Railway infrastructure maintenance resource and plan associated scheduling method and system

Abstract

The invention relates to the technical field of railways and rail transit, and provides a railway infrastructure maintenance resource and plan associated scheduling method and system, and the method comprises the following steps: S1, analyzing the maintenance plan compilation optimization and maintenance resource scheduling problems of a railway service depot, and summarizing the problems as maintenance resources and maintenance resource combined scheduling problems; s2, integrating the organization structure, the disease type, the maintenance resource information and the maintenance plan information of the railway work depot, designing a combined dispatching system database of the maintenance resources and plans of the railway infrastructure, and developing a data interface in a railway work management system; s3, on the basis of actual requirements, a software framework and a system interface of the combined dispatching system for the maintenance and repair resources and plans of the railway infrastructure are developed, and the system interface comprises a resource management module, a resource configuration module and an intelligent work order module; according to the invention, association scheduling can be well carried out.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of railway and rail transit technology, in particular to a railway infrastructure maintenance resource and plan association scheduling method and system. BACKGROUND

[0002] Maintenance of railway infrastructure is the core work of railway maintenance section, which is directly related to railway safety, efficiency and economic benefit, and is indispensable for railway operation. The current railway maintenance plan of the maintenance section relies on manual preparation, and it is difficult to accurately assess the resources (such as personnel, equipment and materials) required for each maintenance task. This may lead to insufficient resources for some tasks, delay the construction period, or excessive resources, resulting in waste. In addition, the maintenance plan and resource scheduling involve detection data, inventory information, personnel scheduling and other aspects, but the manual method is difficult to realize real-time sharing of cross-department information, leading to delayed or inaccurate scheduling decisions. When manually determining the priority of maintenance tasks, it relies on experience and may overlook the potential risks of some key equipment, or allocate too many resources to secondary tasks, affecting overall efficiency. When multiple maintenance tasks are carried out at the same time, manual scheduling is difficult to effectively coordinate the resource demand between tasks, and problems such as repeated allocation of personnel or equipment and congestion at the work site are prone to occur. These problems make the maintenance plan and resource scheduling prepared manually lack efficient linkage, which restricts the overall effectiveness of the maintenance work. With the development of information technology, related technologies have gradually increased in the field of railways. For example, a railway maintenance site operation monitoring and command method and system can generate and adjust the operation process according to the project plan, guide the site operation, and monitor the operation execution in real time, realizing flexible configuration of operation process and automatic adjustment of special circumstances. In addition, there is a virtual assistant system for railway bridge inspection and maintenance, which provides a convenient information query and problem solving service for users through the architecture of data layer, business layer and application layer using knowledge graph. At the same time, there is a method for formulating a comprehensive maintenance scheme for high-speed railways, which acquires high-speed railway network data, establishes a site selection collaborative optimization model, divides feasible maintenance schemes, and selects the optimal scheme according to the evaluation system, thereby improving the maintenance efficiency and resource utilization rate.

[0003] However, the existing railway infrastructure maintenance plan and resource scheduling system has the following problems: lack of resource allocation refinement, low data integration and resource utilization rate, limited business coverage, lack of real-time performance, limited operation process adaptability, and insufficient system integration.

[0004] For example, in the method for formulating a comprehensive maintenance scheme for high-speed railways disclosed in CN110210631B, the following problems exist:

[0005] Insufficient resource allocation refinement: Although this method divides the feasible maintenance scheme of the comprehensive maintenance section by establishing a site collaborative optimization model, and uses the TOPSIS theory combined with gray correlation degree analysis to screen the optimal scheme, the refinement degree of resource allocation in actual operation still needs to be improved. For example, for the specific maintenance needs and priorities of different line points, the existing model may not be able to make real-time dynamic adjustments, leading to unreasonable resource allocation, and there may be resource waste or delayed maintenance of key line points.

[0006] Low data integration and resource utilization: The comprehensive analysis and mining of historical maintenance data, real-time operation data, etc. is not enough, and the potential value of data cannot be fully utilized to further optimize maintenance decisions. The joint scheduling of maintenance plans and maintenance resources in each workshop of the maintenance section is not considered. The maintenance plan may not be well prepared.

