Artificial Intelligence-Based Intelligent Train Operation Control Method and System

Abstract

The present invention provides an intelligent control method and system for train operation based on artificial intelligence, which relates to the field of artificial intelligence technology. By dividing the train state perception data into blocks and generating graph-encoded vector data of multiple train state perception paths, secondly, through the feature focusing process, key features are effectively extracted from the complex train state perception data, generating accurate target state perception features, further improving the recognition accuracy of the train operation state. Moreover, the present invention realizes the dynamic and real-time optimization of the train operation state by determining the collaborative chain of states to be optimized, ensuring the stability and safety of train operation. Finally, intelligent control of the target train is performed based on the collaborative chain of states to be optimized, which not only improves the efficiency of train operation but also reduces the risks brought by human operations.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to an intelligent train operation control method and system based on artificial intelligence. Background Art

[0002] As an important part of the modern transportation system, the operation efficiency and safety of trains have always been the focus of the industry. Traditional train control methods are mainly based on manual operation and fixed rule systems, which can play a good role when the train operation state is relatively simple and stable. However, with the increasing complexity of the train operation environment and the continuous improvement of operation requirements, traditional control methods gradually expose limitations such as lagging response and insufficient control accuracy.

[0003] In order to improve the accuracy and real-time performance of train operation control, the industry has been exploring new technical means. Among them, the full utilization of train state perception data has become an important research direction. A large amount of state data is generated during the operation of trains, including various information such as speed, position, acceleration, and braking force. These data contain the real-time state and future trends of train operation, which are of great significance for optimizing train control strategies.

[0004] However, existing train state data processing methods are often relatively simple and difficult to comprehensively and accurately extract the key information in these data. Some methods only rely on a single train state data, ignoring the correlation and complementarity between different state data; while other methods consider multiple state data, but lack effective data processing and analysis means, resulting in the inability to accurately reflect the actual situation of train operation. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide an intelligent train operation control method based on artificial intelligence, and the method includes:

[0006] Taking multiple train state perception data in the train state perception data sequence of the target train as candidate train state perception data respectively, and performing chunking on the candidate train state perception data to generate multiple train state perception chunks, each train state perception chunk including multiple train state perception paths;

[0007] Performing graph encoding representation on the multiple train state perception paths respectively to generate graph encoding vector data of the multiple train state perception paths;

[0008] Taking the multiple train state perception blocks as candidate blocks respectively, performing feature focusing on the graph-encoded vector data of multiple train state perception paths in the candidate blocks to generate the feature-focused data of the candidate blocks, and fusing the feature-focused data of the multiple train state perception blocks to generate the target state perception features of the candidate train state perception data;

[0009] Performing feature focusing on the target state perception features of multiple train state perception data in the train state perception data sequence to generate the feature-focused data of the train state perception data sequence, and determining the to-be-optimized state collaboration chain of the train state perception data sequence according to the feature-focused data of the train state perception data sequence;

[0010] Performing intelligent control on the target train based on the to-be-optimized state collaboration chain of the train state perception data sequence.

[0011] In a possible implementation manner of the first aspect, the step of taking the multiple train state perception blocks as candidate blocks respectively, performing feature focusing on the graph-encoded vector data of multiple train state perception paths in the candidate blocks to generate the feature-focused data of the candidate blocks includes:

[0012] Taking the multiple train state perception blocks as candidate blocks respectively, performing feature focusing on the graph-encoded vector data of multiple train state perception paths in the candidate blocks to generate first feature-focused data, where the first feature-focused data includes multiple first focused sub-clusters;

[0013] Determining the feature-focused data of the candidate blocks based on the feature-focused data of the multiple first focused sub-clusters;

[0014] The step of performing feature focusing on the target state perception features of multiple train state perception data in the train state perception data sequence to generate the feature-focused data of the train state perception data sequence includes:

[0015] Performing feature focusing on the target state perception features of multiple train state perception data in the train state perception data sequence to generate second feature-focused data, where the second feature-focused data includes multiple second focused sub-clusters;

[0016] Determining the feature-focused data of the train state perception data sequence based on the feature-focused data of the multiple second focused sub-clusters.

[0017] In a possible implementation manner of the first aspect, the step of determining the feature-focused data of the candidate blocks based on the feature-focused data of the multiple first focused sub-clusters includes:

[0018] Fuse the representation vectors of the cluster centers of the multiple first focused clusters to generate the feature focused data of the candidate block;

[0019] Determining the feature focused data of the train state perception data sequence based on the feature focused data of the multiple second focused clusters includes:

[0020] Fuse the representation vectors of the cluster centers of the multiple second focused clusters to generate the feature focused data of the train state perception data sequence.

[0021] In a possible implementation manner of the first aspect, taking the multiple train state perception blocks as candidate blocks respectively, and performing feature focusing on the graph encoding vector data of multiple train state perception paths in the candidate blocks to generate first feature focused data, includes:

[0022] Taking the multiple train state perception blocks as candidate blocks respectively, and performing feature focusing on the graph encoding vector data of multiple train state perception paths in the candidate blocks according to multiple different first clustering numbers, to generate first feature focused data respectively associated with the multiple first clustering numbers, where the first clustering number represents the scale of the first focused clusters in the corresponding first feature focused data; the feature focused data of the candidate block includes the feature focused data respectively associated with the multiple first clustering numbers of the candidate block;

[0023] Fusing the feature focused data of the multiple train state perception blocks to generate the target state perception features of the candidate train state perception data includes:

[0024] Fuse the feature focused data corresponding to the same first clustering number in the feature focused data of the multiple train state perception blocks to generate the target state perception features of the candidate train state perception data respectively corresponding to the multiple first clustering numbers;

[0025] Performing feature focusing on the target state perception features of the multiple train state perception data in the train state perception data sequence to generate second feature focused data includes:

[0026] Perform feature focusing on the target state perception features corresponding to the same first clustering number among the target state perception features of the multiple train state perception data in the train state perception data sequence, to generate second feature focused data respectively corresponding to the multiple first clustering numbers; the feature focused data of the train state perception data sequence includes the feature focused data respectively associated with the multiple first clustering numbers of the train state perception data sequence.

[0027] In a possible implementation of the first aspect, each second feature focusing data includes feature focusing sub-data respectively associated with N different second clustering quantities, and each feature focusing sub-data includes a plurality of second focusing sub-clusters; the second clustering quantity represents the scale of the second focusing sub-clusters in the corresponding second feature focusing data.

[0028] Each feature focusing data of the first clustering quantity in the train state perception data sequence includes a plurality of sub-vector data of the first clustering quantity in the train state perception data sequence respectively corresponding to a plurality of second clustering quantities.

[0029] In a possible implementation of the first aspect, the step of determining the to-be-optimized state collaboration chain of the train state perception data sequence according to the feature focusing data of the train state perception data sequence includes:

[0030] Performing state space mapping on the feature focusing data of the train state perception data sequence to generate state space mapping vector data of the train state perception data sequence;

[0031] Making a decision on the state space mapping vector data of the train state perception data sequence to generate the to-be-optimized state collaboration chain of the train state perception data sequence.

[0032] In a possible implementation of the first aspect, the making a decision on the state space mapping vector data of the train state perception data sequence to generate the to-be-optimized state collaboration chain of the train state perception data sequence includes:

[0033] Based on the train control knowledge expressed in the train control knowledge base, using an artificial intelligence network model to make a decision on the state space mapping vector data of the train state perception data sequence to determine the to-be-optimized state collaboration chain of the train state perception data sequence.

