DAS driving strategy optimization method driven by subway line real-time data

Abstract

The invention relates to the technical field of subway train driving strategy optimization, and discloses a subway line real-time data driven DAS driving strategy optimization method, which comprises the following steps: acquiring multi-dimensional operation data including a train real-time position sequence, station passenger flow data and a trackside signal equipment state; extracting line operation characteristics through a convolution space-time encoder, generating a passenger flow fluctuation prediction map by using a time sequence decomposition algorithm, and performing characteristic fusion on the trackside signal equipment state to generate an equipment health degree evaluation matrix; inputting the multi-dimensional data into the driving strategy model to generate an initial driving instruction sequence, dividing operation control priorities through a regional clustering algorithm, and constructing a dynamic optimization strategy network in combination with a genetic algorithm to generate a final driving strategy scheme; and acquiring an execution state log in real time, and updating the strategy network through the abnormal decision detection model. According to the method, the dynamic optimization of the subway driving strategy is realized, and the operation safety, efficiency and comfort are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of optimizing subway train driving strategies, and specifically to a method for optimizing DAS driving strategies driven by real-time data of subway lines. Background Art

[0002] With the rapid development of urban rail transit, the subway has become a core component of modern urban public transportation. However, with the continuous increase in passenger flow, the improvement of line complexity, and the dynamic changes in the operating status of equipment, traditional subway driving strategies have gradually revealed many deficiencies and are difficult to meet the higher requirements of modern subway operations for safety, efficiency, comfort, and economy.

[0003] Most traditional driving strategies are based on fixed timetables and preset operating parameters, lacking the effective utilization of real-time dynamic data and the ability of adaptive adjustment. In terms of passenger flow, traditional methods cannot accurately perceive the changes in passenger flow at stations in real time, such as the density of inbound passenger flow and the congestion degree of transfer channels, resulting in insufficient transportation capacity during peak hours or waste of transportation capacity during off-peak hours, affecting the passenger experience and operating efficiency. For example, during the morning and evening rush hours, transfer stations often face huge passenger flow pressure, and traditional strategies are difficult to adjust the train stopping time or departure interval according to the real-time congestion degree, easily causing platform congestion and even potential safety hazards.

[0004] In terms of the train operating status, traditional strategies do not process key information such as the speed change gradient and braking distance parameters in the real-time position sequence of the train finely enough, and cannot extract the line operation characteristics in a timely and accurate manner, such as the interval passing delay coefficient, signal response interval time, and track occupancy status label. This makes it possible for the train to fail to respond to sudden situations on the line in a timely manner, such as signal delays caused by track occupancy, thereby affecting the operating efficiency and safety of the train.

[0005] In terms of the on-site signal equipment status management, traditional methods lack the multi-dimensional feature fusion and health assessment of equipment status, and it is difficult to detect potential equipment failures in advance, resulting in sudden equipment failures that may affect the normal operation of the train and even cause operation accidents. For example, if minor faults of signal equipment cannot be detected and processed in time, they may trigger a chain reaction under specific conditions, resulting in large-scale operation delays.

[0006] In addition, the optimization process of traditional driving strategies often relies on manual experience and offline simulation, lacking a dynamic real-time optimization mechanism. In the face of a complex and changeable operating environment, it is impossible to quickly generate driving strategies adapted to real-time situations, resulting in insufficient flexibility and response speed of train dispatching. For example, when a temporary fault occurs in a certain interval, traditional strategies are difficult to quickly adjust the train operation path and speed, resulting in a significant decrease in the operating efficiency of the entire line. Summary of the Invention

[0007] The object of the present invention is to provide a method for optimizing the DAS driving strategy driven by real-time subway line data to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: a method for optimizing the DAS driving strategy driven by real-time subway line data, the method comprising: obtaining a multi-dimensional operation data set; the multi-dimensional operation data includes a train real-time position sequence, station passenger flow data, and trackside signal equipment status; the train real-time position sequence includes a speed change gradient and a braking distance parameter, and the station passenger flow data includes an inbound passenger flow density and a congestion degree of a transfer passage; based on the train real-time position sequence, extracting line operation features through a convolutional spatio-temporal encoder, the line operation features including an interval passing delay coefficient, a signal response interval time, and a track occupancy status label; according to the station passenger flow data, generating a passenger flow fluctuation prediction map through a time series decomposition algorithm, the map including a peak time distribution interval and a heat map of operation conflict probability; performing multi-dimensional feature fusion on the trackside signal equipment status to generate an equipment health assessment matrix; inputting the line operation features, the passenger flow fluctuation prediction map, and the equipment health assessment matrix into a driving strategy model to generate an initial driving instruction sequence; based on the initial driving instruction sequence, dividing operation control priorities through a regional clustering algorithm to output a train scheduling weight distribution table; fusing the train scheduling weight distribution table with the initial driving instruction sequence, and constructing a dynamic optimization strategy network through a genetic algorithm to generate a final driving strategy plan.

[0009] Preferably, the extracting of the line operation features through the convolutional spatio-temporal encoder includes: intercepting the train real-time position sequence by a sliding window to generate spatio-temporal slice data blocks; dividing the spatio-temporal slice data blocks into multiple operation intervals based on a regional clustering rule, and calculating the standard deviation of signal response delay within each interval; extracting the track occupancy status label by using a residual connection algorithm, and generating a signal system response relationship in combination with the interval passing delay coefficient; encoding the signal response interval time, the track occupancy status label, and the signal system response relationship into line operation features.

[0010] Preferably, the generating of the passenger flow fluctuation prediction map through the time series decomposition algorithm includes: separating a trend component from the station passenger flow data to extract periodic fluctuation features; predicting the future passenger flow distribution based on an autoregressive model, and calculating the confidence level of the peak time distribution interval; generating a risk probability model according to a historical conflict event database, and outputting a heat map of operation conflict probability in combination with the confidence level; mapping the peak time distribution interval and the heat map of operation conflict probability into a passenger flow fluctuation prediction map.

[0011] Preferably, the driving strategy model includes a data fusion module and an instruction generation module. The data fusion module includes: standardizing the signal response interval time in the line operation characteristics to obtain a first fusion tensor; performing morphological filtering on the operation conflict probability heat map in the passenger flow fluctuation prediction map to generate a second fusion tensor; performing eigenvalue decomposition on the equipment health assessment matrix to extract the equipment state degradation gradient to obtain a third fusion tensor; and combining the first fusion tensor, the second fusion tensor, and the third fusion tensor into a policy input sequence through a bidirectional long short-term memory network.

[0012] Preferably, the instruction generation module includes: aligning the policy input sequence along the time axis to generate an instruction correlation vector; extracting key control node features through an attention mechanism to generate an instruction node correlation matrix; performing a tensor splicing operation on the instruction correlation vector and the instruction node correlation matrix to generate a candidate instruction set; and selecting an initial driving instruction sequence from the candidate instruction set through a branch and bound algorithm.

