Rail transit multi-system information interaction control method based on train-ground network cooperation

Abstract

The invention provides a rail transit multi-system information interaction control method based on vehicle-ground network cooperation, which relates to the technical field of rail transit control and comprises the following steps: acquiring operation data of vehicle-mounted equipment and ground infrastructure, extracting semantic feature vectors and realizing heterogeneous data standardization, executing time-space alignment of vehicle-ground data to establish a synchronous state set, and establishing a synchronous state set; and constructing a directed graph structure to identify a key state subset, and then constructing a multi-objective optimization function to solve a Pareto optimal solution set to obtain an optimal cooperative control instruction sequence. According to the invention, information interaction and cooperative control among multiple systems are realized, and the rail transit operation safety and the energy utilization efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit control technology, and in particular to a rail transit multi-system information interaction control method based on train-ground network cooperation. BACKGROUND

[0002] As an important part of modern urban transportation system, the safe, efficient and reliable operation of urban rail transit is of great significance to ensuring the smoothness of urban traffic. With the continuous expansion of the scale of urban rail transit and the continuous development of technology, train-ground network cooperative control technology has become a key means to improve the efficiency and safety of rail transit operation. Through information interaction between on-board devices and ground infrastructure, the train-ground network cooperative system realizes coordinated control during train operation, improving the safety and efficiency of train operation.

[0003] Traditional rail transit control systems mainly rely on independent operation of on-board devices and ground facilities for information processing and decision-making, with limited data exchange between trains and grounds, making it difficult to achieve comprehensive cooperative control. With the development of Internet of Things, big data and artificial intelligence technology, train-ground network cooperative control technology has gradually become the development trend of rail transit control systems, realizing precise control and optimized scheduling of train operation through efficient information exchange and intelligent decision-making. SUMMARY

[0004] The embodiment of the present application provides a rail transit multi-system information interaction control method based on train-ground network cooperation, which can solve the problems in the prior art.

[0005] In a first aspect of the embodiment of the present application, a rail transit multi-system information interaction control method based on train-ground network cooperation is provided, comprising:

[0006] Obtaining train on-board device operation data and ground infrastructure operation data;

[0007] Extracting semantic feature vectors of the on-board device operation data and infrastructure operation data, learning domain mapping transformation matrix by minimizing source domain and target domain semantic space distribution distance, projecting heterogeneous data fields to a unified semantic space to obtain standardized operation data;

[0008] According to the timestamp and spatial location information in the standardized operation data, calculating the train-ground data space-time offset and performing space-time alignment to establish a train-ground synchronous state set;

[0009] The train-ground synchronous state set is constructed as a directed graph structure, the directed edges between nodes are established based on time series mutual information, and the topological sorting is performed to obtain a causal sequence, the change rate of the state parameter to the control output is calculated, and a key state subset is selected;

[0010] Based on the key state subset and the safety constraint rule, a multi-objective optimization function of minimum running adjustment time, minimum energy consumption and maximum safety margin is constructed, a particle swarm algorithm is used to solve the multi-objective optimization function, a Pareto optimal solution set is obtained, and an optimal collaborative control instruction sequence is calculated combined with an expert knowledge base rule;

[0011] The optimal collaborative control instruction sequence is issued to the vehicle-ground execution unit, and feedback states are collected.

[0012] The semantic feature vectors of the vehicle-mounted equipment operation data and the infrastructure operation data are extracted, a domain-to-domain mapping transformation matrix is learned by minimizing the semantic space distribution distance between the source domain and the target domain, and the heterogeneous data fields are projected into a unified semantic space to obtain standardized operation data, including:

[0013] For each data field in the vehicle-mounted equipment operation data, a field identifier, a data type descriptor and a service context descriptor are extracted, input into a pre-trained semantic encoder to obtain a source domain semantic feature vector, and the same extraction and encoding operations are performed on each data field in the infrastructure operation data to obtain a target domain semantic feature vector;

[0014] According to the covariance matrix of the source domain semantic feature vector and the target domain semantic feature vector, the Frobenius norm distance between the covariance matrices is calculated, the domain-to-domain mapping transformation matrix is initialized, and the transformed source domain feature vector is obtained by multiplying the source domain semantic feature vector by the domain-to-domain mapping transformation matrix;

[0015] The maximum mean difference between the transformed source domain feature vector and the target domain semantic feature vector is calculated, a loss function is constructed combined with the Frobenius norm distance, and the gradient descent method is used to iteratively update the domain-to-domain mapping transformation matrix until the loss function converges to a preset convergence condition;

[0016] The converged domain-to-domain mapping transformation matrix is used to perform linear transformation on the source domain semantic feature vector to obtain a feature representation in a unified semantic space, and the standardized format and standardized values of the data fields are reconstructed according to the feature representation in the unified semantic space to obtain standardized operation data.

[0017] According to the timestamp and spatial position information in the standardized operation data, a vehicle-ground data space-time offset is calculated and space-time alignment is performed to establish a vehicle-ground synchronous state set, including:

[0018] The timestamp information of the vehicle-mounted equipment and the infrastructure is extracted from the standardized operation data, the vehicle-mounted equipment timestamp is converted into a train local time reference, the infrastructure timestamp is converted into a ground system time reference, and a time offset vector is calculated;

[0019] According to the standardized operation data, track coordinates of a current mileage position of the train and an installation position of the infrastructure are extracted, a spatial distance between the train and the infrastructure at the same time is calculated, a spatial offset vector is obtained, a time offset vector and the spatial offset vector are combined to construct a space-time offset matrix;

[0020] According to the space-time offset matrix, timestamps of the on-board device data are corrected to achieve time alignment, and a spatial reference system of the infrastructure data is converted to achieve spatial registration, train state data after time alignment and device state data after spatial registration are extracted; a train-ground data correlation rule is established using the space-time offset matrix, spatially related ground device states are found based on the train position, and state parameters with the same time and space are combined into a train-ground synchronous state set.

[0021] The train-ground synchronous state set is constructed as a directed graph structure, a directed edge between nodes is established based on time series mutual information, and a topological sorting is performed to obtain a causal sequence, a change rate of a state parameter to a control output is calculated, and a key state subset is selected, including:

[0022] For each state parameter in the train-ground synchronous state set, time series arrangement is performed, a parameter value sequence is extracted within a sliding time window, and a probability density function of the state parameter is calculated through kernel density estimation, time series mutual information is calculated based on the integral of the logarithmic ratio of joint probability density and marginal probability density;

[0023] The time series mutual information under different time delays is calculated by traversing all state parameter pairs, the maximum value of the time series mutual information is taken as the causal correlation strength, a directed edge is established between state parameters when the causal correlation strength exceeds a time series threshold, a directed graph structure is formed; the number of incoming connections of each state parameter is counted, state parameters with zero incoming connection number are sequentially added to the causal sequence, the number of incoming connections of associated parameters is updated, and the process is repeated until the causal sorting of all state parameters is completed, to obtain a causal sequence;

[0024] According to the causal sequence, all causal path sets of each state parameter to the control output are constructed, and a mapping weight matrix of the state to the control is calculated, the change rate of the control output is calculated by applying a disturbance to the state parameter, and the state parameter with a change rate greater than a change threshold is taken as a key state subset.

