Transferable utility-based intelligent automobile lane change evaluation index dynamic construction method

Abstract

The invention belongs to the technical field of intelligent automobile automatic test and evaluation, and particularly relates to an intelligent automobile lane change evaluation index dynamic construction method based on transferable utility, which comprises the following steps: constructing an expert dynamic game framework based on an improved Delphi method, screening core evaluation indexes through three-wheel anonymous feedback iteration, and constructing an expert dynamic game framework; a multi-dimensional evaluation index is constructed; based on the constructed multi-dimensional evaluation indexes, constructing a feature function of the transferable utility alliance by adopting an improved Shapley value algorithm; and identifying a typical scene to which the current vehicle currently belongs, analyzing a driving intention, activating a corresponding priority strategy based on a current scene state, and adjusting the weight of a multi-dimensional evaluation index in real time by reconstructing a transferable utility alliance. Compared with the prior art, the method can effectively coordinate the balance relation between the safety and the passing efficiency, and provides a more accurate quantitative evaluation means for verification of an intelligent automobile lane changing decision algorithm.

Description

Technical Field

[0001] This invention relates to the field of automated testing and evaluation technology for intelligent vehicles, and in particular to a dynamic fusion method for multi-dimensional evaluation indicators of lane-changing behavior of intelligent vehicles in complex traffic scenarios. Background Technology

[0002] With the rapid development of intelligent driving technology, lane changing, as a core function of autonomous driving, directly impacts the road adaptability of intelligent vehicles due to the completeness of its behavior evaluation system. Current technologies for evaluating lane changing behavior still have significant shortcomings:

[0003] ① Limited Evaluation Dimensions: Traditional methods often focus on static threshold determination of a single indicator (such as collision risk or lane change time), lacking a systematic and coordinated assessment of safety, efficiency, comfort, intelligence, and reliability. For example, existing technologies only determine safety based on longitudinal distance thresholds, neglecting the impact of lateral offset dynamics on overall safety; some studies use fixed weights to fuse multi-dimensional indicators, but fail to consider the dynamic coupling relationship of multi-objective optimization in complex traffic scenarios, leading to evaluation results that deviate from actual driving needs.

[0004] ② The static weight allocation mechanism lacks adaptability: Mainstream evaluation systems rely on expert experience to pre-set weights, and their static characteristics make it difficult to adapt to real-time changes in dynamic traffic flow. When encountering sudden scenarios such as emergency vehicles cutting in, evaluation models based on fixed weights will experience decision-making delays due to the increased conflict between safety and comfort indicators, which seriously affects the system's response timeliness.

[0005] ③ Lack of dynamic coordination in multi-objective optimization: Existing methods generally employ linear weighted summation when dealing with conflicting objectives such as safety and traffic efficiency, comfort and response speed, leading to evaluation results falling into local optima. For example, in low-visibility weather, when trajectory prediction accuracy decreases significantly, the system forces a speed reduction to maintain safety, but this causes a sharp drop in traffic efficiency, exposing the limitations of static evaluation logic.

[0006] Despite recent attempts to introduce methods such as fuzzy comprehensive evaluation and neural network fusion, problems remain, including high redundancy of indicators and exponentially increasing computational complexity. Especially in complex traffic scenarios, existing technologies cannot dynamically prioritize multi-dimensional indicators, making it difficult for intelligent vehicle lane-changing decision-making algorithms to meet practical needs in terms of verification efficiency and accuracy. Therefore, constructing a dynamic evaluation system with scene-adaptive capabilities has become a key challenge in overcoming the bottlenecks in the implementation of intelligent driving technology. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology, such as the single evaluation dimension, the inability of static weight allocation to adapt to dynamic scenarios, and the prominent contradictions in multi-objective optimization, and to provide a dynamic fusion method for multiple evaluation indicators of intelligent vehicle lane changing based on transferable utility.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] A method for dynamically constructing lane-changing evaluation indicators for intelligent vehicles based on transferable utility, comprising the following steps:

[0010] An expert dynamic game framework is constructed based on the improved Delphi method. Core evaluation indicators are screened through three rounds of anonymous feedback iteration, and multi-dimensional evaluation indicators are constructed.

[0011] Based on the constructed multidimensional evaluation index, the characteristic function of the transferable utility alliance is constructed using the improved Shapley value algorithm.

[0012] Identify the typical scenario of the current vehicle and analyze the driving intention. Activate the corresponding priority strategy based on the current scenario state, and adjust the weights of multi-dimensional evaluation indicators in real time by reconstructing the transferable utility alliance.

[0013] As a preferred technical solution, the core evaluation indicators are iteratively screened based on the improved Delphi method, and a multi-dimensional evaluation indicator is constructed, as follows:

[0014] Establish a dynamic game framework through an encrypted online collaboration platform;

[0015] The first round of research used a structured questionnaire to collect expert scores on the importance of several preliminary indicators;

[0016] In the second round of anonymous feedback, the entropy weight method was used to calculate the information entropy of each preliminary indicator to quantify the sensitivity of the indicator. The coefficient of variation method was combined to evaluate the data dispersion and select several candidate indicators.