[0007] For example, in a railway bridge inspection and maintenance virtual assistant system with publication number CN110543951B, the following problems exist:

[0008] Limited business coverage: This system mainly focuses on railway bridge inspection and maintenance information query, question answering and information automatic pushing, and has limited coverage of on-site operation monitoring, maintenance scheme development and resource allocation for railway bridge. It cannot meet the all-round needs of railway bridge inspection and maintenance work, and it is difficult to form a complete business closed loop. It does not involve the preparation of inspection and maintenance plans, but only provides information retrieval and pushing as an auxiliary function.

[0009] Lack of real-time performance: It cannot obtain the current state and operation progress of the bridge in real time, nor can it adjust the pushing content and information service in a timely manner according to the on-site situation. This leads to certain limitations in the timeliness and practicality of the system in actual application.

[0010] For example, in a railway maintenance site operation monitoring and command method and system with publication number CN110910048B, the following problems exist:

[0011] Limited adaptability of operation process: Although this method realizes flexible configuration of operation process and automatic adjustment of special cases, its adaptability is still limited for complex and variable railway maintenance site operation environment. It does not jointly schedule plans and resources before optimizing operation. When operation preparation and resource allocation are unreasonable, the system may not be able to provide accurate and effective adjustment schemes, leading to interruption or reduction of operation efficiency.

[0012] Insufficient system collaboration: This system mainly focuses on the monitoring and command of on-site operation, and has poor collaboration with other railway management systems (such as maintenance scheme development system, resource scheduling system, etc.). Data sharing and collaborative work between systems are difficult, forming an information island, which affects the improvement of overall management efficiency of railway maintenance.

[0013] In view of the above, there is an urgent need for a railway infrastructure maintenance resource and plan joint scheduling method and system to solve the above problems. SUMMARY

[0014] The content of the present application is to provide a railway infrastructure maintenance resource and plan joint scheduling method and system, which can overcome some or some defects of the prior art.

[0015] According to the railway infrastructure maintenance resource and plan joint scheduling method of the present application, it comprises the following steps:

[0016] Step S1, analyze the railway maintenance plan optimization and maintenance resource scheduling problem of the maintenance section, and summarize it as a maintenance resource and maintenance resource joint scheduling problem;

[0017] Step S2, comprehensively analyze the organizational structure, disease type, maintenance resource information and maintenance plan information of the railway maintenance section, design the database of the railway infrastructure maintenance resource and plan joint scheduling system, and develop the data interface in the railway maintenance management system;

[0018] Step S3, based on the actual demand, develop the software framework and system interface of the railway infrastructure maintenance resource and plan joint scheduling system, including the resource management module, the resource configuration module and the intelligent work order module;

[0019] The resource management module manages according to the organizational structure of the maintenance section department, integrates the information of personnel and machines, and constitutes a resource pool;

[0020] The resource configuration module jointly schedules the maintenance resource and the maintenance plan according to the maintenance resource and the maintenance plan list, combines the deep reinforcement learning algorithm, automatically realizes the optimal maintenance plan sorting and maintenance resource allocation;

[0021] The intelligent work order module realizes the work order management and tracking, realizes the unified management of the maintenance work order and synchronously updates the maintenance resource information according to the state of the work order.

[0022] As a preferred, in the resource management module, the organizational structure of the maintenance section department is maintenance section-workshop-work area-team; the resource management module can maintain and display the information of the personnel in the organization, and also allows users to quickly retrieve personnel information; the resource management module has personnel information management function and machine information management function, the personnel information management function supports the add, delete and modify operation of information, and ensures the real-time update of personnel data; the machine information management function standardizes the borrowing and returning process of the machine, and automatically updates the machine state; the resource management module allows users to quickly retrieve and update machine information.

[0023] As preferred, the resource configuration module comprises a maintenance resource query function, a resource configuration function, and a backtracking function; the resource configuration function models the maintenance plan and the maintenance resource into an environment state model based on a flexible workshop scheduling model, and realizes joint scheduling of the resource and the maintenance resource in combination with a graph neural network.

[0024] As preferred, the maintenance plan and the maintenance resource are modeled into the environment state model, and a Markov decision process (MDP) modeling is introduced.

