Highway lane control method and system based on control unit

Abstract

The present invention provides a method and system for highway lane control based on control units. This method obtains control configuration data for lanes to be controlled under target control units on a target highway and integrates execution scores based on the control categories of control instructions to generate a first control evaluation result and a second control evaluation result. Furthermore, these two evaluation results are aggregated using a first machine learning model and a second machine learning model, ultimately generating a global execution score for the lanes to be controlled. This method can effectively assess the control restriction level of lanes to be controlled, providing a precise and scientific decision-making basis for highway lane control, thereby improving the intelligence and efficiency of lane control.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a highway lane control method and system based on a control unit. Background Art

[0002] With the rapid development of highways and the continuous increase in traffic volume, the effective control of highway lanes has become an important issue. Traditional highway lane control methods often rely on manual experience and fixed rules, lacking flexibility and intelligence, and are difficult to adapt to complex and changing traffic conditions.

[0003] Existing highway lane control systems typically implement a series of lane control instructions to regulate vehicle behavior. However, these instructions often come in a wide variety, including restrictive instructions (such as speed limits and no-overtaking regulations) and non-restrictive instructions (such as recommendations to maintain safe distance and turn on headlights). Each instruction also has varying degrees of importance and impact on lane control. Traditional approaches often struggle to comprehensively and accurately evaluate and integrate these control instructions, resulting in unsatisfactory lane control results. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a highway lane control method based on a control unit, the method comprising:

[0005] Obtaining control configuration data for a lane to be controlled under a target control unit of a target highway, the control configuration data comprising X control instructions; the control categories of the X control instructions include restrictive instructions and non-restrictive instructions; each control instruction corresponds to an execution score, and the execution score of any control instruction indicates the likelihood that the lane to be controlled meets control instruction requirements associated with the control instruction;

[0006] Integrating the execution scores of the X control instructions based on the control categories of the X control instructions to generate a first control evaluation result and a second control evaluation result, wherein the first control evaluation result is obtained based on the execution scores of the control instructions of the restrictive instructions; and the second control evaluation result is obtained based on the execution scores of the control instructions of the non-restrictive instructions;

[0007] aggregating the first control and assessment results through a first machine learning model to generate first aggregated information, and aggregating the second control and assessment results through a second machine learning model to generate second aggregated information;

[0008] The first convergence information and the second convergence information are accumulated to generate a global execution score of the lane to be controlled, where the global execution score of the lane to be controlled represents a control restriction level of the lane to be controlled.

[0009] In a possible implementation of the first aspect, aggregating the first management and control assessment results by using a first machine learning model to generate first aggregated information includes:

[0010] Obtaining the weight coefficient space and intercept term information of the first machine learning model;

[0011] Performing a non-negative transformation on the weight coefficient space of the first machine learning model to generate a non-negative transformed weight coefficient space;

[0012] Performing a fusion calculation on the first control and assessment result through the non-negatively transformed weight coefficient space and the intercept term information of the first machine learning model to generate a fusion calculation result of the first control and assessment result;

[0013] Perform incentive function mapping on the fusion calculation result of the first control evaluation result to generate first convergence information.

[0014] In a possible implementation of the first aspect, the first machine learning model and the second machine learning model are obtained by performing combined parameter learning on the first initialization model and the second initialization model; and the training steps of the first initialization model and the second initialization model specifically include:

[0015] Acquire a training sample sequence, the training sample sequence including a first template control evaluation result and a second template control evaluation result of a template controlled lane, and acquire a standard control result of the template controlled lane;

[0016] aggregating the first template control and assessment results using the first initialization model to generate a first model derivation result, and aggregating the second template control and assessment results using the second initialization model to generate a second model derivation result;

[0017] Aggregating the first model derivation result and the second model derivation result to generate a global model derivation result of the template controlled lane;

[0018] Based on the loss function value between the global model derivation result of the template-controlled lane and the standard control result of the template-controlled lane, the first initialization model and the second initialization model are parameter optimized to generate the first machine learning model and the second machine learning model.

[0019] In a possible implementation manner of the first aspect, the first initialization model includes initialization weight coefficient space and initialization intercept term information;

[0020] The step of performing parameter optimization on the first initialization model and the second initialization model based on the loss function value between the global model derivation result of the template controlled lane and the standard control result of the template controlled lane to generate the first machine learning model and the second machine learning model includes:

[0021] Based on the loss function value between the global model derivation result of the template-controlled lane and the standard control result of the template-controlled lane, the initialized weight coefficient space and the initialized intercept term information are optimized to generate the weight coefficient space and intercept term information of the first machine learning model.

[0022] In a possible implementation of the first aspect, obtaining the standard control result of the template-controlled lane includes:

[0023] Obtaining prior control data of the template controlled lane under the X control instructions, as well as an influence coefficient of each control instruction;

[0024] If there is at least one target control instruction corresponding to the template controlled lane among the X control instructions, calculating a standard control result of the template controlled lane according to an influence coefficient of the target control instruction;

[0025] Among them, the target control instruction corresponding to the template controlled lane is used to indicate that the prior control data of the template controlled lane under the target control instruction belongs to the effective execution range of the target control instruction.

[0026] In a possible implementation of the first aspect, aggregating the second management and control assessment results by using a second machine learning model to generate second aggregated information includes:

[0027] Obtaining the weight coefficient space and intercept term information of the second machine learning model;

[0028] Performing a fusion calculation on the second control and assessment result using the weight coefficient space of the second machine learning model and the intercept term information of the second machine learning model to generate a fusion calculation result of the second control and assessment result;

[0029] Perform incentive function mapping on the fusion calculation result of the second control evaluation result to generate second convergence information.

[0030] In a possible implementation of the first aspect, accumulating the first convergence information and the second convergence information to generate a global execution score for the lane to be controlled includes:

[0031] performing an addition calculation on the first aggregated information and the second aggregated information to generate an addition calculation result;

[0032] Perform incentive function mapping on the added calculation result to generate a global execution score for the lane to be controlled.

[0033] In a possible implementation of the first aspect, integrating the execution scores of the X control instructions based on the control categories of the X control instructions to generate a first control evaluation result and a second control evaluation result includes:

[0034] Integrating the execution scores of the X control instructions to generate a first control evaluation result, and integrating the control instructions that are non-restrictive instructions among the X control instructions to generate a second control evaluation result; or

[0035] The control instructions that are restrictive instructions among the X control instructions are integrated to generate a first control and management evaluation result, and the control instructions that are non-restrictive instructions among the X control instructions are integrated to generate a second control and management evaluation result.

[0036] In a possible implementation of the first aspect, obtaining control configuration data of a lane to be controlled under a target control unit of a target highway includes:

[0037] Acquiring a priori feature database, wherein the priori feature database includes road condition features of the lane to be controlled;

[0038] integrating the road condition characteristics of the lane to be controlled, the vehicle behavior characteristic data corresponding to the road condition characteristics of the lane to be controlled, and the traffic flow characteristic data to generate integrated characteristic data; the traffic flow characteristic data represents the traffic connection between the vehicle behavior corresponding to the vehicle behavior characteristic data and the lane to be controlled;

[0039] Performing logical knowledge encoding on the integrated feature data through Y logical knowledge models to generate Y logical knowledge encoding vector diagrams of the integrated feature data, where Y is an integer greater than 1;

[0040] Predicting the Y logical knowledge coding vector diagrams to generate execution scores for the X control instructions included in the control configuration data of the lane to be controlled;

[0041] In a possible implementation of the first aspect, the step of performing logical knowledge encoding on the integrated feature data using Y logical knowledge models to generate Y logical knowledge encoding vector diagrams of the integrated feature data includes:

[0042] Extracting a basic logic rule set from the prior feature database, the basic logic rule set including the relationship between lane use rules, vehicle behavior patterns, and traffic flow characteristics;

[0043] Dividing the basic logic rule set into Y logic rule subsets, each logic rule subset corresponds to a construction unit of a logic knowledge model;

[0044] Using a machine learning algorithm, initializing Y logical knowledge models based on the logical rule subset;

[0045] After preprocessing the lane condition characteristics, vehicle behavior characteristic data, and traffic flow characteristic data of the lane to be controlled, a preprocessed integrated feature data matrix is ​​generated, wherein the integrated feature data matrix represents the lane condition characteristics, vehicle behavior characteristic data, and traffic flow characteristic data of the lane to be controlled in a structured form, and each row of the integrated feature data matrix represents a feature set at each time point;

[0046] For each logical knowledge model, traverse each row of feature sets in the integrated feature data matrix and match the feature sets with the logical rules in the logical knowledge model. During the matching process, determine whether the feature sets meet the requirements of the logical rules based on the values ​​of the feature sets and the conditions of the logical rules. If they meet the requirements, record the number of the logical rules that are successfully matched, thereby generating an intermediate logical matching result set corresponding to each logical knowledge model. The intermediate logical matching result set includes the successful matching of the feature sets with the rules of the logical knowledge model at all time points.