[0013] Preferably, constructing a dynamic optimization strategy network through a genetic algorithm includes: initializing a network population according to a train scheduling weight distribution table and generating a fitness evaluation function based on priorities; using the initial driving instruction sequence as chromosome encoding, and the fitness evaluation function consists of an operation energy consumption index and a scheduling priority weight; updating the fitness scores of each chromosome through a crossover mutation operator to adjust the population evolution direction; and generating an optimal driving strategy plan that meets multi-objective constraints according to the optimized fitness scores.

[0014] Preferably, the parameter optimization method for the regional clustering rule includes: calculating an initial clustering radius and a minimum signal delay threshold according to the historical operation data distribution; traversing parameter combinations through a grid verification algorithm and selecting the parameters with the highest matching degree between the partitioning result and the actual operation log; and dynamically adjusting the clustering radius and the minimum signal delay threshold according to the matching degree to optimize the partitioning accuracy of the signal system response relationship.

[0015] Preferably, the construction method of the risk probability model includes: collecting multi-line historical operation conflict data and extracting spatio-temporal correlation features and equipment failure influencing factors; fitting the spatio-temporal correlation features with a Gaussian mixture model to generate a basic risk probability distribution table;

[0016] Performing Bayesian correction on the basic risk probability distribution table according to the equipment failure influencing factors and outputting a risk probability model.

[0017] Preferably, the method for setting parameters of the crossover mutation operator includes: defining the crossover probability between chromosomes as a balance factor of the operating energy consumption index and the scheduling priority weight; initializing the fitness score of each chromosome to a reference value, and the fitness of the starting chromosome to a preset initial value; selecting the chromosome with the highest fitness score in the current population as the parent through the elitist retention strategy; performing multi-point crossover and random bit mutation operations based on the parent chromosome to generate a new generation of population.

[0018] Preferably, the method further includes: collecting the execution status log of the train control system in real time, generating a policy correction instruction set through an anomaly decision detection model, and updating the dynamic optimization policy network; the method for constructing the anomaly decision detection model includes: collecting normal operation samples in the historical execution status log, extracting the acceleration instruction change rate and the braking response time series; constructing a normal operation benchmark model based on one-class support vector machine, and calculating the deviation index between the real-time data and the benchmark model; statistically calculating the cumulative anomaly value of the deviation index through a time decay function, and generating a policy correction instruction set when it exceeds the dynamic threshold; updating the chromosome coding parameters of the dynamic optimization policy network in combination with the correction instruction set.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: In terms of data integration and feature extraction, by obtaining multi-dimensional operation data such as the train real-time position sequence, the passenger flow data at stations, and the status of trackside signal equipment, and using technologies such as convolutional spatio-temporal encoder, time series decomposition algorithm, and multi-dimensional feature fusion, the line operation characteristics can be accurately extracted, the passenger flow fluctuation prediction map and the equipment health assessment matrix can be generated. This enables the system to comprehensively and deeply understand the dynamic changes in the subway operation environment, providing a solid data basis for the optimization of driving strategies. For example, by intercepting the train real-time position sequence with a sliding window and dividing it into regional clusters, the standard deviation of the signal response delay in each interval can be accurately calculated, and the track occupancy status label can be extracted in combination with the residual connection algorithm, so as to generate an accurate signal system response relationship, providing a key basis for the real-time scheduling of trains.

[0020] In terms of driving strategy generation and optimization, the extracted multi-dimensional features are input into the driving strategy model. An initial driving instruction sequence is generated through the data fusion module and the instruction generation module, and the operation control priority is divided by using the regional clustering algorithm, and a train scheduling weight distribution table is output. On this basis, a dynamic optimization strategy network is constructed through the genetic algorithm. Combining fitness evaluation functions such as operation energy consumption indicators and scheduling priority weights, the initial driving instruction sequence is optimized to generate an optimal driving strategy plan that meets multi-objective constraints. This full-process dynamic optimization mechanism from data to strategy enables the driving strategy to adapt to changes in passenger flow, line operation conditions, and equipment status in real time, significantly improving the flexibility and response speed of train operation. For example, in the face of a sudden passenger flow peak, the system can adjust the train stopping time and departure interval in a timely manner according to the passenger flow fluctuation prediction map to relieve platform congestion; when the health of trackside signal equipment deteriorates, the system can adjust the train operation speed and path in advance to avoid operation accidents caused by equipment failures.

[0021] In terms of exception handling and strategy correction, by collecting the execution status logs of the train control system in real time, a strategy correction instruction set is generated by using the exception decision detection model to update the dynamic optimization strategy network. This model constructs a normal operation benchmark model based on one-class support vector machines, which can monitor the deviation degree of the train execution status from normal operations in real time. When an exception is detected, a correction instruction is generated in a timely manner to adjust the driving strategy. This real-time exception detection and strategy correction mechanism can effectively improve the safety and reliability of subway operations and reduce the operation delay and accident risks caused by sudden abnormal situations. For example, when the train braking response time is abnormal, the system can quickly detect it and generate a correction instruction to adjust the braking strategy to ensure the safe stop of the train.

[0022] In terms of comprehensive performance improvement, the method of the present invention realizes multi-objective optimization of subway operations through multi-dimensional data-driven and dynamic optimization mechanisms. It can not only reduce the train operation energy consumption and improve the energy utilization efficiency, but also optimize the train scheduling priority and improve the overall operation efficiency; at the same time, through accurate passenger flow prediction and equipment status management, it can improve the comfort and safety of passengers. For example, by optimizing the driving strategy through the genetic algorithm, on the premise of ensuring the safe operation of the train, it can effectively reduce the operation energy consumption and achieve a win-win situation of economic and environmental benefits; by reasonably adjusting the train operation speed and stopping time, it can reduce the waiting time of passengers on the platform and the congestion in the carriage, improving the passenger experience. Brief Description of the Drawings

[0023] Figure 1 It is the working principle diagram of the DAS driving strategy optimization method driven by real-time data of the subway line described in the present invention;

[0024] Figure 2Flow chart for generating a passenger flow fluctuation prediction map by a time series decomposition algorithm;

[0025] Figure 3 Flow chart for the data fusion module of the driving strategy model;

[0026] Figure 4 Flow chart for the instruction generation module of the driving strategy model. Specific implementation manners

[0027] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0028] Please refer to Figures 1 - 4 , the DAS driving strategy optimization method driven by real-time subway line data involved in the present invention is specifically implemented as follows: Obtain multi-dimensional operation data including train real-time position sequences, station passenger flow data, and trackside signal equipment statuses. Among them, the train real-time position sequence includes speed change gradients and braking distance parameters, which are used to reflect the dynamic operation state of the train on the line; the station passenger flow data includes inbound passenger flow density and transfer channel congestion, which are used to characterize the passenger flow distribution characteristics within the station; the trackside signal equipment status data is used to monitor the real-time working conditions of the signal equipment.