[0025] According to the causal sequence, all causal path sets of each state parameter to the control output are constructed, and a mapping weight matrix of the state to the control is calculated, the change rate of the control output is calculated by applying a disturbance to the state parameter, and the state parameter with a change rate greater than a change threshold is taken as a key state subset, including:

[0026] Based on the sorting position of the state parameters in the causal sequence, breadth-first traversal is performed from the source node of the directed graph, the time series mutual information value of all node sequences of each state parameter to the control output and the connection edges thereof are recorded, the node sequence and edge information are stored as a causal path, all nodes reachable to the control output are traversed to form a complete causal path set;

[0027] The time series mutual information values on each causal path are multiplied to obtain a path propagation coefficient, all path propagation coefficients with the state parameter as the starting node are summed to obtain a direct influence factor, the product sum of the path propagation coefficient and the incoming edge mutual information with the state parameter as the intermediate node is calculated as an indirect influence factor, and the mapping weight matrix of the state to the control is calculated in combination with the direct influence factor;

[0028] The standard deviation of the historical sequence of each state parameter is selected as a disturbance reference, positive and negative disturbances are superimposed on the current value, the change of the control output caused by the disturbance is calculated through the mapping weight, the difference between the positive and negative disturbance influence values is extracted to obtain the bidirectional change rate of the state parameter;

[0029] The interquartile range is calculated based on the bidirectional change rate of all state parameters, the third quartile and the multiple of the interquartile range are set as a screening threshold, the state parameters whose change rates exceed the screening threshold are extracted, and a key state subset is obtained by rearranging the causal sequence.

[0030] Based on the key state subset and the safety constraint rule, a multi-objective optimization function of minimum running adjustment time, energy consumption and maximum safety margin is constructed, a particle swarm algorithm is used to solve the multi-objective optimization function to obtain a Pareto optimal solution set, and an optimal cooperative control instruction sequence is calculated in combination with an expert knowledge base rule, including:

[0031] The deviation amount of the current value and the target value of each state parameter in the key state subset is calculated, the running adjustment time is calculated in combination with the control instruction adjustment rate, the energy consumption coefficient of unit adjustment is obtained, the total energy consumption value is calculated, the minimum interval between the state parameter target value and the safety boundary is calculated as the safety margin, and the running adjustment time, the total energy consumption value and the safety margin are constructed as a multi-objective optimization function;

[0032] An initial particle swarm is generated within the constraint range of the control instruction parameter, the particle position and velocity are iteratively updated, the non-dominated sorting of the multi-objective optimization function value is performed and the crowded distance is calculated, and the final non-dominated solution set is extracted as the Pareto optimal solution;

[0033] The Pareto optimal solutions are substituted into the mapping weight matrix to calculate state responses, the response sensitivity is calculated through disturbance analysis, and the current scene strategy is extracted based on an expert rule base, the node attribute distribution of a key state subset is analyzed, the weight of each target in the multi-objective optimization function is adjusted, the Pareto optimal solution with the lowest response sensitivity and the maximum multi-objective optimization function value is selected as the optimal cooperative control instruction sequence.

[0034] An initial particle swarm is generated within the constraint range of the control instruction parameters, the particle position and velocity are iteratively updated, the non-dominated sorting of the multi-objective optimization function value is performed and the crowded distance is calculated, and the final non-dominated solution set is extracted as the Pareto optimal solution, including:

[0035] The particle position and velocity vector are randomly initialized within the constraint range of the control instruction parameters to form an initial particle swarm, the running adjustment time, total energy consumption value and safety margin target function value of each particle are calculated, and the non-dominated sorting is performed to divide the particles into multiple non-dominated levels;

[0036] For the particles in the first non-dominated level, the interval distance of each particle from adjacent particles in three target dimensions is calculated, and the comprehensive crowded distance is obtained by normalizing and summing, the particles with a comprehensive crowded distance exceeding the density division threshold are marked as sparse areas, and the particles with a comprehensive crowded distance below the density division threshold are marked as dense areas;

[0037] Based on the non-dominated level and area marking of the particles, the inertia weight of the particles in the dense area is dynamically adjusted to enhance global search, and the inertia weight of the particles in the sparse area is reduced to enhance local search, and the velocity update amount is calculated based on the historical optimal position and global optimal position of each particle;

[0038] The velocity update amount and the current velocity are fused and superimposed on the position vector to obtain the new position of the particle, and the non-dominated sorting iteration is repeated after boundary mapping correction until the convergence number limit is reached, and the particles in the final first non-dominated level are extracted as the Pareto optimal solution set.

[0039] The second aspect of the embodiment of the application provides a rail transit multi-system information interaction control system based on train-ground network cooperation, comprising:

[0040] The first unit is used for acquiring train-mounted equipment operation data and ground infrastructure operation data;

[0041] The second unit is used for extracting semantic feature vectors of the train-mounted equipment operation data and the infrastructure operation data, learning a domain mapping transformation matrix by minimizing the semantic space distribution distance between the source domain and the target domain, projecting the heterogeneous data fields to a unified semantic space to obtain standardized operation data;

[0042] The third unit is configured to calculate a time-space offset of train-ground data and perform time-space alignment according to the time stamp and the space position information in the standardized operation data, and establish a train-ground synchronization state set;

[0043] The fourth unit is configured to construct the train-ground synchronization state set into a directed graph structure, establish a directed edge between nodes based on time-series mutual information and perform topological sorting to obtain a causal sequence, calculate a change rate of a state parameter to a control output, and screen a key state subset;

[0044] The fifth unit is configured to construct a multi-objective optimization function of operation adjustment time, minimum energy consumption and maximum safety margin based on the key state subset and a safety constraint rule, solve the multi-objective optimization function by using a particle swarm algorithm, obtain a Pareto optimal solution set, and calculate an optimal cooperative control instruction sequence in combination with an expert knowledge base rule;

[0045] The sixth unit is configured to issue the optimal cooperative control instruction sequence to a train-ground execution unit and collect feedback states.

[0046] The third aspect of the embodiment of the application,

[0047] An electronic device is provided, comprising:

[0048] a processor;

[0049] a memory for storing processor-executable instructions;

[0050] The processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0051] The fourth aspect of the embodiment of the application,

[0052] A computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0053] The beneficial effects of the present application are as follows:

[0054] By extracting semantic feature vectors of operation data of a vehicle-mounted device and infrastructure and minimizing the semantic space distribution distance between a source domain and a target domain to learn a domain mapping transformation matrix, standardized expression of heterogeneous data fields in a unified semantic space is achieved, and the technical problem of non-uniform data formats and large semantic differences of different subsystems in a rail transit system is solved.

[0055] The train-ground synchronization state set is constructed into a directed graph structure, a directed edge between nodes is established based on time-series mutual information and topological sorting is performed to scientifically identify the causal relationship between system parameters, a key state subset is screened through change rate calculation, the calculation complexity is significantly reduced, and the system response speed is improved.

[0056] A multi-objective optimization function including running adjustment time, minimum energy consumption and maximum safety margin is constructed, a particle swarm algorithm is used to solve to obtain a Pareto optimal solution set, an optimal collaborative control instruction sequence is calculated in combination with an expert knowledge base rule, intelligent collaborative control among rail transit multi-systems is realized, and system running efficiency and energy consumption are optimized under the premise of ensuring safety. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 A flowchart of a rail transit multi-system information interaction control method based on vehicle-ground network collaboration of an embodiment of the present application is shown in the figure.

[0058] Figure 2 A flowchart of a standardization running data calculation method of an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0060] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0061] Figure 1 A flowchart of a rail transit multi-system information interaction control method based on vehicle-ground network collaboration of an embodiment of the present application is shown in the figure. Figure 1 As shown in the figure, the method comprises:

[0062] Obtaining train on-board equipment running data and ground infrastructure running data;

[0063] Extracting semantic feature vectors of the on-board equipment running data and the infrastructure running data, learning a domain-to-domain mapping transformation matrix by minimizing the semantic space distribution distance between the source domain and the target domain, projecting the heterogeneous data fields to a unified semantic space to obtain standardized running data;

[0064] According to the time stamp and spatial position information in the standardized running data, calculating a train-ground data space-time offset and performing space-time alignment to establish a train-ground synchronization state set;

[0065] The vehicle-ground synchronization state set is constructed as a directed graph structure, a directed edge between nodes is established based on time series mutual information, and a topological sorting is performed to obtain a causal sequence, a rate of change of a state parameter to a control output is calculated, and a key state subset is screened;

[0066] Based on the key state subset and safety constraint rules, a multi-objective optimization function of minimum running adjustment time, energy consumption and maximum safety margin is constructed, a particle swarm algorithm is used to solve the multi-objective optimization function, a Pareto optimal solution set is obtained, and an optimal cooperative control instruction sequence is calculated in combination with an expert knowledge base rule;

[0067] The optimal cooperative control instruction sequence is issued to a vehicle-ground execution unit, and feedback states are collected.