[0047] For each intermediate logical matching result set, counting the number of successful matches for each logical rule, and calculating a coding value for each logical rule based on a pre-assigned weight coefficient for each logical rule and the weight coefficient and the number of successful matches, wherein the coding value reflects the activation degree or importance of the logical rule in the feature set;

[0048] Combining the code values ​​of all logic rules into a vector to generate a preliminary logic code vector map, thereby generating Y preliminary logic code vector maps corresponding to Y logic knowledge models respectively;

[0049] Normalizing the Y preliminary logical code vector maps, and using a dimensionality reduction algorithm to reduce the dimensionality of the code values ​​in the normalized logical code vector maps to generate Y logical knowledge code vector maps;

[0050] In a possible implementation of the first aspect, predicting the Y logical knowledge coding vector diagrams to generate execution scores for the X control instructions included in the control configuration data of the to-be-controlled lane includes:

[0051] Obtaining influence coefficient information of X control instructions included in the control configuration data of the lane to be controlled;

[0052] Using the influence coefficient information of the X control instructions respectively, performing fusion calculation on the Y logic knowledge coding vector diagrams, and obtaining the fusion calculation results of the X control instructions;

[0053] Predicting the fusion calculation results of each control instruction using the neural network model corresponding to the control instruction to generate execution scores of the X control instructions;

[0054] The step of obtaining the influence coefficient information of the X control instructions included in the control configuration data of the lane to be controlled includes:

[0055] Get the weight coefficient space of the a-th control instruction, where a is not greater than X;

[0056] Performing a fusion calculation on the integrated feature data according to the weight coefficient space of the a-th control instruction to generate target feature data;

[0057] Perform excitation function mapping on the target feature data to generate influence coefficient information corresponding to the a-th control instruction.

[0058] On the other hand, an embodiment of the present invention also provides a highway lane control system based on a control unit, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0059] Based on the above aspects, the embodiment of the present application obtains the control configuration data for the lanes to be controlled under the target control unit of the target highway and integrates the execution scores based on the control category of the control instructions to generate a first control evaluation result and a second control evaluation result. Furthermore, the two evaluation results are respectively aggregated using a first machine learning model and a second machine learning model to ultimately generate a global execution score for the lanes to be controlled. This method can effectively assess the control restriction level of the lanes to be controlled, providing a precise and scientific decision-making basis for lane control on highways, and improving the intelligence and efficiency of lane control. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1It is a schematic diagram of the execution flow of the highway lane control method based on the control unit provided in an embodiment of the present invention.

[0061] Figure 2 Schematic diagram of the hardware architecture of a highway lane control system based on a control unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a highway lane control method based on a control unit provided by an embodiment of the present invention. The highway lane control method based on the control unit is introduced in detail below.

[0063] Step S110: Obtain control configuration data for a lane to be controlled under a target control unit on a target highway. The control configuration data includes X control instructions. The control categories of the X control instructions include restrictive instructions and non-restrictive instructions. Each control instruction corresponds to an execution score. The execution score of any control instruction indicates the likelihood that the lane to be controlled meets the control instruction requirements associated with the control instruction.

[0064] In this example, consider a scenario where the target highway is a busy arterial road connecting multiple major cities. The target control unit can be a specific section of the highway, such as a section where traffic congestion and accidents often occur. The lane to be controlled is assumed to be the leftmost express lane.

[0065] In this scenario, the X control instructions in the control configuration data may include multiple types. In terms of restrictive instructions, for example, the instruction "the speed must not exceed 100 kilometers per hour" is set based on factors such as the road conditions, accident rate, and traffic flow of the section. If the average speed of vehicles on the lane remains between 80 and 90 kilometers per hour for a certain period of time, then the execution score of the instruction "the speed must not exceed 100 kilometers per hour" may be relatively high, assuming it is 0.8, which means that the lane to be controlled is more likely to meet the speed limit requirements.

[0066] For example, consider the restrictive instruction "No Large Trucks Allowed." Monitoring equipment at the lane entrance and analysis of historical data may reveal that if large trucks attempt to enter the lane very rarely during a specific time period, the instruction's enforcement score might be 0.9, indicating good enforcement of the lane's restrictions on large trucks.

[0067] For non-restrictive directives, such as "Encourage vehicles to maintain a safe distance," the safe distance can be set based on relevant traffic regulations and safety standards, such as maintaining a minimum distance of 50 meters. If monitoring of vehicle spacing in a lane reveals that most vehicles are between 40 and 60 meters apart, the directive's compliance score might be 0.7, indicating that vehicles in that lane generally meet the safe distance requirement, but there is room for improvement.

[0068] Another non-restrictive directive, “Adaptive cruise control system recommended”, may receive an implementation score of 0.3 if a survey finds that 30% of vehicles in the lane are equipped with adaptive cruise control systems and frequently use them, indicating that the likelihood of vehicles in the lane complying with this recommendation is low.

[0069] By setting execution scores for different types of control instructions in this way, the execution status of the lane to be controlled under different control requirements can be fully reflected, thereby providing a data basis for subsequent evaluation.

[0070] Step S120: Integrate the execution scores of the X control instructions based on the control categories of the X control instructions to generate a first control evaluation result and a second control evaluation result. The first control evaluation result is obtained based on the execution scores of the control instructions of the restrictive instructions. The second control evaluation result is obtained based on the execution scores of the control instructions of the non-restrictive instructions.

[0071] Let's continue with the scenario of the target highway mentioned above as an example. For restrictive instructions, assume that there are three restrictive instructions: "The speed must not exceed 100 kilometers per hour" with an execution score of 0.8, "Large trucks are not allowed to enter" with an execution score of 0.9, and "Motorcycles are not allowed to enter" with an execution score of 0.95 (assuming that the lane also has this restriction and is well implemented). When integrating the execution scores of these restrictive instructions to generate the first control evaluation result, a weighted average method may be used. Assume that the weights of these three instructions are 0.3, 0.4, and 0.3 respectively (the weights can be set according to factors such as the importance of the instructions). Then the first control evaluation result is: (0.8×0.3+0.9×0.4+0.95×0.3)=0.885. This result comprehensively reflects the overall situation of the lane to be controlled in complying with the restrictive instructions.

[0072] For non-restrictive instructions, assume there is an execution score of 0.7 for "Encourage vehicles to maintain a safe distance," a score of 0.3 for "Recommend use of adaptive cruise control," and a score of 0.6 for "Recommend reduced speed in certain weather conditions" (assuming these instructions exist and are evaluated based on weather conditions). Similarly, a weighted average is used to generate the second control evaluation result, assuming weights of 0.4, 0.3, and 0.3, respectively. The second control evaluation result is: (0.7 × 0.4 + 0.3 × 0.3 + 0.6 × 0.3) = 0.55. This result reflects the overall implementation of the non-restrictive instructions in the controlled lane.

[0073] In this process, the control categories of the instructions are strictly differentiated and integrated, so that the performance of the lanes to be controlled under restrictive and non-restrictive control requirements can be accurately evaluated separately, providing data of different dimensions for subsequent processing through different machine learning models.

[0074] Step S130: Aggregate the first control and assessment results through a first machine learning model to generate first aggregated information, and aggregate the second control and assessment results through a second machine learning model to generate second aggregated information.

[0075] In the above-mentioned target highway scenario, the first control assessment result (0.885) is processed for the first machine learning model. First, the weight coefficient space and intercept term information of the first machine learning model are obtained. Assume that the weight coefficient space of the first machine learning model is [0.2, 0.3, 0.5] and the intercept term information is 0.1.

[0076] The weight coefficient space is non-negatively transformed. Since the weight coefficients are already non-negative, they remain unchanged. The first control assessment results are then fused using the non-negatively transformed weight coefficient space and the intercept term. The calculation is: (0.885 × 0.2 + 0.885 × 0.3 + 0.885 × 0.5 + 0.1) = 1.6515.

[0077] Then, the fusion calculation result is mapped to the excitation function. Assuming that the excitation function used is the sigmoid function, the first converged information is obtained after the sigmoid function mapping. For example, the obtained value is 0.85 (the value after the sigmoid function mapping is between 0 and 1).

[0078] Process the second control assessment result (0.55) for the second machine learning model. Obtain the weight coefficient space of the second machine learning model, assuming it is [0.4, 0.3, 0.3], and the intercept term information is 0.2.

[0079] The second control evaluation result is fused and calculated using this weight coefficient space and intercept term information: (0.55×0.4+0.55×0.3+0.55×0.3+0.2)=0.99.

[0080] Then, the fusion calculation result is mapped to the excitation function, and the sigmoid function is also assumed to be used to obtain the second convergence information, for example, the obtained value is 0.73.

[0081] Through the processing of these two machine learning models, the previously integrated control and assessment results are further converted into aggregated information that is more suitable for subsequent calculations and evaluations, so that the data can better reflect the conditions of the lanes to be controlled under the processing of different models.

