Implicit function driven multi-target path planning optimization method of multi-channel network

Abstract

The invention discloses a multi-channel network implicit function driven multi-objective path planning optimization method, which belongs to the technical field of algorithm optimization, and comprises the following steps: obtaining traffic data and preprocessing the traffic data; carrying out feature extraction on traffic flow density signals in the traffic data from the time domain dimension and the frequency domain dimension; a three-channel network model is constructed and trained; designing a target function; and searching an optimal solution by adopting a genetic algorithm. According to the multi-target path planning optimization method driven by the implicit function of the multi-channel network, different types of data are processed at the same time through three parallel processing channels, integrated output is carried out through the fusion layer, the complexity and dynamic changes of a traffic system are accurately reflected, the adaptability and precision of path planning are improved, and the path planning efficiency is improved. The processing capability and the comprehensive expression capability of the system on different data types are obviously improved; a complex nonlinear relation is processed in an implicit function mode and is embedded into a target function of an optimization algorithm, and the precision of path planning is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of algorithm optimization, and in particular to a multi-objective path planning optimization method driven by an implicit function of a multi-channel network. Background Art

[0002] With the increasing complexity of urban transportation networks, intelligent optimization algorithms, as an indispensable core technology in intelligent transportation systems, undertake important functions such as path generation, traffic flow optimization and real-time adjustment, which directly affect many aspects of path planning, such as accuracy, efficiency, robustness and cost.

[0003] Traditional path planning algorithms are mostly based on static road network models, ignoring the time series characteristics and dynamic changes of traffic flow. Although some algorithms adjust path selection by updating real-time traffic information, this method cannot fully mine data information, resulting in less than optimal path planning results. How to effectively extract key features from multiple dimensions of massive time series data and establish an accurate prediction model will be the key to improving path planning accuracy.

[0004] At the computational level of optimization algorithms, most optimization methods of traditional path planning algorithms are based on classical mathematical models, such as graph theory and shortest path algorithms. However, these methods often lack adaptability and computational efficiency when faced with complex nonlinear and large-scale data features, making it difficult to achieve ideal path planning results in actual traffic management. Modern optimization algorithms, such as genetic algorithms and particle swarm algorithms, have improved the optimization effect to a certain extent, but their computational overhead is large and they are still insufficient when dealing with complex nonlinear optimization problems. To this end, it is necessary to introduce a more advanced algorithm framework that can effectively explore the potential laws of the traffic system. Although deep learning technology can extract features from large amounts of data, it is not suitable to directly use deep learning models for path planning. How to use deep learning models to assist optimization algorithms is another problem that needs to be solved urgently.

[0005] In addition, the factors affecting path planning are extremely complex and diverse, and the data types involved are different. Conventional network models often lead to excessive model complexity, reduced computational efficiency, and weak generalization capabilities when training all features at once. Due to the different distribution forms of different data types, it is difficult for a single network structure to effectively process these heterogeneous data, which in turn affects the stability and accuracy of path planning. Therefore, how to build a network model that can handle multi-dimensional, heterogeneous data and has strong adaptability has become a key challenge to improve the accuracy and efficiency of path planning. Summary of the invention

[0006] The purpose of the present invention is to provide an implicit function-driven multi-objective path planning optimization method for a multi-channel network, which can effectively integrate and process complex multi-source nonlinear data through deep learning technology, optimize the objective function, and improve the accuracy of path planning.

[0007] To achieve the above object, the present invention provides an implicit function driven multi-objective path planning optimization method for a multi-channel network, comprising the following steps: S1, obtain traffic data and pre-process it; S2. Traffic density in traffic data from the time domain and frequency domain dimensions D Extract features from signals; S3, construct a three-channel network model; S4, training the three-channel network model; S5, combine S2 and S4 to design the objective function; S6. Use genetic algorithm to search for the optimal solution.

[0008] Preferably, in S1, the traffic data specifically includes: route type T Include m Category 1; Environmental conditions C Indicates the influence of external factors, including m 2 levels; traffic management mode M Include m 3 types; speed limit V Set to V a to V b Discrete values ​​between D Refers to different road conditions N Traffic density sampling points, sampling is performed at equal intervals in each traffic flow cycle, the unit is vehicle / hour; and the corresponding actual value of road loss L ; The preprocessing of traffic data includes: filling missing values, processing outliers, ensuring the integrity and accuracy of the data set. The final complete data set includes no less than one sample, and each sample includes the route type. T , Environmental conditions C , Traffic Management Mode M , speed limit V , Traffic density D， And the corresponding actual value of road loss L .