[0017] In the third round, a judgment matrix was constructed using fuzzy hierarchical analysis to obtain the correlation matrix among candidate indicators, and the final multidimensional evaluation index was obtained through screening.

[0018] As a preferred technical solution, the constructed multidimensional evaluation index includes:

[0019] In terms of safety, the secondary parameters include collision risk and lateral offset;

[0020] In terms of efficiency, the secondary parameters include lane change time and speed maintenance.

[0021] In terms of comfort, secondary parameters include longitudinal impact and yaw fluctuation;

[0022] In terms of intelligence, the secondary parameters include intent recognition rate and trajectory prediction accuracy;

[0023] In terms of reliability, secondary parameters include system response latency and fault recovery rate.

[0024] As a preferred technical solution, the first round of research on the improved Delphi method introduced a time delay compensation mechanism: when the expert feedback interval exceeds the set time, the system automatically pushes the index weight reference value generated by the LSTM prediction model trained on historical data.

[0025] As a preferred technical solution, the third round of the improved Delphi method introduces a weight adaptive module. For two indicators with implicit conflicts, when the implicit correlation exceeds a threshold indicator, a weight correction factor is calculated to correct the initial weights of the fuzzy hierarchical analysis method (FAHP). The specific steps are as follows:

[0026] Calculate the covariance matrix of latent associations exceeding the threshold index, and perform eigenvalue decomposition on the covariance to obtain the eigenvector matrix and the eigenvalue diagonal matrix;

[0027] The original indices are projected onto the principal component space to eliminate linear correlation.

[0028] Based on the eigenvalues ​​of the eigenvector matrix and the eigenvalue diagonal matrix, the weight correction factor is further obtained:

[0029]

[0030] The initial weights of the fuzzy analytic hierarchy process (FAHP) are corrected based on the aforementioned weight correction factors:

[0031] w′1=w1(1+β),w′2=w2(1+α)

[0032] Where λ1 and λ2 are the eigenvalues ​​of the eigenvector matrix and the eigenvalue diagonal matrix, respectively; α and β are weight correction factors; and w1, w′1, w2, and w′2 are the initial and corrected weights of two indicators with implicit conflicts, respectively.

[0033] As a preferred technical solution, the characteristic function of constructing the transferable utility consortium using the improved Shapley value algorithm is expressed as follows:

[0034] v(S)=∑ i∈S λ(t)x′ i +∑ (i,j)∈S×S η ij CMI ij ·x′ i x′ j

[0035] Where, x′ i 、x′ jη represents the normalized i-th and j-th dimension evaluation index values, respectively; λ(t) is the time-varying correction factor; η ij For nonlinear synergistic gain coefficients; CMI ij Let be the conditional mutual information entropy between multidimensional indices i and j.

[0036] As a preferred technical solution, the characteristic function of the transferable utility alliance includes:

[0037] The time-varying correction factor is expressed as:

[0038]

[0039] Where T is the Delphi round interval;

[0040] The conditional mutual information entropy is obtained by calculating the collaborative relationship of analytical indicators using a dynamic Bayesian network model: A third-order Bayesian network is constructed with multidimensional evaluation indicators as latent variables and vehicle state data as observed variables. The conditional mutual information entropy is then obtained through iterative optimization.

[0041] CMI(X i X j |X k )=H(X i |X k )+H(X j |X k )-H(X i X j |X k )

[0042] Among them, CMI(X) i X j |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i and the j-th dimension evaluation index variable X j The degree of statistical dependence between them; H(X) i X j |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i and the j-th dimension evaluation index variable X j The joint uncertainty of H(X); i |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i The uncertainty is explained by the following formula:

[0043]

[0044] Where p(x) i x k ) is X i and X k The joint probability distribution of p(x); i |x k ) is given X k Time X i The conditional probability distribution;

[0045] The nonlinear synergistic gain coefficient η is expressed as:

[0046]

[0047] CMI represents conditional mutual information entropy.

[0048] As a preferred technical solution, the real-time adjustment of the weights of the multi-dimensional evaluation indicators is as follows:

[0049] A hidden Markov model is constructed to identify typical scenarios, including low-speed lane changing, high-speed lane changing, and emergency obstacle avoidance; a hidden state set is defined using time-series data from vehicle sensors, and driving intentions are analyzed based on the observation symbol probability matrix.

[0050] When identifying trigger mode switching in typical scenarios, the Shapley values ​​of each dimension of the multidimensional evaluation index are calculated in real time based on the currently active index consortium S:

[0051]

[0052] Where S represents the currently active indicator consortium, which changes with the driving scenario; N{i} represents the set of all possible indicators except indicator i; n represents the total number of dimensions of the indicator set N; v(S active ∪{i})-v(S active ) represents the marginal contribution of indicator i, which is the contribution of indicator i to joining the coalition S. active Marginal contribution at time indicates the increase in alliance revenue after indicator i joins;

[0053] The dynamic weight allocation is as follows:

[0054]

[0055] Wherein, γ is the scene sensitivity factor; J represents the overall system performance index in the current scenario. scene Weight W of the i-th dimension evaluation index i The partial derivatives of .