[0082] Step S140: Accumulate the first convergence information and the second convergence information to generate a global execution score of the lane to be controlled, where the global execution score of the lane to be controlled represents the control restriction level of the lane to be controlled.

[0083] Based on the first converged information of 0.85 and the second converged information of 0.73 obtained in the previous step, these two pieces of information are added together to obtain the addition result: 0.85+0.73=1.58.

[0084] Then, an excitation function mapping is performed on this addition calculation result. Assuming that the excitation function used is a sigmoid function, the global execution score of the lane to be controlled is obtained. For example, the value obtained is 0.82.

[0085] This global execution score of 0.82 represents the control restriction level of the lane under control. A high value, such as close to 1, indicates that the lane performs well in following various control instructions (including restrictive and non-restrictive instructions), and the control restrictions are well implemented. A low value, such as close to 0, indicates that the lane has many problems in executing control instructions, the control restriction level is low, and further management or control strategy adjustment may be needed. In this way, the global execution score can comprehensively and comprehensively evaluate the execution of the lane under control within the entire control system.

[0086] Based on the above steps, the embodiment of the present application obtains the control configuration data for the lanes to be controlled under the target control unit of the target highway and integrates the execution scores based on the control category of the control instructions to generate a first control evaluation result and a second control evaluation result. Furthermore, the two evaluation results are respectively aggregated using a first machine learning model and a second machine learning model to ultimately generate a global execution score for the lanes to be controlled. This method can effectively assess the control restriction level of the lanes to be controlled, providing a precise and scientific decision-making basis for lane control on highways, and improving the intelligence and efficiency of lane control.

[0087] In a possible implementation, step S130 includes:

[0088] Step S131, obtain the weight coefficient space and intercept term information of the first machine learning model.

[0089] Step S132: Perform a non-negative transformation on the weight coefficient space of the first machine learning model to generate a weight coefficient space after non-negative transformation.

[0090] Step S133: Perform a fusion calculation on the first control and assessment result through the non-negatively converted weight coefficient space and the intercept term information of the first machine learning model to generate a fusion calculation result of the first control and assessment result.

[0091] Step S134: performing incentive function mapping on the fusion calculation result of the first control evaluation result to generate first converged information.

[0092] In this embodiment, in the target highway scenario, the lane to be controlled is the leftmost express lane, and a first control evaluation result has been previously obtained. The process of aggregating the first control evaluation results using the first machine learning model to generate the first aggregated information will now be described in detail.

[0093] First, obtain the weight coefficient space and intercept term information of the first machine learning model. Assume that in this highway control system, the first machine learning model is a complex model specially built to process the evaluation results related to lane restriction instructions. The weight coefficient space of the model is a set of pre-set values. These values ​​reflect the degree of influence of different parameters in the model on the final result. For example, the weight coefficient space is [0.15, 0.3, 0.25, 0.3], where each value corresponds to the weight of a different feature or factor in the model. The intercept term information is 0.05. The intercept term plays a role in adjusting the overall calculation results in the model, similar to the constant term in a mathematical function.

[0094] Next, the weight coefficient space of the first machine learning model is non-negatively transformed. In this example, since the values ​​in the current weight coefficient space are all non-negative, the weight coefficient space after non-negative transformation is still [0.15, 0.3, 0.25, 0.3]. However, in other possible cases, if negative numbers exist, it is necessary to convert them according to a specific non-negative transformation algorithm, such as by taking absolute values ​​or other appropriate mathematical transformations, to ensure that all values ​​in the weight coefficient space are non-negative. This is done to meet the requirements of model calculation and ensure the rationality of the results.

[0095] The first control assessment result is then fused using the non-negatively transformed weight coefficient space and the intercept term information from the first machine learning model. The previously calculated first control assessment result is 0.885. The fusion calculation process is as follows: First, the first control assessment result is multiplied by each value in the weight coefficient space, namely, 0.885 × 0.15 = 0.13275, 0.885 × 0.3 = 0.2655, 0.885 × 0.25 = 0.22125, and 0.885 × 0.3 = 0.2655. These multiplication results are then added to the intercept term information to obtain 0.13275 + 0.2655 + 0.22125 + 0.2655 + 0.05 = 0.935. This 0.935 is the fused calculation result of the first control assessment result. It combines the weight coefficients, intercept term, and the first control assessment result information, reflecting the adjusted value of the first control assessment result after considering the model parameters.

[0096] Finally, the fusion calculation results of the first control evaluation are mapped to an excitation function to generate the first converged information. In this highway control system, the excitation function used is a specific mathematical function, such as the sigmoid function. The formula of the sigmoid function is f(x) = 1 / (1 + exp(-x)). Substituting the fusion calculation result of 0.935 into this formula, we obtain f(0.935) = 1 / (1 + exp(-0.935)) ≈ 0.715. This 0.715 is the first converged information. The first converged information ranges from 0 to 1 and is the result of the excitation function mapping. Compared with the previous fusion calculation results, this result is more suitable for subsequent processing and evaluation. For example, it can more intuitively reflect the status of the lane to be controlled under the various factors considered by the first machine learning model. It can also be compared with other similar converged information or used to construct a more complex evaluation system, providing important data support for a comprehensive assessment of the control restriction level of the lane to be controlled.

[0097] In one possible implementation, the first machine learning model and the second machine learning model are obtained by learning the combined parameters of the first initialization model and the second initialization model. The training steps of the first initialization model and the second initialization model specifically include:

[0098] Step S101 : obtaining a training sample sequence, wherein the training sample sequence includes a first template control evaluation result and a second template control evaluation result of a template controlled lane, and obtaining a standard control result of the template controlled lane.

[0099] Step S102: using the first initialization model to aggregate the first template control and evaluation results to generate a first model derivation result, and using the second initialization model to aggregate the second template control and evaluation results to generate a second model derivation result.

[0100] Step S103: Aggregating the first model derivation result and the second model derivation result to generate a global model derivation result of the template controlled lane.

[0101] Step S104: Optimize the parameters of the first initialization model and the second initialization model based on the loss function value between the global model derivation result of the template-controlled lane and the standard control result of the template-controlled lane to generate the first machine learning model and the second machine learning model.

[0102] In a possible implementation, the first initialization model includes information on an initialization weight coefficient space and an initialization intercept term.

[0103] Step S104 includes: optimizing the initialization weight coefficient space and the initialization intercept term information based on the loss function value between the global model derivation result of the template-controlled lane and the standard control result of the template-controlled lane, and generating the weight coefficient space and intercept term information of the first machine learning model.

[0104] In a possible implementation, step S101 includes:

[0105] Step S1011: Acquire the priori control data of the template controlled lane under the X control instructions, as well as the influence coefficient of each control instruction.

[0106] Step S1012: If there is at least one target control instruction corresponding to the template controlled lane among the X control instructions, a standard control result of the template controlled lane is calculated according to the influence coefficient of the target control instruction.

[0107] Among them, the target control instruction corresponding to the template controlled lane is used to indicate that the prior control data of the template controlled lane under the target control instruction belongs to the effective execution range of the target control instruction.

[0108] In this embodiment, a specific section of a target highway is used as a template controlled lane. The training example sequence includes the first template control evaluation result, the second template control evaluation result, and the standard control result for the template controlled lane. For a template controlled lane, the first template control evaluation result is an evaluation result related to the restrictive instructions for that lane, such as "Speed ​​must not exceed 100 km / h" or "No entry for large trucks." The first template control evaluation result is obtained by analyzing and integrating the implementation of these restrictive instructions. Assume that within a certain time period, based on speed monitoring data for vehicles in that lane and records of large trucks entering the lane, a series of calculations yield a first template control evaluation result of 0.85. The second template control evaluation result is an evaluation result related to non-restrictive instructions, such as "Encourage vehicles to maintain a safe distance" or "Recommend use of the adaptive cruise control system." Assume that the second template control evaluation result, after analysis and calculation, is 0.6.

[0109] Next, the standard control results for the template-controlled lane are obtained. This process requires obtaining prior control data for the template-controlled lane under X control commands, as well as the influence coefficients of each control command. Prior control data is historical control data for the template-controlled lane under various control commands. For example, for the control command "speed must not exceed 100 km / h," the prior control data might include the speed distribution of vehicles in that lane over different time periods, such as an average speed of 85 km / h during peak hours and 90 km / h during off-peak hours. The influence coefficients of each control command are pre-set based on factors such as the command's importance and its impact on traffic flow and safety. For example, assume the influence coefficient for "speed must not exceed 100 km / h" is 0.3, the influence coefficient for "large trucks prohibited" is 0.25, the influence coefficient for "vehicles encouraged to maintain a safe distance" is 0.2, and the influence coefficient for "adaptive cruise control recommended" is 0.15. If at least one target control instruction among the X control instructions corresponds to a template-controlled lane, meaning the prior control data for the template-controlled lane under that target control instruction falls within the effective execution range of that target control instruction, the standard control result for the template-controlled lane is calculated based on the target control instruction's influence coefficient. For example, if the analysis finds that the prior control data for both the "Speed ​​Must Not Exceed 100 km / h" and "Encourage Vehicles to Maintain a Safe Separation" instructions are within the effective execution range, then the standard control result is: (0.85 × 0.3 + 0.6 × 0.2) = 0.375. Here, 0.85 is the evaluation value associated with "Speed ​​Must Not Exceed 100 km / h," and 0.6 is the evaluation value associated with "Encourage Vehicles to Maintain a Safe Separation." The standard control result is calculated based on their respective influence coefficients.