[0068] Figure 2 The flowchart of the calculation method of the standardized operation data of the embodiment of the application is shown. In an optional implementation, semantic feature vectors of the vehicle-mounted device operation data and the infrastructure operation data are extracted, a domain-to-domain mapping transformation matrix is learned by minimizing the semantic space distribution distance between the source domain and the target domain, and heterogeneous data fields are projected into a unified semantic space to obtain standardized operation data, including:

[0069] For each data field in the vehicle-mounted device operation data, a field identifier, a data type descriptor and a business context descriptor are extracted, a pre-trained semantic encoder is input, a source domain semantic feature vector is obtained, and the same extraction and encoding operations are performed on each data field in the infrastructure operation data to obtain a target domain semantic feature vector;

[0070] According to the covariance matrix of the source domain semantic feature vector and the target domain semantic feature vector, the Frobenius norm distance between the covariance matrices is calculated, the domain-to-domain mapping transformation matrix is initialized, and the transformed source domain feature vector is obtained by multiplying the source domain semantic feature vector by the domain-to-domain mapping transformation matrix;

[0071] The maximum mean difference between the transformed source domain feature vector and the target domain semantic feature vector is calculated, a loss function is constructed in combination with the Frobenius norm distance, and the domain-to-domain mapping transformation matrix is iteratively updated using the gradient descent method until the loss function converges to a preset convergence condition;

[0072] The source domain semantic feature vector is linearly transformed using the converged domain-to-domain mapping transformation matrix to obtain a feature representation in the unified semantic space, and the standardized format and the standardized numerical value of the data field are reconstructed according to the feature representation in the unified semantic space to obtain the standardized operation data.

[0073] In the specific embodiment, semantic information is extracted from each data field in the vehicle-mounted device operation data, including vehicle speed, location information, door status, etc., which all have specific identifiers, type descriptions, and business contexts. For example, for the speed data field, its identifier "SPEED", data type "float", and business context "vehicle real-time running speed" are extracted, and after extraction, the three descriptors are concatenated into a text sequence "SPEED float vehicle real-time running speed", which is input into a pre-trained BERT semantic encoder. The encoder uses a 12-layer Transformer structure, and after the input text sequence is processed by the multi-head attention mechanism, a 768-dimensional vector representation is extracted from the [CLS] position of the last layer as the source domain semantic feature vector.

[0074] The same operation is performed on the infrastructure operation data, including signal system status, power supply parameters, etc. Taking the signal light status as an example, the identifier "SIGNAL_STATUS", data type "enum", and business context "track signal indication state" are extracted, and concatenated into the text "SIGNAL_STATUS enum track signal indication state". Similarly, input the BERT encoder to obtain the target domain semantic feature vector.

[0075] Calculate the inter-domain distribution distance, group all source domain semantic feature vectors into matrix X, and all target domain semantic feature vectors into matrix Y. Calculate their covariance matrices C x and C y The covariance matrix reflects the correlation structure between the semantic features of each field. The Frobenius norm distance D = ||C x -C y ||F between the two covariance matrices is calculated to quantify the distribution difference between the two domains. Initialize the inter-domain mapping transformation matrix W as the identity matrix and multiply it by the source domain feature vector X to obtain the transformed source domain feature vector X' = WX.

[0076] Calculate the maximum mean difference (MMD) between the transformed source domain feature vector X' and the target domain feature vector Y. MMD measures the distribution difference by calculating the mean distance between the two sets of vectors in the reproducing kernel Hilbert space. Use the Gaussian kernel function K(x, y) = exp(-||x-y|| 2 / 2σ 2 ), where σ is the kernel width parameter, set to the median distance between the source domain and target domain feature vectors. MMD can be represented as the average distance between the two domain features in the kernel space.

[0077] Construct the loss function L = MMD(X', Y) + λ||C x '-C y||F, where λ is a balance parameter set to 0.5, the gradient descent is performed using the Adam optimizer with a learning rate of 0.001, the mapping matrix W is updated iteratively, and when the loss changes less than a threshold of 0.0001 for five consecutive iterations, it is considered that the preset convergence condition is reached, and the iteration is stopped.

[0078] Using the converged mapping matrix W, linear transformation is performed on the source domain semantic feature vectors to obtain feature representations in the unified semantic space Z = WX. Based on these unified representations, the data fields are standardized and reconstructed. Specifically, for each transformed feature vector z, the cosine similarity with all target domain semantic feature vectors is calculated, and the target domain field format with the highest similarity is selected as the standardized format.

[0079] For numerical standardization, a value mapping function is constructed based on the field correspondence. For example, for the speed value in the vehicle-mounted device, if the source domain unit is "km / h" and the standard format is "m / s", unit conversion is performed by dividing the value by 3.6. For enumeration types such as door status, a mapping table is established to map the source domain "0 / 1" to the standard format "CLOSED / OPEN".

[0080] Through the above processing, the representation of vehicle-mounted device operation data and infrastructure operation data in the unified semantic space is realized, and standardized operation data is generated. These standardized data maintain the information integrity of the original data, while achieving format unification, facilitating subsequent comprehensive analysis and application.

[0081] In practical applications, for a train operation monitoring system of a subway line, the "vehicleSpeed" field (unit: km / h, float type) recorded by the vehicle-mounted ATP device and the "trainSpeed" field (unit: m / s, float type) recorded by the ground ATS system are mapped to the unified semantic space by this method, identified as fields expressing the same semantics, and standardized as the "vehicleSpeed" field (unit: m / s, float type). In this way, the problem of inconsistent data expression between different systems is solved, providing a unified data basis for rail transit operation analysis.

[0082] In an optional implementation, according to the time stamp and spatial position information in the standardized operation data, the vehicle-ground data space-time offset is calculated and space-time alignment is performed to establish a vehicle-ground synchronous state set, including:

[0083] From the standardized operation data, the time stamp information of the vehicle-mounted device and the infrastructure is extracted respectively, the vehicle-mounted device time stamp is converted to the train local time reference, the infrastructure time stamp is converted to the ground system time reference, and the time offset vector is calculated;

[0084] According to the standardized operation data, track coordinates of the current mileage position of the train and the installation position of the infrastructure are extracted, the spatial distance between the train and the infrastructure at the same time is calculated, a spatial offset vector is obtained, and a time-space offset matrix is constructed by combining the time offset vector and the spatial offset vector;

[0085] According to the time-space offset matrix, the timestamps of the on-board equipment data are corrected to achieve time alignment, and the spatial reference system of the infrastructure data is converted to achieve spatial registration. Train status data after time alignment and equipment status data after spatial registration are extracted. The time-space offset matrix is used to establish a train-ground data correlation rule. Based on the train position, the spatially related ground equipment status is found. Time-synchronous and spatially related status parameters are combined into a train-ground synchronous status set.