[0034] In a possible implementation of the first aspect, the based on the train control knowledge expressed in the train control knowledge base, using an artificial intelligence network model to make a decision on the state space mapping vector data of the train state perception data sequence to determine the to-be-optimized state collaboration chain of the train state perception data sequence includes:

[0035] Fusing the state space mapping vector data of the train state perception data sequence and the collaborative guidance vector data of the train state perception data sequence to generate global vector data of the train state perception data sequence;

[0036] Based on the train control knowledge expressed in the train control knowledge base, using an artificial intelligence network model, make decisions on the global vector data of the train state perception data sequence to determine the state coordination chain to be optimized for the train state perception data sequence;

[0037] Among them, the steps of making decisions on the global vector data of the train state perception data sequence based on the train control knowledge expressed in the train control knowledge base and using an artificial intelligence network model to determine the state coordination chain to be optimized for the train state perception data sequence include:

[0038] Collect and sort out various professional knowledge resources in the field of train operation, and use natural language processing technology and knowledge engineering technology to extract a structured train control knowledge base from the various professional knowledge resources. The train control knowledge base includes train operation rules, safety restrictions, equipment operation logic, and fault handling procedures;

[0039] Represent each knowledge feature in the extracted train control knowledge base in the form of nodes and edges. After generating the corresponding knowledge network, use a feature selection algorithm to extract the key feature set in the global vector data. The key feature set reflects the key change trends of the train operation state;

[0040] Traverse each key feature in the key feature set, use the relationship network in the knowledge network to find the association information of other entities and attributes associated with this key feature, and based on the association information, introduce context information for the original features in the knowledge network to generate an enhanced feature set. The context information includes the states of relevant components, historical fault records, and the impacts of control instructions;

[0041] Input the enhanced feature set into a pre-trained artificial intelligence network model, and output the predicted train state coordination chain. Among them, the pre-trained artificial intelligence network model incorporates a decision rule set formulated in advance based on the train control knowledge base.

[0042] In a possible implementation manner of the first aspect, the generating the state space mapping vector data of the train state perception data sequence by performing state space mapping on the feature-focused data of the train state perception data sequence includes:

[0043] Analyze the structure of the feature-focused data, determine the core data components and auxiliary data components in the feature-focused data, and based on the core data components and auxiliary data components, initialize the relevant parameters for state space mapping. The relevant parameters include the dimension setting of the mapping space and the initial weights of the mapping algorithm;

[0044] For each core data component and auxiliary data component in the feature-focused data, perform feature dimension conversion according to pre-set conversion rules to generate feature-focused data after feature dimension conversion, where the conversion rules are defined based on the physical characteristics of the train running state and the internal logical relationship between the train running state data;

[0045] Construct a state space framework for train state mapping. During the construction process, determine the meanings and value ranges of the coordinate axes of the state space, where the coordinate axes are used to reflect the core state factors of the train state, and use key operating parameters of train-related equipment, etc. as auxiliary coordinate axes, and define the boundaries and characteristics of different state regions in the state space framework;

[0046] Map each core data component and auxiliary data component in the feature-focused data to the constructed state space framework according to the corresponding coordinate axes. During the mapping process, determine the positions of the core data components and auxiliary data components in the state space framework according to the values of the core data components and auxiliary data components and the definitions of the coordinate axes in the state space framework. After the mapping is completed, combine the mapping positions of each core data component and auxiliary data component in the state space framework to generate the state space mapping vector.

[0047] On the other hand, an embodiment of the present invention further provides an intelligent train operation control system based on artificial intelligence, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0048] Based on the above aspects, the embodiment of the present application generates graph encoding vector data of multiple train state perception paths by partitioning the train state perception data. Secondly, through the feature focusing process, key features are effectively extracted from the complex train state perception data, generating accurate target state perception features, further improving the recognition accuracy of the train running state. Moreover, the present invention realizes the dynamic and real-time optimization of the train running state by determining the state collaboration chain to be optimized, ensuring the stability and safety of the train operation. Finally, based on the state collaboration chain to be optimized, intelligent control of the target train is performed, which not only improves the train operation efficiency but also reduces the risks brought by human operations. Description of the Drawings

[0049] Figure 1 It is a schematic flowchart of the execution of the intelligent train operation control method based on artificial intelligence provided by the embodiment of the present invention.

[0050] Figure 2 It is a schematic diagram of the hardware architecture of the intelligent train operation control system based on artificial intelligence provided by an embodiment of the present invention. Specific implementation manners

[0051] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of an intelligent train operation control method based on artificial intelligence provided by an embodiment of the present invention. The intelligent train operation control method based on artificial intelligence will be introduced in detail below.

[0052] Step S110: Respectively use multiple train state perception data in the train state perception data sequence of the target train as candidate train state perception data, and perform block division on the candidate train state perception data to generate multiple train state perception blocks, and each train state perception block includes multiple train state perception paths.

[0053] In this embodiment, the target train is equipped with a large number of sensors, and these sensors continuously collect various state information during the train operation, and this information constitutes the train state perception data sequence. For example, the sensors can collect the train speed, acceleration, temperature and humidity inside the carriage, working state parameters of various key devices (such as braking system, traction system, etc.) (such as braking pressure, current of the traction motor, etc.) and the position information of the train on the track, etc.

[0054] After the server receives this massive train state perception data sequence, it starts to process it. Taking the train state perception data collected in one time period as an example, the server takes each data in the train state perception data sequence in this time period as candidate train state perception data respectively. For example, first take the train speed data at a certain moment as a candidate train state perception data, and then take the temperature data inside the carriage at the same moment as another candidate train state perception data, etc.

[0055] Then, the server performs a chunking operation on each candidate train status perception data. Assume that the server performs chunking according to different subsystems or different functional areas of the train. Taking the traction system of the train as an example, the server can group all status perception data related to the traction system (such as the current, voltage, temperature, etc. of the traction motor) into a train status perception chunk; for the braking system, relevant data such as braking pressure and brake pad wear degree are grouped into another train status perception chunk. And each such train status perception chunk contains multiple train status perception paths. For example, in the train status perception chunk of the traction system, the process from the power supply of the traction motor to the operation of the motor, and then to the conversion of energy into the traction force of the train can be regarded as a train status perception path; and the process from the issuance of the control signal to the adjustment of the current of the traction motor can be regarded as another train status perception path. These different paths together constitute multiple train status perception paths in the train status perception chunk of the train traction system. Similarly, in the train status perception chunk of the braking system, the process from the issuance of the braking command to the adjustment of the braking pressure, and then to the feedback of the actual braking effect of the train is a train status perception path, and the process from the monitoring of the brake pad wear to the adjustment of the braking strategy according to the wear degree is another train status perception path, etc.

[0056] Step S120, respectively perform graph encoding representation on the multiple train status perception paths to generate graph encoding vector data of the multiple train status perception paths.

[0057] In this embodiment, taking a train status perception path in the traction system of the train as an example, this train status perception path is from the power supply of the traction motor to the operation of the motor, and then to the conversion of energy into the traction force of the train. The server first determines each node in this train status perception path, such as the power supply node, the winding node inside the motor (because the operation of the motor is related to the electromagnetic induction of the winding), the coupling node of energy transmission, and the friction force node between the train wheel and the track (because the traction force ultimately acts between the wheel and the track), etc. Then, the server constructs a graph structure based on the relationships between these nodes. For example, there is a relationship of electrical energy transmission between the power supply node and the winding node inside the motor, then this relationship is represented by a directed edge in the graph, and the weight of the edge can be determined according to factors such as the efficiency or power of electrical energy transmission; there is a magnetic field coupling relationship between the winding node inside the motor and the coupling node of energy transmission, and it is also represented by a directed edge and the corresponding weight is determined; there is a force transmission relationship between the coupling node of energy transmission and the friction force node between the train wheel and the track, and the corresponding directed edge and weight are also constructed.

[0058] After constructing the graph structure, the server encodes and represents this graph structure using a specific graph encoding algorithm. Suppose the graph encoding algorithm based on the adjacency matrix is adopted. The server converts the relationship between the nodes and edges in the graph into a graph encoding vector data in matrix form. In this adjacency matrix, the rows and columns of the matrix respectively correspond to the nodes in the graph. If there is an edge connection between two nodes, the weight of the edge is recorded at the corresponding position in the matrix; if there is no connection, it is recorded as 0. In this way, the corresponding graph encoding vector data is generated for this train status perception path. Using the same method, the server performs graph encoding representation on other train status perception paths in the traction system and all train status perception paths in other train status perception blocks such as the braking system and the car body environment control system, generating their respective graph encoding vector data.

[0059] Step S130: Respectively use the multiple train status perception blocks as candidate blocks, perform feature focusing on the graph encoding vector data of the multiple train status perception paths in the candidate blocks to generate the feature focusing data of the candidate blocks, and fuse the feature focusing data of the multiple train status perception blocks to generate the target status perception features of the candidate train status perception data.

[0060] In this embodiment, the multiple previously generated train status perception blocks can be respectively used as candidate blocks, and feature focusing operations are performed on the graph encoding vector data of the multiple train status perception paths in each candidate block.