[0110] Then, the first template control and assessment results are aggregated using the first initialization model to generate the first model derivation result. At the same time, the second template control and assessment results are aggregated using the second initialization model to generate the second model derivation result. The first initialization model contains the initialization weight coefficient space and the initialization intercept term information. Assume that the initialization weight coefficient space of the first initialization model is [0.2, 0.3, 0.1] and the initialization intercept term information is 0.05. Using this first initialization model, the first template control and assessment result 0.85 is aggregated, and the calculation process is: 0.85×0.2+0.85×0.3+0.85×0.1+0.05=0.51. This 0.51 is the first model derivation result. For the second initialization model, assume that its initialization weight coefficient space is [0.3, 0.2, 0.1] and the initialization intercept term information is 0.03. The second template control assessment result, 0.6, is aggregated and calculated as follows: 0.6 × 0.3 + 0.6 × 0.2 + 0.6 × 0.1 + 0.03 = 0.39. This 0.39 is the result derived from the second model.

[0111] The first and second model derivation results are then combined to generate the global model derivation result for the template controlled lane. The first model derivation result (0.51) and the second model derivation result (0.39) are combined in some way, such as by simple addition, to obtain 0.51 + 0.39 = 0.9. This 0.9 is the global model derivation result for the template controlled lane.

[0112] Finally, based on the loss function value between the global model derivation result of the template controlled lane and the standard control result of the template controlled lane, the parameters of the first initialized model and the second initialized model are optimized to generate the first machine learning model and the second machine learning model. The loss function is a function used to measure the difference between the model derivation result and the standard control result. Assuming that the mean square error (MSE) is used as the loss function, its calculation formula is MSE = (predicted value - true value) 2 Here the predicted value is the global model derivation result of 0.9, and the true value is the standard control result of 0.375. The calculated MSE is (0.9-0.375). 2 =0.275625. The initialized weight coefficient space and the initialized intercept term information in the first initialized model are optimized based on this loss function value. The optimization process may involve the use of algorithms such as gradient descent to adjust these parameters based on the partial derivatives of the loss function with respect to the weight coefficient and intercept term. For example, after optimization, the weight coefficient space of the first initialized model may become [0.22, 0.28, 0.12], and the intercept term information becomes 0.04. These optimized parameters constitute the weight coefficient space and intercept term information of the first machine learning model. The same process is applicable to the second initialized model, and the second machine learning model is finally generated by optimizing its initialized weight coefficient space and initialized intercept term information. Through such a combined parameter learning process, the first machine learning model and the second machine learning model can better adapt to the evaluation task of highway lane control and improve the accuracy and effectiveness of the evaluation.

[0113] In a possible implementation, step S130 further includes:

[0114] Step S135, obtain the weight coefficient space and intercept term information of the second machine learning model.

[0115] Step S136: Use the weight coefficient space of the second machine learning model and the intercept term information of the second machine learning model to perform a fusion calculation on the second control evaluation result to generate a fusion calculation result of the second control evaluation result.

[0116] Step S137: Perform incentive function mapping on the fusion calculation result of the second control evaluation result to generate second convergence information.

[0117] In this embodiment, taking the lane to be controlled on the target highway as an example, in the entire control system, the second machine learning model is specially constructed to process the second control evaluation results related to non-restrictive instructions. Assume that the weight coefficient space of the second machine learning model is [0.3, 0.25, 0.15, 0.3]. Each value here corresponds to the weight of different features or factors in the model. These features may be related to different aspects related to non-restrictive instructions. For example, for the instruction of "encouraging vehicles to maintain a safe distance", it may involve the weights of different statistical features of vehicle distance monitoring data; for the instruction of "recommending the use of adaptive cruise control system", it may involve different weights of data related to the proportion of vehicles equipped with the system and the frequency of use. The intercept term information is 0.1. The intercept term plays a role in adjusting the overall calculation results in the model, similar to the constant term in a mathematical function.

[0118] Next, the second control evaluation result is fused and calculated using the weight coefficient space and intercept term information of the second machine learning model. The second control evaluation result has been obtained previously, assuming it is 0.55. The calculation process is as follows: First, the second control evaluation result is multiplied by each value in the weight coefficient space in sequence, that is, 0.55x0.3=0.165, 0.55x0.25=0.1375, 0.55×0.15=0.0825, 0.55x0.3=0.165, and then these product results are added to the intercept term information to obtain 0.165+0.1375+0.0825+0.165+0.1=0.65. This 0.65 is the fused calculation result of the second control evaluation result. It combines the information of the weight coefficient, intercept term, and the second control evaluation result, reflecting the adjusted value of the second control evaluation result taking into account the parameters of the second machine learning model.

[0119] Then, the fusion calculation result of the second control evaluation result is mapped to the excitation function to generate the second converged information. In this highway control system, it is assumed that the excitation function used is the sigmoid function, and its formula is f(x)=1 / (1+exp(-x)). Substitute the fusion calculation result 0.65 into this formula for calculation, and get f((0.65)=1 / (1+exp(-0.65))≈0.659. This 0.659 is the second converged information. The second converged information is between 0 and 1. It is the result after the excitation function mapping. Compared with the previous fusion calculation result, this result has the characteristics that are more suitable for subsequent processing and evaluation. For example, it can more intuitively reflect the situation of the lane to be controlled under the various factors considered by the second machine learning model, and can be compared with other similar converged information or used to build a more complex evaluation system, thereby providing important data support for the comprehensive evaluation of the control restriction level of the lane to be controlled.

[0120] In a possible implementation, step S140 includes:

[0121] Step S141: performing addition calculation on the first aggregated information and the second aggregated information to generate an addition calculation result.

[0122] Step S142: performing incentive function mapping on the sum calculation result to generate a global execution score for the lane to be controlled.

[0123] In this embodiment, the first aggregate information is assumed to be 0.715 and the second aggregate information is 0.659, and the sum of the two is 0.715+0.659=1.374.

[0124] The summed result is then mapped to an activation function, again assuming a sigmoid function. Substituting 1.374 into the formula f(x) = 1 / (1 + exp(-x)) yields f(1.374) = 1 / (1 + exp(-1.374)) ≈ 0.798. This 0.798 is the global execution score for the lane under control. This global execution score represents the level of control restrictions for the lane under control. A high value, such as close to 1, indicates that the lane performs well in complying with various control instructions (both restrictive and non-restrictive), and that control restrictions are being well implemented. A low value, such as close to 0, indicates that the lane has significant issues in implementing control instructions, has a low level of control restrictions, and may require further management enhancements or adjustments to the control strategy. The global execution score provides a comprehensive and integrated assessment of the lane's performance within the overall control system. This score serves as an important basis for subsequent decision-making, such as whether control instructions need to be adjusted or whether traffic management measures for the lane need to be optimized.

[0125] In a possible implementation, step S120 includes:

[0126] Integrate the execution scores of the X control instructions to generate a first control evaluation result, and integrate the control instructions that are non-restrictive instructions among the X control instructions to generate a second control evaluation result. Or,

[0127] The control instructions that are restrictive instructions among the X control instructions are integrated to generate a first control and management evaluation result, and the control instructions that are non-restrictive instructions among the X control instructions are integrated to generate a second control and management evaluation result.

[0128] In this embodiment, in a specific control scenario on a target highway, there are X control instructions for the lanes to be controlled. These control instructions include restrictive instructions and non-restrictive instructions, and each instruction has a corresponding execution score. Taking a lane to be controlled on a specific road section as an example, the X control instructions include restrictive instructions such as "Speed ​​must not exceed 100 km / h," "Large trucks are prohibited from entering," and "Motorcycles are prohibited from entering." Non-restrictive instructions include "Encourage vehicles to maintain a safe distance," "Recommend the use of the adaptive cruise control system," and "Recommend reduced speed in certain weather conditions."

[0129] First, consider an integration approach: integrating the execution scores of X control instructions to generate a first control evaluation result. Simultaneously, integrating the non-restrictive control instructions among the X control instructions to generate a second control evaluation result. The execution score for each control instruction is based on a quantitative assessment of the degree to which the lane's actual operating data complies with the instruction requirements. For example, for the instruction "Speed ​​must not exceed 100 km / h," if speed monitoring devices installed in the lane collect vehicle speed data and statistical analysis show that most vehicles are traveling between 80 and 90 km / h, the execution score for this instruction might be 0.8. For the instruction "No large trucks allowed," if the proportion of large trucks illegally entering the lane is extremely low, based on vehicle type monitoring devices at the entrance and statistics on illegal entries, the execution score for this instruction might be 0.9. For the non-restrictive directive "Encourage vehicles to maintain a safe distance", based on the data from the vehicle distance monitoring equipment, if the average distance between vehicles is within a reasonable range of a safe distance, the implementation score may be 0.7; for the "Recommend the use of adaptive cruise control system", through a survey of vehicle configuration and usage, if the proportion of vehicles using this system is 30%, the implementation score may be 0.3.

