based on a cognitive computing operation paradigm tree, and using a gradual state machine to smoothly switch the driving mode; A multi-layer perception evaluation system is preset to switch driving modes in various test scenarios. Based on each selected driving mode, the multi-layer perception evaluation system is used to evaluate the safety, efficiency, and comfort scores of the intelligent driving system, and to perform visual interaction. Among them, the multi-layer perception evaluation system includes a three-dimensional association network, which forms a multi-level three-dimensional representation by refining the granularity of multiple dimensions, and forms a complete three-dimensional association network through node simulation and directed edge connection; For any node in the three-dimensional association network, the influence evolution propagation mechanism is used to update it. The influence evolution propagation mechanism is designed based on the graph attention network, and the saturation effect of energy propagation, the gating mechanism, the score difference drive of diffusion propagation and the damping effect of energy propagation are introduced to achieve the real diffusion of the influence interaction between evolution dimensions.

[0006] Preferably, the three driving modes of high-speed defense, baseline cruising and environmental penetration speed include: In each driving mode, a weight value is used to quantify the priority of each key dimension, and the weight value corresponding to the main dimension is the largest. Among them, the key dimensions are safety, efficiency and comfort, and the main dimension is the most prioritized key dimension in each driving mode; Hard constraints for key indicators are set for each driving mode to meet the minimum performance requirements in any situation. For the high-speed defense mode, the safety constraints are set to emergency braking response time less than A1, target detection recall rate greater than B1, and braking distance less than C1; the efficiency constraints are set to path length optimization rate greater than D1, and the comfort constraints are set to acceleration change rate less than E1; For the baseline cruise mode, the safety constraints are set as emergency braking response time less than A2, target detection recall rate greater than B2, and braking distance less than C2; the efficiency constraints are set as path length optimization rate greater than D2; and the comfort constraints are set as acceleration change rate less than E2. For the environmental penetration speed mode, the safety constraints are set as the emergency braking response time less than A3, the target detection recall rate greater than B3, and the braking distance less than C3; the efficiency constraints are set as the path length optimization rate less than D3; and the comfort constraints are set as the acceleration change rate less than E3; Among them, A1, A2, A3, B1, B2, B3, C1, C2, C3, D1, D2, D3, E1, E2 and E3 are pre-set values, and the values ​​meet the requirements of A1<A2<A3; B1>B2>B3; C1>C2>C3; D1>D2>D3; E1>E2>E3.

[0007] Preferably, the method of collecting road posture, environmental data and vehicle status parameters during intelligent driving and dynamically selecting three driving modes of high-speed defense, baseline cruising and environmental penetration speed based on a cognitive computing operation paradigm tree includes: Road posture data includes slope, curvature and road friction coefficient, environmental data includes weather data, visibility and traffic density, and vehicle state parameters include vehicle speed and acceleration; The cognitive computing operation paradigm tree is set to consist of a root node and intermediate nodes, and the input is road posture, environmental data and vehicle state parameters, where the root node represents the driving mode selection and the intermediate node is used for subtask decomposition; At each intermediate node, the cognitive computing model is used to reason about the input data and output the priority and confidence of the subtask; The set driving modes include high-speed defense, baseline cruising and environmental penetration speed. Among them, the high-speed defense mode prioritizes safety, and the triggering conditions are that the slope in the road posture is greater than q1, the curvature is greater than s1, and the road friction coefficient is less than d1; the weather data in the environmental data is bad weather, visibility is less than n1, and traffic density is greater than g1; the vehicle state is the speed of the vehicle is greater than c1, and the acceleration change rate is greater than w1; The baseline cruise mode is efficiency-first, and the triggering conditions are that the slope is ≤q1, the curvature is ≤s1, ​​and the road friction coefficient is ≥d1 in the road posture; the weather data in the environmental data is moderate weather, visibility is ≥n1, and traffic density is ∈ (g2, g1); the vehicle speed is ∈ (c2, c1), and the acceleration change rate is ≤w1 in the vehicle state; The environmental penetration speed mode is comfort-first, and the triggering conditions are: slope ≤ q2, curvature > s2 in road posture; visibility < n2, traffic density > g3 in environmental data; vehicle speed < c2, acceleration change rate ≤ w2 in vehicle status; Wherein, q1>q2, s1<s2; d1; n1>n2; g1>g2>g3; c1>c2; w1>w2, and q1, q2, s1, s2, d1, n1, n2, g1, g2, g3, c1, c2, w1 and w2 are all preset values; The vehicle's road posture, environmental data and vehicle status parameters are collected in real time at a fixed frequency hb, and input into the cognitive computing operation paradigm tree to output the selected driving mode. If the driving mode currently output is the same as the previous output, there is no need to switch the driving mode. If the driving mode currently output is different from the previous output, a gradual state machine is used to smoothly switch the driving mode.

[0008] Preferably, the method of using a gradual state machine to smoothly switch the driving mode includes: Smooth switching includes driving mode weight switching and vehicle control parameter switching; For the weight switching of driving modes, the weights of safety, efficiency and comfort in each driving mode are preset, and a switching time window is added. Each weight is gradually adjusted using linear interpolation within the window until the preset weight value in the driving mode is reached; For the vehicle control parameter switching, an intermediate state is introduced in the middle of the switching. The intermediate state is the intermediate value of the control parameters in the two modes. The control parameters include vehicle speed, acceleration and following distance. The RNN algorithm is used to evaluate the scene jump. The lengthening multiple of the switching time window is set according to the scene jump. The adjusted switching time window is obtained and used as the final switching time window and applied to the control parameter switching process.

[0009] Preferably, the design method of the multi-layer perception evaluation system includes: Step L1: construct a three-dimensional association network; With safety, efficiency and comfort as key dimensions, sub-dimensions under the key dimensions are preset, and the sub-dimensions are decomposed to form sub-sub-dimensions. The key dimensions, sub-dimensions and sub-sub-dimensions are distributed in the same level, and so on. All levels are filled. There are dimensions in each level. All dimensions are simulated into nodes. Any two nodes are connected by directed edges, and attribute values ​​are added to each directed edge to form a three-dimensional network structure. Among them, directed edges represent the mutual influence relationship between dimensions, from the influencer to the influenced, and the attribute value is used to represent the intensity of the influence, including positive and negative influences; Step L2: The dimensions of the last level in the stereo association network define quantifiable indicators, and these indicators are used as inputs of the stereo association network; Initialize the attribute value of each directed edge and the energy value of each node, and use the influence evolution propagation mechanism to update the attribute value of the directed edge in the three-dimensional association network, so as to evolve the energy value of the node; Obtain the energy value of the node corresponding to the key dimension in the stereo association network as the final output value.