[0029] Based on the train real-time position sequence, extract line operation characteristics through a convolutional spatio-temporal encoder. The specific process is as follows: Intercept the train real-time position sequence through a sliding window to generate spatio-temporal slice data blocks; divide the spatio-temporal slice data blocks into multiple operation intervals based on regional clustering rules, and calculate the standard deviation of signal response delays within each interval; use a residual connection algorithm to extract track occupancy status labels, and combine the interval passage delay coefficients to generate signal system response relationships; encode the signal response interval time, track occupancy status labels, and signal system response relationships into line operation characteristics, including interval passage delay coefficients, signal response interval times, and track occupancy status labels.

[0030] According to the station passenger flow data, generate a passenger flow fluctuation prediction map through a time series decomposition algorithm. The specific steps are as follows: Separate the trend component of the station passenger flow data and extract periodic fluctuation characteristics; predict the future passenger flow distribution based on an autoregressive model and calculate the confidence level of the distribution interval during peak hours; generate a risk probability model according to the historical conflict event database, and output a heat map of operation conflict probabilities in combination with the confidence level; map the distribution interval during peak hours and the heat map of operation conflict probabilities into a passenger flow fluctuation prediction map, including the distribution interval during peak hours and the heat map of operation conflict probabilities.

[0031] Fuse the multi-dimensional features of the trackside signal equipment status, and generate an equipment health assessment matrix through operations such as data cleaning, feature extraction, and normalization processing, comprehensively reflecting the health status of the equipment.

[0032] Input the line operation characteristics, passenger flow fluctuation prediction map, and equipment health assessment matrix into the driving strategy model, which includes a data fusion module and an instruction generation module. The data fusion module standardizes the signal response interval time in the line operation characteristics to obtain the first fusion tensor; performs morphological filtering on the operation conflict probability heat map in the passenger flow fluctuation prediction map to generate the second fusion tensor; performs eigenvalue decomposition on the equipment health assessment matrix to extract the equipment state degradation gradient to obtain the third fusion tensor; combines the first fusion tensor, the second fusion tensor, and the third fusion tensor into a strategy input sequence through a bidirectional long short-term memory network. The instruction generation module aligns the strategy input sequence along the time axis to generate an instruction correlation vector; extracts the key control node features through an attention mechanism to generate an instruction node correlation matrix; performs a tensor splicing operation on the instruction correlation vector and the instruction node correlation matrix to generate a candidate instruction set; selects an initial driving instruction sequence from the candidate instruction set through a branch and bound algorithm.

[0033] Based on the initial driving instruction sequence, divide the operation control priority through a regional clustering algorithm, calculate the priority weights of each region according to the operation characteristics, passenger flow conditions, and equipment status of different regions, output a train scheduling weight distribution table, and clarify the scheduling priorities of different regions.

[0034] Fuse the train scheduling weight distribution table with the initial driving instruction sequence, and construct a dynamic optimization strategy network through a genetic algorithm. The specific process is as follows: Initialize the network population according to the train scheduling weight distribution table, and generate a fitness evaluation function based on the priority, which consists of the operation energy consumption index and the scheduling priority weight; use the initial driving instruction sequence as the chromosome encoding, update the fitness scores of each chromosome through crossover and mutation operators, and adjust the population evolution direction; generate an optimal driving strategy solution that meets multi-objective constraints according to the optimized fitness scores.

[0035] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1: In the process of extracting line operation characteristics, when processing the real-time position sequence of the train, it is necessary to first determine the interception method of the sliding window. The size of the sliding window is not fixed, but is dynamically adjusted according to the line section length and the train operation speed. For example, when the train is running on a line section with a longer section length and a higher design speed, the size of the sliding window is correspondingly increased to ensure that enough position data points can be included in one window to accurately reflect the change of the train operation state in this section; while in the area with a shorter section and a lower speed limit, the sliding window is appropriately reduced to avoid data redundancy or feature extraction deviation caused by too large a window. Through this dynamic adjustment mechanism, the generated spatio-temporal slice data block can more accurately correspond to the actual operation scenario, providing a reliable data basis for subsequent feature extraction.

[0036] When dividing the spatio-temporal slice data block into multiple operation sections based on the regional clustering rule, it is first necessary to determine the parameters of the regional clustering rule. Specifically, first calculate the initial clustering radius and the minimum signal delay threshold according to the historical operation data distribution. The historical operation data covers the operation records of the train at different time periods and under different line conditions. Through the statistical analysis of these data, the distribution law of the train operation state parameters can be obtained, so as to initially determine the reasonable range of the clustering radius and the signal delay threshold. For example, by analyzing the distribution of the signal response delay in the historical data, an initial minimum signal delay threshold is determined, and the area where the signal response delay is less than this threshold is regarded as the preliminary basis for dividing the same type of operation section.

[0037] Next, use the grid verification algorithm to traverse the parameter combinations. The grid verification algorithm generates different combinations of the clustering radius and the minimum signal delay threshold within a certain range, divides the operation section for each combination, and compares the division result with the actual operation log. The actual operation log details the actual operation of the train in each section, including key information such as signal response time and track occupancy status. By calculating the matching degree between the division result and the actual operation log, such as judging whether the standard deviation of the signal response delay in the divided operation section is consistent with the fluctuation situation in the actual log for this section, the effectiveness of different parameter combinations is evaluated. During the traversal process, a large number of parameter combinations need to be tested to ensure that the combination with the highest matching degree can be selected.

[0038] After selecting the parameter with the highest matching degree, it does not mean that the parameter optimization process is over. Instead, it is necessary to dynamically adjust the clustering radius and the minimum signal delay threshold according to the matching degree results. For example, if it is found that the matching degree between the partitioning result under the current parameter combination and the actual operation log is low during certain special periods (such as morning and evening rush hours), it may be due to factors such as increased train operation density and increased signal system load during peak hours. At this time, it is necessary to appropriately adjust the clustering radius or the minimum signal delay threshold according to the specific matching degree difference to improve the partitioning accuracy in different scenarios. Through this dynamic adjustment mechanism, the partitioning of the signal system response relationship can be continuously optimized to make it more in line with the actual operation situation, laying a foundation for accurately extracting the line operation characteristics in the future.

[0039] When using the residual connection algorithm to extract the track occupancy status label, it is necessary to construct a suitable neural network structure. Specifically, a multi-layer convolutional neural network (CNN) is used to extract features from the spatio-temporal slice data block. The convolutional neural network has a powerful spatial feature extraction ability and can capture the spatial distribution characteristics of the track occupancy status from the spatio-temporal slice data. In the multi-layer convolutional neural network, each convolutional layer is responsible for extracting features at different levels. The shallow convolutional layer extracts relatively basic features, such as local speed change patterns; the deep convolutional layer extracts more complex and abstract features, such as the track occupancy trend of the entire section.