[0130] When generating the first control and management evaluation result, the execution scores of the X control instructions are integrated. The weighted average method is adopted here, and the weight is determined based on factors such as the importance of the instruction and the degree of impact on the traffic management objectives. Assume that the weight of "the speed shall not exceed 100 kilometers per hour" is 0.3, the weight of "large trucks are not allowed to enter" is 0.4, and the weight of "motorcycles are not allowed to enter" is 0.3 (here it is assumed that the instruction exists and has a corresponding execution score). Assuming that the execution score of "motorcycles are not allowed to enter" is 0.95, then the first control and management evaluation result is: (0.8×0.3+0.9×0.4+0.95×0.3)=0.885. This result comprehensively reflects the overall execution status of the lane to be controlled under all control instructions.

[0131] For the second control evaluation, only non-restrictive instructions are integrated. Assuming the weight of "Encourage vehicles to maintain a safe distance" is 0.4, the weight of "Recommend the use of the adaptive cruise control system" is 0.3, and the weight of "Recommend reduced speed in certain weather conditions" (assuming its execution score is 0.6) is also 0.3. The second control evaluation result is: (0.7 × 0.4 + 0.3 × 0.3 + 0.6 × 0.3) = 0.55. This result reflects the comprehensive implementation of non-restrictive instructions in the controlled lane.

[0132] Consider another integration approach: integrating the restrictive instructions among the X control instructions to generate a first control evaluation result, and integrating the non-restrictive instructions among the X control instructions to generate a second control evaluation result. For the restrictive instructions, assuming that "speed must not exceed 100 km / h," "large trucks are prohibited from entering," and "motorcycles are prohibited from entering," according to the previously set execution scores and weights ("speed must not exceed 100 km / h" has a weight of 0.3 and an execution score of 0.8; "large trucks are prohibited from entering" has a weight of 0.4 and an execution score of 0.9; "motorcycles are prohibited from entering" has a weight of 0.3 and an execution score of 0.95), the first control evaluation result is calculated as: (0.8 × 0.3 + 0.9 × 0.4 + 0.95 × 0.3) = 0.885.

[0133] For the non-restrictive instructions, which are also "encourage vehicles to maintain a safe distance", "recommend the use of adaptive cruise control systems", and "recommend reducing the speed in certain weather conditions", according to the previously set weights ("encourage vehicles to maintain a safe distance" weight 0.4, execution score 0.7; "recommend the use of adaptive cruise control systems" weight 0.3, execution score 0.3; "recommend reducing the speed in certain weather conditions" weight 0.3, execution score 0.6), the second control evaluation result is calculated as: (0.7×0.4+0.3×0.3+0.6×0.3)=0.55.

[0134] Through these two integration methods, the first and second control evaluation results can be accurately obtained respectively. These two results reflect the situation of the lanes to be controlled in the execution of control instructions from different angles (all instructions in combination and separately by instruction category), providing basic data for subsequent further processing and evaluation through machine learning models, and helping to more comprehensively and accurately evaluate the control effect of the lanes to be controlled, thereby providing strong support for traffic management decisions on highways, such as whether it is necessary to adjust the requirements of certain control instructions and how to optimize control strategies.

[0135] In a possible implementation, step S110 includes:

[0136] Step S111: Acquire a priori feature database, where the priori feature database includes road condition features of the lane to be controlled.

[0137] In this example, the leftmost express lane on a specific section of the target highway is used as an example of a lane to be controlled. Lane condition characteristics include the lane's road surface condition (e.g., whether it is flat, has potholes, etc.), lane width, lane slope, and surrounding environmental characteristics (e.g., whether it is near a bridge, tunnel, etc.). For example, the lane's road surface condition is good, flat and without potholes, the lane width is 3.75 meters, the lane has a certain slope, and it is near a tunnel.

[0138] Step S112 integrates the road condition characteristics of the lane to be controlled, the vehicle behavior characteristic data corresponding to the road condition characteristics of the lane to be controlled, and the traffic flow characteristic data to generate integrated characteristic data. The traffic flow characteristic data represents the traffic connection between the vehicle behavior corresponding to the vehicle behavior characteristic data and the lane to be controlled.

[0139] In this embodiment, the vehicle behavior characteristic data covers information such as the vehicle's speed, acceleration, braking conditions, and lane change frequency in the lane. For example, the average vehicle speed obtained by the sensors installed on the lane is 90 kilometers per hour, the acceleration fluctuates within the normal range, the braking conditions are relatively stable, and the lane change frequency is low. The traffic flow characteristic data characterizes the traffic connection between vehicle behavior and the lane to be controlled, including traffic volume, vehicle density, and the proportion of different types of vehicles. Assume that the traffic volume on this lane is 1,000 vehicles per hour, the vehicle density is moderate, the proportion of large vehicles is 10%, and the proportion of small vehicles is 90%. These lane road condition characteristics, vehicle behavior characteristic data, and traffic flow characteristic data are integrated to form integrated characteristic data. For example, the integrated characteristic data can be represented in the form of a matrix, where each row represents a combination of characteristic data at a specific time point, and each column corresponds to a specific parameter in the lane road condition characteristics, vehicle behavior characteristic data, and traffic flow characteristic data.

[0140] Step S113 , performing logical knowledge encoding on the integrated feature data through Y logical knowledge models to generate Y logical knowledge encoding vector diagrams of the integrated feature data, where Y is an integer greater than 1.

[0141] Here, we assume that Y = 3, meaning there are three logical knowledge models. A basic logical rule set is extracted from the prior feature database. This basic logical rule set encompasses the relationships between lane usage rules, vehicle behavior patterns, and traffic flow characteristics. For example, lane usage rules stipulate that large vehicles should avoid driving in the fast lane, vehicle behavior patterns indicate that the normal speed range is related to traffic volume and density, and traffic flow characteristics indicate that the proportion of different vehicle types affects overall traffic flow. This basic logical rule set is divided into three logical rule subsets, each corresponding to a building block of a logical knowledge model. Using a machine learning algorithm, three logical knowledge models are initialized based on these three logical rule subsets.

[0142] After preprocessing the integrated feature data, a preprocessed integrated feature data matrix is ​​generated. For example, the data is normalized so that the data value range is within a specific interval to facilitate the processing of the logical knowledge model. For each logical knowledge model, each row of the feature set in the integrated feature data matrix is ​​traversed, and the feature set is matched with the logical rule in the logical knowledge model. During the matching process, based on the value of the feature set and the conditions of the logical rule, it is determined whether the feature set meets the requirements of the logical rule. If it does, the number of the successfully matched logical rule is recorded, thereby generating an intermediate logical matching result set corresponding to each logical knowledge model. For example, for the logical rule "large vehicles should reduce their speed when the traffic volume is large" in a logical knowledge model, if the traffic volume value in a certain feature set is large and the driving speed of large vehicles is low, then this feature set successfully matches the logical rule, and the number of this logical rule is recorded.

[0143] For each intermediate logical matching result set, the number of successful matches for each logical rule is counted. Based on the pre-assigned weight coefficient and the number of successful matches for each logical rule, a code value is calculated for each logical rule. This code value reflects the activation level or importance of the logical rule in the feature set. For example, for a logical rule with a pre-assigned weight coefficient of 0.2, 50 successful matches, and 100 total matches, the code value for this logical rule is 0.2 × (50 / 100) = 0.1. The code values ​​of all logical rules are combined into a vector to generate a preliminary logical code vector map. This generates three preliminary logical code vector maps corresponding to the three logical knowledge models. These three preliminary logical code vector maps are normalized, and a dimensionality reduction algorithm is used to reduce the code values ​​in the normalized logical code vector maps, generating three logical knowledge code vector maps.

[0144] Step S114 : predicting the Y logical knowledge coding vector maps to generate execution scores of the X control instructions included in the control configuration data of the lane to be controlled.

[0145] For example, assuming X = 5, the five control instructions are "Speed ​​must not exceed 100 km / h," "Large trucks are prohibited from entering," "Encourage vehicles to maintain a safe distance," "Recommend the use of the adaptive cruise control system," and "Recommend reducing speed in certain weather conditions." For each control instruction, its weight coefficient space is obtained. For example, for the instruction "Speed ​​must not exceed 100 km / h," its weight coefficient space is [0.1, 0.2, 0.3]. Based on this weight coefficient space, the integrated feature data is fused and calculated to generate the target feature data. For example, the corresponding parameters in the integrated feature data are multiplied by the coefficients in the weight coefficient space, and then added to obtain the target feature data. The target feature data is mapped to an activation function to generate the influence coefficient information corresponding to the control instruction. Assuming the activation function is a sigmoid function, the influence coefficient information for the instruction "Speed ​​must not exceed 100 km / h" is calculated to be 0.6.