[0010] Preferably, the method for initializing the attribute value of each directed edge and the energy value of each node comprises: Pre-train machine learning models in different modes, and use historical data to input into the models in the corresponding modes to obtain the attribute values ​​of each directed edge; The expert scoring method is used to obtain the initial score of each node and normalize it.

[0011] Preferably, the method of using the influence evolution propagation mechanism to update the attribute value of the directed edge in the stereo association network, thereby evolving the energy value of the node, includes: The indicators of the last level in the stereo association network are collected as input, the evaluation period is preset, and it is discretized into time steps for dynamic iteration; At the current time step, the update direction is passed from the lowest level to the highest level in sequence. For any two nodes connected by a directed edge, the influence evolution propagation mechanism is used to update until the entire three-dimensional association network is updated. The iteration is repeated at the next time step until the preset evaluation cycle time is used up. The iteration stops and the energy value of each node and the directed edge attribute value are updated. Then the energy values ​​of the nodes corresponding to safety, efficiency and comfort on the first level are output as the scores of the key dimensions; Construct a comprehensive scoring function to evaluate the overall performance of the system, including normalizing the energy values ​​of the nodes corresponding to safety, efficiency, and comfort, and performing weighted summation according to the corresponding weights under the selected driving mode to obtain a comprehensive score; The scores of key dimensions and the comprehensive scores are output as the evaluation scores of the intelligent driving system.

[0012] Preferably, the influence evolution propagation mechanism is designed based on a graph attention network, including: For any two nodes in the stereo association network, at each time step, according to the direction of the directed edge, the graph attention network is used to calculate the attention weights of the corresponding nodes of the influencer and the influenced as the new attribute value of the directed edge, and the energy propagation and saturation effect are added. The new energy value received by the corresponding node of the influenced is the weighted sum of the attention weights of all predecessor nodes, the initial attribute value of the directed edge, and the energy value; In the process of influence transmission, a gating mechanism is introduced to control the fusion ratio of the new energy value and the current energy value. The specific method is that the update gate is the Sigmoid function value calculated by weighted sum of the current energy value and the new energy value; the candidate state is the hyperbolic tangent function value calculated by weighted sum of the current energy value and the new energy value, and the node energy value is updated using the weighted sum of the current energy value and the candidate state, with the weight determined by the update gate; On the basis of updating the node energy value, the score difference drive of diffusion propagation is introduced, including obtaining the energy value difference between the nodes corresponding to the influencer and the influenced, weighted summing up the initial energy value of all predecessor nodes corresponding to the node of the influenced and the energy value difference, and then multiplying them by the diffusion coefficient to obtain the diffusion term, and using the updated energy value and the diffusion term to accumulate to get the corrected energy value of the node; On the basis of the correction of the node energy value, the damping effect of energy propagation is introduced, including the introduction of the damping coefficient. The damping coefficient is subtracted from 1 to obtain the absolute value. For the current time step, the final state of the energy value on the node in the next time step is the product of the corrected energy value and the absolute value.

[0013] Preferably, the saturation effect of energy propagation includes calculating the Sigmoid function value of the energy value of the node corresponding to the influencer and mapping it to the interval (0, 1).

[0014] Preferably, the visual interaction method comprises: Use graph visualization tools to draw a three-dimensional correlation network structure. The nodes in each level are the same size. Starting from the first level, the node shapes become smaller. The thickness of the directed edges is drawn according to the size of the attribute value. Directed edges are drawn with different colors, and positive and negative effects are set according to the color. Use dynamic visualization tools to demonstrate the updating process of the stereo correlation network structure; The entire evaluation process is arranged on a visual interface for interaction with users.

[0015] The technical effects and advantages of the multi-dimensional fine-grained intelligent driving system evaluation method of the present invention are as follows: A multi-level three-dimensional association network is designed. Through the three-dimensional association network, the evaluation dimensions are decomposed into a multi-level structure, and the dynamic influence relationship between dimensions is represented by directed edges. This not only realizes the fine-grained decomposition of the evaluation dimensions, but also can truly simulate the complex interactions between dimensions, thereby improving the scientificity and accuracy of the evaluation.

[0016] Based on the graph attention network, an influence evolution propagation mechanism is designed. In the update of the three-dimensional association network, the influence evolution propagation mechanism is introduced, and the saturation effect (Sigmoid mapping), gating mechanism (update gate and candidate state), score difference drive (diffusion propagation) and damping effect are combined to dynamically simulate the propagation and evolution of influence between dimensions.

[0017] Dynamic mode selection is achieved through cognitive computing operation paradigm tree, and dynamic selection of driving mode is achieved by combining multi-source data (road posture, environmental data, vehicle status parameters). The tree structure decomposes mode selection into subtasks, and infers priority and confidence through cognitive computing model. Flexible switching of driving strategies according to real-time scenarios can meet diverse driving needs.

[0018] In the driving mode switching, a smooth switching mechanism of the gradient state machine is introduced, weight switching is achieved through linear interpolation, control parameter switching is achieved through intermediate states, and the RNN algorithm is used to evaluate the scene jump to dynamically adjust the switching time window. This significantly reduces the risks in the switching process (such as safety hazards or reduced comfort), which is of great value in the field of intelligent driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of the steps of a multi-dimensional fine-grained intelligent driving system evaluation method of the present invention; Figure 2 It is a structural schematic diagram of a multi-dimensional fine-grained intelligent driving system evaluation method of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example 1

[0021] See also Figure 1 As shown, the multi-dimensional fine-grained intelligent driving system evaluation method described in this embodiment includes: When evaluating intelligent driving systems, it is generally necessary to evaluate them from the perspectives of safety (whether the system can avoid accidents and respond to emergencies or adverse situations, such as sudden rain or snow, driving at night, or pedestrians suddenly crossing the road), efficiency (the system's performance in terms of time, energy consumption, and path optimization), and comfort (user experience, including acceleration and deceleration smoothness, lane change strategies, etc.). However, there are often trade-offs or conflicts between different evaluation dimensions. For example: Safety vs. Efficiency: Overly conservative driving strategies (such as maintaining excessive distance between vehicles) may improve safety but reduce traffic efficiency.

[0022] Comfort vs. Efficiency: To improve comfort (e.g., smooth acceleration), the system may need to sacrifice efficiency (e.g., missing a green light).

[0023] Safety vs. Comfort: In certain emergency situations, emergency accident avoidance may result in reduced comfort (e.g., safety is difficult to ensure due to the impact of comfort, and accidents cannot be avoided).