[0040] To alleviate the problem of gradient disappearance that may occur in the deep network, a residual connection structure is introduced into the convolutional neural network. The core idea of the residual connection is to establish an identity mapping path in the network, enabling the gradient to be propagated back more smoothly, thus allowing the training of deeper neural networks. Through the residual connection, the network can more easily optimize the complex feature extraction process during the learning process to ensure the accurate extraction of the track occupancy status label. For example, in a certain convolutional layer, the input data is directly added to the output data of the convolutional operation and activation function processing of this layer after passing through the convolutional operation and activation function, which enables the network to more effectively capture the subtle changes in the track occupancy status during the learning process.

[0041] After extracting the track occupancy status label, it is necessary to generate the signal system response relationship in combination with the section passing delay coefficient. The section passing delay coefficient reflects the passing efficiency of the train in this section, and its calculation is based on the ratio of the actual running time of the train in this section to the theoretical shortest running time. By performing correlation analysis on the track occupancy status label and the section passing delay coefficient, the response characteristics of the signal system under different track occupancy statuses can be determined. For example, when the track occupancy status label shows that a certain section is in a busy state, if the passing delay coefficient of this section is large, it indicates that the response of the signal system in this state may be delayed, and it is necessary to further optimize the signal system scheduling strategy.

[0042] When encoding the signal response interval time, track occupancy status tags, and signal system response relationships into line operation characteristics, a multi-dimensional vector encoding method is adopted. Specifically, the signal response interval time is converted into a dimension in the vector, and the value of this dimension directly reflects the response speed of the signal system; the track occupancy status tags are converted into multiple dimensions through one-hot encoding, with each dimension corresponding to a specific track occupancy status (such as idle, occupied, about to be occupied, etc.); the signal system response relationship is converted into other dimensions in the vector by extracting key feature parameters (such as mean response delay, variance, etc.). Through this multi-dimensional vector encoding method, different types of feature parameters can be integrated into a unified vector space, facilitating subsequent data fusion processing and input to the driving strategy model.

[0043] During the entire process of extracting line operation characteristics, there are close logical connections and data transmissions between various steps. The spatio-temporal slice data blocks generated by sliding window interception are the basis for all subsequent processing. The parameter optimization of the regional clustering rule ensures the rationality of the operation interval division. The residual connection algorithm guarantees the accuracy of track occupancy status tag extraction, and the multi-dimensional vector encoding facilitates the effective utilization of features. Through this series of rigorous processing procedures, comprehensive and accurate line operation characteristics can be extracted from the train's real-time position sequence, providing key input information for the optimization of subway driving strategies, enabling the driving strategies to better adapt to the actual operation conditions of the line, and improving the safety and efficiency of subway operation.

[0044] Example 2: When generating the passenger flow fluctuation prediction map, multi-stage processing of the passenger flow data at the station is required to extract effective features and complete the prediction mapping. First, for the trend component separation link in the data, the seasonal decomposition algorithm (STL) is used to process the original flow data. This algorithm iteratively decomposes the flow sequence into a trend term, a seasonal term, and a remainder term: the trend term reflects the overall change trend of passenger flow over a long time range, such as the annual increase in passenger flow at a station in a certain area with the development of the city; the seasonal term captures the fluctuating characteristics with periodic patterns, such as the difference in passenger flow between weekdays and weekends, and the fixed time distribution of morning and evening rush hours; the remainder term contains the random fluctuation part in the data that cannot be explained by the trend term and the seasonal term. Through this decomposition process, the periodic fluctuation characteristics of passenger flow data can be clearly extracted, such as clearly identifying the fixed passenger flow peak period patterns during the morning rush hour (7:30 - 9:00) and evening rush hour (17:30 - 19:00) on weekdays, and the rule that the passenger flow distribution is relatively flat throughout the weekend but may have a small peak at noon.

[0045] When predicting the future passenger flow distribution based on an autoregressive model, it is first necessary to determine the order of the model according to historical data. By analyzing the passenger flow data sequence over a past period (such as the passenger flow data every 15 minutes in the past 30 days), the partial autocorrelation function (PACF) is used to judge the reasonable order of the autoregressive model. For example, if the PACF truncates after a lag of 3, a 3-order autoregressive model (AR(3)) is selected as the prediction model. The model parameters are estimated using the least squares method. By minimizing the sum of the squared errors between the predicted values and the actual values, the coefficients of each lag term in the model are determined, thus establishing a quantitative relationship between the passenger flow data and the historical values. For example, a prediction model in the form of is established (only for illustration here, not involving formula derivation), where is the passenger flow value at the current moment, , , are the passenger flow values at historical moments, , , are the model coefficients, and is the random error term. Through this model, the passenger flow distribution in future periods can be predicted. At the same time, the confidence level of the distribution interval during peak hours is calculated based on historical prediction errors. For example, a confidence interval (such as a 90% confidence level) is determined by calculating the standard deviation of the predicted values to quantify the reliability of the prediction results and clarify the possible time range and probability level during peak hours.

[0046] When constructing a risk probability model, it is first necessary to collect historical operation conflict data of multiple lines. These data cover spatio-temporal correlation features such as the time and location of conflicts, the running speeds of the trains involved, and the status of signal equipment, as well as equipment failure impact factors such as equipment failure types (such as signal machine failures, track circuit failures) and failure durations. When analyzing the spatio-temporal correlation features, the Gaussian mixture model (GMM) is used for fitting. The Gaussian mixture model assumes that the data is composed of multiple Gaussian distributions mixed together. The means, variances, and weights of each Gaussian component are estimated through the expectation maximization (EM) algorithm, thus generating a basic risk probability distribution table. For example, through analysis, it is found that the probability of operation conflicts on a certain line is relatively high during morning and evening peak hours and near transfer stations. These areas and periods can be classified as high-risk areas, and relatively high initial probability values are assigned in the basic risk probability distribution table.

[0047] Subsequently, Bayesian correction is performed on the basic risk probability distribution table according to the influencing factors of equipment failures. Bayesian correction adjusts the basic risk probability by introducing prior knowledge of equipment failures. For example, if it is known that after a failure of a certain type of signal equipment (such as abnormal signal display), the probability of operation conflicts in this area will increase significantly, then based on the basic risk probability, the posterior risk probability when the equipment failure occurs is calculated using Bayes' formula, so as to obtain a risk probability model that better fits the actual situation. In specific operations, first determine the conditional probability relationship between equipment failures and operation conflicts, and then dynamically adjust the risk probability of the corresponding area and time period according to the real-time monitored equipment status (such as whether a failure occurs), so that the risk probability model can reflect the impact of equipment status on operation conflicts in real time.

[0048] When mapping the peak-hour distribution interval and the operation conflict probability heat map into a passenger flow fluctuation prediction map, Geographic Information System (GIS) technology is used to realize the visualization of spatio-temporal information. First, digitize the subway line map and the locations of each station to establish a spatial geographic coordinate system. For the peak-hour distribution interval, taking the time axis as the dimension, mark the time range and passenger flow density levels (such as high density, medium density, low density) of the peak hours at the corresponding station locations. For example, during the morning peak hours, transfer stations are marked in red (representing high passenger flow density), and ordinary stations are marked in orange (representing medium passenger flow density). For the operation conflict probability heat map, according to the conflict probability values of each area calculated by the risk probability model, it is visualized through color gradients: areas with lower conflict probabilities are shown in green, medium probability areas are shown in yellow, and high probability areas are shown in red. The darker the color, the higher the conflict probability, intuitively presenting the risk levels of different areas.