[0146] For the "Speed ​​must not exceed 100 km / h" instruction, its impact coefficient information is fused and calculated with the corresponding code values ​​in the three logical knowledge coding vector diagrams. The neural network model corresponding to each control instruction predicts the fused calculation results of that control instruction, generating execution scores for X control instructions. For example, the fused calculation results of the "Speed ​​must not exceed 100 km / h" instruction are input into the corresponding neural network model, and the model calculates the execution score for that instruction. The same method is used to perform corresponding calculations for the other four control instructions, ultimately obtaining the execution scores for these five control instructions. These execution scores constitute the control configuration data for the lane to be controlled, providing the data foundation for subsequent control evaluation.

[0147] In a possible implementation, step S113 includes:

[0148] Step S1131 : extracting a basic logic rule set from the prior feature database, where the basic logic rule set includes the relationship between lane usage rules, vehicle behavior patterns, and traffic flow characteristics.

[0149] Step S1132: Divide the basic logic rule set into Y logic rule subsets, each logic rule subset corresponding to a construction unit of a logic knowledge model.

[0150] Step S1133: Using a machine learning algorithm, initialize Y logical knowledge models based on the logical rule subset.

[0151] Step S 1134, after preprocessing the lane condition characteristics, vehicle behavior characteristic data and traffic flow characteristic data of the lane to be controlled, a preprocessed integrated feature data matrix is ​​generated. The integrated feature data matrix represents the lane condition characteristics, vehicle behavior characteristic data and traffic flow characteristic data of the lane to be controlled in a structured form, and each row of the integrated feature data matrix represents a feature set at each time point.

[0152] Step S1135: For each logical knowledge model, traverse each row of feature sets in the integrated feature data matrix and match the feature sets with the logical rules in the logical knowledge model. During the matching process, determine whether the feature sets meet the requirements of the logical rules based on the values ​​of the feature sets and the conditions of the logical rules. If they meet the requirements, record the number of the successfully matched logical rules, thereby generating an intermediate logical matching result set corresponding to each logical knowledge model. The intermediate logical matching result set includes the successful matching of the feature sets and the rules of the logical knowledge model at all time points.

[0153] Step S1136: For each intermediate logical matching result set, the number of successful matches of each logical rule is counted, and based on the weight coefficient and the weight coefficient pre-assigned to each logical rule and the number of successful matches, the encoding value of each logical rule is calculated, which reflects the activation degree or importance of the logical rule in the feature set.

[0154] Step S1137 , combining the coding values ​​of all the logical rules into a vector to generate a preliminary logical coding vector map, thereby generating Y preliminary logical coding vector maps corresponding to the Y logical knowledge models respectively.

[0155] Step S1138, normalize the Y preliminary logical code vector maps, and use a dimensionality reduction algorithm to reduce the dimensionality of the code values ​​in the normalized logical code vector maps to generate Y logical knowledge code vector maps.

[0156] In this embodiment, taking a specific section of a target highway as an example, a basic logical rule set is first extracted from a priori feature database. This priori feature database contains a wealth of information related to highway traffic. The basic logical rule set covers the relationship between lane usage rules, vehicle behavior patterns, and traffic flow characteristics. Lane usage rules, for example, specify the permitted driving conditions of different types of vehicles (e.g., small cars, large trucks, motorcycles, etc.) in different lanes (e.g., fast lanes, slow lanes, etc.). For example, the restriction of large trucks in fast lanes is based on a combination of lane design objectives (fast lanes are designed to ensure rapid vehicle passage), vehicle performance (large trucks are relatively slow and have poor maneuverability), and safety factors (avoiding interference from large trucks with small passenger cars in fast lanes). Vehicle behavior patterns include rules for the normal speed range of vehicles and the reasonable ranges for acceleration and deceleration. For example, during normal driving, vehicle acceleration should be within a certain range, which is determined by the vehicle's power performance, safety factors, and traffic flow requirements. There are also corresponding rules for the reasonable braking distance of vehicles at different speeds, which are related to the vehicle's braking system performance, road conditions, and driver reaction time. Traffic flow characteristics include the relationship between traffic volume and vehicle density, and the impact of the proportion of different vehicle types on overall traffic flow. For example, when the proportion of a certain type of vehicle (such as large vehicles) is too high, it may reduce the average speed of the entire traffic flow and affect traffic smoothness.

[0157] This basic logical rule set is divided into Y logical rule subsets, each corresponding to a building block of the logical knowledge model. Assuming Y = 3, the basic logical rule set is divided according to specific rules and logic. For example, the first logical rule subset may primarily contain logical rules related to lane usage and vehicle types. These rules focus on the proper distribution of different types of vehicles on lanes and the basic restrictions on lane usage. The second logical rule subset may focus on speed and acceleration-related rules in vehicle behavior patterns. These rules are more related to the dynamic driving behavior of vehicles. The third logical rule subset may focus on traffic flow characteristics such as vehicle volume, vehicle density, and their interrelationships with vehicle behavior.

[0158] Using a machine learning algorithm, three logical knowledge models are initialized based on these three logical rule subsets. For the logical knowledge model initialized based on the first logical rule subset, the machine learning algorithm constructs the model structure based on the lane usage rules and vehicle type-related logical rules within that subset. For example, if the rules emphasize the restriction of large trucks in the fast lane, the model construction will use vehicle type and lane position as important input features. The learning algorithm then determines how to use these features to determine whether a vehicle violates the lane usage rule. For the second logical knowledge model, based on the logical rules related to speed and acceleration, the algorithm focuses on building a model structure that can identify vehicle speed change patterns and the rationality of acceleration. Techniques such as time series analysis may be used to process continuous vehicle speed and acceleration data. The third logical knowledge model focuses on rules related to traffic flow characteristics. By analyzing data such as vehicle volume and vehicle density, the algorithm constructs a model structure that can understand traffic flow conditions and predict their changing trends. For example, cluster analysis can be used to classify different traffic flow conditions.

[0159] After preprocessing the lane condition characteristics, vehicle behavior characteristics, and traffic flow characteristics of the controlled lanes, a preprocessed integrated feature data matrix is ​​generated. Lane condition characteristics include lane width, slope, and road surface conditions; vehicle behavior characteristics include vehicle speed, acceleration, and lane change frequency; and traffic flow characteristics include traffic volume, vehicle density, and the proportion of different vehicle types. The preprocessing process may include operations such as data cleaning (removing outliers and erroneous data) and data normalization (unifying data of different ranges into specific intervals). For example, lane width data in lane condition characteristics may be between 2.5 and 3.75 meters, which is normalized to the range of 0-1; vehicle speed data originally has a wide range, but is normalized to between -1 and 1 to facilitate model processing. The generated integrated feature data matrix represents these feature data of the controlled lanes in a structured form, with each row of the matrix representing a feature set at each time point. For example, a row of data may represent a set of features such as the lane width of 3.5 meters (normalized value), the vehicle speed of 80 kilometers per hour (normalized value), and the traffic volume of 800 vehicles per hour (normalized value) at a specific moment.

[0160] For each logical knowledge model, each row of the integrated feature data matrix is ​​traversed and matched against the logical rules in the logical knowledge model. Taking the first logical knowledge model as an example, when traversing the integrated feature data matrix, for each row of the feature set, the vehicle type and lane position information is determined to be consistent with the lane usage rules in the model. For example, if the vehicle type in a row of the feature set is a large truck and the lane position is in the fast lane, according to the model's logical rule (large trucks should not be in the fast lane), the rule is not satisfied. If the vehicle type is a small car and is in the fast lane, the rule is satisfied, and the logical rule number of the successful match is recorded. In this way, each row of the integrated feature data matrix is ​​evaluated, generating an intermediate logical matching result set corresponding to the first logical knowledge model. This result set contains all instances of successful feature set matches against the rules of the logical knowledge model at all time points. The same process is applied to the second and third logical knowledge models, matching the feature sets according to the logical rules in each model and generating the corresponding intermediate logical matching result sets.

[0161] For each intermediate logical matching result set, the number of successful matches for each logical rule is counted, and the encoding value for each logical rule is calculated based on the pre-assigned weight coefficient and the number of successful matches for each logical rule. For example, in the intermediate logical matching result set of the first logical knowledge model, suppose there is a logical rule "Large trucks are in violation of the fast lane" with a pre-assigned weight coefficient of 0.3. Statistics show that at all time points, this rule has been successfully matched 50 times, with a total of 100 matches. Therefore, the encoding value of this logical rule is 0.3 × (50 / 100) = 0.15. Following the same method, encoding values ​​are calculated for each logical rule, reflecting the activation level or importance of the logical rule in the feature set.