[0086] In actual application, the timestamps of the on-board equipment and the infrastructure need to be extracted from the standardized operation data. The timestamp of the on-board equipment is usually based on the on-board clock system, while the timestamp of the infrastructure is based on the ground system clock. Due to differences in hardware and synchronization mechanisms, there is a deviation in the time reference between the two clock systems. Therefore, the on-board equipment timestamp needs to be converted to the train local time reference, and the infrastructure timestamp needs to be converted to the ground system time reference.

[0087] During the conversion process, a time reference mapping function is used for processing. For the on-board equipment, the deviation parameter between the equipment local clock and the standard time source is used for correction. For example, if the on-board system clock is five seconds slower than the standard time source, a five-second deviation value needs to be added during the conversion. Similarly, the infrastructure timestamp is also converted in a similar manner to ensure that the time references of the two systems are comparable. The two converted time sequences are compared to calculate the time offset vector, which reflects the time difference distribution between the on-board system and the ground system.

[0088] According to the standardized operation data, track coordinates of the current mileage position of the train and the installation position of the infrastructure are extracted. The train position is usually obtained by the on-board positioning system, such as the track odometer, GNSS positioning, or multi-sensor fusion positioning system. The position of the infrastructure is a fixed installation coordinate point. The train position and the infrastructure position at the same time are compared to calculate the spatial distance between them, thereby obtaining the spatial offset vector.

[0089] In an actual application scenario, it is assumed that the train is passing through a certain section, and the on-board data shows that the current train mileage position is K256+500. The signal machine status data recorded at the same time comes from the signal machine installed at K256+680. At this time, the spatial offset is 180 meters, indicating the actual distance between the train and the signal machine.

[0090] The time offset vector and the space offset vector obtained above are combined to construct a space-time offset matrix, which contains the correspondence between the time dimension and the space dimension, and each element represents the space-time relationship between the train and the specific location infrastructure at a specific time. The matrix is constructed using a pairing mapping method, which matches the time synchronization point with the space reference point to form a complete space-time mapping relationship table.

[0091] According to the constructed space-time offset matrix, the timestamps of the on-board device data are corrected to achieve time alignment. The correction process uses a time offset compensation algorithm to adjust the original timestamps based on the time offset values in the offset matrix, so that the on-board device data and the ground system remain consistent in the time dimension. At the same time, the spatial reference system of the infrastructure data is converted to achieve spatial registration. This process ensures the comparability of spatial positions by using a coordinate transformation function to unify spatial data in different reference systems into the same coordinate system.

[0092] After alignment, the time-aligned train state data and the spatially registered device state data are extracted. The train state data includes parameters such as speed, acceleration, and running mode, while the device state data includes information such as switch position, signal display state, and track circuit occupancy state. These data, after space-time alignment processing, can accurately reflect the overall system state under the same space-time conditions.

[0093] The space-time offset matrix is used to establish a train-ground data association rule, which defines how to find the spatially related ground device state based on the train position. The association rule uses a spatial nearest neighbor search algorithm to set a reasonable spatial window range (such as 2000 meters ahead and 500 meters behind) centered on the train's current position, and filters out all infrastructure within this range and obtains their real-time state data.

[0094] For example, when the train runs to a certain section, according to the space-time association rule, it can automatically identify the switch state, signal display, and track occupancy within one kilometer ahead, providing necessary environmental perception information for train operation control. The same time and spatially related state parameters are combined into a train-ground synchronous state set, and each record in this set contains a timestamp, train state parameters (position, speed, etc.), and corresponding spatially related ground device state parameters, forming a complete train-ground collaborative state snapshot.

[0095] Through the above space-time alignment method of train-ground data, the space-time reference difference problem between different data sources is solved, providing a reliable data foundation for train-ground collaborative analysis and decision-making, effectively supporting railway intelligent operation control and safety monitoring applications.

[0096] In an alternative embodiment, the vehicle-ground synchronous state set is constructed as a directed graph structure, the directed edges between nodes are established based on the time series mutual information and topological sorting is performed to obtain the causal sequence, the rate of change of the state parameters to the control output is calculated, and the key state subset is screened, including:

[0097] For each state parameter in the vehicle-ground synchronous state set, time series arrangement is performed, the parameter value sequence is extracted within a sliding time window, and the probability density function of the state parameter is calculated through kernel density estimation. The time series mutual information is calculated based on the integral of the logarithmic ratio of joint probability density and marginal probability density.

[0098] The time series mutual information under different time delays is calculated by traversing all state parameter pairs, the maximum value of the time series mutual information is taken as the causal correlation strength, and a directed edge is established between the state parameters when it exceeds the time series threshold, forming a directed graph structure. The number of incoming connections of each state parameter is counted, and the state parameters with zero incoming connection number are sequentially added to the causal sequence, and the incoming connection number of the associated parameters is updated. Repeat the process until the causal sorting of all state parameters is completed to obtain the causal sequence.

[0099] According to the causal sequence, all causal path sets of each state parameter to the control output are constructed, and the mapping weight matrix of the state to the control is calculated. The rate of change of the control output is calculated by applying a disturbance to the state parameter, and the state parameter with a rate of change greater than the change threshold is taken as the key state subset.

[0100] In this specific embodiment, the vehicle-ground synchronous state set in the vehicle-mounted control system is collected, which includes vehicle system state parameters and ground control system state parameters. These state parameters include but are not limited to vehicle speed, position, acceleration, steering angle, battery capacity, ground control instructions, communication delay, etc. After collection, time series arrangement is performed on these state parameters. By setting an appropriate sampling frequency (such as ten times per second), the changes of state parameters within a period of time (such as thirty minutes) are recorded. Based on the time series data, a sliding time window (such as ten seconds) is set, and the value sequence of each state parameter within the window is extracted from the time series data.

[0101] For each state parameter value sequence within each sliding window, the kernel density estimation method is used to calculate the probability density function, specifically, Gaussian kernel function is used to smooth each data point, and the bandwidth parameter is automatically determined by Silverman rule, so as to obtain the marginal probability density function of each state parameter, for any two state parameters, the joint probability density function is calculated, based on the obtained probability density function, the time series mutual information between the state parameters is calculated, the time series mutual information is calculated by integrating the logarithmic ratio of joint probability and marginal probability, for example, for state parameters X and Y, the time series mutual information under time delay τ can be expressed as the integral of the logarithmic ratio of joint probability P(X(t),Y(t+τ)) and marginal probability P(X(t)) and P(Y(t+τ)).

[0102] All possible state parameter pairs are traversed, and the time series mutual information under different time delays (such as from 0 to 5 seconds, with a step of 0.1 seconds) is calculated, for each pair of parameters, the maximum value of the time series mutual information is recorded as the causal correlation strength, and the corresponding optimal time delay is recorded, when the causal correlation strength exceeds the preset time series threshold (set according to system characteristics, such as 0.3), it is considered that there is a causal relationship between the two state parameters, and a directed edge is established between the corresponding state parameters, thereby forming a directed graph structure.

[0103] After the directed graph is constructed, topological sorting is performed to obtain the causal sequence, the number of incoming connections of each state parameter is counted, and the state parameter with zero incoming connection number (i.e. the source node not affected by other parameters) is identified, which is added to the causal sequence, the parameter is removed from the directed graph, and the incoming connection number of its associated parameters is updated, the above process is repeated, and the parameters with zero incoming connection number are sequentially added to the causal sequence until the sorting of all state parameters is completed.

[0104] According to the obtained causal sequence, all possible causal path sets from each state parameter to the control output are constructed, for example, if the causal sequence shows that parameter A affects parameter B, and parameter B affects control output C, there is a causal path A→B→C from A to C, by identifying all such paths, the influence of state parameters on control output can be fully understood.