[0009] Preferably, in S2, the traffic density DThe time domain features of the signal include maximum value, minimum value, sharpness, symmetry, skewness, peak-to-peak value, absolute peak-to-average ratio, average value and standard deviation. The time domain features reflect the amplitude, fluctuation degree and distribution characteristics of the signal; the frequency domain features include main frequency, spectrum energy, spectrum center, frequency bandwidth, harmonic energy ratio and harmonic energy concentration; there are 15 time domain features and frequency domain features in total.

[0010] Preferably, in S3, the construction of the three-channel network model is specifically as follows: S31, the 15 time domain features and frequency domain features in S2 and the speed limit in S1 V Normalized as a numerical feature, the calculation formula is: ; in, X represents the original data, X’ is the normalized value, the normalization interval is [0,1], and the feature vector input to the numerical channel is ,in They are the 15 normalized time domain features and frequency domain features and the speed limit in S1 V ; S32, Traffic Management Mode M , Path Type T and environmental conditions C As a categorical feature, one-hot encoding is used to convert the value of each category into a binary vector, where each category is represented by an independent dimension; Path Type T Include m 1 type, after encoding, a length of m Binary vector of 1s; environmental conditions C Include m 2 levels, after encoding, a length of m Binary vector of 2; traffic management mode M Include m 3 types, after encoding, a length of m 3; concatenate the above three binary vectors to obtain the feature vector of the input classification channel X cat ， ; S33, Traffic density D It has obvious temporal characteristics and is classified as a temporal feature. The long short-term memory network is used to capture the long-term dependencies in the time series. The feature vector input to the temporal channel X seq By traffic density D ofN Sampling points composition, ; S34, the numerical channel and the classification channel process their respective input features through a fully connected layer. ; ; in, W num It is a numerical channel weight matrix with a dimension of 16×64; b num It is a numerical channel bias term with a dimension of 64, indicating the number of neurons in the output layer; ReLU is the activation function; W cat is the classification channel weight matrix, the dimension is ( m 1+ m 2+ m 3) ×64; b cat It is the classification channel bias item with a dimension of 64; Y num and Y cat is the output of the fully connected layer of each channel, in the form of a numerical vector, representing the feature information after ReLU transformation; S35, the time series channel is processed by the long short-term memory network LSTM to obtain a hidden state sequence at each time step t , the output hidden state of LSTM h t Depends on the current input d t and the state at the previous moment h t-1 , specifically: ; in, c t is the cell state, responsible for preserving long-term memory; the output length of the LSTM layer is N The hidden state sequence of , set the size of the hidden layer to 64, and select the hidden state of the last time step of the LSTM output sequence h N As a summary representation of the entire time series data; The hidden state of the last time step of LSTM is connected through a fully connected layer h N Mapped to the same feature space dimension as the numerical channel and the categorical channel, specifically: ; in, Y seq is the output of the fully connected layer of the temporal channel, W lstmfc and b lstmfc They are the weight matrix and bias term of the fully connected layer of the temporal channel respectively; S36, introduce the cross-channel attention mechanism, calculate the importance of the output features of the numerical channel, classification channel and time series channel, and fuse them in a weighted manner. The fused features are: ; in, Y fusion Represents the final feature in the form of a numerical vector, ɑ x are the attention weights of the three channels, calculated through a scoring network with learning capabilities; S37, fused features Y fusion Through a final fully connected layer, it is mapped to the final output space to obtain the final road loss prediction value : ; in, W fusion is the weight matrix of the final fusion layer, with a dimension of 192×1, b fusion is the bias term of the final fusion layer.

[0011] Preferably, in S4, specifically: the data set is divided into a training set and a test set at an 8:2 ratio, and cross-validation is used to verify the performance of the three-channel network model. During the training process, the training set is used to update the parameters of the three-channel network model, and the loss function is used to measure the difference between the predicted value of the three-channel network model and the actual target value. The loss function is defined as follows: ; in, U is the number of samples. By minimizing the loss function, the three-channel network model can gradually adjust the parameters to make the road loss prediction value The actual value of road loss L The error between them is reduced.