[0024] These conflicts make it a challenge to scientifically balance and comprehensively evaluate intelligent driving when designing the evaluation system. Moreover, in the evaluation process, safety, comfort, and efficiency are relatively general and vague concepts, and how to refine the evaluation dimensions into quantifiable indicators is also a technical challenge. For example, the performance of the perception module can be decomposed into object recognition rate, false detection rate, missed detection rate, etc., but these indicators need to be defined and measured.

[0025] Therefore, design and innovation are needed to meet actual needs when conducting multi-dimensional and fine-grained evaluation of intelligent driving systems.

[0026] A multi-dimensional fine-grained intelligent driving system evaluation method includes: collecting road posture and environmental data during intelligent driving, dynamically selecting three driving modes of high-speed defense, baseline cruise and environmental penetration speed based on a cognitive computing operation paradigm tree, and using a gradual state machine to smoothly switch the driving mode; A multi-layer perception evaluation system is preset to switch driving modes in various test scenarios. Based on each selected driving mode, the multi-layer perception evaluation system is used to evaluate the safety, efficiency, comfort and comprehensive score of the intelligent driving system. Among them, the multi-layer perception evaluation system includes a three-dimensional association network, which forms a multi-level three-dimensional representation by refining the granularity of multiple dimensions, and forms a complete three-dimensional association network through node simulation and directed edge connection; For any node in the three-dimensional association network, the influence evolution propagation mechanism is used to update it. The influence evolution propagation mechanism is designed based on the graph attention network, and the saturation effect of energy propagation, the gating mechanism, the score difference drive of diffusion propagation and the damping effect of energy propagation are introduced to achieve the real diffusion of the influence interaction between evolution dimensions.

[0027] Based on the problems in the background technology, the solution is specifically analyzed, and the specific contents are as follows: 1. Traditional systems may select driving modes based on only a single dimension or fixed rules and cannot dynamically adapt to complex scenarios.

[0028] In response to the above problems, this solution defines three driving modes (high-speed defense, baseline cruise, and environmental penetration speed), with safety, efficiency, and comfort as the priority dimensions. This design ensures that the system can select the most appropriate mode in different scenarios, rather than a single mode to deal with all scenarios. By introducing a cognitive computing model, it can dynamically reason and select a driving mode based on real-time collected road posture (slope, curvature, road friction coefficient), environmental data (weather, visibility, traffic density), and vehicle state parameters (vehicle speed, acceleration). It overcomes the limitations of static rules and can adapt to complex and changing driving scenarios. By setting clear trigger conditions for each driving mode (such as slope>q1, visibility<n1, etc.), and realizing fine division of modes through parameter threshold differences (such as q1>q2, n1>n2), it ensures that the mode selection is highly matched with the scene requirements.

[0029] 2. Traditional systems lack quantitative evaluation and dynamic priority adjustment mechanisms for multiple dimensions such as safety, efficiency and comfort.

[0030] In response to the above problems, this solution quantifies the priorities of safety, efficiency and comfort through weight values ​​in each driving mode, and ensures that the main dimension (such as safety in high-speed defense mode) has the maximum weight. This quantification method makes multi-dimensional trade-offs more scientific and controllable. Hard constraints are set for key indicators of each driving mode (such as emergency braking response time, target detection recall rate, braking distance, etc.), and different thresholds are set according to the characteristics of the mode (such as A1<A2<A3, B1>B2>B3, etc.). This ensures that the system can meet the minimum performance requirements under any circumstances and avoids performance shortcomings caused by improper trade-offs. Through the cognitive computing model to infer the priority and confidence of subtasks in the operation paradigm tree, the system can dynamically adjust the priority of each dimension and further optimize the multi-dimensional trade-off.

[0031] 3. Traditional systems may cause sudden changes in control parameters when switching driving modes, affecting safety, comfort and efficiency.

[0032] To address the above issues, this solution introduces a gradual state machine to achieve smooth switching of driving modes. Specifically, it includes weight switching and control parameter switching, and dynamically adjusts the length of the switching time window according to the jump size. This method ensures smooth transition of control parameters such as vehicle speed, acceleration, and following distance, and reduces the impact of mode switching on driving experience and safety.

[0033] Therefore, this solution solves the problems of singleness and staticness of mode selection through multi-mode design and cognitive computing model; solves the difficult problem of multi-dimensional trade-off through weight quantification and hard constraints; solves the abruptness of mode switching through gradual state machine and switching time window; solves the problem of insufficient adaptability in complex scenarios through multi-dimensional data input and parameter threshold design; solves the coarse-grained problem of evaluation method through fine-grained indicator design and dynamic evaluation. This makes the evaluation and operation of intelligent driving system more scientific, precise and efficient, and can provide safe, efficient and comfortable driving experience in various complex scenarios.

[0034] The three driving modes of high-speed defense, baseline cruise and environmental penetration speed include: In each driving mode, a weight value is used to quantify the priority of each key dimension, and the weight value corresponding to the main dimension is the largest. Among them, the key dimensions are safety, efficiency and comfort, and the main dimension is the most prioritized key dimension in each driving mode; For example, in high-speed defense mode, safety is the primary goal, so the weight value of safety is assigned the highest, such as 0.8, and efficiency and comfort are second, which can be set to 0.1 and 0.1.

[0035] Hard constraints for key indicators are set for each driving mode to meet minimum performance requirements (such as safety indicators) under any circumstances. For the high-speed defense mode, the safety constraint is set to the emergency braking response time being less than A1, for example, A1=300ms, the target detection recall rate (an important indicator for measuring the model's detection capability, indicating the model's ability to correctly identify all real targets) being greater than B1, for example, B1=98%, and the braking distance (calculated based on vehicle speed and road conditions) being less than C1, for example, C1=10m; the efficiency constraint is set to the path length optimization rate being greater than D1, for example, D1=80% (a certain degree of efficiency sacrifice is allowed); and the comfort constraint is set to the acceleration change rate being less than E1, for example, E1=3.0 m / s³ (an appropriate reduction in comfort is allowed).

[0036] For the baseline cruise mode, the safety constraints are set as emergency braking response time less than A2, for example, A2=500ms, target detection recall rate greater than B2, for example, B2=95%, and braking distance less than C2, for example, C2=15m; the efficiency constraints are set as path length optimization rate greater than D2, for example, D2=90%, and average travel time less than 1.1 times the baseline time; the comfort constraints are set as acceleration change rate less than E2, for example, E2=2.0 m / s³, and lateral acceleration can be introduced, for example, lateral acceleration less than 2.0m / s²; For the environmental penetration speed mode, the safety constraints are set as the emergency braking response time is less than A3, for example, A3=700ms, the target detection recall rate is greater than B3, for example, B3=90%, and the braking distance is less than C3, for example, C3=5m; the efficiency constraint is set as the path length optimization rate is less than D3, for example, D3=70% (lower efficiency requirement); the comfort constraint is set as the acceleration change rate is less than E3, for example, E3=1.5 m / s³, and lateral acceleration can be introduced, for example, lateral acceleration is less than 1.0 m / s².