[0105] In order to quantify the influence of state parameters on control output, a small perturbation (such as increasing or decreasing its value by 5%) is applied to each state parameter, and the change of control output is observed, specifically, for state parameter x i The perturbation Δx i is applied, the change Δy of control output y is measured, and the change rate r i =Δy / Δx iThe rate of change reflects the sensitivity of the state parameter to the control output. The state parameter with a rate of change greater than a preset change threshold (such as 0.1) is identified as a key state parameter, and a key state subset is formed. These parameters have a significant impact on the control output and should be prioritized in the control algorithm design.

[0106] In practical applications, by analyzing the relationship between the vehicle lateral position, lateral velocity, lateral acceleration, steering wheel angle, lane deviation, and the steering control output, it is found that the steering wheel angle and lateral acceleration are the two most critical state parameters affecting steering control. This result can be used to optimize the automatic driving control algorithm and improve the accuracy and response speed of lateral control. This method is also applicable to intelligent connected vehicle and roadside device cooperative control scenarios. By analyzing the causal relationship between vehicle state, roadside perception data, and control instructions, it is found that communication delay and roadside predicted trajectory are the key state parameters affecting cooperative control effect, so the vehicle-road cooperative control strategy can be optimized accordingly.

[0107] In an optional implementation, according to the causal sequence, a set of all causal paths of each state parameter to the control output is constructed, and a mapping weight matrix of the state to the control is calculated. The rate of change of the control output is calculated by applying a disturbance to the state parameter. The state parameter with a rate of change greater than a change threshold is included in the key state subset, including:

[0108] Based on the sorted position of the state parameter in the causal sequence, breadth-first traversal is performed from the source node of the directed graph. The time series mutual information value of each state parameter to the control output and its connection edge is recorded. The node sequence and its edge information are stored as a causal path. All reachable control output nodes are traversed to form a complete causal path set.

[0109] The time series mutual information value on each causal path is calculated, and the path propagation coefficient is obtained. The sum of all path propagation coefficients with the state parameter as the starting node is calculated to obtain the direct influence factor. The sum of the path propagation coefficient and the edge mutual information with the state parameter as the intermediate node is calculated as the indirect influence factor. The mapping weight matrix of the state to the control is calculated by combining the direct influence factor.

[0110] The standard deviation of the historical sequence of each state parameter is selected as the disturbance reference. Positive and negative disturbances are added to the current value. The change in control output caused by the disturbance is calculated by the mapping weight. The difference between the positive and negative disturbance influence values is extracted to obtain the bidirectional rate of change of the state parameter.

[0111] Based on the bidirectional rate of change of all state parameters, the interquartile range is calculated. The third quartile and the multiple of the interquartile range are set as the screening threshold. The state parameters with a rate of change greater than the screening threshold are extracted, and the key state subset is obtained by rearranging according to the causal sequence.

[0112] In the specific embodiment, a causal directed graph model of the train-ground network system is obtained, wherein nodes represent state parameters and control outputs, and edges represent the influence relationship between nodes. The state parameters include train running speed, acceleration, position information, electromechanical equipment working state, track state, etc., and the control outputs include train traction force, braking force, signal system control command, etc.

[0113] According to the ordering position of the state parameters in the causal sequence, breadth-first traversal is performed from the source node of the directed graph, and during the traversal process, the time sequence mutual information value of all node sequences and connecting edges from each state parameter to the control output is recorded. The time sequence mutual information value is calculated by analyzing historical data and represents the influence degree of the state parameter change on the subsequent node;

[0114] Taking the train speed state parameter as an example, the following paths are obtained by performing breadth-first traversal: train speed→braking distance→braking force control output, and train speed→energy consumption estimation→traction force control output. For each path, the time sequence mutual information value of each edge on the path is recorded, such as the mutual information value of speed and braking distance is 0.82, and the mutual information value of braking distance and braking force is 0.91. The node sequence and edge information obtained by traversal are stored as a causal path set. For each state parameter in the rail transit system, all possible paths from it to each control output are counted to form a complete causal path set. For example, there are multiple paths from the track state parameter to the speed control output, including direct influence paths and indirect paths through other state parameters.

[0115] The path propagation coefficient of each causal path is calculated, which is the time sequence mutual information value of all edges on the path multiplied together, for example, the propagation coefficient of the train speed→braking distance→braking force control output path is 0.82×0.91=0.746. For a certain state parameter, the sum of the propagation coefficients of all paths with it as the starting node is obtained, which is the direct influence factor, such as the direct influence factors of the train speed parameter on the braking force and the traction force are 0.746 and 0.623, respectively. The indirect influence of each state parameter as an intermediate node is calculated, for example, when the current speed is used as an intermediate node to affect the braking control, the sum of the propagation coefficients of all paths passing through the speed node to the braking control output is calculated, which is multiplied by the mutual information of the incoming edge, and is used as the indirect influence factor. The direct influence factor and the indirect influence factor are combined to obtain the mapping weight of the state to the control.

[0116] For each state parameter in the train-ground collaborative system, the standard deviation of its historical sequence data is calculated as the parameter disturbance reference, for example, the historical data standard deviation of the train speed parameter is 2.5 km / h, and the disturbance reference is set to 2.5 km / h. Positive and negative disturbances are added to the current speed value, and the changes in control outputs caused by the disturbances are obtained through the mapping weight calculated in the foregoing.

[0117] Assuming the current speed of the train is 60 km / h, the positive disturbance is 62.5 km / h, and the negative disturbance is 57.5 km / h, the mapping weight calculation shows that the positive disturbance causes the braking force to increase by 8.2%, and the negative disturbance causes the braking force to decrease by 7.9%. By extracting the difference between the positive and negative disturbance influence values, the bidirectional change rate of the speed parameter on the braking control is 16.1%. Sort the bidirectional change rates of all state parameters in the train-ground network, calculate the interquartile range, for example, the first quartile of all parameter change rates is 3.2%, and the third quartile is 12.8%, with an interquartile range of 9.6%. Set the screening threshold to be 1.5 times the third quartile plus the interquartile range, i.e. 12.8% + 9.6% x 1.5 = 27.2%, extract state parameters with change rates exceeding the screening threshold, and rearrange them according to the causal sequence to obtain the key state subset.

[0118] In rail transit applications, the key state subset includes parameters such as train position, speed, track curvature, slope, turnout state, etc., which have a significant impact on control output and need to be monitored and processed first. By screening the key state subset, the train-ground communication load can be reduced and the control system response speed can be improved. Based on the key state subset obtained by screening, precise control of multiple systems in rail transit can be achieved, for example, when a change in the key state parameter of track curvature is detected, the system can quickly adjust the train speed control strategy to ensure safe operation. At the same time, non-critical parameters can use low-frequency acquisition and transmission strategies to reduce the occupation of train-ground communication resources.

[0119] Through the above method, the key state parameters affecting the rail transit control system can be effectively identified, the system response speed and control accuracy can be improved, and technical support can be provided for rail transit multi-system information interaction control in train-ground network cooperation. This method is applicable to various rail transit scenarios such as intercity railways, high-speed railways, urban rail transit, etc., and has wide application prospects.

[0120] In an optional implementation, based on the key state subset and safety constraint rules, a multi-objective optimization function is constructed to minimize the operation adjustment time, energy consumption, and maximize the safety margin. The particle swarm algorithm is used to solve the multi-objective optimization function to obtain a set of Pareto optimal solutions. The optimal cooperative control instruction sequence is calculated in combination with the expert knowledge base rules, including:

[0121] The deviation of the current value of each state parameter in the key state subset from the target value is calculated, the operation adjustment time is calculated in combination with the control instruction adjustment rate, the energy consumption coefficient per unit adjustment is obtained, the total energy consumption value is calculated, and the minimum interval between the target value of the state parameter and the safety boundary is calculated as the safety margin. The operation adjustment time, total energy consumption value, and safety margin are constructed into a multi-objective optimization function;

[0122] Within the constraints of the control command parameters, an initial particle swarm is generated, the particle position and velocity are iteratively updated, the multi-objective optimization function value is non-dominated and the crowding distance is calculated, and the final non-dominated solution set is extracted as the Pareto optimal solution.