[0025] As preferred, the Markov decision process (MDP) comprises the following aspects:

[0026] State: The state is a description of the environment, which is the basis for the agent to make decisions; the set of all states is called the state space, denoted by S;

[0027] Action: The action is a description of the behavior of the agent, which is the result of the agent's decision; the set of all possible actions is called the action space, denoted by A;

[0028] Policy: The policy is a conditional probability distribution of actions given states, which is used to guide the agent to choose actions in different states, denoted by π(a|s);

[0029] State transition probability: After performing an action in a certain state, the probability of the environment transitioning to the next state is denoted by P(s ′ |s,a), where s is the current state, a is the action performed, and s' is the next state;

[0030] Reward function: The immediate feedback given by the environment after the agent performs an action in a certain state is denoted by R(s,a);

[0031] Discount factor: A measure of the importance of future rewards, usually between 0 and 1, denoted by γ;

[0032] State value function: The state value function V π (s) represents the expected cumulative return that can be obtained in state s by following the policy π;

[0033] Action value function: The action value function Q(s,a) represents the expected cumulative return that can be obtained by performing action a in state s and then following the policy π.

[0034] As preferred, a heterogeneous graph is used to represent the relationship between processes and device resources, and a graph neural network is used to extract the relationship features, and the process nodes and device resource nodes are embedded in stages, and finally a state representation of uniform size is output; the graph neural network extracts the relationship features in the heterogeneous graph through the following steps:

[0035] 1) Feature embedding: First, feature embedding is performed on the process nodes and equipment resource nodes respectively;

[0036] 2) Graph attention mechanism: the influence between different nodes is quantified using the graph attention mechanism; by calculating the attention coefficient and normalizing it using the Softmax function, the new features of the nodes are obtained;

[0037] The calculation method of the graph attention mechanism is as follows:

[0038] A) Attention coefficient calculation:

[0039] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0040] Where e ij is the attention coefficient between node i and node j, a is a learnable parameter vector, W is a weight matrix, h i and h j are the feature vectors of nodes i and j respectively;

[0041] B) Attention coefficient normalization:

[0042]

[0043] Where α ij is the normalized attention coefficient, N i is the neighbor node set of node i;

[0044] C) Node feature update:

[0045]

[0046] Where h ′ i is the updated feature vector of node i, and σ is the activation function.

[0047] As a preferred, the output of the graph neural network is taken as the input of the policy network and the value network, the policy network adjusts the policy according to the feedback provided by the value network, and the value network updates the value estimate according to the reward feedback from the environment; at the same time, the proximal policy optimization PPO algorithm is used to stabilize the training of the entire decision network, and PPO is as follows:

[0048] Clipping mechanism of policy update: PPO limits the amplitude of policy update by clipping the policy probability ratio; the expression of the policy probability ratio r t (θ) is as follows:

[0049]

[0050] where π θ (a t | s t ) is the probability of the new policy selecting action a t in state s t , is the probability of the old policy; PPO clips this ratio to keep it within [1-∈,1+∈];

[0051] Estimation of advantage function: in PPO, the estimation of advantage function A t is realized by generalized advantage estimation GAE, the formula is as follows:

[0052]

[0053] where γ is the discount factor, λ is the parameter of GAE, used to control the trade-off between the bias and variance of the estimated advantage function, is the time series difference TD residual at time step t+l;

[0054] Update of policy network and value network: PPO uses the clipped objective function to update the policy network, and uses the mean square error MSE loss function to update the value network; specifically, the update of the policy network is realized by maximizing the following objective function L CLIP (θ) as follows:

[0055]

[0056] where represents the expected value calculation at time step t;

[0057] The update of the value network is realized by minimizing the following loss function L VF as follows:

[0058]

[0059] where represents the size of the training data set of the kth batch, T represents the length of each trajectory, τ represents a complete trajectory, V φ (s t ) is the value estimation of the value network for state s t , is the actual return estimation;

[0060] Complete objective function L PPO : the formula is as follows:

[0061]

[0062] wherein c1 and c2 are coefficients that control the weight of each part, is the entropy of the policy, which is used to measure the uncertainty of the policy, and increasing the entropy can improve the exploration ability of the policy.