[0037] Among them, A1, A2, A3, B1, B2, B3, C1, C2, C3, D1, D2, D3, E1, E2 and E3 are pre-set values, and the values ​​meet the requirements of A1<A2<A3; B1>B2>B3; C1>C2>C3; D1>D2>D3; E1>E2>E3; The above path length optimization rate = (theoretical optimal path length / actual driving path length) × 100%. The theoretical optimal path length refers to the shortest path length from the starting point to the end point calculated by the path planning algorithm (such as Dijkstra, A*) under ideal conditions (no obstacles, no traffic rules restrictions, no dynamic environment interference); the actual driving path length refers to the total length of the track recorded by the vehicle during the actual driving process (obtained by high-precision odometer or GPS track integration) In high-speed defense mode, it is allowed to sacrifice 20% path efficiency in exchange for safety (such as actively bypassing potential risk areas); in the baseline cruise mode, safety and efficiency are balanced, and the path deviation from the optimal value does not exceed 10%. In the environmental penetration speed mode, more than 30% path redundancy is accepted (such as multiple adjustments to posture in narrow areas and refined obstacle avoidance).

[0038] The method of collecting road posture, environmental data and vehicle status parameters during intelligent driving and dynamically selecting three driving modes of high-speed defense, baseline cruising and environmental penetration speed based on the cognitive computing operation paradigm tree includes: Road posture data includes slope, curvature, and road friction coefficient; environmental data includes weather data, visibility, and traffic density; vehicle state parameters include vehicle speed and acceleration; In the evaluation process of multi-dimensional fine-grained intelligent driving systems, the road posture and environmental data during intelligent driving are the basis for system decision-making and evaluation. The specific collection content and methods are as follows: Road posture data is used to describe the geometric and physical characteristics of the vehicle's driving path, and mainly includes the following parameters: Slope: Indicates the longitudinal inclination of the road in percentage (%). Obtained through the vehicle-mounted IMU (Inertial Measurement Unit) or high-precision map data, the measurement accuracy must reach ±0.1%.

[0039] Curvature: In units of 1 / m, it indicates the curvature of the road. It is calculated using on-board cameras, LiDAR, or high-precision map data. The curvature calculation must take into account the continuity of the road centerline and the accuracy must reach ±0.01 / m.

[0040] Road friction coefficient: A dimensionless parameter that represents the friction characteristics between the road surface and the tire. It is estimated by on-board sensors (such as tire pressure sensors, acceleration sensors) combined with road material recognition algorithms, with an accuracy of ±0.05.

[0041] Environmental data is used to describe the external conditions of vehicle driving, which directly affects the selection of driving mode and the evaluation of system performance. It mainly includes the following parameters: Weather data: divided into sunny, rainy, snowy, foggy, etc. Weather information is obtained through vehicle-mounted meteorological sensors (such as rain sensors, temperature sensors) or V2X (vehicle-to-everything) communications. Weather classification must achieve an accuracy rate of more than 95%.

[0042] Visibility: In meters (m), it indicates the visible distance of the current environment. It is estimated by the on-board laser radar or camera combined with the image processing algorithm. The measurement error must be controlled within ±10m.

[0043] Traffic density: measured in vehicles / km, indicating the density of vehicles on the road. The number of surrounding vehicles is detected through V2X communication, vehicle-mounted radar or camera, and combined with the road length estimation, with an accuracy of ±5 vehicles / km.

[0044] Vehicle status data is used to describe the dynamic behavior of the vehicle and is an important basis for mode selection and evaluation. It mainly includes the following parameters: Vehicle speed: In km / h, it indicates the real-time speed of the vehicle. It is obtained through the vehicle speed sensor or GPS module, and the measurement error must be controlled within ±0.5km / h.

[0045] Acceleration: In m / s², it indicates the longitudinal and lateral acceleration of the vehicle. It is measured by the onboard IMU with an accuracy of ±0.01m / s².

[0046] The collected data can be pre-processed by using Kalman filtering or low-pass filtering algorithms to remove sensor noise and ensure the smoothness and consistency of the data. Multi-source data can also be aligned through timestamps to ensure the time consistency of data from different sensors. The synchronization error must be controlled within ±10ms. At the same time, abnormal data can be identified and eliminated through statistical methods (such as Z scores) or machine learning algorithms (such as isolation forests). The proportion of abnormal data must be controlled below 1%.

[0047] The cognitive computing operation paradigm tree is a decision-making framework based on cognitive computing. It decomposes complex driving scenarios into multiple subtasks by building a multi-level decision tree and dynamically selects the optimal driving mode based on input parameters. The specific design is as follows: The cognitive computing operation paradigm tree is set to consist of a root node and intermediate nodes, and the input is road posture, environmental data and vehicle state parameters. The root node represents the driving mode selection, and the intermediate node is used for subtask decomposition, that is, the intermediate node includes subtasks such as road posture analysis, environmental assessment, and vehicle state judgment.

[0048] At each intermediate node, cognitive computing models (such as Bayesian networks, fuzzy logic, or deep neural networks) are used to reason about the input data and output the priority and confidence of the subtask; The set driving modes include high-speed defense, baseline cruise and environmental penetration speed. Among them, the high-speed defense mode prioritizes safety, and the efficiency and comfort weights are relatively low. That is, the goal is to prioritize safety in high-risk scenarios (such as steep slopes and slippery roads) to minimize the risk of accidents. The trigger conditions are that the slope in the road posture is greater than q1 (for example, q1=10%), the curvature is greater than s1 (for example, s1=0.1 / m), and the road friction coefficient is less than d1 (for example, d1=0.5); the weather data in the environmental data is bad weather (such as rain, snow, fog), visibility is less than n1 (for example, n1=100m), and traffic density is greater than g1 (for example, g1=50 vehicles / km); the vehicle speed in the vehicle state is greater than c1 (for example, c1=80km / h), and the acceleration change rate is greater than w1 (w1=3.0m / s³); The baseline cruise mode prioritizes efficiency, safety and comfort, in order to optimize travel time and energy consumption while maintaining basic safety and comfort. The triggering conditions are that the slope is ≤q1 (e.g., q1=10%), the curvature is ≤s1 (e.g., s1=0.1 / m), and the road friction coefficient is ≥d1 (e.g., d1=0.5) in the road posture; the weather data in the environmental data is moderate weather (e.g., clear weather or light rain), visibility is ≥n1 (e.g., n1=100m), traffic density∈(g2,g1) (e.g., g1=50 vehicles / km, g2=20 vehicles / km); the vehicle speed is ∈(c2,c1) (e.g., c1=80km / h, c2=30km / h), and the acceleration change rate is ≤w1 (w1=3.0m / s³) in the vehicle state; The environmental penetration speed mode prioritizes comfort, safety and efficiency, in order to improve user experience (such as stability). The triggering conditions are that the slope in the road posture is ≤q2 (such as q2=5%), the curvature is greater than s2 (such as s2=0.2 / m, parking lot or narrow road); the visibility in the environmental data is less than n2 (such as n2=50m), and the traffic density is greater than g3 (such as congestion); the vehicle speed is less than c2, and the acceleration change rate is less than w2 (w2=2m / s³) in the vehicle state; Among them, q1, q2, s1, s2, d1, n1, n2, g1, g2, g3, c1, c2, w1 and w2 are all pre-set values; q1>q2, s1<s2; d1; n1>n2; g1>g2>g3; c1>c2; w1>w2, which can be calculated by averaging or weighted averaging based on real-time data analysis, or predicted by a machine model (such as deep learning).