[0012] Preferably, in S5, the objective function design includes the following steps: S51, the three-channel network model trained in S4 is combined with the traffic density D Feature extraction module, regarded as the road loss calculation value L ’Function F: L ’ =F( V , D , M , T , C ); S52. Calculate traffic mobility E ’ , traffic mobility E ’ Refers to the degree of mobility of vehicles in the traffic system and the speed limit on the road section V and traffic density D Peak value within the cycle D peak Related, the calculation formula is: E ’ = V · D peak ; S53. Make traffic mobility E ’ and road loss calculation value L ’ Perform weighted combination to provide comprehensive evaluation criteria for the optimization algorithm and comprehensive objective function Z The design is: Z = w · E ’ - (1- w ) L ’ ; in, w is the weight coefficient used to control traffic mobility E ’ and road loss calculation value L ’ The relative importance of the route planning optimization is adjusted according to the traffic situation.

[0013] Preferably, in S6, the specific steps are: S61. Randomly generate a set of path planning solutions and calculate the objective function Z , and evaluate fitness; S62, based on fitness, select excellent individuals as parents, and through crossover operation, reorganize the genetic information of the parents to generate new offspring individuals; S63, randomly mutate some individuals, replace the old population with the newly generated offspring, and form the next generation population; S64. When one of the two conditions, namely, the maximum number of iterations is reached and the fitness reaches a preset threshold, is met, the algorithm terminates and outputs the optimal solution, that is, the optimal combination of the values ​​of each feature.

[0014] Therefore, the present invention adopts the above-mentioned implicit function-driven multi-objective path planning optimization method of a multi-channel network, which has the following beneficial effects: (1) Multi-dimensional feature extraction: The present invention performs multi-dimensional feature extraction on the traffic density signal to ensure that the extracted features are highly representative and effective. Feature extraction not only covers the basic statistical information of the signal, but also includes more advanced morphological features, time series changes and fluctuation patterns. These multi-dimensional features provide reliable input data for subsequent optimization algorithms, directly affect the accuracy and stability of the path planning system, improve the model prediction ability and optimization performance, and provide a more forward-looking decision-making basis for path planning.

[0015] (2) Multi-channel network model construction: By designing three parallel processing channels, including a numerical channel, a classification channel, and a time series channel, it is possible to process data from different types simultaneously. Each channel independently processes a specific type of data, and then integrates the results of each channel in the fusion layer to obtain a comprehensive output. This multi-channel structure can more accurately reflect the complexity and dynamic changes of the traffic system and improve the adaptability and accuracy of path planning. Compared with the traditional single-channel neural network model, this method significantly improves the system's processing capabilities and comprehensive expression capabilities for different types of data.

[0016] (3) Implicit function driven target optimization: During the optimization process, the neural network model handles complex nonlinear relationships through implicit functions and embeds this implicit nonlinear relationship into the objective function of the optimization algorithm. This can effectively make up for the shortcomings of traditional optimization algorithms in dealing with complex nonlinear problems in traffic systems, thereby improving the accuracy of path planning and fully tapping the potential regularity of the traffic system.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a scheme flow chart of a multi-objective path planning optimization method driven by an implicit function of a multi-channel network of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Embodiment 1 The present invention provides a multi-objective path planning optimization method driven by an implicit function of a multi-channel network, the process is as follows Figure 1 As shown, the specific steps include: S1, obtain traffic data and pre-process it; In the process of optimizing path planning, road loss is very important. Road loss is affected by many factors, including speed limit. V , Traffic density D , Management Model M , Path Type T and environmental conditions C These factors interact and overlap, presenting a complex nonlinear relationship. T Include m Category 1; Environmental conditions C Indicates the influence of external factors (weather, temperature, etc.), including m 2 levels; traffic management mode M Include m 3 types; speed limit V Directly affects the energy efficiency and traffic capacity of the road, set as V a to V b Discrete values ​​between D It is a dynamic variable that changes with time, reflecting the road usage in different time periods. During peak hours, the traffic density will increase significantly, affecting traffic flow and road loss. D Including different road conditions N Traffic density sampling points are sampled at equal intervals in each traffic flow cycle, with the unit being vehicles per hour. The corresponding actual value of road loss can be obtained through the above traffic data. L .