[0056] As a preferred technical solution, the priority strategy activated based on the current scene state is as follows:

[0057] When an emergency obstacle avoidance situation is detected, the safety-intelligence indicator priority strategy is activated, and the emergency obstacle avoidance indicator coalition is set to S. EOA ={S a , E, I, R}; Feature function calls include: security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of intelligence I is v(S∪I)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S); the Shapley values ​​of each dimension index are calculated.

[0058] When a low-speed lane change is detected, the comfort-first strategy is activated, and the low-speed lane change scenario indicator coalition is set to S. LLC ={S a ,E,C,R}; Feature function calls include: security S a Marginal contribution v(S∪S) a )-v(S), marginal contribution of efficiency E v(S∪E)-v(S), marginal contribution of comfort C v(S∪C)-v(S), marginal contribution of reliability R v(S∪R)-v(s);

[0059] When a high-speed lane-changing situation is detected, the efficiency-first strategy is activated, and the high-speed lane-changing scenario metric coalition is S. HLC ={S a , R, C, I, R}; Feature function calls include: security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of comfort C is v(S∪C)-v(S), the marginal contribution of intelligence is v(S∪I)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S).

[0060] As a preferred technical solution, during the real-time adjustment of the weights of the multidimensional evaluation indicators, a second-order inertial filter is used to avoid oscillations in the evaluation model caused by sudden changes in the Shapley value.

[0061] α t+1 =K1α t +K2(α target -α t )+K3α t-1

[0062] K1+K2+K3=I,||K2||2<1

[0063] Where, αt Let α represent the weight vector of the multidimensional evaluation index at the current time t. t+1 This represents the filtered weight vector at the next time step t+1, α t-1 K1 represents the historical weight vector at the previous time t-1; K2 represents the historical weight inertia coefficient, K3 represents the target weight tracking coefficient, and K3 represents the second-order historical inertia coefficient.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1) This invention constructs a dynamic fusion framework for multiple evaluation indicators based on transferable utility theory. At the indicator processing level, an improved Delphi method is used to select 12 core evaluation indicators, constructing a five-dimensional evaluation system encompassing safety, efficiency, comfort, intelligence, and reliability. Furthermore, the transferable utility calculation method in alliance game theory is utilized to achieve real-time dynamic fusion of the multi-dimensional evaluation indicators. Verified on a hardware-in-the-loop testing platform for intelligent vehicle algorithms, it effectively coordinates the balance between safety and traffic efficiency, providing a more accurate quantitative evaluation method for verifying intelligent vehicle lane-changing decision-making algorithms.

[0066] 2) This invention improves upon the Delphi method by introducing a time delay compensation mechanism in the first round of research. When the expert feedback interval exceeds a set interval, the reference values ​​of the indicator weights generated by the prediction model are used to avoid cognitive lag caused by asynchronous interaction. In the final round, a weight adaptive module is introduced. When an implicit correlation is detected between trajectory prediction accuracy in the intelligence dimension and system response delay in the reliability dimension, an orthogonalization correction factor is automatically injected for compensation, thus balancing the intelligence and reliability indicators.

[0067] 3) This invention employs an improved Shapley value algorithm to construct the TU-games feature function. A time-varying correction factor is introduced to compensate for the iteration delay of the Delphi method; subsequently, a dynamic Bayesian network (DBN) model is established to quantify the synergistic effect between indicators, and conditional mutual information entropy is used to identify the positive coupling relationship between efficiency and comfort, and between security and reliability. Furthermore, an alliance cooperation gain coefficient is defined, breaking through the traditional linear superposition assumption.

[0068] 4) This invention constructs a dynamic evaluation scheme with scene adaptation capabilities. Scene recognition is achieved through a Hidden Markov Model (HMM), and driving intentions are analyzed based on the observed symbolic probability matrix. An indicator consortium is activated based on the current driving scene, and the Shapley values ​​of each indicator are calculated in real time, with weights dynamically allocated. When an emergency obstacle avoidance state is detected, a safety-first strategy is activated. The objective function is dynamically calibrated using a forward-backward algorithm, and the five-dimensional weights are adjusted in real time by reconstructing a transferable utility consortium. This enables dynamic priority reconstruction of multi-dimensional indicators, improving the verification efficiency and accuracy of the intelligent vehicle lane-changing decision algorithm. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the dynamic fusion method of multiple evaluation indicators for lane changing in intelligent vehicles based on transferable utility, as proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0071] Example 1