[0049] The vehicle's road posture, environmental data and vehicle status parameters are collected in real time at a fixed frequency hb, and input into the cognitive computing operation paradigm tree to output the selected driving mode. If the driving mode currently output is the same as the previous output, there is no need to switch the driving mode. If the driving mode currently output is different from the previous output, a gradual state machine is used to smoothly switch the driving mode.

[0050] To avoid sudden changes when switching driving modes (such as sudden acceleration, sudden deceleration, or steering wheel shaking), the system uses a gradual state machine to achieve smooth switching of driving modes. The method of using a gradual state machine to smoothly switch driving modes includes: Smooth switching includes driving mode weight switching and vehicle control parameter switching; For the weight switching of driving modes, the weights of safety, efficiency and comfort in each driving mode are preset, and a switching time window T (for example, T = 5 seconds) is added. Each weight is gradually adjusted using linear interpolation within the window until the preset weight value in the driving mode is reached; For the control parameter switching of the vehicle, an intermediate state is introduced in the middle of the switching, and the intermediate state is the intermediate value of the control parameters in the two modes. The control parameters include vehicle speed, acceleration and following distance. For example, when the high-speed defense switches to the reference cruise, the vehicle speed switches from 10km / h to 30km / h, and the intermediate value of the intermediate state can be set to 20km / h. When the RNN algorithm is used to evaluate the scene jump (that is, the complexity of the scene, which can be determined by the evaluation value output by the RNN algorithm. The larger the evaluation value, the more complex the scene. When using the RNN algorithm, the road posture and environmental data can be used as input data, and the RNN algorithm is pre-trained using historical data), the lengthening multiple of the switching time window is set according to the scene jump, and the adjusted switching time window is obtained. It is used as the final switching time window and applied to the control parameter switching process. For example, assuming that the scene jump is 5, the preset scene jump is 3 times longer when the jump is 4~6, that is, the final switching time window is 3 times the length of the initial switching time window.

[0051] The above solution is to solve the problem that direct switching may cause large control parameter jumps in some complex scenarios. To this end, an intermediate state is introduced as a transition, and the switching time window is doubled to ensure smoothness.

[0052] Each mode switching event (including trigger conditions, switching time, weight changes, control parameter adjustments, etc.) is recorded in the log. The log format includes timestamp, input parameters, mode changes, switching smoothness score, etc.

[0053] Through the cognitive computing operation paradigm tree, combined with the real-time collected road posture, environmental data and vehicle status parameters, three driving modes (high-speed defense, baseline cruise, and environmental penetration speed) are dynamically selected. Each mode sets different weights and hard constraints for different scenarios (such as bad weather, high-speed cruise, and comfortable driving) to ensure the performance of the system in various complex scenarios. Maintain robustness in a variety of actual driving scenarios to meet the needs of different users (such as users who prioritize safety or users who prioritize efficiency).

[0054] The smooth switching of driving modes is achieved through the gradual state machine, including weight switching (linear interpolation) and control parameter switching (introduction of intermediate states). The length of the switching time window is dynamically adjusted according to the scene jump evaluated by the RNN algorithm to ensure a smooth switching process without mutations. The impact of driving mode switching on vehicle control is reduced, driving safety and comfort are improved, and the user experience is enhanced.

[0055] The design method of the multi-layer perception evaluation system includes: Step L1: construct a three-dimensional association network; With safety, efficiency and comfort as key dimensions, sub-dimensions under the key dimensions are preset, and the sub-dimensions are decomposed to form sub-sub-dimensions. The key dimensions, sub-dimensions and sub-sub-dimensions are distributed in the same level, and so on. All levels are filled. There are dimensions in each level. All dimensions are simulated into nodes. Any two nodes are connected by directed edges, and attribute values ​​are added to each directed edge to form a three-dimensional network structure. Among them, directed edges represent the mutual influence relationship between dimensions, from the influencer to the influenced. For example, if perception directly affects decision-making ability, then there is an edge. The direction of the edge is represented by perception pointing to decision-making ability, and the attribute value is used to represent the intensity of the influence, including positive influence and negative influence (that is, the positive influence intensity is represented by a positive number, or the negative influence intensity is represented by a negative number); For example, the multi-layer perception evaluation system includes three layers. The first layer is the key dimension node, including safety, efficiency, and comfort. The second layer is the sub-dimension node. For example, the sub-dimensions under safety include perception ability, decision-making ability, control ability, and fault tolerance. The sub-dimensions under perception ability are on the third layer, including object detection ability and dynamic target tracking ability. Among them, it is necessary to design the calculation method for each dimension. For example, perception ability refers to the accuracy and robustness of the system's perception of the environment. The object detection ability is reflected by the object recognition rate (the proportion of correctly identified targets), the false detection rate (the proportion of incorrectly identified targets), and the missed detection rate (the proportion of unrecognized targets). Specifically, in the scene, use labeled data sets (such as pedestrians, vehicles, and obstacles) for testing and calculate the above indicators.

[0056] Step L2: The last level of dimensions in the stereo association network defines quantifiable indicators, but these indicators are no longer measured independently, but are used as inputs in the network to calculate the performance of the entire system. For example, the object recognition rate (the proportion of correctly identified targets), false positive rate (the proportion of incorrectly identified targets), and missed detection rate (the proportion of unrecognized targets) in object detection capabilities are quantifiable indicators that are defined and used as inputs to the stereo association network; Initialize the attribute value of each directed edge and the energy value of each node, and use the influence evolution propagation mechanism to update the attribute value of the directed edge in the three-dimensional association network, so as to evolve the energy value of the node; The energy values ​​of the nodes corresponding to the key dimensions in the three-dimensional association network are obtained as the final output values, that is, the scores of safety, efficiency and comfort are obtained.