[0162] The code values ​​of all logical rules are combined into a vector to generate a preliminary logical code vector map. This generates three preliminary logical code vector maps corresponding to the three logical knowledge models. For example, for the first logical knowledge model, assume there are five logical rules and the calculated code values ​​are 0.15, 0.2, 0.1, 0.3, and 0.25, respectively. These code values ​​are combined into a vector [0.15, 0.2, 0.1, 0.3, 0.25], which is the preliminary logical code vector map corresponding to the first logical knowledge model. Similarly, preliminary logical code vector maps are generated for the second and third logical knowledge models.

[0163] These three preliminary logical code vectors are normalized, and a dimensionality reduction algorithm is used to reduce the dimensionality of the code values ​​in the normalized logical code vectors to generate three logical knowledge code vectors. Normalization can be performed using a variety of methods, such as dividing each code value by the sum of all code values ​​in the vector, so that each code value is between 0 and 1, and the sum of all code values ​​is 1. For example, the first preliminary logical code vector [0.15, 0.2, 0.1, 0.3, 0.25] may become [0.136, 0.182, 0.091, 0.273, 0.318] after normalization. A dimensionality reduction algorithm, such as principal component analysis (PCA), is then used to reduce the dimensionality of the normalized code values. Assuming that the 5-dimensional code values ​​are reduced to 3 dimensions, after PCA processing, the logical knowledge code vector corresponding to the first logical knowledge model is [0.2, 0.3, 0.1]. The same operation is applied to the second and third preliminary logic coding vector maps, ultimately resulting in three logic knowledge coding vector maps. These vector maps provide feature data encoded through logic knowledge for subsequent prediction operations, helping to more accurately generate execution scores for the control instructions contained in the control configuration data of the lane to be controlled.

[0164] In a possible implementation, step S114 includes:

[0165] Step S1141: Obtain influence coefficient information of X control instructions included in the control configuration data of the lane to be controlled.

[0166] Step S1142 : Using the influence coefficient information of the X control instructions respectively, performing fusion calculation on the Y logic knowledge coding vector maps, and obtaining the fusion calculation result of the X control instructions.

[0167] Step S1143 : predicting the fusion calculation result of each control instruction by using the neural network model corresponding to the control instruction to generate execution scores of the X control instructions.

[0168] Step S1141 includes:

[0169] Step S1141-1, obtain the weight coefficient space of the a-th control instruction, where a is not greater than X.

[0170] Step S1141 - 2 , performing fusion calculation on the integrated feature data according to the weight coefficient space of the a th control instruction to generate target feature data.

[0171] Step S1141-3: Perform excitation function mapping on the target feature data to generate influence coefficient information corresponding to the a-th control instruction.

[0172] In this embodiment, taking a lane to be controlled on a specific section of a target highway as an example, it is assumed that the control configuration data includes X control instructions such as "vehicle speed must not exceed 100 km / h", "large trucks are prohibited from entering", "vehicles are encouraged to maintain a safe distance", "adaptive cruise control system is recommended", and "reduced speed is recommended in certain weather conditions" (this is only an example and may actually include more or different instructions).

[0173] To obtain the influence coefficient information for each control instruction, taking the ath control instruction as an example (here assuming a = 1, i.e., the first control instruction, "Speed ​​must not exceed 100 km / h"), we first obtain its weight coefficient space. This weight coefficient space is determined based on a comprehensive consideration of multiple factors, including lane control objectives, traffic flow characteristics, and vehicle performance. For example, for the control instruction "Speed ​​must not exceed 100 km / h," its weight coefficient space might be [0.2, 0.3, 0.1, 0.4]. Each coefficient here corresponds to the weight of a different parameter or feature combination in the integrated feature data. These parameters or feature combinations include lane slope (corresponding to a weight of 0.2) in the lane road condition characteristics, average vehicle speed (corresponding to a weight of 0.3) in the vehicle behavior characteristics data, traffic volume (corresponding to a weight of 0.1) in the traffic flow characteristics data, and vehicle type ratio (corresponding to a weight of 0.4).

[0174] The integrated feature data is fused and calculated based on the weight coefficient space of the ath control instruction ("the vehicle speed must not exceed 100 km / h") to generate the target feature data. The integrated feature data includes lane condition characteristics (such as lane width, slope, and road surface condition), vehicle behavior characteristics (such as vehicle speed, acceleration, and lane change frequency), and traffic flow characteristics (such as traffic volume, vehicle density, and the proportion of different vehicle types). Taking the integrated feature data at a certain moment as an example, assuming the lane slope is 3% (normalized to 0.1), the average vehicle speed is 90 km / h (normalized to 0.8), the traffic volume is 800 vehicles per hour (normalized to 0.6), and the proportion of small vehicles is 80% (normalized to 0.8). The fusion calculation is performed according to the weight coefficient space [0.2, 0.3, 0.1, 0.4]: 0.1×0.2+0.8×0.3+0.6×0.1+0.8×0.4=0.64. This 0.64 is the target feature data generated through fusion calculation for the control instruction "the vehicle speed must not exceed 100 kilometers per hour."

[0175] The target feature data is mapped to an excitation function to generate the influence coefficient information corresponding to the ath control instruction ("Speed ​​must not exceed 100 km / h"). Assuming a sigmoid function as the excitation function, its formula is f(x) = 1 / (1 + exp(-x)). Substituting the target feature data 0.64 into this formula, the following calculation results: f(0.64) = 1 / (1 + exp(-0.64)) ≈ 0.65. This 0.65 is the influence coefficient information corresponding to the control instruction "Speed ​​must not exceed 100 km / h." Using the same method, the weight coefficient space is obtained for each of the other x-1 control instructions (such as "Large trucks are prohibited from entering," "Vehicles are encouraged to maintain a safe distance," "Adaptive cruise control is recommended," "Reduced speed is recommended in certain weather conditions," etc.), and a fusion calculation and excitation function mapping are performed to obtain the influence coefficient information corresponding to each control instruction.

[0176] After obtaining the influence coefficient information for X control instructions, the influence coefficient information for each of the X control instructions is used to perform a fusion calculation on Y logical knowledge coding vectors, resulting in a fusion calculation result for the X control instructions. Assuming Y = 3 (i.e., there are three logical knowledge coding vectors), taking the control instruction "vehicle speed must not exceed 100 km / h" as an example, its influence coefficient information is 0.65. For the three logical knowledge coding vectors, assume that the first logical knowledge coding vector is [0.2, 0.3, 0.1], the second is [0.1, 0.2, 0.3], and the third is [0.3, 0.1, 0.2]. The fusion calculation process is as follows: for the first logical knowledge coding vector, calculate 0.65 × 0.2 = 0.13; for the second logical knowledge coding vector, calculate 0.65 × 0.1 = 0.065; and for the third logical knowledge coding vector, calculate 0.65 × 0.3 = 0.195. Combining these results yields the fused calculation result [0.13, 0.065, 0.195] for the control instruction "Speed ​​must not exceed 100 km / h." Using the same method, perform fusion calculations on the other X-1 control instructions to obtain their respective fusion calculation results.

[0177] The neural network model corresponding to each control instruction predicts the fused calculation results of that control instruction, generating execution scores for X control instructions. Each control instruction has its own corresponding neural network model, which is built and trained based on a large amount of historical data, simulation data, and a deep understanding of lane control logic. For example, for the control instruction "Speed ​​must not exceed 100 km / h," its fused calculation results [0.13, 0.065, 0.195] are input into the corresponding neural network model. This neural network model may consist of an input layer, hidden layers, and an output layer. The input layer receives the fused calculation results, while the hidden layer performs complex nonlinear transformations on the input data through a series of neurons and activation functions, such as neurons using ReLU (Rectified Linear Unit) as the activation function. The output layer ultimately outputs a single value, which is the execution score for the control instruction "Speed ​​must not exceed 100 km / h." Assume that the neural network model calculates the execution score for this instruction to be 0.8. Similarly, the fused calculation results of the other X-1 control instructions are fed into their corresponding neural network models for prediction, resulting in execution scores for all X control instructions. These execution scores accurately reflect the execution of each control instruction in the controlled lane, providing a crucial data foundation for subsequent control evaluation and decision-making. For example, these execution scores can be used to determine whether specific control instructions need to be adjusted or other control measures need to be implemented to optimize lane traffic management.

[0178] Figure 2 The hardware structure of the highway lane control system 100 based on the control unit provided in the embodiment of the present invention is shown as follows: Figure 2 As shown, the highway lane control system 100 based on the control unit may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0179] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store data and / or instructions that the control unit-based highway lane control system 100 executes or uses to implement the exemplary methods described herein.

[0180] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the highway lane control method based on the control unit in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0181] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned highway lane control system 100 based on the control unit. The implementation principles and technical effects are similar and will not be repeated here in this embodiment.

[0182] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned highway lane control method based on the control unit is implemented.