[0061] Taking the traction system of the train as an example of this candidate block, it contains multiple train status perception paths for which graph encoding vector data has been previously generated. The server first performs a clustering operation on these graph encoding vector data according to a certain feature focusing algorithm, such as the density-based spatial clustering algorithm, to generate the first feature focusing data. In this process, the server clusters the similar graph encoding vector data together according to the feature space distribution of the graph encoding vector data to form multiple first focusing clusters. For example, those graph encoding vector data related to the normal operating state of the traction motor (such as graph encoding vector data related to normal power supply, normal winding electromagnetic induction, etc.) may be clustered into one first focusing cluster, while those graph encoding vector data related to the special operating state of the traction motor under high load (such as graph encoding vector data related to high current, high magnetic field intensity, etc.) may be clustered into another first focusing cluster.

[0062] Then, the server determines the feature focus data of this candidate block (traction system) based on the feature focus data of these multiple first focus clusters. Specifically, the server regards each first focus cluster as a whole and calculates the cluster center of each first focus cluster. For example, for a first focus cluster containing multiple graph encoding vector data, the representative vector of the cluster center is determined by calculating the mean vector of these vector data. Finally, the server fuses the representative vectors of the cluster centers of these first focus clusters to generate the feature focus data of the traction system, which is this candidate block.

[0063] For other candidate blocks (such as braking systems, car body environment control systems, etc.), the server operates in the same way as above to obtain the feature focus data of each candidate block. Then, the server fuses the feature focus data of multiple train state perception blocks (such as traction systems, braking systems, car body environment control systems, etc.). For example, the server performs weighted summation on the corresponding elements in the feature focus data of different candidate blocks according to a certain weight assignment rule to generate the target state perception features of the candidate train state perception data. Suppose the weight of the feature focus data of the traction system in the target state perception features is 0.4, that of the braking system is 0.3, and that of the car body environment control system is 0.3. Then, the feature focus data of each candidate block is fused according to this weight.

[0064] Step S140: Feature focus the target state perception features of multiple train state perception data in the train state perception data sequence to generate the feature focus data of the train state perception data sequence, and determine the state collaboration chain to be optimized of the train state perception data sequence according to the feature focus data of the train state perception data sequence.

[0065] In this embodiment, after obtaining the target state perception features of each candidate train state perception data in the train state perception data sequence, the feature focus operation is started on these target state perception features to generate the feature focus data of the train state perception data sequence.

[0066] Taking the multiple candidate train state perception data of a train within a period of time (such as the target state perception features of speed, braking pressure, car body temperature, etc.) as an example, the server uses a clustering algorithm similar to that in step S130 before to cluster these target state perception features and generate second feature focused data. During this process, multiple second focused clusters will be formed. For example, those target state perception features related to the safe operation of the train (such as the target state perception features related to normal braking pressure, reasonable speed range, etc.) may be clustered into a second focused cluster, while those target state perception features related to the comfort of train passengers (such as the target state perception features related to appropriate car body temperature, humidity, etc.) may be clustered into another second focused cluster.

[0067] Then, the server determines the feature focused data of the train state perception data sequence based on the feature focused data of these multiple second focused clusters. The specific operation is to fuse the representation vectors of the cluster centers of each second focused cluster together, for example, by weighted summation.

[0068] After obtaining the feature focused data of the train state perception data sequence, the server begins to determine the state coordination chain to be optimized for the train state perception data sequence according to this feature focused data. First, the server performs a state space mapping operation on the feature focused data of the train state perception data sequence. For example, the server analyzes the structure of the feature focused data to determine the core data components therein (such as the key operation parameters of the train, such as speed, braking pressure, etc.) and the auxiliary data components (such as the environmental parameters inside the car body, etc.). Then, based on these core data components and auxiliary data components, the server initializes the relevant parameters for state space mapping. For example, if the core data component is the train speed, the server determines the speed dimension setting of the mapping space according to the designed speed range of the train; if the auxiliary data component is the car body temperature, the relevant parameters in the state space mapping are determined according to the normal temperature range.

[0069] Next, for each core data component and auxiliary data component in the feature-focused data, perform feature dimension transformation according to the pre-set transformation rules. For example, for the core data component of train speed, if the original data is the actual speed value (such as 100 kilometers per hour), perform transformation according to the pre-set transformation rules (such as mapping the speed value to a value between 0 and 1, assuming that 100 kilometers per hour corresponds to 0.5). Then, the server constructs a state space framework for train state mapping. In this framework, determine the meanings of the coordinate axes of the state space. For example, take the train speed as one coordinate axis, the braking pressure as another coordinate axis, the carriage temperature as the third coordinate axis, etc., and determine the value range of each coordinate axis. At the same time, take the key operating parameters of train-related equipment, etc. as auxiliary coordinate axes, and define the boundaries and characteristics of different state regions in this state space framework. For example, define the region corresponding to the normal operating speed, braking pressure, and carriage temperature range of the train as the normal operating region, and define the region corresponding to parameters such as speed and braking pressure close to the safety limit as the warning region, etc.

[0070] Finally, the server maps each core data component and auxiliary data component in the feature-focused data to the constructed state space framework according to the corresponding coordinate axes to generate state space mapping vector data. Then, based on the train control knowledge expressed in the train control knowledge base, use an artificial intelligence network model to make decisions on this state space mapping vector data. For example, the server obtains knowledge such as the train operation rules (such as speed limits on different sections, braking distance requirements under different load conditions, etc.), safety limits (such as maximum braking pressure, maximum operating speed, etc.), equipment operation logic (such as the cooperative operation logic of the traction system and the braking system), and fault handling procedures (such as the emergency handling procedures when the braking system fails) from the train control knowledge base. Combine this knowledge with the state space mapping vector data, and use a pre-trained artificial intelligence network model (such as a deep neural network model), through a series of calculations and inferences, to determine the to-be-optimized state collaboration chain of the train state perception data sequence. For example, the artificial intelligence network model determines how the train's traction system and braking system should cooperate in the next period of time (such as when approaching a curve, appropriately reducing the speed, reducing the power output of the traction system, and preparing the braking system, etc.) based on the current train state (represented by the state space mapping vector data) and the train control knowledge. This is the to-be-optimized state collaboration chain.

[0071] Step S150, perform intelligent control on the target train based on the to-be-optimized state collaboration chain of the train state perception data sequence.

[0072] In this embodiment, taking the case where a train is about to enter a speed-limited section as an example, the collaborative chain in the state to be optimized indicates that the train needs to reduce its speed before entering this section. The server first sends a control command to the traction system of the train. According to the current speed of the train and the distance to the speed-limited section, it calculates the amount of power output reduction required by the traction system, and then sends this precise power adjustment command to the controller of the traction system, causing the traction motor to gradually reduce its power output, thereby making the train speed start to decrease.

[0073] Meanwhile, the server will also perform collaborative control with the braking system of the train. Although mainly the traction system is used to reduce the speed, the braking system also needs to be prepared in case the speed decreases too slowly or unexpected situations occur. The server sends a pre-command to the braking system according to the information in the collaborative chain in the state to be optimized, informing the braking system of the current speed adjustment situation of the train and the possible braking assistance required. For example, if the train speed is still higher than the limit speed at a certain distance from the speed-limited section, the server will send a more specific braking command to the braking system, such as a command to increase the braking pressure, causing the braking system to start working to ensure that the train can reach the specified speed limit when entering the speed-limited section.

[0074] Throughout the process, the server will also continuously monitor the sequence of train state perception data, such as continuously obtaining data such as the real-time speed, braking pressure, and car body temperature of the train. If it is found that there is a deviation between the actual running state of the train and the expected state in the collaborative chain in the state to be optimized, for example, the train speed decreases too fast or too slow, the server will recalculate and adjust the control command according to the new state perception data. For example, if the train speed decreases too fast, the server will reduce the braking pressure of the braking system or increase the power output of the traction system to maintain the smooth running of the train and ensure that the train runs safely and efficiently according to the requirements of the collaborative chain in the state to be optimized.

[0075] Based on the above steps, in the embodiment of the present application, by dividing the train state perception data into blocks and generating graph-encoded vector data of multiple train state perception paths, secondly, through the feature focusing process, key features are effectively extracted from the complex train state perception data, generating accurate target state perception features, further improving the recognition accuracy of the train running state. Moreover, the present invention realizes the dynamic and real-time optimization of the train running state by determining the collaborative chain in the state to be optimized, ensuring the stability and safety of the train running. Finally, based on the collaborative chain in the state to be optimized, intelligent control of the target train is performed, which not only improves the train running efficiency but also reduces the risks brought by manual operation.