[0063] As preferred, the intelligent work order module has a work order information query function, which provides users with detailed viewing and management of current and historical work orders; the work order information query function allows users to display and filter the current work order list, and retrieve according to the work order number, assigner, and creation time conditions; users can feed back on the processed work orders, and confirm the completion status of the work orders through the processed button.

[0064] The present application provides a railway infrastructure maintenance resource and plan associated scheduling system, which adopts the above-mentioned railway infrastructure maintenance resource and plan associated scheduling method, and comprises a data layer, a business layer and an application layer;

[0065] The data layer and the data interface in the railway maintenance management system are connected to obtain maintenance resource data, maintenance plan data and maintenance work order data; the business layer provides query, management of maintenance resources and maintenance plans, joint scheduling function of maintenance list and maintenance plan; through the deep reinforcement learning engine, the resource data and the maintenance guide manual are combined to automatically adjust the execution order of the maintenance plan and the corresponding resource allocation, and the resource data is automatically updated according to the execution state of the maintenance plan and the maintenance work order, so as to realize the optimal scheduling of the maintenance plan and the maintenance resource; the application layer is used to provide query and display functions based on the maintenance plan, the maintenance resource and the maintenance work order.

[0066] The beneficial effects of the present application are as follows:

[0067] I. Dynamic optimization of resources solves the problem of insufficient allocation refinement: through the flexible workshop scheduling model, the maintenance tasks are mapped as "processes", and the resource pool (personnel, equipment) is modeled as "machines", and through the DRL agent, the resource state (such as equipment state, personnel idle rate) is perceived in real time, and the task execution order and resource allocation are dynamically adjusted. According to the indicators such as rail damage grade and window period limit, resources are automatically allocated to high-risk tasks to avoid priority misjudgment caused by traditional manual experience.

[0068] II. Data fusion improves collaborative decision-making capability: through dynamic integration of the data layer, the maintenance plan, resource inventory, and work order execution state data are integrated through unified interface, a cross-department resource pool is constructed, and information barriers are eliminated. Through real-time feedback of the intelligent work order, the work order state (such as completion, delay) will automatically trigger the dynamic adjustment of the resource configuration module, forming a "plan-execution-feedback" closed loop, and solving the problem of data update lag in traditional systems.

[0069] III. Automated scheduling, enhanced adaptability to complex scenarios: Through the DRL autonomous decision engine, based on Markov Decision Process (MDP), using reward function to drive policy optimization, automatically realize the optimal maintenance plan scheduling and maintenance resource allocation. Use Graph Neural Network (GNN) to model relationship constraints, dynamically avoid resource conflicts (such as device resource overlap in parallel tasks).

[0070] IV. Modular system architecture, full coverage of business: Through the resource management module (step S4): Structuring resource data by hierarchy, facilitating information filtering and management. Through the intelligent work order module: Unified management of work order life cycle (creation, assignment, feedback), and real-time synchronization with resource status, forming full-process coverage from plan compilation to execution tracking, breaking through the limitations of traditional system function fragmentation. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 A flowchart of a railway infrastructure maintenance resource and plan associated scheduling method in the embodiment. DETAILED DESCRIPTION

[0072] For a further understanding of the present application, reference will be made to the following description taken in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely explanatory of the present application and are not limiting.

[0073] Embodiments

[0074] As shown in the figure, the present embodiment provides a railway infrastructure maintenance resource and plan associated scheduling method, which includes the following steps: Figure 1 Step S1, analyze the railway maintenance plan compilation optimization and maintenance resource scheduling problem of the railway maintenance section, and summarize it as a joint scheduling problem of maintenance resources and maintenance resources;

[0075] Step S2, comprehensively analyze the organizational structure, disease type, maintenance resource information and maintenance plan information of the railway maintenance section, design a joint scheduling system database of railway infrastructure maintenance resources and plans, and develop a data interface in the railway maintenance management system;

[0076] Step S3, based on actual demand, develop the software framework and system interface of the joint scheduling system of railway infrastructure maintenance resources and plans, including resource management module, resource configuration module and intelligent work order module;

[0077]

[0078] ​The resource management module manages according to the department organization structure of the maintenance section, integrates the information of personnel and machines, and constitutes a resource pool; the department organization structure of the maintenance section is maintenance section-workshop-work area-team; the resource management module can maintain and display the information of personnel within the organization, and also allows users to quickly search for personnel information; the resource management module has personnel information management function and machine information management function, the personnel information management function supports information adding, deleting and modifying operations, and ensures real-time updating of personnel data; the machine information management function standardizes the borrowing and returning process of machines, and automatically updates the machine state; the resource management module allows users to quickly search and update machine information, including name, category, license plate number and belonging workshop, etc. The machine borrowing and returning management function standardizes the borrowing and returning process of machines, and automatically updates the machine state. The feedback mechanism after using machines ensures timely maintenance of machines.