[0183] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A highway lane control method based on a control unit, characterized in that: The method comprises: Obtaining control configuration data for a lane to be controlled under a target control unit of a target highway, the control configuration data comprising X control instructions; the control categories of the X control instructions include restrictive instructions and non-restrictive instructions; each control instruction corresponds to an execution score, and the execution score of any control instruction indicates the likelihood that the lane to be controlled meets control instruction requirements associated with the control instruction; Integrating the execution scores of the X control instructions based on the control categories of the X control instructions to generate a first control evaluation result and a second control evaluation result, wherein the first control evaluation result is obtained based on the execution scores of the control instructions of the restrictive instructions; and the second control evaluation result is obtained based on the execution scores of the control instructions of the non-restrictive instructions; aggregating the first control and assessment results through a first machine learning model to generate first aggregated information, and aggregating the second control and assessment results through a second machine learning model to generate second aggregated information; Accumulating the first convergence information and the second convergence information to generate a global execution score for the lane to be controlled, where the global execution score for the lane to be controlled represents a control restriction level of the lane to be controlled; The aggregating the first control and assessment results by using the first machine learning model to generate first aggregated information includes: Obtaining the weight coefficient space and intercept term information of the first machine learning model; Performing a non-negative transformation on the weight coefficient space of the first machine learning model to generate a non-negative transformed weight coefficient space; Performing a fusion calculation on the first control and assessment result through the non-negatively transformed weight coefficient space and the intercept term information of the first machine learning model to generate a fusion calculation result of the first control and assessment result; Performing an incentive function mapping on the fusion calculation result of the first control evaluation result to generate first converged information; The aggregating the second control and assessment results by the second machine learning model to generate second aggregated information includes: Obtaining the weight coefficient space and intercept term information of the second machine learning model; Performing a fusion calculation on the second control and assessment result using the weight coefficient space of the second machine learning model and the intercept term information of the second machine learning model to generate a fusion calculation result of the second control and assessment result; Perform incentive function mapping on the fusion calculation result of the second control evaluation result to generate second convergence information.

2. The highway lane control method based on the control unit according to claim 1 is characterized in that: The first machine learning model and the second machine learning model are obtained by performing combined parameter learning on the first initialization model and the second initialization model; The training steps of the first initialization model and the second initialization model specifically include: Acquire a training sample sequence, the training sample sequence including a first template control evaluation result and a second template control evaluation result of a template controlled lane, and acquire a standard control result of the template controlled lane; aggregating the first template control and assessment results using the first initialization model to generate a first model derivation result, and aggregating the second template control and assessment results using the second initialization model to generate a second model derivation result; Aggregating the first model derivation result and the second model derivation result to generate a global model derivation result of the template controlled lane; Based on the loss function value between the global model derivation result of the template-controlled lane and the standard control result of the template-controlled lane, the first initialization model and the second initialization model are parameter optimized to generate the first machine learning model and the second machine learning model.

3. The highway lane control method based on the control unit according to claim 2 is characterized in that: The first initialization model includes initialization weight coefficient space and initialization intercept term information; The step of performing parameter optimization on the first initialization model and the second initialization model based on the loss function value between the global model derivation result of the template controlled lane and the standard control result of the template controlled lane to generate the first machine learning model and the second machine learning model includes: Based on the loss function value between the global model derivation result of the template-controlled lane and the standard control result of the template-controlled lane, the initialized weight coefficient space and the initialized intercept term information are optimized to generate the weight coefficient space and intercept term information of the first machine learning model.

4. The highway lane control method based on the control unit according to claim 2 is characterized in that: The obtaining of the standard control result of the template controlled lane includes: Obtaining prior control data of the template controlled lane under the X control instructions, as well as an influence coefficient of each control instruction; If there is at least one target control instruction corresponding to the template controlled lane among the X control instructions, calculating a standard control result of the template controlled lane according to an influence coefficient of the target control instruction; Among them, the target control instruction corresponding to the template controlled lane is used to indicate that the prior control data of the template controlled lane under the target control instruction belongs to the effective execution range of the target control instruction.

5. The highway lane control method based on the control unit according to claim 1 is characterized in that: The accumulating the first convergence information and the second convergence information to generate a global execution score for the lane to be controlled includes: performing an addition calculation on the first aggregated information and the second aggregated information to generate an addition calculation result; Perform incentive function mapping on the added calculation result to generate a global execution score for the lane to be controlled.

6. The highway lane control method based on the control unit according to claim 1 is characterized in that: Integrating the execution scores of the X control instructions based on the control categories of the X control instructions to generate a first control evaluation result and a second control evaluation result includes: Integrating the execution scores of the X control instructions to generate a first control evaluation result, and integrating the control instructions that are non-restrictive instructions among the X control instructions to generate a second control evaluation result; or The control instructions that are restrictive instructions among the X control instructions are integrated to generate a first control and management evaluation result, and the control instructions that are non-restrictive instructions among the X control instructions are integrated to generate a second control and management evaluation result.

7. The highway lane control method based on the control unit according to claim 1 is characterized in that: The step of obtaining the control configuration data of the lane to be controlled under the target control unit of the target highway includes: Acquiring a priori feature database, wherein the priori feature database includes road condition features of the lane to be controlled; integrating the road condition characteristics of the lane to be controlled, the vehicle behavior characteristic data corresponding to the road condition characteristics of the lane to be controlled, and the traffic flow characteristic data to generate integrated characteristic data; the traffic flow characteristic data represents the traffic connection between the vehicle behavior corresponding to the vehicle behavior characteristic data and the lane to be controlled; Performing logical knowledge encoding on the integrated feature data through Y logical knowledge models to generate Y logical knowledge encoding vector diagrams of the integrated feature data, where Y is an integer greater than 1; Predicting the Y logical knowledge coding vector diagrams to generate execution scores for the X control instructions included in the control configuration data of the lane to be controlled; The step of performing logical knowledge encoding on the integrated feature data through Y logical knowledge models to generate Y logical knowledge encoding vector diagrams of the integrated feature data includes: Extracting a basic logic rule set from the prior feature database, the basic logic rule set including the relationship between lane use rules, vehicle behavior patterns, and traffic flow characteristics; Dividing the basic logic rule set into Y logic rule subsets, each logic rule subset corresponds to a construction unit of a logic knowledge model; Using a machine learning algorithm, initializing Y logical knowledge models based on the logical rule subset; After preprocessing the lane condition characteristics, vehicle behavior characteristic data, and traffic flow characteristic data of the lane to be controlled, a preprocessed integrated feature data matrix is ​​generated, wherein the integrated feature data matrix represents the lane condition characteristics, vehicle behavior characteristic data, and traffic flow characteristic data of the lane to be controlled in a structured form, and each row of the integrated feature data matrix represents a feature set at each time point; For each logical knowledge model, traverse each row of feature sets in the integrated feature data matrix and match the feature sets with the logical rules in the logical knowledge model. During the matching process, determine whether the feature sets meet the requirements of the logical rules based on the values ​​of the feature sets and the conditions of the logical rules. If they meet the requirements, record the number of the logical rules that are successfully matched, thereby generating an intermediate logical matching result set corresponding to each logical knowledge model. The intermediate logical matching result set includes the successful matching of the feature sets with the rules of the logical knowledge model at all time points. For each intermediate logical matching result set, counting the number of successful matches for each logical rule, and calculating a coding value for each logical rule based on a pre-assigned weight coefficient for each logical rule and the number of successful matches, wherein the coding value reflects the activation degree or importance of the logical rule in the feature set; Combining the code values ​​of all logic rules into a vector to generate a preliminary logic code vector map, thereby generating Y preliminary logic code vector maps corresponding to Y logic knowledge models respectively; Normalizing the Y preliminary logical code vector maps, and using a dimensionality reduction algorithm to reduce the dimensionality of the code values ​​in the normalized logical code vector maps to generate Y logical knowledge code vector maps; The predicting of the Y logical knowledge coding vector diagrams to generate execution scores of the X control instructions included in the control configuration data of the lane to be controlled includes: Obtaining influence coefficient information of X control instructions included in the control configuration data of the lane to be controlled; Using the influence coefficient information of the X control instructions respectively, performing fusion calculation on the Y logic knowledge coding vector diagrams, and obtaining the fusion calculation results of the X control instructions; Predicting the fusion calculation results of each control instruction using the neural network model corresponding to the control instruction to generate execution scores of the X control instructions; The step of obtaining the influence coefficient information of the X control instructions included in the control configuration data of the lane to be controlled includes: Get the weight coefficient space of the a-th control instruction, where a is not greater than X; Performing a fusion calculation on the integrated feature data according to the weight coefficient space of the a-th control instruction to generate target feature data; Perform excitation function mapping on the target feature data to generate influence coefficient information corresponding to the a-th control instruction.

8. A highway lane control system based on a control unit, characterized in that: The highway lane control system based on the control unit includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the highway lane control method based on the control unit as described in any one of claims 1 to 7 above.

Images