[0076] In a possible implementation manner, step S130 includes:

[0077] Step S131: Take the multiple train status perception sub - blocks as candidate sub - blocks respectively, and perform feature focusing on the graph - encoded vector data of multiple train status perception paths in the candidate sub - blocks to generate first - feature - focused data, where the first - feature - focused data includes multiple first - focused clusters.

[0078] Step S132: Determine the feature - focused data of the candidate sub - blocks based on the feature - focused data of the multiple first - focused clusters.

[0079] In this embodiment, work is carried out for each candidate sub - block first. For example, the traction system of a train is taken as a train status perception sub - block. The train status perception paths in this sub - block cover the paths corresponding to processes from power supply to motor operation and then to traction force generation. The graph - encoded vector data of these paths contains information about the relationships between nodes and relevant parameters. The server performs feature focusing on this graph - encoded vector data and uses a clustering algorithm to generate first - feature - focused data. During this process, according to the distribution characteristics of the graph - encoded vector data in the feature space, similar data is clustered together to form multiple first - focused clusters. For example, the graph - encoded vector data related to the normal and stable operation of the motor, such as data related to normal voltage and current supply and stable magnetic field generation, forms a first - focused cluster; while the graph - encoded vector data related to the motor under special working conditions, such as high current and high magnetic field changes during overload, forms another first - focused cluster.

[0080] After obtaining these first - focused clusters, the server determines the feature - focused data of the candidate sub - blocks based on their feature - focused data. For each first - focused cluster, the server calculates its cluster center. Taking a first - focused cluster containing multiple graph - encoded vector data as an example, through specific mathematical calculations on these vector data, such as calculating the mean vector, a representative vector of the cluster center is obtained. Then, the representative vectors of the cluster centers of each first - focused cluster are fused, for example, by weighted summation, and the fused result is the feature - focused data of the candidate sub - block (such as the traction system).

[0081] Step S140 includes:

[0082] Step S141: Perform feature focusing on the target status perception features of multiple train status perception data in the train status perception data sequence to generate second - feature - focused data, where the second - feature - focused data includes multiple second - focused clusters.

[0083] Step S142: Determine the feature - focused data of the train status perception data sequence based on the feature - focused data of the multiple second - focused clusters.

[0084] In this embodiment, when performing feature focusing on the target state perception features of multiple train state perception data in the train state perception data sequence to generate the feature-focused data of the train state perception data sequence, the operation process is similar. Taking the target state perception features such as train speed, braking pressure, and car body temperature in the train state perception data sequence as an example, the server uses a clustering algorithm to perform feature focusing on these target state perception features, thereby generating the second feature-focused data. In this process, multiple second focusing clusters are formed according to the feature space distribution of the target state perception features. For example, the target state perception features related to train safety assurance, such as the target state perception features related to a reasonable braking pressure range, speed values within the normal speed limit, etc., form a second focusing cluster; while the target state perception features related to train passenger comfort, such as the target state perception features related to a suitable car body temperature, etc., form another second focusing cluster.

[0085] After that, the server determines the feature-focused data of the train state perception data sequence based on the feature-focused data of multiple second focusing clusters. For each second focusing cluster, the representation vector of its cluster center is also calculated, for example, by calculating the mean value of the target state perception features within the focusing cluster. Then, the representation vectors of the cluster centers of these second focusing clusters are fused, and the fusion method can be weighted summation according to specific weights. The final result is the feature-focused data of the train state perception data sequence. This data lays a foundation for subsequently determining the state collaboration chain to be optimized in the train state perception data sequence. The server can further analyze the running state of the train based on this data and make corresponding control decisions.

[0086] In a possible implementation manner, step S132 includes:

[0087] Fusing the representation vectors of the cluster centers of the multiple first focusing clusters to generate the feature-focused data of the candidate block.

[0088] Step S142 includes:

[0089] Fusing the representation vectors of the cluster centers of the multiple second focusing clusters to generate the feature-focused data of the train state perception data sequence.

[0090] In a possible implementation manner, step S131 includes:

[0091] Taking the multiple train state perception blocks as candidate blocks respectively, according to multiple different first clustering numbers, perform feature focusing on the graph encoding vector data of multiple train state perception paths in the candidate blocks to generate first feature focusing data respectively associated with the multiple first clustering numbers, where the first clustering number represents the scale of the first focused cluster in the corresponding first feature focusing data. The feature focusing data of the candidate blocks includes the feature focusing data respectively associated with the multiple first clustering numbers of the candidate blocks.

[0092] Step S130 may further include:

[0093] Step S133, fusing the feature focusing data corresponding to the same first clustering number in the feature focusing data of the multiple train state perception blocks to generate target state perception features respectively corresponding to the multiple first clustering numbers of the candidate train state perception data.

[0094] Step S141 may include: performing feature focusing on the target state perception features corresponding to the same first clustering number in the multiple train state perception data in the train state perception data sequence to generate second feature focusing data respectively corresponding to the multiple first clustering numbers. The feature focusing data of the train state perception data sequence includes the feature focusing data respectively associated with the multiple first clustering numbers of the train state perception data sequence.

[0095] In a possible implementation manner, each second feature focusing data includes feature focusing sub-data respectively associated with N different second clustering numbers, and each feature focusing sub-data includes multiple second focused clusters. The second clustering number represents the scale of the second focused cluster in the corresponding second feature focusing data. The feature focusing data of each first clustering number in the train state perception data sequence includes multiple sub-vector data of the first clustering number in the train state perception data sequence respectively corresponding to the multiple second clustering numbers.

[0096] In this embodiment, taking the traction system of the train as an example of the train state perception block. In this traction system block, first feature focusing data including multiple first focused clusters has been generated by performing feature focusing on the graph encoding vector data of multiple train state perception paths therein. Each first focused cluster has a representation vector of its cluster center, and this representation vector is obtained by performing specific calculations on the graph encoding vector data within the focused cluster. For example, for a first focused cluster containing several graph encoding vector data, it may be to calculate the mean vector of these vector data as the representation vector of the cluster center.

[0097] The server fuses the representation vectors of the cluster centers of each first focused cluster. Suppose there are three first focused clusters in this candidate block of the traction system, namely the first focused cluster related to the normal operation of the motor, the first focused cluster related to the overload protection of the motor, and the first focused cluster related to the starting process of the motor. For the first focused cluster related to the normal operation of the motor, the representation vector of its cluster center contains the mean information of the graph-encoded vector data features such as voltage, current, and magnetic field strength during normal operation; the representation vector of the cluster center of the first focused cluster related to the overload protection of the motor contains the mean information of the graph-encoded vector data features such as special current and temperature during overload; the representation vector of the cluster center of the first focused cluster related to the starting process of the motor contains the mean information of the graph-encoded vector data features such as voltage change and current rise during starting. The server fuses the representation vectors of these three cluster centers according to certain rules, such as weighted summation according to pre-set weights. If the weight of the first focused cluster related to normal operation is set to 0.5, the weight related to overload protection is set to 0.3, and the weight related to the starting process is set to 0.2, then the representation vectors of each cluster center are weighted and summed according to this weight, and the result is the feature focused data of this candidate block of the traction system. This feature focused data synthesizes the key feature information under different operating states of the traction system and can comprehensively reflect the overall state of the traction system.

[0098] Similarly, when the server determines the feature focused data of the train state perception data sequence based on the feature focused data of multiple second focused clusters, taking the target state perception features such as the train speed, braking pressure, and car body temperature in the train state perception data sequence as an example. These target state perception features form multiple second focused clusters through previous clustering operations. Each second focused cluster also has a representation vector of its cluster center, which is obtained by calculating the target state perception features within the focused cluster. For example, for the second focused cluster related to train safety guarantee, the representation vector of its cluster center may be obtained by calculating the mean of the target state perception features such as the reasonable braking pressure range and the speed values within the normal speed limit within the focused cluster; for the second focused cluster related to train passenger comfort, the representation vector of its cluster center may be obtained by calculating the mean of the target state perception features such as the appropriate car body temperature.

[0099] The server fuses the representation vectors of the cluster centers of these second focused clusters. Suppose there are two second focused clusters, the weight of the second focused cluster related to safety assurance is set to 0.6, and the weight related to passenger comfort is set to 0.4. The representation vectors of the cluster centers are weighted and summed according to this weight, and the result obtained is the feature focused data of the train state perception data sequence. This data integrates key information in different aspects during train operation and is of great significance for subsequent analysis of the overall train state and making control decisions.