[0123] The Pareto optimal solution is substituted into the mapping weight matrix to calculate the state response. The response sensitivity is calculated through perturbation analysis. The current scenario strategy is extracted based on the expert rule base. The node attribute distribution of the key state subset is analyzed. The weights of each objective in the multi-objective optimization function are adjusted. The Pareto optimal solution with the lowest response sensitivity and the largest multi-objective optimization function value is selected as the optimal cooperative control command sequence.

[0124] In this specific embodiment, it is necessary to calculate the deviation between the current value and the target value of each parameter in the critical state subset, that is, for the critical state subset S={s1, s2, ..., s... n Each state parameter s in} i Let n be the total number of state parameters in the key state subset, and calculate its current value s. i_current With target value s i_target Deviation δ i =|s i_current -s i_target | Then, the rate r is adjusted in conjunction with control commands. i Calculate the running adjustment time T i =δ i / r i The maximum value among all parameter adjustment times is taken as the overall system adjustment time T = max(T1, T2, ..., T). n ).

[0125] Obtain the energy consumption coefficient e of unit adjustment i This represents the energy consumption required for each unit of state parameter adjustment. The total energy consumption value E is calculated as the product of the state parameter adjustment and the energy consumption coefficient: E = Σ(δ). i ×e i Meanwhile, the minimum interval between the target value of the statistical state parameter and the safety boundary is used as the safety margin S=min(|s i_target -s i_boundary |), where s i_boundary Represents the state parameter s i The safety boundary value.

[0126] Based on the above calculation results, a multi-objective optimization function F=[f1(T), f2(E), f3(S)] is constructed, wherein f1(T) represents an adjustment time target function, and takes T as a value; f2(E) represents an energy consumption target function, and takes E as a value; f3(S) represents a safety margin target function, and takes -S (a negative sign indicates that the greater the safety margin is, the better) as a value; the goal of multi-objective optimization is to minimize f1(T) and f2(E), and maximize f3(S).

[0127] In the multi-objective optimization solving process, an initial particle swarm is generated within the control instruction parameter constraint range, and for an N-dimensional control instruction parameter space, M initial particles are randomly generated, each particle being represented as an N-dimensional vector X i =(x i1 , x i2 ,..., x iN ), wherein i=1,2,...,M, and each parameter satisfies the constraint condition x ij ∈[x min_j , x max_j ], and an initial velocity vector V i =(v i1 , v i2 ,..., v iN ) is allocated to each particle, and the velocity range is limited within [-v max , v max ].

[0128] The particle swarm algorithm iteration optimization process includes: updating the position and velocity of each particle, evaluating the multi-objective function value, performing non-dominated sorting and calculating the crowded distance, and specifically, in the t+1 iteration, the velocity and position update formula of particle i is: V i (t+1)=w×V i (t)+c1×r1×[p besti -X i (t)]+c2×r2×[g best -X i (t)], X i (t+1)=X i (t)+V i (t+1), wherein w is an inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers in the interval [0,1], p besti is the historical optimal position of particle i, and g best is the global optimal position.

[0129] The updated particle swarm is evaluated for multi-objective function values, the particles are layered using a non-dominated sorting method, a Pareto front is determined, for particles in the same non-dominated layer, a crowding distance is calculated to maintain the diversity of solutions, the crowding distance calculation is based on the normalized range of each objective function value, and represents the distribution density of the particle in the solution space, after a predetermined number of iterations, the final non-dominated solution set is extracted as the Pareto optimal solution.

[0130] Each Pareto optimal solution is substituted into the mapping weight matrix W to calculate the state response vector R = W x X, where X is the control command vector and R is the expected state response. Through perturbation analysis, a small perturbation δ X is applied to the control command X R , the response change amount δ X = W x δ r is calculated, and the response sensitivity S R = |δ X | / |δ X | is obtained, which reflects the sensitivity of the system to changes in the control command.

[0131] At the same time, based on the expert rule base, the decision strategy applicable to the current scene is extracted, such as "high safety margin priority", "low energy consumption priority" or "fast response priority", etc. According to the extracted strategy and the node attribute distribution characteristics of the key state subset, the weights α1, α2 and α3 of each objective in the multi-objective optimization function are dynamically adjusted to form the weighted objective function F w = α1 x f1(T) + α2 x f2(E) + α3 x f3(S), from the Pareto optimal solution set, the solution with the lowest response sensitivity S r and the optimal weighted objective function value F w is selected as the final control scheme, which is converted into a specific control command sequence output. This command sequence not only meets the multi-objective optimization requirements, but also has a low response sensitivity, which can maintain stable performance when the system parameters fluctuate slightly.

[0132] In an alternative embodiment, an initial particle swarm is generated within the control command parameter constraint range, the particle position and velocity are iteratively updated, the non-dominated sorting of the multi-objective optimization function value is performed and the crowding distance is calculated, and the final non-dominated solution set is extracted as the Pareto optimal solution, comprising:

[0133] The particle position and velocity vectors are randomly initialized within the control command parameter constraint range to form an initial particle swarm, the running adjustment time, total energy consumption value and safety margin objective function value of each particle are calculated, and the non-dominated sorting is performed to divide the particles into multiple non-dominated levels;

[0134] For particles in the first non-dominated level, the interval distance of each particle in three objective dimensions with adjacent particles is calculated, and the comprehensive crowding distance is obtained by normalizing and summing, the particles with comprehensive crowding distance exceeding the density division threshold are marked as sparse areas, and the particles below the density division threshold are marked as dense areas;

[0135] Based on the non-dominated level and area marking of the particles, the inertia weight of the particles in the dense area is dynamically adjusted to enhance global search, and the inertia weight of the particles in the sparse area is reduced to enhance local search, and the speed update amount is calculated by combining the historical optimal position and the global optimal position of each particle;

[0136] The speed update amount is fused and superimposed on the position vector to obtain the new position of the particle, and the non-dominated sorting iteration is repeated after boundary mapping correction until the convergence number limit is reached, and the final first non-dominated level particles are extracted as the Pareto optimal solution set.

[0137] In the specific embodiment, the particle swarm is randomly initialized within the control instruction parameter constraint range, for a population of N particles, each particle is represented as a D-dimensional vector, representing different parameters of the control instruction, for example, if the control instruction contains three parameters of speed, acceleration and time, then D=3, the position vector of particle i can be represented as X i =[x i1 , x i2 ,...,x iD ], where the value range of each dimension is limited by the physical constraints of the system, and the velocity vector V i =[v i1 , v i2 ,..., v iD ] represents the moving speed of the particle in each dimension. The initial position is randomly generated within the parameter constraint [L min , L max ], and the initial speed is set to a random value within the range [-V max , V max ], where V max is usually set to 10%-20% of the parameter range.

[0138] After initialization, the values of multiple objective functions corresponding to each particle are calculated, for control system optimization, three objective functions are mainly considered: running adjustment time f1(X i ), which represents the time required for the system to reach the target state from the initial state; total energy consumption f2(X i ), which represents the energy consumption during the entire control process; safety margin f3(X i ), which represents the distance between the system operating state and the dangerous boundary. These three objectives are usually conflicting, for example, reducing the adjustment time will usually increase the energy consumption or reduce the safety margin.

[0139] After the objective function values are calculated, non-dominated sorting is performed to divide the particles into different levels. A particle a is said to dominate a particle b if a is not worse than b in all objectives and better than b in at least one objective. The dominance relationship comparison is performed for all particles. The particles that are not dominated by any other particle form the first non-dominated level. The non-dominated particles are searched again from the remaining particles to form the second non-dominated level, and so on, until all particles are assigned to the corresponding non-dominated levels.