[0057] By building a multi-layer perception evaluation system and a three-dimensional association network, the technical solution can start from the three key dimensions of safety, efficiency and comfort, and gradually refine them into sub-dimensions and sub-sub-dimensions until they reach quantifiable indicators (such as emergency braking response time, path length optimization rate, etc.). This fine-grained decomposition makes the evaluation results more comprehensive and specific, avoiding the problem of fuzzy or overly general dimensions in traditional evaluation methods. Then, based on fine-grained indicators, performance bottlenecks (such as insufficient target detection recall rate of the perception module) can be located, thereby optimizing the system in a targeted manner.

[0058] The method of initializing the attribute value of each directed edge and the energy value of each node includes: Pre-train machine learning models (such as neural networks) in different modes, and use historical data to input into the models in the corresponding modes to obtain the attribute values ​​of each directed edge; The expert scoring method is used to obtain the initial score of each node and normalize it.

[0059] The influence evolution propagation mechanism is used to update the attribute values ​​of the directed edges in the stereo association network, thereby evolving the energy values ​​of the nodes. The method includes: Since the actual driving environment is not independent, the change of each dimension will cause ripples and exert influence on the surrounding dimensions. Therefore, when the attribute values ​​of the directed edges and the energy values ​​of the nodes are updated, a comprehensive propagation model needs to be designed. This design is based on the powerful expression ability of the graph neural network (GNN), and combines the functions of the gated propagation model, discrete time dynamic propagation model, energy propagation model and diffusion propagation model. Specifically, on the basis of the graph attention mechanism (GAT), the gated mechanism (used to dynamically adjust the influence intensity of the propagation to avoid excessive propagation or accumulation of negative influences), discrete time dynamics (used to simulate the dynamic evolution process of propagation and ensure the convergence of cyclic dependencies), energy propagation (used to introduce damping and saturation effects to simulate the energy loss in the propagation process) and diffusion propagation (used to introduce score difference driven smooth propagation to simulate the balance process between dimensions) will be integrated to form a comprehensive influence evolution propagation mechanism. The specific contents are as follows: The indicators of the last level in the stereo association network are collected as input, the evaluation period is preset, and it is discretized into time steps for dynamic iteration; At the current time step, the update direction is passed from the lowest level to the highest level in sequence. For any two nodes connected by a directed edge, the influence evolution propagation mechanism is used to update until the entire three-dimensional association network is updated. The iteration is repeated at the next time step until the preset evaluation cycle time is used up. The iteration stops and the energy value of each node and the directed edge attribute value are updated. Then the energy values ​​of the nodes corresponding to safety, efficiency and comfort on the first level are output as the scores of the key dimensions; Construct a comprehensive scoring function to evaluate the overall performance of the system, including normalizing the energy values ​​of the nodes corresponding to safety, efficiency, and comfort, and performing weighted summation according to the corresponding weights under the selected driving mode to obtain a comprehensive score; The scores of key dimensions and the comprehensive scores are output as the evaluation scores of the intelligent driving system.

[0060] Set up a variety of test scenarios, such as high-risk scenarios (such as pedestrians crossing, bad weather), regular scenarios (such as urban roads, highways) and low-speed scenarios (such as parking lots, congested roads), calculate scores for each test scenario, and analyze the performance differences of the system in different scenarios. For example, the system may perform well in highway scenarios, but perform poorly in congested roads.

[0061] The influence evolution propagation mechanism is designed based on the graph attention network, including: For any two nodes in the stereo association network, at each time step, according to the direction of the directed edge, the attention weights of the corresponding nodes of the influencer and the influenced are calculated using the graph attention network (GAT) (the specific calculation process is to use a learnable linear transformation matrix to map the energy value of the node to the high-order feature space, and then use the softmax function to calculate the attention weight of the node), indicating the importance of influence propagation as the new attribute value of the directed edge, and adding the saturation effect of energy propagation, the new energy value received by the corresponding node of the influenced is the weighted sum of the attention weights of all predecessor nodes (such as all corresponding influencer nodes of the influenced node), the initial attribute value of the directed edge, and the energy value; In the process of influence transmission, a gating mechanism is introduced to dynamically adjust the influence intensity of the transmission. The gating mechanism is similar to the update gate in GRU, which is used to control the fusion ratio of the new energy value and the current energy value. The specific method is that the update gate is the Sigmoid function value calculated by weighted sum of the current energy value and the new energy value, which is used to indicate the degree of acceptance of the new energy value; the candidate state is the hyperbolic tangent function value calculated by weighted sum of the current energy value and the new energy value, which is used to indicate the state after the new energy value is fully received. The weighted sum of the current energy value and the candidate state is used to update the node energy value. The weight is determined by the update gate to achieve the update of the node energy value. On the basis of updating the node energy value, the score difference drive of diffusion propagation is introduced to simulate the smooth propagation between dimensions, including obtaining the energy value difference between the nodes corresponding to the influencer and the influenced, weighted summing up the initial energy value of all predecessor nodes corresponding to the node of the influenced and the energy value difference, and then multiplying them with the diffusion coefficient (used to control the diffusion speed) to obtain the diffusion term, and using the updated energy value and the diffusion term to accumulate to get the corrected energy value of the node; On the basis of the node energy value correction, the damping effect of energy propagation is introduced to simulate the energy loss in the propagation process, including the introduction of the damping coefficient to control the energy attenuation speed. The damping coefficient is subtracted from 1 to obtain the absolute value. For the current time step, the final state of the energy value on the node in the next time step is the product of the corrected energy value and the absolute value. The saturation effect of energy propagation includes calculating the Sigmoid function value of the energy value of the corresponding node of the influencer and mapping it to the interval (0, 1) to simulate the saturation effect and avoid infinite growth of the energy value.

[0062] In the update of the three-dimensional association network, an influence evolution propagation mechanism based on the graph attention network is introduced, combined with saturation effect (sigmoid mapping), gating mechanism (update gate and candidate state), score difference drive (diffusion propagation) and damping effect. This mechanism can dynamically simulate the propagation and evolution of influence between dimensions.

[0063] The graph attention network captures the important influence relationships between dimensions through attention weights, avoiding the limitations of traditional equal propagation.

[0064] The saturation effect and damping effect prevent the energy value from being infinitely amplified or rapidly decayed, thus improving the stability of the model.

[0065] The gating mechanism and score difference drive enhance the nonlinear expression ability of the propagation process, making the evaluation results closer to the real scene.