[0100] When the server takes multiple train state perception chunks as candidate chunks respectively and performs feature focusing on the graph encoding vector data of multiple train state perception paths in the candidate chunks according to multiple different first clustering numbers to generate first feature focused data associated with each of the multiple first clustering numbers, still taking the train state perception chunk of the train's traction system as an example. Suppose three different first clustering numbers are set, which are 5, 8, and 10 respectively.

[0101] For the case where the first clustering number is 5, the server clusters the graph encoding vector data of multiple train state perception paths in the traction system according to a certain clustering algorithm, so as to finally form first feature focused data with a scale of 5 first focused clusters. In this process, the server will gather the data together according to features such as the similarity between the graph encoding vector data to form these 5 first focused clusters. The graph encoding vector data within each first focused cluster has high similarity in some features. For example, it may be the graph encoding vector data related to specific sub - functions or specific working conditions in the traction system that are gathered together.

[0102] For the cases where the first clustering numbers are 8 and 10, the server also performs clustering operations on the graph encoding vector data according to their respective clustering requirements, and obtains first feature focused data with scales of 8 and 10 first focused clusters respectively. In this way, the feature focused data of the traction system candidate chunk includes the feature focused data respectively associated with these three different first clustering numbers (5, 8, and 10). The first feature focused data under each different clustering number reflects the state information of the traction system from different clustering granularities.

[0103] When the server fuses the feature focused data of multiple train state perception chunks to generate the target state perception features of the candidate train state perception data, suppose in addition to the train state perception chunk of the traction system, there are also two train state perception chunks, namely the braking system and the carriage environment control system. Each chunk has its own feature focused data corresponding to different first clustering numbers (such as 5, 8, and 10).

[0104] For the case where the number of the first clusters is 5, the server fuses the feature focus data of the three blocks, namely the traction system, the braking system, and the car body environment control system, when the number of the first clusters is 5. For example, the feature focus data of the traction system when the number of the first clusters is 5 includes some key operating state information of the traction motor, the feature focus data of the braking system under this number of clusters includes key information related to braking such as braking pressure and brake pad wear, and the feature focus data of the car body environment control system under this number of clusters includes environmental information such as car body temperature and humidity. The server fuses the feature focus data of these three blocks when the number of the first clusters is 5 according to certain fusion rules, such as weighted summation or other specific algorithms, to generate the target state perception features of the candidate train state perception data when the number of the first clusters is 5.

[0105] Similarly, for the cases where the number of the first clusters is 8 and 10, the corresponding feature focus data of the three blocks are fused in the same way to generate the target state perception features of the candidate train state perception data when the number of the first clusters is 8 and 10 respectively.

[0106] When the server generates the second feature focus data by performing feature focus on the target state perception features of multiple train state perception data in the train state perception data sequence, taking the target state perception features such as train speed, braking pressure, and car body temperature in the train state perception data sequence as examples, and considering the previously generated target state perception features corresponding to different numbers of the first clusters (5, 8, and 10).

[0107] For the case where the number of the first clusters is 5, the server performs feature focus on the target state perception features corresponding to the number of the first clusters being 5 among the target state perception features such as train speed, braking pressure, and car body temperature. Assume that each target state perception feature contains rich information. For example, the target state perception feature of train speed includes speed information under different sections and different loads, etc. The server performs clustering operations on these target state perception features corresponding to the number of the first clusters being 5 through a clustering algorithm to generate the second feature focus data corresponding to the number of the first clusters being 5. This second feature focus data contains multiple feature focus sub-data related to different numbers of the second clusters, and each feature focus sub-data contains multiple second focus sub-groups. For example, there may be a feature focus sub-data with the number of the second clusters being 3 related to the key parameters of train operation safety, and the three second focus sub-groups in it respectively focus on information such as the safety threshold of train speed, the safety range of braking pressure, and the reasonable interval of car body temperature.

[0108] Similarly, for the cases where the number of the first clusters is 8 and 10, similar feature focusing operations are also performed on the corresponding target state perception features to generate second feature focusing data respectively corresponding to the number of the first clusters being 8 and 10. The feature focusing data for each number of the first clusters (5, 8, and 10) of the train state perception data sequence contains multiple sub-vector data corresponding to their respective multiple numbers of the second clusters. These data reflect the feature information of the train state perception data sequence from different levels and perspectives, providing a rich data basis for subsequently determining the collaborative chain of the states to be optimized of the train state perception data sequence.

[0109] In a possible implementation manner, step S140 further includes:

[0110] Step S143, performing state space mapping on the feature focusing data of the train state perception data sequence to generate state space mapping vector data of the train state perception data sequence.

[0111] Step S144, making a decision on the state space mapping vector data of the train state perception data sequence to generate a collaborative chain of the states to be optimized of the train state perception data sequence.

[0112] In a possible implementation manner, step S144 includes: Based on the train control knowledge expressed in the train control knowledge base, using an artificial intelligence network model to make a decision on the state space mapping vector data of the train state perception data sequence to determine the collaborative chain of the states to be optimized of the train state perception data sequence.

[0113] In this embodiment, taking the previously obtained feature focusing data of the train state perception data sequence as an example, it contains key information integrated from various subsystems of the train (such as the traction system, the braking system, the carriage environment control system, etc.). The server starts to analyze the structure of the feature focusing data to determine the core data components and the auxiliary data components therein. The core data components may be key parameters directly related to the train operation, such as the train speed, the braking pressure, the power output of the traction system, etc.; the auxiliary data components may be data such as the temperature and humidity inside the carriage, which do not directly affect the train operation but have a certain correlation with the overall state. Based on these core data components and auxiliary data components, the server initializes the relevant parameters for state space mapping. For example, for the core data component of the train speed, if the designed speed range of the train is 0 - 200 kilometers per hour, then in the state space mapping, the value range of the speed dimension is set to this interval; for the auxiliary data component of the carriage temperature, if the normal carriage temperature range is 18 - 26 degrees Celsius, the relevant parameter range in the state space mapping is set.

[0114] Next, for each core data component and auxiliary data component in the feature-focused data, perform feature dimension conversion according to pre-set conversion rules. Taking the train speed as an example, if the original speed data is 120 km / h, according to the conversion rules, it may be converted to 0.6 (assuming that the speed value is mapped to a value between 0 and 1, and the proportion corresponding to 120 km / h in the range of 0 - 200 km / h is 0.6). Then the server constructs a state space framework for train state mapping. In this framework, determine the meanings of the coordinate axes of the state space, take the train speed as one coordinate axis, the braking pressure as another coordinate axis, the carriage temperature as the third coordinate axis, etc. At the same time, determine the value range of each coordinate axis. For example, the value range of the speed coordinate axis is 0 - 1 (the converted range), and the value range of the braking pressure coordinate axis is set according to the pressure range of the braking system. Also, take the key operating parameters of train-related equipment, etc. as auxiliary coordinate axes, and define the boundaries and characteristics of different state regions in this state space framework. For example, define the region corresponding to the value ranges of the train speed, braking pressure, and carriage temperature during normal train operation as the normal operation region, and define the region corresponding to the value ranges when the speed is close to the maximum value, the braking pressure is at a high value, and the carriage temperature is abnormal as the danger warning region, etc.

[0115] Finally, the server maps each core data component and auxiliary data component in the feature-focused data to the constructed state space framework according to the corresponding coordinate axes. During the mapping process, determine its position in the state space framework based on the numerical values of the core data components and auxiliary data components and the definitions of the coordinate axes in the state space framework. After the mapping is completed, combine the mapping positions of each core data component and auxiliary data component in the state space framework to generate state space mapping vector data.

[0116] After obtaining the state space mapping vector data of the train state perception data sequence, the server needs to make a decision on it to generate a co-optimized state collaboration chain of the train state perception data sequence. This process is realized by using an artificial intelligence network model based on the train control knowledge expressed in the train control knowledge base.

[0117] The server first obtains relevant train control knowledge from the train control knowledge base. This knowledge base contains information such as train operation rules, safety restrictions, equipment operation logic, and fault handling procedures. For example, the train operation rules stipulate the speed limits in different sections, and the speed should be reduced to a certain value at bends; the safety restrictions clarify the maximum braking pressure of the braking system, the maximum operating speed of the train, etc.; the equipment operation logic describes how the traction system and the braking system work together. For example, when decelerating, first reduce the power output through the traction system, and then start the braking system according to the situation; the fault handling procedure includes the emergency operation procedure when the braking system fails, etc.