[0140] For the particles in the first non-dominated level, the crowding distance is calculated to maintain the diversity of solutions. For each objective function, the particles are first sorted according to the objective value. The distance between adjacent particles in the objective dimension is calculated. For boundary particles (i.e., the particles with the maximum or minimum objective value), the crowding distance is set to infinity. For intermediate particles, the crowding distance in objective j is calculated as the normalized difference between the adjacent two particles in the objective. The crowding distances in the three objective dimensions are summed to obtain the comprehensive crowding distance. A density distinction threshold D is set (usually 0.8-1.2 times the average crowding distance). The particles with a crowding distance greater than D are marked as sparse regions, and the particles with a crowding distance less than D are marked as dense regions. c c c

[0141] Based on the non-dominated levels of the particles and the region marking, the optimization strategy is dynamically adjusted. For the particles in the dense region, the inertia weight w is increased from a typical value of 0.5 to 0.7-0.9 to enhance the global search ability. For the particles in the sparse region, the inertia weight is reduced to 0.2-0.4 to enhance the local search. The individual learning factor c1 and the social learning factor c2 are usually set to 1.5-2.5. The velocity update formula of particle i considers its historical optimal position P and the nearest position G in the global non-dominated particles. The updated velocity is limited in the range of [-V, V]. besti besti max max

[0142] After the position is updated, it is checked whether it exceeds the parameter constraint boundary. If it does, the boundary mapping method is used to map the particle position back to the effective region, and the velocity direction is adjusted in the opposite direction to avoid the particle continuously rushing out of the boundary. After the position is updated, the objective function value is recalculated and the non-dominated sorting is performed.

[0143] The iteration process continues until the preset maximum iteration number (usually 100-300 times) or the change of the particles in the first non-dominated level is small for a large number of consecutive iterations (such as 20 times), indicating that the algorithm has basically converged. Finally, all particles in the first non-dominated level are extracted to form a Pareto optimal solution set, providing a series of non-dominated optional solutions for the decision maker.​​​​​​​

[0144] In practical applications, the algorithm parameters can be adjusted according to the specific system characteristics, for example, for a control system with high response time requirements, the population size can be increased to 50-100 to improve search efficiency; for a complex nonlinear system, the maximum number of iterations can be increased to more than 500 to ensure that the algorithm converges fully; the density threshold can also be dynamically adjusted according to the characteristics of the problem, and a smaller value is set in the early search to expand the search range, and gradually increased later to refine the optimal solution distribution.

[0145] The method can effectively balance global search and local refinement, improve search efficiency while maintaining solution diversity by dynamically adjusting particle behavior characteristics, and provides a practical and effective solution for multi-objective optimization of complex industrial control systems.

[0146] The track traffic multi-system information interaction control system based on vehicle-ground network cooperation according to the embodiment of the application comprises:

[0147] The first unit is used for acquiring train on-board equipment operation data and ground infrastructure operation data.

[0148] The second unit is used for extracting semantic feature vectors of the on-board equipment operation data and the infrastructure operation data, learning a domain-to-domain mapping transformation matrix by minimizing the semantic space distribution distance between the source domain and the target domain, projecting the heterogeneous data fields to a unified semantic space to obtain standardized operation data.

[0149] The third unit is used for calculating a time-space offset of the train-ground data and performing time-space alignment according to the time stamp and the spatial position information in the standardized operation data, and establishing a train-ground synchronization state set.

[0150] The fourth unit is used for constructing the train-ground synchronization state set into a directed graph structure, establishing a directed edge between nodes based on time-series mutual information and performing topological sorting to obtain a causal sequence, calculating the change rate of the state parameters to the control output, and screening a key state subset.

[0151] The fifth unit is used for constructing a multi-objective optimization function of minimum operation adjustment time, minimum energy consumption and maximum safety margin based on the key state subset and safety constraint rules, solving the multi-objective optimization function by using a particle swarm algorithm, obtaining a Pareto optimal solution set, and calculating an optimal cooperative control instruction sequence in combination with an expert knowledge base rule.

[0152] The sixth unit is used for issuing the optimal cooperative control instruction sequence to a train-ground execution unit and collecting feedback states.

[0153] In a third aspect, the embodiment of the application provides an electronic device, comprising:

[0154] a processor;

[0155] Memory used to store processor-executable instructions;

[0156] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0157] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0158] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for track traffic multi-system information interaction control based on train-ground network cooperation, characterized in that, The method comprises the following steps: acquiring train-mounted equipment operation data and ground infrastructure operation data; extracting semantic feature vectors of the train-mounted equipment operation data and the infrastructure operation data, learning an inter-domain mapping transformation matrix by minimizing the semantic space distribution distance between a source domain and a target domain, projecting the heterogeneous data fields into a unified semantic space to obtain standardized operation data; calculating a train-ground data space-time offset and performing space-time alignment according to the time stamp and spatial position information in the standardized operation data, and establishing a train-ground synchronous state set; constructing the train-ground synchronous state set into a directed graph structure, establishing directed edges between nodes based on time-series mutual information and performing topological sorting to obtain a causal sequence, calculating the rate of change of state parameters on control output, and screening a key state subset; based on the key state subset and safety constraint rules, constructing a multi-objective optimization function of minimum operation adjustment time, minimum energy consumption and maximum safety margin, solving the multi-objective optimization function by using a particle swarm algorithm to obtain a Pareto optimal solution set, and calculating an optimal cooperative control instruction sequence in combination with an expert knowledge base rule; downloading the optimal cooperative control instruction sequence to a train-ground execution unit, and collecting feedback states.

2. The method of claim 1, wherein, extracting semantic feature vectors of the train-mounted equipment operation data and the infrastructure operation data, learning an inter-domain mapping transformation matrix by minimizing the semantic space distribution distance between a source domain and a target domain, projecting the heterogeneous data fields into a unified semantic space to obtain standardized operation data, comprising: for each data field in the train-mounted equipment operation data, extracting a field identifier, a data type descriptor and a business context descriptor, inputting a pre-trained semantic encoder to obtain a source domain semantic feature vector, and performing the same extraction and encoding operation on each data field in the infrastructure operation data to obtain a target domain semantic feature vector; calculating the Frobenius norm distance between the covariance matrices according to the covariance matrices of the source domain semantic feature vectors and the target domain semantic feature vectors, initializing the inter-domain mapping transformation matrix, and multiplying the source domain semantic feature vectors by the inter-domain mapping transformation matrix to obtain transformed source domain feature vectors; calculating the maximum mean difference between the transformed source domain feature vectors and the target domain semantic feature vectors, constructing a loss function in combination with the Frobenius norm distance, and iteratively updating the inter-domain mapping transformation matrix by using a gradient descent method until the loss function converges to a preset convergence condition; using the converged inter-domain mapping transformation matrix to perform linear transformation on the source domain semantic feature vectors to obtain feature representations in a unified semantic space, reconstructing the standardized format and standardized values of the data fields according to the feature representations in the unified semantic space, and obtaining standardized operation data.

3. The method of claim 1, wherein, calculating a train-ground data space-time offset and performing space-time alignment according to the time stamp and spatial position information in the standardized operation data, and establishing a train-ground synchronous state set, comprising: extracting the time stamp information of the train-mounted equipment and the infrastructure from the standardized operation data, converting the train-mounted equipment time stamp into a train local time reference, converting the infrastructure time stamp into a ground system time reference, and calculating a time offset vector; According to the standardized operation data, track coordinates of the current mileage position of the train and the installation position of the infrastructure are extracted, the spatial distance between the train and the infrastructure at the same time is calculated, a spatial offset vector is obtained, and a time-space offset matrix is constructed by combining the time offset vector and the spatial offset vector; According to the time-space offset matrix, the timestamps of the on-board device data are corrected to achieve time alignment, and the spatial reference system of the infrastructure data is converted to achieve spatial registration. Train status data after time alignment and device status data after spatial registration are extracted. A train-ground data correlation rule is established using the time-space offset matrix. Based on the train position, the spatially related ground device status is searched. Time-synchronous state sets are combined.