[0049] In the specific implementation process, it is necessary to ensure that the time dimension and the spatial dimension of the peak-hour distribution interval and the operation conflict probability heat map are completely aligned. For example, for the morning peak hours (7:30 - 9:00) of a certain station, the operation conflict probability distribution within this time period for this station and the surrounding line sections is synchronously displayed in the heat map, so that the two form an interrelated overall map. Through this mapping method, the driving strategy model can intuitively obtain the spatio-temporal correlation characteristics between passenger flow fluctuations and operation risks. For example, it can be identified that during the morning peak hours and in the area near transfer stations, not only is the passenger flow density high, but also the operation conflict probability increases significantly. Therefore, when formulating driving strategies, more attention is paid to this area, and corresponding dispatching measures (such as increasing the train stop time, adjusting the departure frequency) are taken to reduce the conflict risk and improve the operation efficiency.

[0050] In the entire process of generating the passenger flow fluctuation prediction map, from data decomposition, model prediction to risk modeling and visualization mapping, each link is closely connected and mutually verified. The seasonal decomposition algorithm ensures the accurate extraction of the periodic characteristics of passenger flow. The autoregressive model realizes the quantitative prediction of future passenger flow. The combination of the Gaussian mixture model and Bayesian correction constructs a dynamically adjustable risk probability model. And the geographic information system technology transforms the abstract data analysis results into an intuitive and easy-to-understand map form. Through this series of rigorous processing procedures, a passenger flow fluctuation prediction map containing the distribution interval of peak hours and the heat map of operation conflict probability can be generated, providing comprehensive passenger flow information support for the optimization of subway driving strategies, enabling the driving strategies to respond to passenger flow fluctuations and potential risks in advance, and realizing more intelligent and safer train dispatching and operation control.

[0051] Embodiment 3: In the data fusion module of the driving strategy model, the processing of multi-source data needs to go through multiple links such as standardization, filtering, feature decomposition and time series modeling to achieve the effective fusion of data in different dimensions. First, standardize the signal response interval time in the line operation characteristics. The Z-score standardization method is adopted. Its core idea is to convert the original data into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of different feature dimensions. The specific formula is: Among them, represents the original signal response interval time data, is the mean of this data, is the standard deviation, is the first fusion tensor after standardization. Through this processing, the signal response interval times of different lines and different time periods can be unified to the same dimension scale, facilitating subsequent tensor operations and feature fusion with other data.

[0052] For the heat map of operation conflict probability in the passenger flow fluctuation prediction map, morphological filtering processing is required to remove noise and smooth the edges. Morphological filtering is based on the theory of mathematical morphology. By designing a structural element, operations such as erosion and dilation are performed on the pixel matrix of the heat map. For example, a square structural element with a size of is used. First, perform an erosion operation on the heat map to remove isolated noise points on the edge, and then perform a dilation operation to fill small holes inside, thereby generating a smoother and more reliable second fusion tensor. This process can effectively reduce outliers caused by sensor errors or data acquisition fluctuations, making the spatial distribution characteristics of the operation conflict probability clearer and providing more accurate input for subsequent data fusion.

[0053] For the processing of the equipment health assessment matrix, the eigenvalue decomposition method is used to extract the equipment state degradation gradient. The equipment health assessment matrix is a square matrix composed of multi-dimensional features (such as voltage, current, temperature, working hours of signal equipment, etc.), denoted as Through the Singular Value Decomposition (SVD) algorithm, a matrix can be decomposed into , where and are orthogonal matrices, is a diagonal matrix, and its diagonal elements are the singular values of the matrix . The singular values are arranged from largest to smallest, reflecting the importance degree of each feature and the energy distribution in the matrix. By analyzing the changing trend of the singular values over time, the degradation gradient of the device state can be extracted, that is, the descent rate or change slope of the singular values, as the third fusion tensor. For example, if the singular values of a certain device show an obvious downward trend recently, it indicates that the health state of the device may be degrading, and its potential impact on the signal system response needs to be considered in the driving strategy.

[0054] After the generation of the three fusion tensors is completed, data merging is performed through a Bidirectional Long Short-Term Memory network (Bi-LSTM). The Bi-LSTM network consists of a forward and a backward LSTM layer, which can capture both the forward and backward temporal dependencies in the data and is suitable for processing multi-dimensional data with time series characteristics. In specific implementation, the first fusion tensor (the standardized signal response interval time), the second fusion tensor (the filtered heat map of the running conflict probability), and the third fusion tensor (the degradation gradient of the device state) are aligned along the time dimension to form an input sequence , where each time step contains the corresponding dimensional features of the three tensors.

[0055] Each LSTM cell in the Bi-LSTM network contains a forget gate, an input gate, and an output gate, and controls the retention and forgetting of information through a gating mechanism. The forward LSTM layer processes data sequentially from the starting time step to the ending time step of the sequence, capturing the context information of past moments; the backward LSTM layer processes data from the ending time step to the starting time step, capturing the context information of future moments. The outputs of the two layers are concatenated in the hidden layer to generate a policy input sequence containing bidirectional temporal information. For example, for an abnormal prolongation of the signal response interval time at a certain moment, the forward LSTM layer can combine the previous signal delay history to judge whether it is a persistent problem, and the backward LSTM layer can predict whether this abnormality will trigger a chain reaction through the subsequent changes in the device health state, so that the policy input sequence can comprehensively reflect the temporal correlation and causal relationship in the data.

[0056] In the actual operation of data fusion, attention should be paid to the dimension matching problem of different tensors. For example, the first fusion tensor of the signal response interval time may be a one-dimensional vector (each time step corresponds to a scalar value), the second fusion tensor of the running conflict probability heat map may be a two-dimensional matrix (probability distribution corresponding to spatial positions), and the third fusion tensor of the device state degradation gradient may be a multi-dimensional vector (including degradation characteristics of multiple devices). To achieve effective merging, the low-dimensional tensor needs to be lifted to the same number of dimensions as the highest-dimensional tensor through dimension expansion operations. For example, expand a one-dimensional tensor into a two-dimensional or three-dimensional tensor to align it with other tensors on the time axis and spatial axis. Specifically, dimension expansion can be performed by filling with zero values or repeating eigenvalue to ensure that when input into the Bi-LSTM network, the dimension structures of all tensors are consistent, avoiding calculation errors caused by dimension mismatch.