[0118] Then, the server combines the state space mapping vector data of the train status perception data sequence with this train control knowledge and inputs it into a pre-trained artificial intelligence network model. This artificial intelligence network model may be a deep neural network model that has been trained with a large amount of train operation data and corresponding control strategy data. Based on the input data, the artificial intelligence network model determines the to-be-optimized state coordination chain of the train status perception data sequence through a series of calculations and inferences. For example, according to the state information such as the current train speed, braking pressure, and car body temperature (represented by the state space mapping vector data), combined with the train control knowledge such as the speed limit and safety requirements on the current section of the train, the model determines how the traction system and braking system of the train should cooperate in the next period of time. If the train is approaching a speed limit section, the model may determine that the traction system should gradually reduce the power output, and the braking system should be ready. When the train is still a certain distance from the speed limit section and the speed is still higher than the limit speed, the braking system starts to apply appropriate braking pressure to ensure that the train can reach the specified speed limit when entering the speed limit section. This is the determined to-be-optimized state coordination chain.

[0119] In a possible implementation manner, using the train control knowledge expressed based on the train control knowledge base, an artificial intelligence network model makes a decision on the state space mapping vector data of the train status perception data sequence to determine the to-be-optimized state coordination chain of the train status perception data sequence, including:

[0120] Step S1441, fusing the state space mapping vector data of the train status perception data sequence and the collaborative guidance vector data of the train status perception data sequence to generate the global vector data of the train status perception data sequence.

[0121] Step S1442, based on the train control knowledge expressed by the train control knowledge base, using the artificial intelligence network model, making a decision on the global vector data of the train status perception data sequence to determine the to-be-optimized state coordination chain of the train status perception data sequence.

[0122] Among them, step S1442 includes:

[0123] Step S1442-1, collecting and organizing various professional knowledge resources in the field of train operation, and using natural language processing technology and knowledge engineering technology to extract a structured train control knowledge base from the various professional knowledge resources. The train control knowledge base includes train operation rules, safety restrictions, equipment operation logic, and fault handling procedures.

[0124] Step S1442-2: Represent each knowledge feature in the extracted train control knowledge base in the form of nodes and edges. After generating the corresponding knowledge network, use a feature selection algorithm to extract the key feature set from the global vector data, where the key feature set reflects the key change trends of the train operation status.

[0125] Step S1442-3: Traverse each key feature in the key feature set, use the relationship network in the knowledge network to find the association information of other entities and attributes associated with this key feature, and based on the association information, introduce context information for the original features in the knowledge network to generate an enhanced feature set, where the context information includes the status of related components, historical fault records, and the impact of control instructions.

[0126] Step S1442-4: Input the enhanced feature set into a pre-trained artificial intelligence network model to output the predicted train state collaboration chain, where the pre-trained artificial intelligence network model incorporates a decision rule set formulated in advance based on the train control knowledge base.

[0127] In this embodiment, taking the state space mapping vector data of the previously obtained train state perception data sequence as an example, it contains the state information of various aspects of the train after state space mapping, such as the mapping values of speed, braking pressure, car body temperature, etc. in the state space. And the collaborative guidance vector data of the train state perception data sequence may contain some specific information for guiding the train operation collaboration, such as in the scenario of multi-train collaborative operation, the guidance information related to the distance between the front and rear trains, or the priority information of the collaborative operation between different subsystems of the train, etc. The server fuses these two types of vector data according to a specific fusion algorithm. For example, the weighted summation method can be adopted, and weights are set according to the importance of the state space mapping vector data and the collaborative guidance vector data in reflecting the overall state of the train. If the state space mapping vector data is more important in reflecting the current self-state of the train, the weight may be set to 0.7, and the weight of the collaborative guidance vector data in reflecting the collaborative relationship between the train and the outside is set to 0.3. Add the corresponding elements of the two types of vector data according to this weight to generate the global vector data of the train state perception data sequence. This global vector data synthesizes the information of both the train's own state and the collaborative relationship, providing a more comprehensive data basis for subsequent decision-making.

[0128] Next, the server needs to make a decision on the global vector data of the train state perception data sequence based on the train control knowledge expressed in the train control knowledge base using the artificial intelligence network model to determine the to-be-optimized state collaboration chain of the train state perception data sequence.

[0129] First, various professional knowledge resources in the field of train operation are collected and sorted, and a structured train control knowledge base is extracted from these resources using natural language processing technology and knowledge engineering technology. The professional knowledge resources in the field of train operation are very rich, including various technical documents, operation manuals, safety specifications, etc. formulated by the railway department. For example, in terms of train operation rules, there are speed limit regulations for different lines. For example, the speed limit at certain bends on mountainous lines is 80 km / h, and the speed limit on straight sections of plain areas is 160 km / h; there are also operation priority rules for different types of trains (such as passenger trains and freight trains). In terms of safety limits, the maximum braking pressure of the train braking system is specified. For example, the maximum braking pressure of a certain type of train braking system is 1000 kPa, and the maximum axle load limit of the train is also specified. The equipment operation logic includes how the traction system adjusts the power output according to instructions, and how the braking system precisely controls the braking pressure according to braking instructions and other operation processes. The fault handling process covers how to switch to standby equipment if the traction motor of the train fails, and how to take emergency braking measures when the braking system fails.

[0130] The server uses natural language processing technology to parse these documents, identify the key information, and then uses knowledge engineering technology to structure this information and build a train control knowledge base. For example, the speed limit rules are classified and stored according to line types, train types, etc.; the safety limit parameters of the braking system are indexed according to equipment models.

[0131] After building the train control knowledge base, each knowledge feature is represented in the form of nodes and edges to generate the corresponding knowledge network. Taking the speed limit in train operation rules as an example, the speed limit values for different lines are used as nodes, and the corresponding relationship between the line type and the speed limit is used as an edge; for the equipment operation logic, the different operation states of the traction system are used as nodes, and the conversion conditions between the operation states are used as edges. Then, a feature selection algorithm is used to extract the key feature set from the global vector data. This feature selection algorithm may be based on principles such as information gain or correlation analysis. For example, in the global vector data, data such as train speed, braking pressure, and the distance from the front and rear trains have a greater impact on the change trend of the train operation state. Through the feature selection algorithm, these features reflecting the key change trends of the train operation state are screened out to form the key feature set.

[0132] Next, the server traverses each key feature in the key feature set and uses the relationship network in the knowledge network to find the association information of other entities and attributes associated with this key feature. Taking the key feature of train speed as an example, entities and attributes related to it are searched in the knowledge network. It may be found that entities and attributes associated with train speed include the slope of the line where the train is located (because the slope affects the control of train speed), the load of the current train (different load has different speed control strategies), etc. According to this association information, context information is introduced for the original features in the knowledge network to generate an enhanced feature set. For example, for the original feature of train speed, in addition to the speed value itself, context information such as the current line slope and train load is added. If the train speed is 120 km / h, the line slope is 5%, and the load is 500 tons, then these information are combined to form a more rich enhanced feature.

[0133] Finally, the enhanced feature set is input into a pre-trained artificial intelligence network model to output the predicted train state coordination chain. This pre-trained artificial intelligence network model incorporates a decision rule set formulated in advance based on the train control knowledge base. When training this artificial intelligence network model, a large amount of train operation historical data is used, which includes train state information, control instructions, and actual operation results, etc. At the same time, according to the rules in the train control knowledge base, such as the braking distance requirements at a specific speed, the traction power limits under different loads, etc., a decision rule set is formulated and incorporated into the training process of the model. When the enhanced feature set is input, the model predicts the future operation state of the train and determines the train state coordination chain based on the existing training results and the incorporated decision rules. For example, based on information such as the current train speed, load, line slope, and the distance between the front and rear trains, combined with knowledge such as speed limits and safety requirements in the train control knowledge base, the model predicts how the traction system of the train should adjust the power output, how the braking system should cooperate, and how to maintain the distance from the front and rear trains in the next period of time, etc., which is the train state coordination chain. This train state coordination chain can guide the intelligent control of the train to ensure the safe and efficient operation of the train.

[0134] In a possible implementation manner, step S143 includes:

[0135] Step S1431, parse the structure of the feature focus data to determine the core data components and auxiliary data components in the feature focus data, and initialize the relevant parameters for state space mapping based on the core data components and auxiliary data components. The relevant parameters include the dimension setting of the mapping space and the initial weights of the mapping algorithm.