[0079] The resource configuration module combines deep reinforcement learning algorithm to jointly schedule maintenance resources and maintenance plans according to the maintenance resources and maintenance plan list, automatically realizes the optimal maintenance plan sorting and maintenance resource allocation; the resource configuration module includes maintenance resource query function, resource configuration function and backtracking function; the resource configuration function models the maintenance plan and maintenance resource as an environment state model based on the flexible job shop scheduling model, and realizes the joint scheduling of resources and maintenance resources by combining graph neural network.

[0080] The intelligent work order module realizes work order management and tracking, realizes unified management of maintenance work orders and synchronous updating of maintenance resource information according to the state of work order. The intelligent work order module has work order information query function, which provides users with detailed viewing and management of current and historical work orders; the work order information query function allows users to display and filter the current work order list, and searches according to the work order number, assigner and creation time conditions; users can feed back the processed work order, and confirm the completion state of the work order through the "processed" button.

[0081] Railway maintenance scheduling problem (RMSP) has an essential mapping relationship with flexible job-shop scheduling (FJSP) in structural characteristics. FJSP, as a classic scheduling model, realizes multi-dimensional optimization objectives (such as minimizing the completion time, balancing equipment load, etc.) through the flexibility of machine selection and the flexibility of timing scheduling. Its core elements can accurately correspond to the scheduling requirements in the maintenance scene: each maintenance plan can be abstracted as a job in FJSP, the sequence chain of the plan corresponds to the operation with strict timing constraints in FJSP, and the equipment resource is mapped to the machine in FJSP. In the FJSP framework, each operation can choose an execution unit from any idle equipment in the machine group, and the same equipment can only serve a single operation, which is completely consistent with the exclusive use rule of the same tool in the maintenance scene. We transform the maintenance scheduling problem into a process of step-by-step decision-making to gradually approach the global optimal solution. The core idea is to decompose the complex global scheduling scheme into a series of simple local decisions: starting from the initial state where all operations are unscheduled, the system selects an executable operation and allocates appropriate equipment resources for it at each step based on the current progress (such as completed operations, equipment occupancy, etc.), gradually builds a complete scheduling scheme, and finally arranges all operations. To formally describe this step-by-step decision-making mechanism, we introduce Markov decision process (MDP) modeling, which accurately corresponds to the core elements of MDP in each step of the maintenance scheduling decision-making process. MDP includes the following aspects:

[0082] State: State is a description of the environment, which is the basis for the agent to make decisions; the set of all states is called the state space, denoted by S;

[0083] Action: Action is a description of the agent's behavior, which is the result of the agent's decision; the set of all possible actions is called the action space, denoted by A;

[0084] Policy: Policy is given by the conditional probability distribution of actions according to states, which is used to guide the agent to choose actions in different states, denoted by π(a|s);

[0085] State transition probability: After performing a certain action in a certain state, the environment transitions to the next state with a certain probability, denoted by P(s'|s,a), where s is the current state, a is the executed action, and s' is the next state;

[0086] Reward function: The immediate feedback given by the environment after the agent performs a certain action in a certain state, denoted by R(s,a);

[0087] Discount factor: Used to measure the importance of future rewards, usually between 0 and 1, denoted by γ;

[0088] State value function: State value function V π (s) represents the expected cumulative reward that can be obtained by acting according to policy π in state s;

[0089] Action value function: The action value function Q(s,a) represents the expected cumulative reward that can be obtained by performing action a in state s and then acting according to policy π.

[0090] The MDP modeling of maintenance scheduling can achieve dynamic interaction through a reinforcement learning framework. The agent, based on the current state S... t Select action A t The environment generates a new state S based on the state transition rule P(s'|s,a). t+1 And return reward R t Ultimately, the optimal global scheduling scheme is approximated through strategy optimization.