[0066] Therefore, the complex interactions between modules in the intelligent driving system can be better simulated, significantly improving the authenticity and robustness of the evaluation.

[0067] Use graph visualization tools (such as Gephi, NetworkX) to draw a three-dimensional correlation network structure, display the relationship between nodes and edges and the performance of each dimension, and help developers and users intuitively understand system performance. The nodes in each level are the same size. Starting from the first level, the node shape becomes smaller. The thickness of the directed edge is drawn according to the size of the attribute value. Different colors are used to draw the directed edge, and the positive and negative effects are set according to the color; Use dynamic visualization tools to demonstrate the updating process of the stereo correlation network structure; The entire evaluation process is arranged on a visual interface for interaction with users.

[0068] Visual tools can be used to visually identify key trade-offs between dimensions. For example, if the attribute value of an edge is high and negative, it means that the improvement of security will lead to a significant decrease in efficiency, and the system design needs to be optimized to alleviate this conflict. If a node has a high impact weight on multiple nodes, but a low score, it means that the node is the bottleneck of the system and needs to be optimized first. If the attribute value of an edge is negative and large, it means that there is a significant conflict between the two, and the conflict needs to be alleviated through algorithm or policy adjustments. Example 2

[0069] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a multi-dimensional fine-grained intelligent driving system evaluation system, including: Data acquisition module: used to collect road posture, environmental data and vehicle status parameters during intelligent driving; Driving mode dynamic switching module: Dynamically selects three driving modes, namely high-speed defense, baseline cruise and environmental penetration speed, based on the cognitive computing operation paradigm tree, and uses a gradual state machine to smoothly switch the driving mode; Multi-dimensional fine-grained evaluation module: Based on each driving mode selected by switching, a multi-layer perception evaluation system is used to evaluate the intelligent driving system in terms of safety, efficiency, comfort and comprehensive score, and conduct visual interaction. Example 3

[0070] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the multi-dimensional fine-grained intelligent driving system evaluation method provided above is implemented.

[0071] Since the electronic device introduced in this embodiment is an electronic device used to implement a multi-dimensional fine-grained intelligent driving system evaluation method in the embodiment of the present application, based on the multi-dimensional fine-grained intelligent driving system evaluation method introduced in the embodiment of the present application, the technical personnel of the field can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not introduced in detail here. As long as the technical personnel of the field implement the electronic device used in the multi-dimensional fine-grained intelligent driving system evaluation method in the embodiment of the present application, it belongs to the scope of protection of this application.

[0072] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0073] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A multi-dimensional fine-grained intelligent driving system evaluation method, characterized in that: include: Collect road posture, environmental data and vehicle status parameters during intelligent driving, dynamically select three driving modes: high-speed defense, baseline cruise and environmental penetration speed based on cognitive computing operation paradigm tree, and use gradual state machine to smoothly switch driving modes; A multi-layer perception evaluation system is preset to switch driving modes in various test scenarios. Based on each selected driving mode, the multi-layer perception evaluation system is used to evaluate the safety, efficiency, and comfort scores of the intelligent driving system, and to perform visual interaction. Among them, the multi-layer perception evaluation system includes a three-dimensional association network, which forms a multi-level three-dimensional representation by refining the granularity of multiple dimensions, and forms a complete three-dimensional association network through node simulation and directed edge connection; For any node in the three-dimensional association network, the influence evolution propagation mechanism is used to update it. The influence evolution propagation mechanism is designed based on the graph attention network, and the saturation effect of energy propagation, the gating mechanism, the score difference drive of diffusion propagation and the damping effect of energy propagation are introduced to achieve the real diffusion of the influence interaction between evolution dimensions.

2. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 1, characterized in that: The three driving modes of high-speed defense, baseline cruising and environmental penetration speed include: In each driving mode, a weight value is used to quantify the priority of each key dimension, and the weight value corresponding to the main dimension is the largest. Among them, the key dimensions are safety, efficiency and comfort, and the main dimension is the most prioritized key dimension in each driving mode; Hard constraints for key indicators are set for each driving mode to meet the minimum performance requirements in any situation. For the high-speed defense mode, the safety constraints are set to emergency braking response time less than A1, target detection recall rate greater than B1, and braking distance less than C1; the efficiency constraints are set to path length optimization rate greater than D1, and the comfort constraints are set to acceleration change rate less than E1; For the baseline cruise mode, the safety constraints are set as emergency braking response time less than A2, target detection recall rate greater than B2, and braking distance less than C2; the efficiency constraints are set as path length optimization rate greater than D2; and the comfort constraints are set as acceleration change rate less than E2. For the environmental penetration speed mode, the safety constraints are set as the emergency braking response time less than A3, the target detection recall rate greater than B3, and the braking distance less than C3; the efficiency constraints are set as the path length optimization rate less than D3; and the comfort constraints are set as the acceleration change rate less than E3; Among them, A1, A2, A3, B1, B2, B3, C1, C2, C3, D1, D2, D3, E1, E2 and E3 are pre-set values, and the values ​​meet the requirements of A1<A2<A3; B1>B2>B3; C1>C2>C3; D1>D2>D3; E1>E2>E3.

3. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 2, characterized in that: The method for collecting road posture, environmental data and vehicle status parameters during intelligent driving and dynamically selecting three driving modes of high-speed defense, baseline cruising and environmental penetration speed based on a cognitive computing operation paradigm tree includes: Road posture data include slope, curvature and road friction coefficient, environmental data include weather data, visibility and traffic density, and vehicle state parameters include vehicle speed and acceleration; The cognitive computing operation paradigm tree is set to consist of a root node and intermediate nodes, and the input is road posture, environmental data and vehicle state parameters, where the root node represents the driving mode selection and the intermediate node is used for subtask decomposition; At each intermediate node, the cognitive computing model is used to reason about the input data and output the priority and confidence of the subtask; The set driving modes include high-speed defense, baseline cruising and environmental penetration speed. Among them, the high-speed defense mode prioritizes safety, and the triggering conditions are that the slope in the road posture is greater than q1, the curvature is greater than s1, and the road friction coefficient is less than d1; the weather data in the environmental data is bad weather, visibility is less than n1, and traffic density is greater than g1; the vehicle state is the speed of the vehicle is greater than c1, and the acceleration change rate is greater than w1; The baseline cruise mode is efficiency-first, and the triggering conditions are that the slope is ≤q1, the curvature is ≤s1, ​​and the road friction coefficient is ≥d1 in the road posture; the weather data in the environmental data is moderate weather, visibility is ≥n1, and traffic density is ∈ (g2, g1); the vehicle speed is ∈ (c2, c1), and the acceleration change rate is ≤w1 in the vehicle state; The environmental penetration speed mode is comfort-first, and the triggering conditions are: slope ≤ q2, curvature > s2 in road posture; visibility < n2, traffic density > g3 in environmental data; vehicle speed < c2, acceleration change rate ≤ w2 in vehicle status; Wherein, q1>q2, s1<s2; d1; n1>n2; g1>g2>g3; c1>c2; w1>w2, and q1, q2, s1, s2, d1, n1, n2, g1, g2, g3, c1, c2, w1 and w2 are all preset values; The vehicle's road posture, environmental data and vehicle status parameters are collected in real time at a fixed frequency hb, and input into the cognitive computing operation paradigm tree to output the selected driving mode. If the driving mode currently output is the same as the previous output, there is no need to switch the driving mode. If the driving mode currently output is different from the previous output, a gradual state machine is used to smoothly switch the driving mode.

4. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 3, characterized in that: The method of using a gradual state machine to smoothly switch the driving mode includes: Smooth switching includes switching the weight of driving modes and switching the control parameters of the vehicle; For the weight switching of driving modes, the weights of safety, efficiency and comfort in each driving mode are preset, and a switching time window is added. Each weight is gradually adjusted using linear interpolation within the window until the preset weight value in the driving mode is reached; For the vehicle control parameter switching, an intermediate state is introduced in the middle of the switching. The intermediate state is the intermediate value of the control parameters in the two modes. The control parameters include vehicle speed, acceleration and following distance. The RNN algorithm is used to evaluate the scene jump. The lengthening multiple of the switching time window is set according to the scene jump. The adjusted switching time window is obtained and used as the final switching time window and applied to the control parameter switching process.

5. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 4, characterized in that: The design method of the multi-layer perception evaluation system includes: Step L1: construct a three-dimensional association network; With safety, efficiency and comfort as key dimensions, sub-dimensions under the key dimensions are preset, and the sub-dimensions are decomposed to form sub-sub-dimensions. The key dimensions, sub-dimensions and sub-sub-dimensions are distributed in the same level, and so on. All levels are filled. There are dimensions in each level. All dimensions are simulated into nodes. Any two nodes are connected by directed edges, and attribute values ​​are added to each directed edge to form a three-dimensional network structure. Among them, directed edges represent the mutual influence relationship between dimensions, from the influencer to the influenced, and the attribute value is used to represent the intensity of the influence, including positive and negative influences; Step L2: The dimensions of the last level in the stereo association network define quantifiable indicators, and these indicators are used as inputs of the stereo association network; Initialize the attribute value of each directed edge and the energy value of each node, and use the influence evolution propagation mechanism to update the attribute value of the directed edge in the three-dimensional association network, so as to evolve the energy value of the node; Obtain the energy value of the node corresponding to the key dimension in the stereo association network as the final output value.

6. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 5, characterized in that: The method for initializing the attribute value of each directed edge and the energy value of each node includes: Pre-train machine learning models in different modes, and use historical data to input into the models in the corresponding modes to obtain the attribute values ​​of each directed edge; The expert scoring method is used to obtain the initial score of each node and normalize it.

7. The multi-dimensional fine-grained intelligent driving system evaluation method according to claim 6 is characterized in that: The method of using the influence evolution propagation mechanism to update the attribute value of the directed edge in the stereo association network, thereby evolving the energy value of the node, includes: The indicators of the last level in the stereo association network are collected as input, the evaluation period is preset, and it is discretized into time steps for dynamic iteration; At the current time step, the update direction is passed from the lowest level to the highest level in sequence. For any two nodes connected by a directed edge, the influence evolution propagation mechanism is used to update until the entire three-dimensional association network is updated. The iteration is repeated at the next time step until the preset evaluation cycle time is used up. The iteration stops and the energy value of each node and the directed edge attribute value are updated. Then the energy values ​​of the nodes corresponding to safety, efficiency and comfort on the first level are output as the scores of the key dimensions; Construct a comprehensive scoring function to evaluate the overall performance of the system, including normalizing the energy values ​​of the nodes corresponding to safety, efficiency, and comfort, and performing weighted summation according to the corresponding weights under the selected driving mode to obtain a comprehensive score; The scores of key dimensions and the comprehensive scores are output as the evaluation scores of the intelligent driving system.

8. The multi-dimensional fine-grained intelligent driving system evaluation method according to claim 7 is characterized in that: The influence evolution propagation mechanism is designed based on the graph attention network, including: For any two nodes in the stereo association network, at each time step, according to the direction of the directed edge, the graph attention network is used to calculate the attention weights of the corresponding nodes of the influencer and the influenced as the new attribute value of the directed edge, and the energy propagation and saturation effect are added. The new energy value received by the corresponding node of the influenced is the weighted sum of the attention weights of all predecessor nodes, the initial attribute value of the directed edge, and the energy value; In the process of influence transmission, a gating mechanism is introduced to control the fusion ratio of the new energy value and the current energy value. The specific method is that the update gate is the Sigmoid function value calculated by weighted sum of the current energy value and the new energy value; the candidate state is the hyperbolic tangent function value calculated by weighted sum of the current energy value and the new energy value, and the node energy value is updated using the weighted sum of the current energy value and the candidate state, with the weight determined by the update gate; On the basis of updating the node energy value, the score difference drive of diffusion propagation is introduced, including obtaining the energy value difference between the nodes corresponding to the influencer and the influenced, weighted summing up the initial energy value of all predecessor nodes corresponding to the node of the influenced and the energy value difference, and then multiplying them by the diffusion coefficient to obtain the diffusion term, and using the updated energy value and the diffusion term to accumulate to get the corrected energy value of the node; On the basis of the correction of the node energy value, the damping effect of energy propagation is introduced, including the introduction of the damping coefficient. The damping coefficient is subtracted from 1 to obtain the absolute value. For the current time step, the final state of the energy value on the node in the next time step is the product of the corrected energy value and the absolute value.

9. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 8, characterized in that: The saturation effect of energy propagation includes calculating the Sigmoid function value of the energy value of the node corresponding to the influencer and mapping it to the interval (0, 1).

10. A multi-dimensional fine-grained intelligent driving system evaluation method according to claim 9, characterized in that: The visual interaction method comprises: Use graph visualization tools to draw a three-dimensional correlation network structure. The nodes in each level are the same size. Starting from the first level, the node shapes become smaller. The thickness of the directed edges is drawn according to the size of the attribute value. Directed edges are drawn with different colors, and positive and negative effects are set according to the color. Use dynamic visualization tools to demonstrate the updating process of the stereo correlation network structure; The entire evaluation process is arranged on a visual interface for interaction with users.

Images