[0136] Step S1432: For each core data component and auxiliary data component in the feature-focused data, perform feature dimension transformation according to a preset transformation rule to generate the feature-focused data after feature dimension transformation, where the transformation rule is defined based on the physical characteristics of the train operation state and the internal logical relationship between the train operation state data.

[0137] Step S1433: Construct a state space framework for train state mapping. During the construction process, determine the meanings and value ranges of the coordinate axes of the state space. The coordinate axes are used to reflect each core state factor of the train state. In addition, use the key operation parameters of train-related equipment, etc. as auxiliary coordinate axes, and define the boundaries and characteristics of different state regions in the state space framework.

[0138] Step S1434: Map each core data component and auxiliary data component in the feature-focused data to the constructed state space framework according to the corresponding coordinate axes. During the mapping process, determine the positions of the core data components and auxiliary data components in the state space framework according to the values of the core data components and auxiliary data components and the definitions of the coordinate axes in the state space framework. After the mapping is completed, combine the mapping positions of each core data component and auxiliary data component in the state space framework to generate the state space mapping vector.

[0139] In this embodiment, taking train operation data as an example, the feature-focused data is comprehensive data obtained through a series of previous processes. The core data components may include key parameters that directly affect the train operation state, such as the speed of the train, the braking pressure of the braking system, and the power output of the traction system. For example, the train speed is an important indicator of the train operation state. Its numerical value is directly related to the train's driving process, safety, and the distance control from other trains, etc.; the braking pressure is a key indicator of the effectiveness of the braking system and directly determines the train's deceleration ability; the power output of the traction system affects the train's acceleration and climbing ability, etc. The auxiliary data components may be data such as the temperature and humidity inside the carriage, which do not directly determine the train operation but are somewhat related to the overall state of the train. For example, too high a carriage temperature may affect the passenger comfort and indirectly affect the train operation efficiency.

[0140] Based on these core data components and auxiliary data components, the server initializes the relevant parameters for state space mapping. In terms of setting the dimension of the mapping space, if there are three core data components: train speed, braking pressure, and traction power output, then a three-dimensional mapping space is set. For the train speed, according to the designed speed range of the train, such as 0 - 300 kilometers per hour, the value range corresponding to the dimension in the mapping space is determined; for the braking pressure, it is determined according to the designed pressure range of the braking system, for example, 0 - 1000 kilopascals; for the traction power output, it is determined according to the power range of the traction system, such as 0 - 5000 kilowatts. At the same time, for the initial weights of the mapping algorithm, they are set according to the relative importance of each core data component in reflecting the overall state of the train. Suppose the train speed has the most crucial impact on the train operation state, the weight is set to 0.5, the braking pressure weight is 0.3, and the traction power output weight is 0.2.

[0141] Next, for each core data component and auxiliary data component in the feature-focused data, feature dimension conversion is performed according to the pre-set conversion rules. These conversion rules are defined based on the physical characteristics of the train operation state and the internal logical relationship between the train operation state data. Taking the train speed as an example, if the original train speed data is the actual speed value in kilometers per hour, such as 150 kilometers per hour, and in the state space mapping, it is desired to map the speed to the value range between 0 and 1. According to the train speed design range of 0 - 300 kilometers per hour, through a simple linear conversion rule, 150 kilometers per hour can be converted to 0.5 (i.e., 150 / 300 = 0.5). For the braking pressure, if the original data is the actual pressure value in kilopascals, assumed to be 300 kilopascals, according to the mapping requirement of the braking pressure in the state space, based on its value range of 0 - 1000 kilopascals, it is converted to 0.3 (i.e., 300 / 1000 = 0.3). A similar method is used for the traction power output conversion. Suppose the original power is 1500 kilowatts, according to the value range of 0 - 5000 kilowatts, it is converted to 0.3 (i.e., 1500 / 5000 = 0.3). For the auxiliary data component such as the carriage temperature, if the original temperature is 25 degrees Celsius, and the mapping range of the carriage temperature in the state space is 18 - 28 degrees Celsius, it is converted to a value between 0 and 1 through the corresponding conversion rules. Suppose it is converted to 0.7 (the calculation method is based on the pre-set conversion rules related to the 18 - 28 degrees Celsius range). In this way, the feature-focused data after feature dimension conversion is generated.

[0142] Then, the server constructs a state space framework for train state mapping. During the construction process, the meanings and value ranges of the coordinate axes of the state space are determined. The core state factors of the train are used as the coordinate axes. For example, the train speed, braking pressure, and traction power output mentioned earlier are used as the three coordinate axes respectively. The meaning of the train speed coordinate axis is the running speed of the train, and the value range is 0 - 1 (the converted range) determined previously; the meaning of the braking pressure coordinate axis is the pressure exerted by the braking system, and the value range is 0 - 1; the meaning of the traction power output coordinate axis is the power provided by the traction system, and the value range is 0 - 1. At the same time, the key operating parameters of train-related equipment, etc., are used as auxiliary coordinate axes. For example, the carriage temperature coordinate axis, whose meaning is the temperature condition inside the carriage, and the value range is 0 - 1 after conversion. And the boundaries and characteristics of different state regions are defined in this state space framework. For example, the region corresponding to the value ranges of the train speed, braking pressure, traction power output, and carriage temperature during normal operation of the train is defined as the normal operation region. Suppose the range corresponding to the train speed during normal operation is 0.3 - 0.7, the braking pressure range is 0.2 - 0.6, the traction power output range is 0.1 - 0.5, and the carriage temperature range is 0.4 - 0.8. Then, in this state space framework, the region enclosed by these value ranges is the normal operation region. When the train speed approaches 1, the braking pressure approaches 1, and the traction power output approaches 0, it may be defined as the emergency braking region, which indicates that the train is performing high-intensity braking operations.

[0143] Finally, the server maps each core data component and auxiliary data component in the feature focus data to the constructed state space framework according to the corresponding coordinate axes. During the mapping process, the position in the state space framework is determined based on the numerical values of the core data components and auxiliary data components and the definitions of the coordinate axes in the state space framework. For example, for the previously converted train speed value of 0.5, it is at the 0.5 position on the train speed coordinate axis; the braking pressure value of 0.3 is at the 0.3 position on the braking pressure coordinate axis; the traction power output value of 0.3 is at the 0.3 position on the traction power output coordinate axis; the carriage temperature value of 0.7 is at the 0.7 position on the carriage temperature coordinate axis. After the mapping is completed, the mapping positions of each core data component and auxiliary data component in the state space framework are combined to generate a state space mapping vector. This state space mapping vector can be understood as a specific coordinate point of the current state of the train in this constructed state space framework, which synthesizes the state information of all aspects of the train and provides an important data basis for subsequent decision-making based on the train state, such as determining the collaborative chain of the states to be optimized in the train state perception data sequence, etc.

[0144] Figure 2The figure shows the hardware structure diagram of the artificial intelligence-based train operation intelligent control system 100 provided by the embodiments of the present invention for implementing the above-mentioned artificial intelligence-based train operation intelligent control method, as Figure 2 shown, the artificial intelligence-based train operation intelligent control system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0145] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store the data and / or instructions used by the artificial intelligence-based train operation intelligent control system 100 to execute or use to complete the exemplary methods described in the present invention.

[0146] In a specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the artificial intelligence-based train operation intelligent control method of the above method embodiment. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the transceiver actions of the communication unit 140.

[0147] For the specific implementation process of the processor 110, reference may be made to the various method embodiments executed by the artificial intelligence-based train operation intelligent control system 100 above. The implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0148] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based train operation intelligent control method is implemented.

[0149] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. An intelligent train operation control method based on artificial intelligence, characterized in that: The method comprises: taking multiple train state perception data in the train state perception data sequence of the target train as candidate train state perception data respectively, dividing the candidate train state perception data into blocks to generate multiple train state perception blocks, each of the train state perception blocks including multiple train state perception paths; Performing graph encoding on the multiple train state perception paths respectively to generate graph encoding vector data of the multiple train state perception paths; The multiple train state perception blocks are respectively used as candidate blocks, feature focusing is performed on the graph coded vector data of multiple train state perception paths in the candidate blocks to generate feature focused data of the candidate blocks, and the feature focused data of the multiple train state perception blocks are fused to generate target state perception features of the candidate train state perception data; Feature focusing is performed on target state perception features of multiple train state perception data in the train state perception data sequence to generate feature focused data of the train state perception data sequence, and a state coordination chain to be optimized of the train state perception data sequence is determined according to the feature focused data of the train state perception data sequence; The target train is intelligently controlled based on the coordinated chain of states to be optimized of the train state perception data sequence.