[0072] This invention proposes a dynamic weight allocation method for multi-dimensional evaluation indicators of lane-changing behavior in intelligent vehicles, integrating the Delphi method and transferable utility. At the indicator level, an improved Delphi method is used to select 12 core evaluation indicators, constructing a five-dimensional evaluation system encompassing safety, efficiency, comfort, intelligence, and reliability. Furthermore, a dynamic fusion framework for multiple evaluation indicators based on transferable utility theory is built, utilizing transferable utility calculation methods from alliance game theory to achieve real-time dynamic fusion of multi-dimensional evaluation indicators. Theoretical closed-loop verification is completed in the decision verification scenario of an intelligent vehicle algorithm hardware-in-the-loop testing platform. The process of this invention is as follows: Figure 1 As shown, the specific steps are as follows:

[0073] Step 1: Test Scenario Setup and Data Collection

[0074] A multi-objective game-theoretic test matrix was built based on a hardware-in-the-loop testing platform for intelligent vehicle algorithms. Three typical scenarios were set up, covering low-speed lane changing, high-speed lane alternation, and emergency obstacle avoidance, encompassing dynamic operating conditions with vehicle speeds of 30-120 km / h and adhesion coefficients of 0.3-0.9. Key vehicle state variables such as steering wheel angle, yaw rate, acceleration, and velocity were configured, and vehicle dynamic response data were synchronously collected at a sampling frequency of 100Hz. A data acquisition module was also integrated to obtain real-time trajectory information of surrounding traffic participants, constructing a raw database containing intelligent vehicle testing and evaluation data.

[0075] Step 2: Selecting core evaluation indicators

[0076] An expert dynamic game framework was constructed based on the improved Delphi method, integrating a time delay compensation mechanism and a weight adaptive module. Cognitive bias was eliminated through three rounds of anonymous feedback iteration. In collaboration with 15 experts in the field of intelligent driving (including automotive R&D, university scholars, and traffic regulation drafters), the entropy weight method was used to quantify the sensitivity of indicators, and the coefficient of variation method was combined to evaluate data dispersion. Ultimately, five primary indicators and 12 secondary parameters covering safety (collision risk, lateral drift), efficiency (lane change time, speed maintenance), comfort (longitudinal impact, yaw fluctuation), intelligence (intent recognition rate, trajectory prediction accuracy), and reliability (system response delay, fault recovery rate) were selected. The Pearson correlation between the indicators was all below 0.35, satisfying the independence constraint.

[0077] Step 3: Determine the characteristic function in the coalition game

[0078] Based on the selected five-dimensional evaluation indexes, an improved Shapley value algorithm is used to construct the characteristic function of the transferable utility alliance game TU-games. First, the independent value of the indexes is normalized by the range method, and a time-varying correction factor (λ = 0.82 ± 0.05) is introduced to compensate for the iteration delay of the Delphi method. Then, a dynamic Bayesian network (DBN) model is established to quantify the synergistic effect between the indexes. The conditional mutual information entropy (CMI > 0.65) is used to identify the positive coupling relationship between effectiveness and comfort, and between security and reliability. The alliance cooperation gain coefficient η = 1.3-2.1 is defined, breaking through the traditional linear superposition assumption.

[0079] Step 4: Dynamic Weight Update Mechanism

[0080] A Hidden Markov Model (HMM) is constructed to achieve scene recognition. A set of hidden states (low-speed lane change, high-speed lane change, emergency obstacle avoidance, etc.) is defined using time-series data from onboard sensors (steering wheel angle, yaw rate, etc.), and driving intentions are analyzed based on the observed symbolic probability matrix. When an emergency obstacle avoidance state is detected, a safety-first strategy is activated. The objective function is dynamically calibrated using a forward-backward algorithm, and the weights of the five-dimensional evaluation index are adjusted in real time through transferable utility (TU) alliance reconstruction: safety weights are significantly enhanced, intelligence weights are simultaneously improved, and efficiency and comfort weights are moderately reduced, ensuring a smooth transition during strategy switching.

[0081] The specific implementation process of each step is as follows:

[0082] Step 1: Test scenario setup and data collection.

[0083] On the intelligent vehicle algorithm hardware-in-the-loop testing platform, a multi-objective game-theoretic testing framework was first built based on a real-time simulation system, integrating an industrial control computer, vehicle dynamics model interface, and sensor model interface. In terms of test matrix design, three typical scenarios were constructed using the SCANeR scene editor: low-speed lane changing (initial following distance 15m, target vehicle speed 30-50km / h), high-speed lane changing (cut-in angle rate 0.15-0.35rad / s, vehicle speed 80-120km / h), and emergency obstacle avoidance (obstacle generation response time ≤0.3s). The adhesion coefficient was dynamically adjusted using the tire magic formula (0.3 corresponds to icy / snowy roads, 0.9 corresponds to dry asphalt) to achieve high-precision simulation of the vehicle-environment coupling effect.

[0084] In the data acquisition phase, a NIPXIe-1082 chassis equipped with a PCI-CAN module was used, along with a steering torque sensor, an accelerometer, and virtual scene software output information. Key vehicle state variables such as steering wheel angle, yaw rate, acceleration, and speed were simultaneously captured at a 100Hz sampling rate, along with traffic environment information surrounding the test vehicle, including traffic vehicle position, speed, and yaw rate. In the database construction phase, a sliding window-based outlier removal algorithm (window length 50ms) and a Kalman filter data fusion module were developed, ultimately forming a structured dataset containing timestamps, scene labels, vehicle dynamics states, and environmental topology information.