[0091] Heterogeneous graphs are used to represent the relationships between processes and equipment resources. Graph neural networks are used to extract the relationship features, embedding features into process nodes and equipment resource nodes in stages, and finally outputting a state representation of uniform size. The graph neural network extracts the relationship features from the heterogeneous graph through the following steps:

[0092] 1) Feature embedding: First, feature embedding is performed on process nodes and equipment resource nodes respectively;

[0093] 2) Graph Attention Mechanism: The graph attention mechanism is used to quantify the influence between different nodes; by calculating the attention coefficient and normalizing it using the Softmax function, new features of the nodes are obtained;

[0094] The calculation method for the graph attention mechanism is as follows:

[0095] A) Attention coefficient calculation:

[0096] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0097] Among them, e ij Here, is the attention coefficient between node i and node j, a is the learnable parameter vector, W is the weight matrix, and h is the weight matrix. i and h j These are the feature vectors of node i and node j, respectively;

[0098] B) Attention coefficient normalization:

[0099]

[0100] where α ij is the normalized attention coefficient, N i is the neighbor node set of node i;

[0101] C) Node feature update:

[0102]

[0103] where h ′ i is the updated feature vector of node i, and σ is the activation function.

[0104] The output of the graph neural network is used as the input of the policy network and the value network. The policy network adjusts the policy according to the feedback provided by the value network, while the value network updates the value estimate according to the reward feedback from the environment. At the same time, the proximal policy optimization (PPO) algorithm is used to stabilize the training of the entire decision network. PPO effectively avoids large fluctuations in the training process by limiting the magnitude of policy updates, ensuring smooth transition of policy updates. PPO is as follows:

[0105] Clipping mechanism of policy update: PPO limits the magnitude of policy updates by clipping the policy probability ratio. The policy probability ratio r t (θ) is expressed as follows:

[0106]

[0107] where π θ (a t |s t ) is the probability of the new policy selecting action a t in state s t , is the corresponding probability of the old policy. PPO clips this ratio to keep it within the interval [1-∈,1+∈];

[0108] Estimation of advantage function: In PPO, the estimation of the advantage function A t is realized by generalized advantage estimation (GAE). The formula is as follows:

[0109]

[0110] where γ is the discount factor, λ is the parameter of GAE, used to control the trade-off between the bias and variance of the estimated advantage function, is the time series difference TD residual at time step t+l;

[0111] The update of the policy network and the value network: PPO uses a clipped objective function to update the policy network and uses a mean square error (MSE) loss function to update the value network; specifically, the update of the policy network is achieved by maximizing the following objective function L CLIP( θ) is as follows:

[0112]

[0113] wherein, represents the expected value calculation at time step t;

[0114] The update of the value network is achieved by minimizing the following loss function L VF :

[0115]

[0116] wherein, represents the size of the training data set of the kth batch, T represents the length of each trajectory (how many time steps each trajectory contains), τ represents a complete trajectory (i.e., the complete process experienced by a sample, including a series of state, action, reward, etc.), V φ (s t ) is the value estimate of the value network for the state s t , is the actual reward estimate;

[0117] The complete objective function L PPO is as follows:

[0118]

[0119] wherein, c1 and c2 are coefficients for controlling the weights of the respective parts, is the entropy of the policy, which is used to measure the uncertainty of the policy, and increasing the entropy can improve the exploration ability of the policy.

[0120] Such a design enables the model to adaptively learn the optimal maintenance plan and resource allocation strategy in complex and variable railway maintenance scenarios, thereby improving the overall management efficiency and decision-making quality.

[0121] The scheduling method trained using the above technology can optimize the actual maintenance plan and automatically generate a resource adaptation scheme; by dynamically analyzing the process constraints in the maintenance technical list, the system will generate executable maintenance task work orders based on the optimized scheduling scheme. For different types of maintenance plans (such as line repair and equipment maintenance), the system realizes intelligent list matching through the pre-set maintenance technical knowledge base, ensuring accurate alignment of process requirements and resource allocation.