[0057] In addition, the time resolution differences of different data sources also need to be considered during the data fusion process. For example, the acquisition frequency of the train's real-time position sequence may be once per second, while the acquisition frequency of the passenger flow data at the station may be once per minute. To solve this problem, interpolation processing such as linear interpolation or cubic spline interpolation needs to be performed on the data with low time resolution first to increase its time resolution to be consistent with that of the high-frequency data, ensuring the alignment accuracy on the time axis. For example, interpolate the passenger flow data once per minute into data points once per second to make its time stamps exactly match those of the train position data, so as to ensure that the correlation relationship between multi-source data at the same moment can be accurately captured in the Bi-LSTM network.

[0058] The processing flow of the entire data fusion module closely revolves around the characteristics of multi-source data. The data quality is improved through preprocessing steps such as standardization, filtering, and eigen-decomposition, and the deep extraction and fusion of temporal features are achieved through the Bi-LSTM network. Standardization processing solves the problem of dimension difference, morphological filtering enhances the reliability of spatial features, eigenvalue decomposition discovers the potential change rules of device states, and the Bi-LSTM network captures the long-distance dependence relationships in the data through bidirectional temporal modeling. These steps cooperate with each other to transform the line operation characteristics, passenger flow fluctuation prediction maps, and device health assessment matrices into a strategy input sequence containing rich spatio-temporal information and temporal correlations, providing comprehensive and accurate input data for the subsequent instruction generation module, ensuring that the driving strategy model can make scientific and reasonable decisions based on multi-dimensional information, and realizing the optimal control of subway train operation.

[0059] Embodiment 4: In the instruction generation module of the driving strategy model, the processing of the strategy input sequence goes through multiple steps such as timeline alignment, key feature extraction, tensor concatenation, and instruction selection to generate a reasonable initial driving instruction sequence. First, regarding the timeline alignment problem of the strategy input sequence, since there may be differences in the collection frequencies or timestamps of the line operation characteristics, passenger flow fluctuation prediction maps, and equipment health assessment matrices, the dynamic time warping (DTW) algorithm is needed to solve the problem of inconsistent sequence lengths. For example, assume that the signal response interval time data is collected at a frequency of once per second, while the operation conflict probability heat map data is updated at a frequency of once every 5 seconds. The time series lengths of the two are different and the time points are not completely aligned. At this time, the DTW algorithm calculates the similarity distance between the two sequences, searches for the optimal alignment path on the timeline, stretches or compresses the short sequence in time so that its time points correspond one by one with those of the long sequence, and generates an instruction correlation vector with timeline alignment. This process ensures the consistency of data from different sources in the time dimension, facilitating the subsequent analysis of the mutual influence of various feature parameters at the same moment.

[0060] When extracting key control node features through the attention mechanism, the scaled dot-product attention mechanism is used. Its core idea is to assign different attention weights to each feature node in the strategy input sequence to highlight the key nodes that have a greater impact on the driving strategy. For example, in the strategy input sequence, there are features such as an extended signal response interval time in a certain period, an increased passenger flow density at the transfer station, and a degradation in the health status of a certain signal device. The attention mechanism automatically calculates the weights of these feature nodes: if the extended signal response interval time may lead to a signal delay risk for subsequent trains, the weight of this feature will be significantly increased; if the increased passenger flow density at the transfer station may affect the train stop time, its weight will also increase accordingly. In this way, an instruction node correlation matrix is generated, where each element represents the attention weight of the corresponding feature node, intuitively reflecting the importance level of each node in driving control. For example, in the transfer station area during the morning rush hour, the node weights related to passenger flow density and signal response delay may be higher than those of other ordinary nodes, indicating that these nodes are the control factors that need to be focused on currently.

[0061] When performing a tensor concatenation operation on the instruction association vector and the instruction node association matrix, it is necessary to first clarify the dimensional structures of the two. The instruction association vector is a one-dimensional sequence after alignment along the time axis, and each element corresponds to the comprehensive eigenvalue at a time point; the instruction node association matrix is a two-dimensional matrix, where the rows represent time points, the columns represent different feature nodes, and the elements are the attention weights of each node. The tensor concatenation operation combines the one-dimensional vector and the two-dimensional matrix into a three-dimensional tensor by expanding and merging in dimensions. For example, the instruction association vector is expanded into a two-dimensional matrix in dimensions (the vector values at each time point are repeatedly arranged into multiple rows), and then concatenated with the instruction node association matrix in the third dimension to form a three-dimensional structure of "time-node-feature". This concatenation method can retain both the global features (the overall operating state reflected by the instruction association vector) and the local key node features (the key control factors highlighted by the instruction node association matrix), providing a richer feature representation for generating the candidate instruction set.

[0062] When selecting the initial driving instruction sequence from the candidate instruction set through the branch and bound algorithm, it is necessary to first set a multi-objective optimization function, such as taking running energy consumption, punctuality rate, and passenger comfort as optimization objectives. The candidate instruction set contains various possible combinations of driving instructions, such as different acceleration / braking strategies, stop time adjustment plans, etc. The branch and bound algorithm decomposes the problem into multiple sub-problems (branching process) and calculates the upper or lower bounds of the objective function for each sub-problem (bounding process), gradually excluding the sub-problems that cannot contain the optimal solution and narrowing the search range. For example, in a certain operating interval, the candidate instruction set includes three options: "maintain the current speed", "accelerate through", and "decelerate to avoid". The algorithm first calculates the estimated energy consumption and the impact on the punctuality rate of each option: "accelerate through" may reduce energy consumption but increase the risk of punctuality, and "decelerate to avoid" may improve safety but increase energy consumption. By comparing the objective function values of each option, the option with better comprehensive performance is selected as part of the initial driving instruction sequence. During the search process, the algorithm will give priority to processing the sub-problems with better objective function values to ensure finding a feasible solution close to the optimal solution within a limited time.

[0063] In practical applications, the processing of the instruction generation module needs to be dynamically adjusted in combination with specific scenarios. For example, when it is detected that the congestion degree of the transfer passage at a certain station has increased significantly, the corresponding passenger flow eigenvalue in the instruction association vector will increase, and the attention mechanism will automatically increase the weight of this feature node, causing the weight of this node in the instruction node association matrix to rise. After tensor concatenation, instructions such as extending the stop time at this station and slowing down the inbound speed may be generated in the candidate instruction set to cooperate with the station staff to divert the passenger flow. When the branch and bound algorithm selects instructions, it will comprehensively consider the impact of extending the stop time on the subsequent train operation time (such as may cause subsequent trains to be late) and the impact of slowing down the inbound speed on energy consumption, and select the optimal instruction combination through multi-objective trade-off.

[0064] In addition, the instruction generation module also needs to handle data noise and uncertainty. For example, a certain characteristic parameter in the equipment health assessment matrix may have an abnormal value due to a temporary sensor failure, resulting in interference information in the policy input sequence. At this time, the attention mechanism will reduce the impact of this abnormal characteristic node through weight allocation to avoid generating incorrect driving instructions. At the same time, when calculating the objective function, the branch and bound algorithm will introduce robustness constraint conditions to ensure that the selected instruction sequence can still maintain stable performance under a certain degree of data fluctuation.