[0085] Step 2: Selecting core evaluation indicators

[0086] In improving the implementation of the Delphi method, a cross-disciplinary expert group of 15 people, consisting of R&D engineers from car companies, autonomous driving research teams from universities, and experts from traffic management departments, was first formed. A dynamic game framework was established through an encrypted online collaboration platform. In the first round of research, a structured questionnaire was used to collect experts' importance scores for 62 preliminary indicators (a Likert scale of 1-9). A time delay compensation mechanism was simultaneously activated—when the interval between expert feedback exceeded 72 hours, the system automatically pushed reference values ​​of indicator weights generated by an LSTM prediction model trained on historical data, avoiding cognitive lag caused by asynchronous interaction. In the second round of anonymous feedback, the entropy weight method was used to calculate the information entropy of each indicator (H≥0.75 was judged as a highly sensitive parameter), and the coefficient of variation method (CV<0.15 was judged as controllable data dispersion) was used to screen out 18 candidate indicators. A heatmap of expert disagreement (disputed items with a coefficient of variation >0.3 were marked in red) was attached to guide targeted discussions. In the final round, a judgment matrix is ​​constructed using fuzzy hierarchical analysis (FAHP). The Group Decision Support System (GDSS) then displays the Pearson correlation matrix between indicators in real time (threshold set r < 0.35). When an implicit correlation (r = 0.41) is detected between "trajectory prediction accuracy" in the intelligence dimension and "system response delay" in the reliability dimension, the weight adaptive module automatically injects an orthogonalization correction factor. The weight adaptive module resolves implicit indicator conflicts. For example, improving "trajectory prediction accuracy" in the intelligence dimension relies on more complex artificial intelligence algorithms, which requires more computation, leading to an increase in "system response delay" in the reliability dimension. Directly assigning weights based on expert scores might ignore objective engineering constraints. The weight adaptive module explicitly decouples such conflicts, as shown below: First, based on expert score data, the covariance matrix of indicators with implicit correlations exceeding the threshold is calculated:

[0087] σ 12 =r·σ 21

[0088] Where σ1 represents the score variance of trajectory prediction accuracy, σ2 represents the score variance of system response delay, and P represents the covariance of the two indicators. Eigenvalue decomposition is performed on this covariance to obtain the eigenvector matrix U and the eigenvalue diagonal matrix Λ. Then, the original indicators are projected onto the principal component space to eliminate linear correlation. The eigenvalues ​​of the eigenvector matrix U and the eigenvalue diagonal matrix Λ are λ1 and λ2, respectively, and the weight correction factors are further obtained as follows:

[0089]

[0090] The initial weights w of the fuzzy analytic hierarchy process (FAHP) are corrected based on the aforementioned weight correction factors:

[0091] w′1=w1(1+β), w′2=w2(1+α)

[0092] Wherein, w1 and w′1 are the initial and corrected weights for trajectory prediction accuracy, respectively; w2 and w′2 are the initial and corrected weights for system response delay, respectively. Finally, five primary indicators and twelve secondary parameters covering safety (collision risk, lateral drift), efficiency (lane change time, speed maintenance), comfort (longitudinal impact, yaw fluctuation), intelligence (intent recognition rate, trajectory prediction accuracy), and reliability (system response delay, fault recovery rate) were selected. The effectiveness of the expert consensus was confirmed by the Kendall coordination coefficient test (W = 0.86), laying the foundation for subsequent alliance game modeling.

[0093] Step 3: Determine the characteristic function in the coalition game

[0094] In constructing the characteristic function of the coalition game, the five-dimensional evaluation indicators are first preprocessed and their value quantified. Based on the working condition data in the database, the range method is used to normalize the 12 secondary indicators: for positive indicators (such as trajectory prediction accuracy), the following is used:

[0095]

[0096] Negative indicators are used as follows:

[0097]

[0098] Where x is the original evaluation index, x′ is the normalized index, and x max x is the maximum value of the indicator in the database. min The minimum value of the indicator in the database is used to eliminate dimensional differences.

[0099] To compensate for the time delay deviation caused by multiple iterations of the Delphi method, a time-varying correction factor is designed.

[0100]

[0101] Where T is the Delphi round interval, the normalization results are dynamically weighted by a sliding time window (window length Δt = 2 rounds) so that the decay rate of early expert opinions matches the data update cycle.

[0102] Subsequently, a dynamic Bayesian network model was established to analyze the collaborative relationships of the indicators. Using the five-dimensional evaluation index as the latent variable and vehicle state data as the observed variable, a third-order Bayesian network was constructed. After iterative optimization using the EM algorithm, the conditional mutual information entropy was calculated.