4. The method of claim 1, wherein, The train-ground synchronous state set is constructed as a directed graph structure. A directed edge between nodes is established based on the time-series mutual information, and a topological sort is performed to obtain a causal sequence. The rate of change of the state parameter to the control output is calculated, and a key state subset is selected, including: For each state parameter in the train-ground synchronous state set, perform time-series arrangement, extract parameter value sequences within a sliding time window, and calculate the probability density function of the state parameter through kernel density estimation. The time-series mutual information is calculated based on the integral of the logarithmic ratio of joint probability density and marginal probability density. The time-series mutual information under different time delays is calculated by traversing all state parameter pairs. The maximum time-series mutual information is taken as the causal correlation strength. When it exceeds the time-series threshold, a directed edge is established between the state parameters, forming a directed graph structure. The number of incoming connections of each state parameter is counted. The state parameters with zero incoming connections are sequentially added to the causal sequence, and the incoming connection number of their associated parameters is updated. Repeat the process until the causal sorting of all state parameters is completed to obtain the causal sequence. According to the causal sequence, all causal path sets of each state parameter to the control output are constructed, and a state-to-control mapping weight matrix is calculated. The rate of change of the control output is calculated by applying a disturbance to the state parameter. The state parameters with a change rate greater than the change threshold are taken as the key state subset.

5. The method of claim 4, wherein, According to the causal sequence, all causal path sets of each state parameter to the control output are constructed, and a state-to-control mapping weight matrix is calculated. The rate of change of the control output is calculated by applying a disturbance to the state parameter. The state parameters with a change rate greater than the change threshold are taken as the key state subset, including: Based on the sorting position of the state parameter in the causal sequence, breadth-first traversal is performed from the source node of the directed graph. The time-series mutual information value of all node sequences and their connection edges from each state parameter to the control output is recorded. The node sequence and its edge information are stored as a causal path. All reachable control outputs are traversed to form a complete causal path set. The path propagation coefficient is obtained by calculating the product of the time-series mutual information values on each causal path. The direct influence factor is obtained by summing all path propagation coefficients with the state parameter as the starting node. The indirect influence factor is obtained by counting the product of the path propagation coefficient and the incoming edge mutual information with the state parameter as the intermediate node. The state-to-control mapping weight matrix is calculated by combining the direct influence factor. The standard deviation of the historical sequence of each state parameter is selected as a disturbance reference, positive and negative disturbances are superimposed on the current numerical value, the change of the control output caused by the disturbance is calculated through the mapping weight, the difference between the positive and negative disturbance influence values is extracted, and the bidirectional change rate of the state parameter is obtained; The quartile range is calculated based on the bidirectional change rate of all state parameters, the third quartile and the multiple of the quartile range are set as the screening threshold, the state parameters whose change rates exceed the screening threshold are extracted, and the key state subset is rearranged in the causal sequence.

6. The method of claim 1, wherein, Based on the key state subset and the safety constraint rule, a multi-objective optimization function of minimum running adjustment time, energy consumption and maximum safety margin is constructed, the particle swarm optimization algorithm is used to solve the multi-objective optimization function, the Pareto optimal solution set is obtained, and the optimal cooperative control instruction sequence is calculated combined with the expert knowledge base rule, including: The deviation amount of the current value and the target value of each state parameter in the key state subset is calculated, the running adjustment time is calculated combined with the control instruction adjustment rate, the energy consumption coefficient of unit adjustment is obtained, the total energy consumption value is calculated, the minimum interval between the target value of the state parameter and the safety boundary is calculated as the safety margin, and the running adjustment time, the total energy consumption value and the safety margin are constructed as a multi-objective optimization function; An initial particle swarm is generated within the constraint range of the control instruction parameters, the particle position and velocity are iteratively updated, the non-dominated sorting of the multi-objective optimization function value is performed and the crowded distance is calculated, and the final non-dominated solution set is extracted as the Pareto optimal solution; Each Pareto optimal solution is substituted into the mapping weight matrix to calculate the state response, the response sensitivity is calculated through disturbance analysis, the current scene strategy is extracted based on the expert rule base, the node attribute distribution of the key state subset is analyzed, the weights of each target in the multi-objective optimization function are adjusted, the Pareto optimal solution with the lowest response sensitivity and the maximum multi-objective optimization function value is selected as the optimal cooperative control instruction sequence.

7. The method of claim 6, wherein, An initial particle swarm is generated within the constraint range of the control instruction parameters, the particle position and velocity are iteratively updated, the non-dominated sorting of the multi-objective optimization function value is performed and the crowded distance is calculated, and the final non-dominated solution set is extracted as the Pareto optimal solution, including: The particle position and velocity vector are randomly initialized within the constraint range of the control instruction parameters to form an initial particle swarm, the running adjustment time, total energy consumption value and safety margin objective function value of each particle are calculated, and the particles are divided into multiple non-dominated levels through non-dominated sorting; For the particles in the first non-dominated level, the interval distance of each particle in the three target dimensions and the adjacent particles is calculated, and the comprehensive crowded distance is obtained by normalizing the sum, the particles with a comprehensive crowded distance exceeding the density division threshold are marked as sparse areas, and the particles with a comprehensive crowded distance below the density division threshold are marked as dense areas; Based on the non-dominated level and area marking of the particles, the inertia weight of the particles in the dense area is dynamically adjusted to enhance global search, and the inertia weight of the particles in the sparse area is reduced to enhance local search, the velocity update amount is calculated combined with the historical optimal position and the global optimal position of each particle; The velocity update amount is fused and superimposed with the current velocity and a position vector to obtain a new position of the particle, and the non-dominant sorting iteration is repeated after boundary mapping correction until a convergence time limit is reached, and the final first non-dominant level particles are extracted as a Pareto optimal solution set.

8. A track traffic multi-system information interaction control system based on train-ground network cooperation, for implementing the method of any one of claims 1-7, characterized in that, The method comprises the following steps: A first unit is configured to acquire train-mounted equipment operation data and ground infrastructure operation data. A second unit is configured to extract semantic feature vectors of the train-mounted equipment operation data and the infrastructure operation data, learn a domain mapping transformation matrix by minimizing the semantic space distribution distance between source and target domains, project heterogeneous data fields into a unified semantic space to obtain standardized operation data. A third unit is configured to calculate train-ground data space-time offset and perform space-time alignment according to time stamp and spatial position information in the standardized operation data, and establish a train-ground synchronization state set. A fourth unit is configured to construct the train-ground synchronization state set into a directed graph structure, establish directed edges between nodes based on time series mutual information, perform topological sorting to obtain a causal sequence, calculate the change rate of state parameters to control output, and select a key state subset. A fifth unit is configured to construct a multi-objective optimization function of minimum operation adjustment time, minimum energy consumption and maximum safety margin based on the key state subset and safety constraint rules, solve the multi-objective optimization function by using a particle swarm algorithm, obtain a Pareto optimal solution set, and calculate an optimal cooperative control instruction sequence in combination with an expert knowledge base rule. A sixth unit is configured to issue the optimal cooperative control instruction sequence to a train-ground execution unit and collect feedback states.

9. An electronic device, comprising: The method comprises the following steps: A processor; A memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 7.

Images