2. The train operation intelligent control method based on artificial intelligence according to claim 1 is characterized in that: The method of taking the plurality of train state perception blocks as candidate blocks respectively, performing feature focusing on the graph coded vector data of the plurality of train state perception paths in the candidate blocks, and generating feature focused data of the candidate blocks includes: The plurality of train state perception blocks are respectively used as candidate blocks, and feature focusing is performed on the graph encoding vector data of the plurality of train state perception paths in the candidate blocks to generate first feature focusing data, wherein the first feature focusing data includes a plurality of first focusing clusters; Determine the feature focusing data of the candidate block based on the feature focusing data of the plurality of first focusing clusters; The step of performing feature focusing on target state perception features of a plurality of train state perception data in the train state perception data sequence to generate feature focused data of the train state perception data sequence includes: Performing feature focusing on target state perception features of a plurality of train state perception data in the train state perception data sequence to generate second feature focused data, wherein the second feature focused data includes a plurality of second focused subgroups; The feature focused data of the train state perception data sequence is determined based on the feature focused data of the plurality of second focused clusters.

3. The train operation intelligent control method based on artificial intelligence according to claim 2 is characterized in that: The determining the feature focusing data of the candidate blocks based on the feature focusing data of the plurality of first focusing clusters comprises: Fusing the representation vectors of the cluster centers of the multiple first focused clusters to generate feature focused data of the candidate blocks; The determining the feature focused data of the train state perception data sequence based on the feature focused data of the plurality of second focused subgroups comprises: The characterization vectors of the cluster centers of the multiple second focused clusters are fused to generate feature focused data of the train state perception data sequence.

4. The train operation intelligent control method based on artificial intelligence according to claim 2 is characterized in that: The method of taking the plurality of train state perception blocks as candidate blocks respectively and performing feature focusing on the graph coded vector data of the plurality of train state perception paths in the candidate blocks to generate first feature focused data comprises: The plurality of train state perception blocks are respectively used as candidate blocks, and according to a plurality of different first clustering quantities, the graph encoding vector data of the plurality of train state perception paths in the candidate blocks are feature focused to generate first feature focused data associated with each of the plurality of first clustering quantities, wherein the first clustering quantity represents the size of the first focused cluster in the corresponding first feature focused data; the feature focused data of the candidate blocks include the feature focused data associated with each of the plurality of first clustering quantities of the candidate blocks; The step of fusing the feature focused data of the plurality of train state perception blocks to generate the target state perception features of the candidate train state perception data includes: Fusion of feature focused data corresponding to the same first cluster quantity in the feature focused data of the plurality of train state perception blocks to generate target state perception features of the candidate train state perception data corresponding to a plurality of first cluster quantities respectively; The step of performing feature focusing on target state perception features of a plurality of train state perception data in the train state perception data sequence to generate second feature focused data includes: Among the target state perception features of multiple train state perception data in the train state perception data sequence, feature focusing is performed on the target state perception features corresponding to the same first cluster quantity to generate second feature focused data corresponding to multiple first cluster quantities respectively; the feature focused data of the train state perception data sequence includes feature focused data associated with each of the multiple first cluster quantities of the train state perception data sequence.

5. The train operation intelligent control method based on artificial intelligence according to claim 3 is characterized in that: Each second feature focused data includes feature focused sub-data associated with N different second cluster numbers, each feature focused sub-data includes a plurality of second focused sub-clusters; the second cluster number represents the size of the second focused sub-clusters in the corresponding second feature focused data; The feature focused data of each first cluster number of the train state perception data sequence includes a plurality of sub-vector data of the first cluster number of the train state perception data sequence respectively corresponding to a plurality of second cluster numbers.

6. The train operation intelligent control method based on artificial intelligence according to any one of claims 1 to 5, characterized in that: The step of determining the state coordination chain to be optimized of the train state perception data sequence according to the feature focused data of the train state perception data sequence comprises: Performing state-space mapping on the feature-focused data of the train state perception data sequence to generate state-space mapping vector data of the train state perception data sequence; A decision is made on the state space mapping vector data of the train state perception data sequence to generate a state coordination chain to be optimized of the train state perception data sequence.

7. The train operation intelligent control method based on artificial intelligence according to claim 6 is characterized in that: The making a decision on the state space mapping vector data of the train state perception data sequence to generate a state coordination chain to be optimized of the train state perception data sequence includes: Based on the train control knowledge expressed by the train control knowledge base, an artificial intelligence network model is used to make decisions on the state space mapping vector data of the train state perception data sequence, and determine the state coordination chain to be optimized of the train state perception data sequence.

8. The train operation intelligent control method based on artificial intelligence according to claim 7 is characterized in that: The train control knowledge expressed by the train control knowledge base is used to make a decision on the state space mapping vector data of the train state perception data sequence by using an artificial intelligence network model to determine the state coordination chain to be optimized of the train state perception data sequence, including: fusing the state space mapping vector data of the train state perception data sequence and the collaborative guidance vector data of the train state perception data sequence to generate global vector data of the train state perception data sequence; Based on the train control knowledge expressed by the train control knowledge base, using an artificial intelligence network model, making decisions on the global vector data of the train state perception data sequence, and determining a state coordination chain to be optimized of the train state perception data sequence; Among them, the step of making a decision on the global vector data of the train state perception data sequence based on the train control knowledge expressed by the train control knowledge base and determining the state coordination chain to be optimized of the train state perception data sequence by using an artificial intelligence network model includes: Collect and organize various professional knowledge resources in the field of train operation, and use natural language processing technology and knowledge engineering technology to extract a structured train control knowledge base from the various professional knowledge resources. The train control knowledge base includes train operation rules, safety restrictions, equipment operation logic, and fault handling procedures; Representing each knowledge feature in the extracted train control knowledge base in the form of nodes and edges, generating a corresponding knowledge network, and extracting a key feature set in the global vector data using a feature selection algorithm, wherein the key feature set reflects a key change trend of the train running state; Traversing each key feature in the key feature set, using the relationship network in the knowledge network, searching for association information of other entities and attributes associated with the key feature, and introducing context information for the original features in the knowledge network according to the association information to generate an enhanced feature set, wherein the context information includes the status of related components, historical fault records, and the impact of control instructions; The enhanced feature set is input into a pre-trained artificial intelligence network model to output a predicted train state coordination chain, wherein the pre-trained artificial intelligence network model incorporates a decision rule set pre-developed based on the train control knowledge base.

9. The train operation intelligent control method based on artificial intelligence according to claim 6 is characterized in that: The performing state space mapping on the feature focused data of the train state perception data sequence to generate state space mapping vector data of the train state perception data sequence includes: Parsing the structure of the feature-focused data, determining the core data component and the auxiliary data component in the feature-focused data, and initializing relevant parameters for state-space mapping based on the core data component and the auxiliary data component, wherein the relevant parameters include the dimension setting of the mapping space and the initial weight of the mapping algorithm; For each core data component and auxiliary data component in the feature focused data, feature dimension conversion is performed according to a preset conversion rule to generate feature focused data after feature dimension conversion, wherein the conversion rule is defined based on the physical characteristics of the train running state and the inherent logical relationship between the train running state data; Constructing a state space framework for train state mapping, during which the meaning and value range of the coordinate axes of the state space are determined, the coordinate axes are used to reflect the core state factors of the train state, and the key operating parameters of the train-related equipment are used as auxiliary coordinate axes, and the boundaries and characteristics of different state areas are defined in the state space framework; Each core data component and auxiliary data component in the feature focused data is mapped to the constructed state space framework according to the corresponding coordinate axes. During the mapping process, the positions of the core data components and the auxiliary data components in the state space framework are determined according to the numerical values ​​of the core data components and the auxiliary data components and the definitions of the coordinate axes in the state space framework. After the mapping is completed, the mapping positions of each core data component and the auxiliary data component in the state space framework are combined to generate the state space mapping vector.

10. An intelligent train operation control system based on artificial intelligence, characterized in that: The artificial intelligence-based train operation intelligent control system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based train operation intelligent control method described in any one of claims 1 to 9 above.

Images