[0103] CMI(X i X j |X k )=H(X i |X k)+H(X j |X k )-H(X i X j |X k )

[0104] Among them, CMI(X) i X j |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i and the j-th dimension evaluation index variable X j The degree of statistical dependence between them; H(X) i X j |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i and the j-th dimension evaluation index variable X j The joint uncertainty of H(X); i |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i The uncertainty is explained by the following formula:

[0105]

[0106] Where p(x) i ,x k ) is X i and X k The joint probability distribution of p(x); i |x k ) is given X k Time X i The conditional probability distribution.

[0107] Strong synergistic pairs between performance and comfort, as well as cross-domain couplings between safety and reliability, are identified. Based on this, a nonlinear synergistic gain coefficient η is defined:

[0108]

[0109] The hyperbolic tangent function constrains the gain range to 1.3-2.1, and the gain sensitivity is verified by Monte Carlo simulation.

[0110] Finally, the feature function of the Shapley value algorithm is constructed:

[0111] v(S)=∑ i∈S λ(t)x′ i +∑ (i，j)∈S×S ηij CMI ij ·x′ i x′ j

[0112] Where, x′ i 、x′ j η represents the normalized i-th and j-th dimension evaluation index values, respectively; λ(t) is the time-varying correction factor; η ij For nonlinear synergistic gain coefficients; CMI ij Let be the conditional mutual information entropy between multidimensional indices i and j.

[0113] By exhaustively enumerating 31 non-empty alliance combinations, the parameters are optimized using gradient descent (learning rate 0.01, 500 iterations) to make the alliance payoff matrix satisfy superadditivity (success rate > 92%).

[0114] Step 4: Dynamic Weight Update Mechanism

[0115] A Hidden Markov Model (HMM) is constructed to recognize typical scenarios. A set of hidden states (low-speed lane change, high-speed lane change, emergency obstacle avoidance, etc.) is defined using time-series data from onboard sensors (steering wheel angle, yaw rate, etc.), and driving intentions are analyzed based on the observed symbolic probability matrix. When an emergency obstacle avoidance state is detected, a safety-first strategy is activated; when a low-speed lane change is detected, a comfort-first strategy is activated, as drivers are more sensitive to jerking at low speeds; when a high-speed lane change is detected, an efficiency-first strategy is activated to quickly complete the lane change and minimize the impact on traffic flow. Then, the weights of the five-dimensional evaluation index are adjusted in real time through transferable utility (TU) alliance reconstruction. This adjustment process is shown below:

[0116] 4.1. When identifying trigger mode switching in typical scenarios, based on the currently active indicator alliance Real-time calculation of Shapley values ​​for each dimension of the multidimensional evaluation index:

[0117]

[0118] Among them, S active This indicates the currently active indicator coalition, which changes with the driving scenario, such as S in a low-speed lane-changing scenario. LLC Includes {safety dimension, efficiency dimension, comfort dimension, intelligence dimension, reliability dimension}; N\{i} represents the set of all possible indicators except indicator i; n represents the total number of dimensions in the complete set of indicators N; v(S active ∪{i})-v(S active ) represents the marginal contribution of indicator i, which is the contribution of indicator i to joining the coalition S. active Marginal contribution at time indicates the increase in alliance revenue after indicator i joins.

[0119] Application examples of emergency obstacle avoidance scenarios:

[0120] Alliance Structure: When an emergency obstacle avoidance scenario is identified, a safety-intelligence priority strategy is activated. Considering the limited lane-changing safety space in emergency obstacle avoidance scenarios—meaning a shorter reversible collision time between the vehicle and other road users—this scenario prioritizes safety and intelligence metrics while neglecting comfort metrics. Therefore, the Emergency Obstacle Avoidance Index Alliance S... EOA ={S a ,E,I,R}.

[0121] Feature function call: Security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of intelligence I is v(S∪I)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S).

[0122] Shapley value output: φ I =0.29, φ E =0.12, φ C =0.11, φ R =0.10.

[0123] Application examples of low-speed lane changing scenarios:

[0124] Alliance Composition: When a low-speed lane change is detected, a comfort-first strategy is activated. Considering that traffic participants in low-speed lane change scenarios have lower speeds, leading to reduced trajectory prediction accuracy and lane change intention recognition success rate, and that comfort metrics are prioritized over intelligence metrics in this scenario, the low-speed lane change scenario metric alliance is S. LLC ={S a ,E,C,R}.

[0125] Feature function call: Security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of comfort C is v(S∪C)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S).

[0126] Shapley value output: φ E =0.18, φ C =0.22, φ R =0.28.

[0127] Application examples of high-speed lane changing scenarios:

[0128] Alliance Composition: When a high-speed lane-changing scenario is identified, an efficiency-first strategy is activated. Considering the high speed of the vehicle in a high-speed lane-changing scenario, a safe lane-changing space needs to be reserved. To reduce the impact of the vehicle's lane-changing on the efficiency of highway traffic flow, the indicator evaluation alliance needs to balance five-dimensional evaluation indicators, with greater emphasis on efficiency indicators. Therefore, the high-speed lane-changing scenario indicator alliance S... HLC ={S a ,E,C,I,R}.