[0065] The processing flow of the entire instruction generation module is based on time-axis alignment, realizes feature weighting through the attention mechanism, fuses multi-dimensional information using tensor splicing, and finally completes instruction optimization through the branch and bound algorithm. Time-axis alignment ensures the spatio-temporal consistency of data, the attention mechanism simulates the focus process in human decision-making, tensor splicing realizes the organic combination of global and local features, and the branch and bound algorithm provides an efficient multi-objective optimization method. These links cooperate with each other to generate an initial driving instruction sequence with strong pertinence and excellent comprehensive performance according to the multi-dimensional data collected in real time, providing a reliable decision-making basis for subsequent driving strategy optimization, enabling the subway train to make intelligent and reasonable control responses in a complex operating environment, and balancing various requirements such as safety, efficiency, and passenger experience.

[0066] Example 5: In the process of constructing a dynamic optimization policy network through the genetic algorithm, key links such as population initialization, fitness evaluation, crossover and mutation operations, and population update need to be experienced, and at the same time, dynamic correction of the policy network is realized by combining real-time feedback. Taking the driving strategy optimization during the morning peak period of a certain subway line as an example, the initial population is constructed based on the train scheduling weight distribution table. This table divides the line into high-priority scheduling areas (such as transfer station intervals) and ordinary scheduling areas according to factors such as passenger flow density and signal equipment load during the morning peak period. Each chromosome corresponds to a set of initial driving instruction sequences. For example, a certain chromosome is encoded as "limit the train speed to 50 km / h in the high-priority area and extend the stop time to 40 seconds; maintain a speed of 60 km / h and a stop time of 25 seconds in the ordinary area", where parameters such as speed and stop time constitute the gene segments of the chromosome.

[0067] When initializing the population, the crossover probability between chromosomes is defined as the balance factor of the operating energy consumption index and the scheduling priority weight. For example, during the morning peak period, the scheduling priority weight is relatively high (such as set to 0.6), and the operating energy consumption weight is relatively low (such as 0.4). The crossover probability can be set to 0.7 to promote the gene exchange of the scheduling strategy in high-priority areas and accelerate the population evolution to adapt to the complex scenarios during the morning peak. The fitness scores of each chromosome are initialized to a benchmark value (such as 100 points), and the fitness of the starting chromosome (i.e., the initial driving instruction sequence) is determined according to a preset initial value (such as the score of 85 points based on historical strategies), providing an initial reference standard for population evolution.

[0068] When selecting parental chromosomes through the elite retention strategy, the individual with the highest fitness score is screened from the current population. For example, if a chromosome during the simulation operation in the morning peak period not only meets the scheduling requirements of high-priority areas (such as no departure delay due to crowded passenger flow) but also maintains a relatively low energy consumption level, with a fitness score of 95 points, higher than the population average score of 88 points, it is selected as a parental chromosome. The parental chromosomes generate a new generation of population through multi-point crossover and random bit mutation operations: multi-point crossover selects multiple gene segment exchange points. For example, crossover points are set at the speed control gene and the stop time gene respectively, and the high-priority area speed gene (50 km / h) of parent 1 is exchanged with the normal area speed gene (65 km / h) of parent 2 to generate a new chromosome combination; random bit mutation randomly modifies a certain gene position in the chromosome, such as mutating the normal area stop time gene of a chromosome from 25 seconds to 28 seconds to increase the genetic diversity of the population and avoid falling into local optimal solutions.

[0069] During the fitness evaluation process, the fitness function consists of the operating energy consumption index and the scheduling priority weight. The operating energy consumption index is determined by calculating the energy consumption model of the train at different speeds and accelerations. For example, when the speed increases from 50 km / h to 60 km / h, the energy consumption per unit distance increases by about 15%; the scheduling priority weight is scored according to whether the scheduling tasks in high-priority areas are completed on time (such as whether the train departs on time during the peak period), and corresponding scores are deducted if not completed on time. Through iterative evolution, the fitness scores of the chromosomes in the population are gradually improved. For example, after 5 generations of evolution, the fitness score of the optimal chromosome increases from the initial 85 points to 98 points, and the corresponding driving strategy reduces the energy consumption by about 10% compared with the initial strategy while ensuring the scheduling efficiency in high-priority areas (only for logical description, without actual data).

[0070] The real-time correction mechanism is implemented through an abnormal decision-making detection model. For example, when a train suddenly encounters a signal equipment failure during operation, resulting in a delay in the actual execution of the braking instruction compared to the planned time in the initial driving instruction sequence, the execution status log of the train control system will record this abnormal event, including the acceleration instruction change rate (such as the actual acceleration suddenly drops from 0.3 m / s² to 0.1 m / s²) and the braking response time series (such as the planned braking time is 10:05:00, and the actual execution time is 10:05:03). The abnormal decision-making detection model constructs a one-class support vector machine benchmark model based on historical normal operation samples (such as the instruction execution data without failures in the past week), and calculates the deviation index between the real-time data and the benchmark model. If the cumulative abnormal value of the deviation index exceeds the dynamic threshold (such as the braking response is delayed by more than 2 seconds for 3 consecutive times), a set of policy correction instructions will be generated, such as "reduce the train speed by 10 km / h in the next 3 sections and increase the braking distance margin".

[0071] The correction instruction set adjusts the strategy by updating the chromosome coding parameters of the dynamic optimization strategy network. For example, encode the speed reduction parameter in the above correction instruction into a new gene segment, replace the corresponding gene position in the original chromosome, and form a new chromosome population. This process enables the genetic algorithm to respond to sudden abnormal situations in real time and adjust the optimization direction. For example, during the period before the signal equipment failure is repaired, the chromosomes in the population generally carry the genes of "reduce speed and increase braking margin" until the abnormal event is eliminated, and the benchmark model re-learns the normal operation data and gradually eliminates this correction gene.

[0072] In practical applications, the parameter settings of the crossover and mutation operators need to be dynamically adjusted according to the operation characteristics of different periods. For example, during off-peak hours, the scheduling priority weight is reduced (such as 0.4), the operation energy consumption weight is increased (such as 0.6), the crossover probability can be adjusted to 0.5 to balance energy consumption optimization and regular scheduling requirements; the mutation probability is appropriately increased (such as from 0.01 to 0.03) to explore more low-energy consumption driving strategy combinations. The application ratio of the elite retention strategy can also be adjusted according to the population diversity. For example, when there are multiple high-fitness chromosomes in the population, retain the top 20% of the elite individuals; if the population diversity is insufficient, expand the elite retention range to 30% to avoid the loss of high-quality genes.