[0022] The preprocessing of traffic data includes: filling missing values, processing outliers, ensuring the integrity and accuracy of the data set. The final complete data set includes no less than one sample, and each sample includes the route type. T , Environmental conditions C , Traffic Management Mode M , speed limit V , Traffic density D And the corresponding road loss L .

[0023] S2. Traffic density from the time domain and frequency domain dimensions D The signal is extracted to more comprehensively reflect the traffic density through the combination of multi-dimensional features. D The dynamic characteristics of the signal.

[0024] Time domain features include maximum value, minimum value, sharpness, symmetry, skewness, peak-to-peak value, absolute peak-to-average ratio, average value and standard deviation. Time domain features reflect the amplitude, fluctuation degree and distribution characteristics of the signal; frequency domain features include main frequency, spectrum energy, spectrum center, frequency bandwidth, harmonic energy ratio and harmonic energy concentration; there are 15 time domain features and frequency domain features in total.

[0025] S3. Construct a three-channel network model, which specifically includes the following steps: First, the traffic data is processed and converted to meet the requirements of the three-channel network model input.

[0026] S31, 15 time domain features and frequency domain features in S2 and speed limit in S1 V They are all discrete values ​​within a large range and are normalized as numerical features to prevent the gradient descent from converging or converging slowly during training. The calculation formula is: ; in, X represents the original data, X’ is the normalized value, the normalization interval is [0,1], and the feature vector input to the numerical channel is ,in They are the 15 normalized time domain features and frequency domain features and the speed limit in S1 V .

[0027] S32, Traffic Management Mode M , Path Type T and environmental conditions CFor finite discrete values, as categorical features, use one-hot encoding to convert the value of each category into a binary vector, where each category is represented by an independent dimension; Path Type T Include m 1 type, then after encoding, a length of m Binary vector of 1s; environmental conditions C Include m 2 levels, then after encoding, a length of m Binary vector of 2; traffic management mode M Include m 3, then after encoding, a length of m 3; concatenate the above three binary vectors to obtain the feature vector input to the classification channel X cat ， .

[0028] S33, Traffic density D It has obvious temporal characteristics and is classified as a temporal feature. The long short-term memory network is used to capture the long-term dependencies in the time series. The feature vector input to the temporal channel X seq By traffic density N Sampling points composition, .

[0029] The different types of data are then integrated into different input channels of the three-channel model.

[0030] S34, the numerical channel and the classification channel process their respective input features through a fully connected layer. ; ; in, W num It is a numerical channel weight matrix with a dimension of 16×64; b num It is a numerical channel bias term with a dimension of 64, indicating the number of neurons in the output layer; ReLU is the activation function; W cat is the classification channel weight matrix, the dimension is ( m 1+ m 2+ m 3) ×64; b cat It is the classification channel bias item with a dimension of 64; Ynum and Y cat It is the output of the fully connected layer of each channel in the form of a numerical vector, representing the feature information after ReLU transformation; because the data of the input numerical channel is normalized, and the data of the input categorical channel is processed by one-hot encoding, there is no structural relationship and natural order relationship between the data features. Therefore, it is efficient and simple to use a fully connected layer to process numerical features and categorical features, and a fully connected layer is used to map the numerical features and categorical features to the appropriate space.

[0031] S35, the time series channel is processed by the long short-term memory network LSTM to obtain a hidden state sequence at each time step t , the output hidden state of LSTM h t Depends on the current input d t and the state at the previous moment h t-1 , specifically: ; in, c t is the cell state, responsible for preserving long-term memory; the output length of the LSTM layer is N The hidden state sequence of , set the size of the hidden layer to 64, and select the hidden state of the last time step of the LSTM output sequence h N As a summary representation of the entire time series data, it carries all the information of the time series data; In order to obtain a unified feature representation, the hidden state of the last time step of LSTM is connected through a fully connected layer. h N Mapped to the same feature space dimension as the numerical channel and the categorical channel, specifically: ; in, Y seq is the output of the fully connected layer of the temporal channel, W lstmfc and b lstmfc They are the weight matrix and bias term of the fully connected layer of the temporal channel, respectively.