[0122] The embodiment provides a railway infrastructure maintenance resource and plan associated scheduling system, which adopts the railway infrastructure maintenance resource and plan associated scheduling method and comprises a data layer, a business layer and an application layer;

[0123] The data layer and a data interface in the railway maintenance management system are connected to obtain maintenance resource data, maintenance plan data and maintenance work order data; the business layer provides query and management of the maintenance resource and the maintenance plan, joint scheduling functions of the maintenance list and the maintenance plan; through a deep reinforcement learning engine, in combination with resource data and a maintenance guide manual, the maintenance plan execution sequence and corresponding resource allocation are automatically adjusted, the resource data is automatically updated according to the execution state of the maintenance plan and the maintenance work order, and optimal scheduling of the maintenance plan and the maintenance resource is realized; and the application layer is used for providing query and display functions based on the maintenance plan, the maintenance resource and the maintenance work order.

[0124] The above description of the present application and its embodiments is illustrative and not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, similar structural modes and embodiments can be designed without creativity, which should belong to the protection scope of the present application.

Claims

1. A method of scheduling the correlation of railway infrastructure maintenance resources and plans, characterized in that: Comprise the following steps: Step S1, analyze the railway maintenance plan optimization and maintenance resource scheduling problem of the maintenance section, and reduce it to a joint scheduling problem of maintenance resources and maintenance resources; Step S2, comprehensively consider the organizational structure, disease type, maintenance resource information and maintenance plan information of the railway maintenance section, design a joint scheduling system database of railway infrastructure maintenance resources and plans, and develop a data interface in the railway maintenance management system; Step S3, based on actual demand, develop a software framework and system interface of the joint scheduling system of railway infrastructure maintenance resources and plans, including a resource management module, a resource allocation module and an intelligent work order module; The resource management module manages the departmental organization structure of the maintenance section, integrates personnel and machine information, and forms a resource pool; The resource allocation module jointly schedules the maintenance resources and maintenance plans based on the maintenance resource and maintenance plan list, and automatically realizes the optimal maintenance plan sorting and maintenance resource allocation by combining a deep reinforcement learning algorithm; The intelligent work order module realizes work order management and tracking, realizes unified management of maintenance work orders, and synchronously updates the maintenance resource information according to the state of the work order.

2. The method of claim 1, wherein: In the resource management module, the departmental organization structure of the maintenance section is maintenance section-workshop-work area-team; the resource management module can maintain and display the information of the personnel within the organization, and also allows users to quickly retrieve personnel information; the resource management module has personnel information management function and machine information management function, the personnel information management function supports the add, delete and modify operations of information, and ensures the real-time update of personnel data; the machine information management function standardizes the borrowing and returning process of machines, and automatically updates the machine state; The resource management module allows users to quickly retrieve and update machine information.

3. The method of claim 1, wherein: The resource allocation module includes maintenance resource query function, resource allocation function and backtracking function; the resource allocation function models the maintenance plan and maintenance resource as an environment state model based on a flexible job shop scheduling model, and realizes the joint scheduling of resources and maintenance resources by combining a graph neural network.

4. The method of claim 3, wherein: In the modeling of the maintenance plan and the maintenance resource as an environment state model, a Markov decision process (MDP) is introduced.

5. The method of claim 4, wherein: The Markov decision process (MDP) includes the following aspects: State: The state is a description of the environment, which is the basis for the agent to make decisions; the set of all states is called the state space, denoted by S; Action: Action is a description of the behavior of the agent, which is the result of the agent's decision; the set of all possible actions is called the action space, denoted by A; Strategy: Strategy is given according to state, conditional probability distribution of action, used to guide the agent to choose action in different states, symbol is π(a|s); State transition probability: In a certain state, after executing a certain action, the environment transfers to the next state with a certain probability, denoted by P(s'|s,a), where s is the current state, a is the executed action, and s' is the next state; Reward function: The immediate feedback given by the environment to the agent after executing a certain action in a certain state, denoted by R(s,a); Discount factor: used to measure the importance of future rewards, usually between 0 and 1, denoted by γ; State-Value Function: The state-value function V(s) is the expected cumulative reward that can be obtained by following policy π from state s. π (s) represents the expected cumulative reward that can be obtained by following policy π from state s. Action-value function: The action-value function Q(s, a) represents the expected cumulative reward that can be obtained by performing action a in state s and then acting according to policy π.