[0129] Feature function call: Security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of comfort C is v(S∪C)-v(S), the marginal contribution of intelligence is v(S∪I)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S).

[0130] Shapley value output: φ E =0.25, φ C =0.18, φ I =0.10, φ R =0.17.

[0131] 4.2. Reconstruction of Transferable Utility Alliances and Weight Mapping:

[0132] Dynamic weight allocation rules:

[0133]

[0134] The scene sensitivity factor γ∈[0.2, 0.5] is dynamically adjusted based on the current vehicle state identified by the Hidden Markov Model (HMM). During emergency obstacle avoidance, safety and intelligence weights are increased; for example, the scene sensitivity factor γ increases from 0.3 to... J represents the overall system performance index in the current scenario. scene Weight W of the i-th dimension evaluation index i The partial derivatives of .

[0135] 4.3. Inertial smoothing and stability assurance

[0136] To avoid oscillations in the evaluation model caused by sudden changes in the Shapley value, a second-order inertial filter is designed:

[0137] α t+1 =K1α t +K2(α target -α t )+K3α t-1

[0138] The coefficient matrix is ​​constrained by the Lyapunov stability condition:

[0139] K1+K2+K3=I,||K2||2<1。

[0140] Where, α t Let α represent the weight vector of the multidimensional evaluation index at the current time t. t+1 This represents the filtered weight vector at the next time step t+1, α t-1 K1 represents the historical weight vector at the previous time t-1; K2 represents the historical weight inertia coefficient, K3 represents the target weight tracking coefficient, and K3 represents the second-order historical inertia coefficient.

[0141] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for dynamically constructing lane-changing evaluation indicators for intelligent vehicles based on transferable utility, characterized by the following steps: include: An expert dynamic game framework is constructed based on the improved Delphi method. Core evaluation indicators are screened through three rounds of anonymous feedback iteration, and multi-dimensional evaluation indicators are constructed. Based on the constructed multidimensional evaluation index, the characteristic function of the transferable utility alliance is constructed using the improved Shapley value algorithm. Identify the typical scenario of the current vehicle and analyze the driving intention. Activate the corresponding priority strategy based on the current scenario state, and adjust the weights of multi-dimensional evaluation indicators in real time by reconstructing the transferable utility alliance.

2. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 1, is characterized in that... The core evaluation indicators are iteratively screened based on the improved Delphi method, and a multidimensional evaluation indicator is constructed, as follows: Establish a dynamic game framework through an encrypted online collaboration platform; The first round of research used a structured questionnaire to collect expert scores on the importance of several preliminary indicators; In the second round of anonymous feedback, the entropy weight method was used to calculate the information entropy of each preliminary indicator to quantify the sensitivity of the indicator. The coefficient of variation method was combined to evaluate the data dispersion and select several candidate indicators. In the third round, a judgment matrix was constructed using fuzzy hierarchical analysis to obtain the correlation matrix among candidate indicators, and the final multidimensional evaluation index was obtained through screening.

3. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 2, is characterized in that... The constructed multidimensional evaluation includes: In terms of safety, the secondary parameters include collision risk and lateral offset; In terms of efficiency, the secondary parameters include lane change time and speed maintenance. In terms of comfort, secondary parameters include longitudinal impact and yaw fluctuation; In terms of intelligence, the secondary parameters include intent recognition rate and trajectory prediction accuracy; In terms of reliability, secondary parameters include system response latency and fault recovery rate.

4. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 2, is characterized in that... The first round of research on the improved Delphi method introduced a time delay compensation mechanism: when the expert feedback interval exceeds the set time, the system automatically pushes the index weight reference value generated by the LSTM prediction model trained on historical data.

5. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 3, is characterized in that... In the third round of the improved Delphi method, a weight adaptive module is introduced. For two indicators with implicit conflicts, when the implicit correlation exceeds a threshold indicator, a weight correction factor is calculated to correct the initial weights of the fuzzy hierarchical analysis method (FAHP). The specific steps are as follows: Calculate the covariance matrix of latent associations exceeding the threshold index, and perform eigenvalue decomposition on the covariance to obtain the eigenvector matrix and the eigenvalue diagonal matrix; The original indices are projected onto the principal component space to eliminate linear correlation. Based on the eigenvalues ​​of the eigenvector matrix and the eigenvalue diagonal matrix, the weight correction factor is further obtained: The initial weights of the fuzzy analytic hierarchy process (FAHP) are corrected based on the aforementioned weight correction factors: w′1=w1(1+β), w′2=w2(1+α) Where λ1 and λ2 are the eigenvalues ​​of the eigenvector matrix and the eigenvalue diagonal matrix, respectively; α and β are weight correction factors; and w1, w′1, w2, and w′2 are the initial and corrected weights of two indicators with implicit conflicts, respectively.

6. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 1, is characterized in that... The characteristic function of the transferable utility consortium constructed using the improved Shapley value algorithm is expressed as follows: v(S)=∑ i∈S λ(t)x′ i +∑ (i，j)∈S×s η ij ·CMI ij ·x′ i x′ j Where, x′ i 、x′ j η represents the normalized i-th and j-th dimension evaluation index values, respectively; λ(t) is the time-varying correction factor; η ij For nonlinear synergistic gain coefficients; CMI ij Let be the conditional mutual information entropy between multidimensional indices i and j.

7. The method for dynamically constructing an evaluation index for lane changing of intelligent vehicles based on transferable utility, as described in claim 6, is characterized in that... In the characteristic function of the transferable utility alliance: The time-varying correction factor is expressed as: Where T is the Delphi round interval; The conditional mutual information entropy is obtained by calculating the collaborative relationship of analytical indicators using a dynamic Bayesian network model: A third-order Bayesian network is constructed with multidimensional evaluation indicators as latent variables and vehicle state data as observed variables. The conditional mutual information entropy is then obtained through iterative optimization. CMI(X i ,X j |X k )=H(X i |X k )+H(X j |X k )-H(X i ,X j |X k ) Among them, CMI(X) i ,X j |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i and the j-th dimension evaluation index variable X j The degree of statistical dependence between them; H(X) i X j |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i and the j-th dimension evaluation index variable X j The joint uncertainty of H(X); i |X k ) represents the condition where, given a k-dimensional evaluation index variable X, k Under the condition that the i-th dimension evaluation index variable X i The uncertainty is explained by the following formula: Where p(x) i x k ) is X i and X k The joint probability distribution of p(x); i | xk ) is given X k Time X i The conditional probability distribution; The nonlinear synergistic gain coefficient η is expressed as: CMI represents conditional mutual information entropy.

8. The method for dynamically constructing an evaluation index for lane changing of intelligent vehicles based on transferable utility as described in claim 1, characterized in that, The real-time adjustment of the weights of the multi-dimensional evaluation indicators is as follows: A hidden Markov model is constructed to identify typical scenarios, including low-speed lane changing, high-speed lane changing, and emergency obstacle avoidance; a hidden state set is defined using time-series data from vehicle sensors, and driving intentions are analyzed based on the observation symbol probability matrix. When identifying trigger mode switching in typical scenarios, the Shapley values ​​of each dimension of the multidimensional evaluation index are calculated in real time based on the currently active index consortium S: Where S represents the currently active indicator consortium, which changes with the driving scenario; N{i} represents the set of all possible indicators except indicator i; n represents the total number of dimensions of the indicator set N; v(S active ∪{i})-v(S active ) represents the marginal contribution of indicator i, which is the contribution of indicator i to joining the coalition S. active Marginal contribution at time indicates the increase in alliance revenue after indicator i joins; The dynamic weight allocation is as follows: Wherein, γ is the scene sensitivity factor; J represents the overall system performance index in the current scenario. scene Weight W of the i-th dimension evaluation index i The partial derivatives of .

9. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 8, is characterized in that... The priority strategy for activating the corresponding strategy based on the current scene state is as follows: When an emergency obstacle avoidance situation is detected, the safety-intelligence indicator priority strategy is activated, and the emergency obstacle avoidance indicator coalition is set to S. EOA ={S a , E, I, R}; Features Function calls include: security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of intelligence I is v(S∪I)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S); the Shapley values ​​of each dimension index are calculated. When a low-speed lane change is detected, the comfort-first strategy is activated, and the low-speed lane change scenario indicator coalition is set to S. LLC ={S a , E, C, R}; Feature function calls include: security S a Marginal contribution v(S∪S) a )-v(S), marginal contribution of efficiency E v(S∪E)-v(S), marginal contribution of comfort C v(S∪C)-v(S), marginal contribution of reliability R v(S∪R)-v(S). When a high-speed lane-changing situation is detected, the efficiency-first strategy is activated, and the high-speed lane-changing scenario metric coalition is S. HLC ={S a , E, C, I, R}; Feature function calls include: security S a Marginal contribution v(S∪S) a The marginal contribution of efficiency E is v(S∪E)-v(S), the marginal contribution of comfort C is v(S∪C)-v(S), the marginal contribution of intelligence is v(S∪I)-v(S), and the marginal contribution of reliability R is v(S∪R)-v(S).

10. The method for dynamically constructing an evaluation index for intelligent vehicles based on transferable utility, as described in claim 8, is characterized in that... During the real-time adjustment of the weights of the multidimensional evaluation indicators, a second-order inertial filter is used to avoid oscillations in the evaluation model caused by sudden changes in the Shapley value. a t+1 =K1a t +K2(a target -a t )+K3a t-1 K1+K2+K3=I,||K2||2<1 Where, α t Let α represent the weight vector of the multidimensional evaluation index at the current time t. t+1 This represents the filtered weight vector at the next time step t+1, α t-1 K1 represents the historical weight vector at the previous time t-1; K2 represents the historical weight inertia coefficient, K3 represents the target weight tracking coefficient, and K3 represents the second-order historical inertia coefficient.

Images