[0073] Throughout the implementation process, the global search ability of the genetic algorithm is combined with the local correction ability of the real-time feedback mechanism to achieve the dynamic optimization of the driving strategy. Population initialization and fitness evaluation ensure the goal orientation of strategy optimization. Crossover and mutation operations maintain population diversity. The elite retention strategy accelerates the evolution process. The abnormal decision detection and correction mechanism endows the system with the adaptive ability to handle emergencies. Taking the morning rush hour as an example, in complex scenarios such as sudden changes in passenger flow and temporary equipment failures, this mechanism can generate the optimal driving strategy that can not only meet the high-priority scheduling requirements but also take into account energy consumption efficiency through continuous iterative evolution and real-time correction, ensuring the safe and efficient operation of subway trains in a dynamically changing operating environment.

[0074] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0075] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing the DAS driving strategy driven by real-time data of a subway line, characterized in that, Including: Obtain a multi-dimensional operation data set; the multi-dimensional operation data includes a train real-time position sequence, station passenger flow data, and trackside signal equipment status; the train real-time position sequence includes a speed change gradient and a braking distance parameter, and the station passenger flow data includes an inbound passenger flow density and a transfer channel congestion degree; based on the train real-time position sequence, extract line operation features through a convolutional spatio-temporal encoder, and the line operation features include an interval passing delay coefficient, a signal response interval time, and an occupancy status label of the track; According to the station passenger flow data, generate a passenger flow fluctuation prediction map through a time series decomposition algorithm, and the map includes a peak period distribution interval and a heat map of operation conflict probability; Perform multi-dimensional feature fusion on the trackside signal equipment status to generate an equipment health assessment matrix; input the line operation features, the passenger flow fluctuation prediction map, and the equipment health assessment matrix into a driving strategy model to generate an initial driving instruction sequence; based on the initial driving instruction sequence, divide the operation control priority through a regional clustering algorithm, and output a train scheduling weight distribution table; fuse the train scheduling weight distribution table with the initial driving instruction sequence, and construct a dynamic optimization strategy network through a genetic algorithm to generate a final driving strategy plan.

2. The DAS driving strategy optimization method driven by real-time data of a subway line according to claim 1, characterized in that, The extraction of line operation features through the convolutional spatio-temporal encoder includes: intercepting the train real-time position sequence through a sliding window to generate spatio-temporal slice data blocks; dividing the spatio-temporal slice data blocks into multiple operation intervals based on regional clustering rules, and calculating the standard deviation of signal response delays within each interval; using a residual connection algorithm to extract an occupancy status label of the track, and combining the interval passing delay coefficient to generate a signal system response relationship; encoding the signal response interval time, the occupancy status label of the track, and the signal system response relationship into line operation features.

3. The DAS driving strategy optimization method driven by real-time data of a subway line according to claim 1, wherein The generation of a passenger flow fluctuation prediction map through the time series decomposition algorithm includes: separating the trend component of the station passenger flow data to extract periodic fluctuation features; predicting the future passenger flow distribution based on an autoregressive model, and calculating the confidence level of the peak period distribution interval; generating a risk probability model according to the historical conflict event database, and outputting a heat map of operation conflict probability in combination with the confidence level; mapping the peak period distribution interval and the heat map of operation conflict probability into a passenger flow fluctuation prediction map.

4. A method for optimizing the DAS driving strategy driven by real-time data of a subway line according to claim 1, characterized in that, The driving strategy model includes a data fusion module and an instruction generation module. The data fusion module includes: performing standardization processing on the signal response interval time in the line operation features to obtain a first fusion tensor; performing morphological filtering on the heat map of operation conflict probability in the passenger flow fluctuation prediction map to generate a second fusion tensor; performing eigenvalue decomposition on the equipment health assessment matrix to extract the equipment state degradation gradient to obtain a third fusion tensor; combining the first fusion tensor, the second fusion tensor, and the third fusion tensor into a strategy input sequence through a bidirectional long short-term memory network.

5. A method for optimizing the DAS driving strategy driven by real-time data of a subway line according to claim 4, characterized in that The instruction generation module includes: aligning the policy input sequence along the time axis to generate an instruction correlation vector; extracting key control node features through an attention mechanism to generate an instruction node correlation matrix; performing a tensor concatenation operation on the instruction correlation vector and the instruction node correlation matrix to generate a candidate instruction set; and selecting an initial driving instruction sequence from the candidate instruction set through a branch and bound algorithm.

6. The DAS driving strategy optimization method driven by real-time data of a subway line according to claim 1, wherein The construction of the dynamic optimization policy network through a genetic algorithm includes: initializing the network population according to the train scheduling weight distribution table and generating a fitness evaluation function based on priorities; using the initial driving instruction sequence as chromosome encoding, where the fitness evaluation function consists of the running energy consumption index and the scheduling priority weight; updating the fitness scores of each chromosome through a crossover and mutation operator to adjust the population evolution direction; and generating an optimal driving policy solution that meets multi-objective constraints according to the optimized fitness scores.

7. A method for optimizing the DAS driving strategy driven by real-time data of a subway line according to claim 2, characterized in that, The parameter optimization method for the regional clustering rule includes: calculating the initial clustering radius and the minimum signal delay threshold according to the historical operation data distribution; traversing parameter combinations through a grid verification algorithm and selecting the parameters with the highest matching degree between the partitioning result and the actual operation log; dynamically adjusting the clustering radius and the minimum signal delay threshold according to the matching degree to optimize the partitioning accuracy of the signal system response relationship.

8. The optimized DAS driving strategy method driven by real-time data of a subway line according to claim 3, characterized in that, The construction method of the risk probability model includes: collecting historical operation conflict data of multiple lines, extracting spatio-temporal correlation features and equipment failure influencing factors; fitting the spatio-temporal correlation features with a Gaussian mixture model to generate a basic risk probability distribution table; and performing Bayesian correction on the basic risk probability distribution table according to the equipment failure influencing factors to output the risk probability model.

9. The real-time data-driven DAS driving strategy optimization method for subway lines according to claim 6, characterized in that, The parameter setting method for the crossover and mutation operator includes: defining the crossover probability between chromosomes as a balance factor of the running energy consumption index and the scheduling priority weight; initializing the fitness scores of each chromosome as a reference value, and the fitness of the starting chromosome as a preset initial value; selecting the chromosome with the highest fitness score in the current population as the parent through an elitist retention strategy; and generating a new generation of population through multi-point crossover and random bit mutation operations based on the parent chromosome.

10. A method for optimizing the DAS driving strategy driven by real-time data of a subway line according to claim 9, characterized in that, The method further includes: collecting the execution status log of the train control system in real time, generating a policy correction instruction set through an abnormal decision detection model, and updating the dynamic optimization policy network; the construction method of the abnormal decision detection model includes: collecting normal operation samples in the historical execution status log, extracting the acceleration instruction change rate and the braking response time series; constructing a normal operation benchmark model based on a one-class support vector machine, calculating the deviation index between the real-time data and the benchmark model; statistically calculating the cumulative abnormal value of the deviation index through a time decay function, and generating a policy correction instruction set when it exceeds the dynamic threshold; and updating the chromosome encoding parameters of the dynamic optimization policy network in combination with the correction instruction set.

Images