6. The method of claim 3, wherein: The relationship between the process and the equipment resource is represented by a heterogeneous graph, and the relationship features are extracted by a graph neural network. The process nodes and equipment resource nodes are embedded in stages, and the state representation of a unified size is finally output. The graph neural network extracts the relationship features in the heterogeneous graph through the following steps: 1) Feature embedding: First, the process nodes and equipment resource nodes are embedded respectively; 2) Graph attention mechanism: The graph attention mechanism is used to quantify the influence between different nodes; by calculating the attention coefficient and normalizing it using the Softmax function, the new features of the nodes are obtained; The calculation method of the graph attention mechanism is as follows: A) Attention coefficient calculation: e ij = LeakyReLU(a T [Wh i || Wh j ]) where e ij is the attention coefficient between node i and node j, a is a learnable parameter vector, W is a weight matrix, h i and h j are the feature vectors of nodes i and j, respectively; B) Attention coefficient normalization: wherein a ij is the normalized attention coefficient, N i is the set of neighbor nodes of node i; C) Node feature update: where h' = h + h i is the updated feature vector of node i, and σ is an activation function.

7. A method of scheduling maintenance resources and plans for railway infrastructure according to claim 6, characterized in that: The output of the graph neural network is used as the input of the policy network and the value network. The policy network adjusts the policy according to the feedback provided by the value network, while the value network updates the value estimate according to the reward feedback from the environment. At the same time, the proximal policy optimization (PPO) algorithm is used to stabilize the training of the entire decision-making network. The PPO algorithm is as follows: Clipping mechanism of policy update: PPO limits the magnitude of policy updates by clipping the policy probability ratios; policy probability ratio r t (θ) is given by the expression: where π θ (a t | s t ) is the probability of the new policy selecting action a t in state s t , is the corresponding probability of the old policy; PPO clips this ratio to keep it within [1-∈,1+∈] Estimation of advantage function: In PPO, the estimation of the advantage function A t is achieved by generalized advantage estimation (GAE), which is given by the formula: where γ is a discount factor, and λ is a parameter of GAE, used to control the trade-off between bias and variance of the estimated advantage function, is the time-difference TD residual at time step t + 1; Update of policy network and value network: PPO uses a clipped objective function to update the policy network and uses a mean squared error (MSE) loss function to update the value network; specifically, the update of the policy network is achieved by maximizing the following objective function L CLIP (0) wherein, denotes the expected value computation at time step t; The update of the value network is achieved by minimizing the following loss function L VF L = ∑i∈V (vi - vi+1)2 where, denotes the size of the k-th batch of training data, T denotes the length of each trajectory, τ denotes a complete trajectory, V φ (s t ) is the value network's estimate of the value of state s t , is the actual return estimate; The complete objective function L PPO : is given by the formula: wherein c1 and c2 are coefficients that control the weight of each part, is the entropy of the policy, which measures the uncertainty of the policy.

8. The method of claim 1, wherein: The intelligent work order module has a work order information query function, which provides users with detailed viewing and management of current and historical work orders. The work order information query function allows users to display and filter the current work order list, and search according to the work order number, assignee, and creation time conditions. Users can provide feedback on processed work orders and confirm the completion status of the work order through the processed button.

9. A system for coordinating scheduling of maintenance resources and plans for railway infrastructure, characterized by: It adopts the association scheduling method of railway infrastructure maintenance resources and plans as claimed in any one of claims 1-8, and comprises a data layer, a business layer and an application layer; The data layer and the data interface in the railway maintenance management system are connected to obtain maintenance resource data, maintenance plan data and maintenance work order data; the business layer provides query, management, joint scheduling of maintenance resource and maintenance plan, maintenance list and maintenance plan; through the deep reinforcement learning engine, combined with resource data and maintenance guide manual, the execution order of maintenance plan and corresponding resource allocation are automatically adjusted, and the resource data is automatically updated according to the execution state of maintenance plan and maintenance work order, realizing the optimal scheduling of maintenance plan and maintenance resource; the application layer is used to provide query and display functions based on maintenance plan, maintenance resource and maintenance work order.

Images