[0032] S36. In order to make the three-channel model adaptively distribute information from the three channels and help the three-channel model adjust the contribution according to the importance of each channel, a cross-channel attention mechanism is introduced to calculate the importance of the output features of the numerical channel, the classification channel and the temporal channel, and fuse them in a weighted manner. The fused features are: ; in, Y fusion Represents the final feature in the form of a numerical vector, ɑ x are the attention weights of the three channels, calculated through a scoring network with learning capabilities.

[0033] S37, fused features Through a final fully connected layer, it is mapped to the final output space to obtain the final road loss prediction value : ; in, W fusion is the weight matrix of the final fusion layer, with a dimension of 192×1, b fusion is the bias term of the final fusion layer.

[0034] S4, training the three-channel network model; The data set is divided into a training set and a test set at an 8:2 ratio, and cross-validation is used to verify the performance of the three-channel network model. During the training process, the training set is used to update the parameters of the three-channel network model, and the loss function is used to measure the difference between the predicted value of the three-channel network model and the actual target value. The loss function is defined as follows: ; in, U is the number of samples. By minimizing the loss function, the three-channel network model can gradually adjust the parameters to make the road loss prediction value The actual value of road loss L The error between them is reduced, and the Adam optimizer is used to optimize the parameters of the three-channel network model to effectively process high-dimensional data and accelerate convergence.

[0035] S5, combine S2 and S4 to design the objective function; Traffic mobility in the design of objective function E ’ and road loss calculation value L ’ These are the two core indicators that affect path selection, directly affecting the direction and accuracy of the optimization results, and the calculated value of road loss. L ’ It indicates that the road is consumed or has reduced traffic capacity due to various road conditions during use, which is related to the speed limit. V , Traffic density D , Management Model M , Path Type Tand environmental conditions C Closely related, the objective function design includes the following steps: S51, the three-channel network model trained in S4 is combined with the traffic density D Feature extraction module, regarded as the road loss calculation value L ’ Function F: L ’ =F( V , D , M , T , C ); The deep learning model F is used as an implicit function and embedded into the objective function to process different types of traffic data and calculate the road loss value with high accuracy. L ’ , thus improving the accuracy of the optimization target.

[0036] S52. Calculate traffic mobility E ’ , traffic mobility E ’ Refers to the degree of mobility of vehicles in the traffic system and the speed limit on the road section V and traffic density D Peak value within the cycle D peak Related, improve traffic mobility E ’ It helps to reduce congestion, improve traffic efficiency, and speed up the passage speed of the path. The calculation formula is: E ’ = V · D peak .

[0037] S53. Improve traffic mobility E ’ It helps improve the efficiency of path planning, but over-optimization or under-optimization may lead to the calculated value of road loss. L ’ The increase in traffic mobility E ’ and road loss calculation value L ’ Perform weighted combination to provide comprehensive evaluation criteria for the optimization algorithm and comprehensive objective function Z The design is: Z = w · E ’ -(1- w ) L ’ ; in, w is the weight coefficient used to control traffic mobility E ’ The relative importance of the calculated value of road loss L' in path planning optimization is adjusted according to traffic conditions.

[0038] S6, using genetic algorithm to search for the optimal solution; In path planning optimization, the goal is to maximize traffic flow simultaneously E ’ And minimize the calculated value of road loss L', that is, maximize the comprehensive objective function Z ,The constraints are mainly the value range restrictions of each feature, and the genetic algorithm is used to search for the optimal solution.

[0039] S61. Randomly generate a set of path planning solutions and calculate the objective function Z , and evaluate the fitness. The higher the fitness, the closer the path planning scheme is to the optimal solution.

[0040] S62. Based on fitness, excellent individuals are selected as parents, and through crossover operations, the genetic information of the parents is reorganized to generate new offspring individuals.

[0041] S63. Randomly mutate some individuals. The mutation operation changes some gene values ​​of the individuals, thereby exploring a wider solution space and replacing the old population with newly generated offspring to form the next generation population.

[0042] S64. When at least one of the two conditions, namely, the maximum number of iterations and the fitness reaching the preset threshold, is met, the algorithm terminates and outputs the optimal solution, that is, the optimal combination of the values ​​of each feature.

[0043] Through S6, under given constraints, the comprehensive objective function is obtained Z The maximum feature combination is obtained to achieve the best path planning. V , Management Model M and environmental conditions C Usually these are a priori values, which can be set according to actual conditions or collected through system collection. D It can be obtained in real time through the roadside unit. Therefore, by fixing the prior features, the optimization system can be used to search for the path type that optimizes the core indicators. T , to achieve optimal path planning.

[0044] Therefore, the present invention adopts the above-mentioned multi-objective path planning optimization method driven by implicit functions of a multi-channel network, processes different types of data simultaneously through three parallel processing channels, and integrates and outputs them through a fusion layer, accurately reflects the complexity and dynamic changes of the traffic system, improves the adaptability and accuracy of path planning, and significantly improves the system's processing capabilities and comprehensive expression capabilities for different data types. During the parameter optimization process, complex nonlinear relationships are processed by implicit functions and embedded into the objective function of the optimization algorithm, effectively making up for the shortcomings of traditional optimization algorithms in dealing with complex nonlinear problems in traffic systems, improving the accuracy of path planning, and fully tapping the potential regularity of the traffic system.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-objective path planning optimization method driven by an implicit function for a multi-channel network, characterized in that: The following steps are involved: S1, obtain traffic data and pre-process it; S2, extracting features of the vehicle density signal in the traffic data from the time domain and frequency domain dimensions; S3, construct a three-channel network model; S4, training the three-channel network model; S5, combine S2 and S4 to design the objective function; S6. Use genetic algorithm to search for the optimal solution.

2. The implicit function driven multi-objective path planning optimization method for a multi-channel network according to claim 1, characterized in that: In S1, traffic data specifically includes: route type T Include m Category 1; Environmental conditions C Indicates the influence of external factors, including m 2 levels; traffic management mode M Include m 3 types; speed limit V Set to V a to V b Discrete values ​​between D Refers to different road conditions N Traffic density sampling points, sampling is performed at equal intervals in each traffic flow cycle, the unit is vehicle / hour; and the corresponding actual value of road loss L ; The preprocessing of traffic data includes: filling missing values, processing outliers, ensuring the integrity and accuracy of the data set. The final complete data set includes no less than one sample, and each sample includes the route type. T , Environmental conditions C , Traffic Management Mode M , speed limit V , Traffic density D， And the corresponding actual value of road loss L .

3. The implicit function driven multi-objective path planning optimization method for a multi-channel network according to claim 2, characterized in that: In S2, the traffic density D The time domain features of the signal include maximum value, minimum value, sharpness, symmetry, skewness, peak-to-peak value, absolute peak-to-average ratio, average value and standard deviation. The time domain features reflect the amplitude, fluctuation degree and distribution characteristics of the signal; Frequency domain characteristics include main frequency, spectrum energy, spectrum center, frequency bandwidth, harmonic energy ratio and harmonic energy concentration; There are 15 time domain features and frequency domain features in total.

4. The implicit function driven multi-objective path planning optimization method for a multi-channel network according to claim 3, characterized in that: In S3, the construction of the three-channel network model is as follows: S31, the 15 time domain features and frequency domain features in S2 and the speed limit in S1 V Normalized as a numerical feature, the calculation formula is: ； in, X represents the original data, X’ is the normalized value, the normalization interval is [0,1], and the feature vector form finally input to the numerical channel is ,in They are the 15 normalized time domain features and frequency domain features and the speed limit in S1 V ; S32, Traffic Management Mode M , Path Type T and environmental conditions C As a categorical feature, one-hot encoding is used to convert the value of each category into a binary vector, where each category is represented by an independent dimension; Path Type T Include m 1 type, after encoding, a length of m Binary vector of 1s; environmental conditions C Include m 2 levels, after encoding, a length of m Binary vector of 2; traffic management mode M Include m 3 types, after encoding, a length of m 3; concatenate the above three binary vectors to obtain the feature vector of the input classification channel X cat ， ； S33, Traffic density D It has obvious temporal characteristics and is classified as a temporal feature. The long short-term memory network is used to capture the long-term dependencies in the time series. The feature vector input to the temporal channel X seq By traffic density D of N Sampling points composition, ; S34, the numerical channel and the classification channel process their respective input features through a fully connected layer. ； ； in, W num It is a numerical channel weight matrix with a dimension of 16×64; b num It is a numerical channel bias term with a dimension of 64, indicating the number of neurons in the output layer; ReLU is the activation function; W cat is the classification channel weight matrix, the dimension is ( m 1+ m 2+ m 3) ×64; b cat It is the classification channel bias item with a dimension of 64; Y num and Y cat is the output of the fully connected layer of each channel, in the form of a numerical vector, representing the feature information after ReLU transformation; S35, the time series channel is processed by the long short-term memory network LSTM to obtain a hidden state sequence at each time step t , the output hidden state of LSTM h t Depends on the current input d t and the state at the previous moment h t-1 , specifically: ； in, c t is the cell state, responsible for preserving long-term memory; the output length of the LSTM layer is N The hidden state sequence of , set the size of the hidden layer to 64, and select the hidden state of the last time step of the LSTM output sequence h N As a summary representation of the entire time series data; The hidden state of the last time step of LSTM is connected through a fully connected layer h N Mapped to the same feature space dimension as the numerical channel and the categorical channel, specifically: ； in, Y seq is the output of the fully connected layer of the temporal channel, W lstmfc and b lstmfc They are the weight matrix and bias term of the fully connected layer of the temporal channel respectively; S36, introduce the cross-channel attention mechanism, calculate the importance of the output features of the numerical channel, classification channel and time series channel, and fuse them in a weighted manner. The fused features are: ； in, Y fusion Represents the final feature in the form of a numerical vector, ɑ x are the attention weights of the three channels, calculated through a scoring network with learning capabilities; S37, fused features Y fusion Through a final fully connected layer, it is mapped to the final output space to obtain the final road loss prediction value : ； in, W fusion is the weight matrix of the final fusion layer, with a dimension of 192×1, b fusion is the bias term of the final fusion layer.

5. The implicit function driven multi-objective path planning optimization method of a multi-channel network according to claim 4, characterized in that: In S4, specifically: the data set is divided into a training set and a test set at an 8:2 ratio, and cross-validation is used to verify the performance of the three-channel network model. During the training process, the training set is used to update the parameters of the three-channel network model, and the loss function is used to measure the difference between the predicted value of the three-channel network model and the actual target value. The loss function is defined as follows: ； in, U is the number of samples. By minimizing the loss function, the three-channel network model can gradually adjust the parameters to make the road loss prediction value The actual value of road loss L The error between them is reduced.

6. The implicit function driven multi-objective path planning optimization method of a multi-channel network according to claim 5, characterized in that: In S5, the objective function design includes the following steps: S51, the three-channel network model trained in S4 is combined with the traffic density D Feature extraction module, regarded as the road loss calculation value L ’ Function F: L ’ =F( V , D , M , T , C )； S52. Calculate traffic mobility E ’ , traffic mobility E ’ Refers to the degree of mobility of vehicles in the traffic system and the speed limit on the road section V and traffic density D Peak value within the cycle D peak Related, the calculation formula is: E ’ = V · D peak ； S53. Make traffic mobility E ’ and road loss calculation value L ’ Perform weighted combination to provide comprehensive evaluation criteria for the optimization algorithm and comprehensive objective function Z The design is: Z = w · E ’ - (1- w )· L ’ ； in, w is the weight coefficient used to control traffic mobility E ’ and road loss calculation value L ’ The relative importance of the route planning optimization is adjusted according to the traffic situation.

7. The implicit function driven multi-objective path planning optimization method of a multi-channel network according to claim 6, characterized in that: In S6, the specific steps are: S61. Randomly generate a set of path planning solutions and calculate the objective function Z , and evaluate fitness; S62, based on fitness, select excellent individuals as parents, and through crossover operation, reorganize the genetic information of the parents to generate new offspring individuals; S63, randomly mutate some individuals, replace the old population with the newly generated offspring, and form the next generation population; S64. When one of the two conditions, namely, the maximum number of iterations is reached and the fitness reaches a preset threshold, is met, the algorithm terminates and outputs the optimal solution, that is, the optimal combination of the values ​​of each feature.

Images