Automatic planning integrated service system based on artificial intelligence

Abstract

The invention provides an automatic planning integrated service system based on artificial intelligence. The automatic planning integrated service system based on artificial intelligence comprises a multi-source heterogeneous data access and processing module used for preprocessing multi-source heterogeneous data from a sensor, a historical business log and a third-party API in real time; the data unified modeling and storage module is used for carrying out semantic modeling storage on the multi-source heterogeneous data through a knowledge representation model; and the intelligent planning and scheduling module is used for reasoning and optimizing the timeliness, the cost and the resource utilization rate. According to the automatic planning integrated service system based on artificial intelligence, a decision strategy can be continuously refined, and it is ensured that an optimal or near-optimal scheduling scheme is always provided in a complex and changeable business environment, so that the overall operation efficiency and the resource utilization level are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and automation system integration, and in particular to an automatic planning integration service system based on artificial intelligence. Background Art

[0002] Most existing business systems and automation tools separate task planning, resource scheduling, and service integration, relying primarily on static interface adaptation and data extraction to connect to heterogeneous data from multiple sources, including real-time sensor data, historical business logs, and third-party APIs. These systems are often limited to relatively stable and single business scenarios. Scheduling algorithms are often based on predefined rule engines or fixed heuristic searches, with limited responsiveness to changes in constraints and optimization objectives. In a microservices architecture, while functional modules are provided as services, there is a lack of unified standards for call protocols, data formats, and authentication and authorization. This results in extensive manual configuration and secondary development required for cross-departmental and cross-platform process orchestration and collaborative calls. Even with some heuristic scheduling capabilities, these capabilities remain limited to a "know-how-to" approach. They struggle to leverage historical data for incremental learning and adaptive optimization, and are unable to perceive external changes in real time and provide closed-loop feedback during execution.

[0003] However, with the increasing complexity and diversity of business scenarios, the above existing technologies have exposed the following shortcomings: the real-time integration and maintenance costs of multi-source heterogeneous data are high. Once the data source or business needs change, manual adjustments to the interface and data processing logic are required frequently; the planning and scheduling algorithms lack flexibility and cannot be dynamically optimized according to multi-dimensional indicators such as resource utilization, timeliness, and cost in different scenarios. Rule modification and algorithm reconstruction are time-consuming and prone to errors; the execution process lacks closed-loop feedback capabilities. When encountering emergencies such as traffic congestion and equipment failure, the system cannot automatically replan and can only rely on manual intervention, resulting in reduced overall efficiency. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an automatic planning and integration service system based on artificial intelligence, which solves the problem of difficult real-time integration of multi-source heterogeneous data and the problem of lack of self-adaptation and closed-loop feedback capabilities in planning and scheduling.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an automatic planning integrated service system based on artificial intelligence, comprising: Multi-source heterogeneous data access and processing module, used for real-time pre-processing of multi-source heterogeneous data from sensors, historical business logs, and third-party APIs; A unified data modeling and storage module is used to perform semantic modeling and storage on the multi-source heterogeneous data through a knowledge representation model; An intelligent planning and scheduling module is used to reason and optimize timeliness, cost, and resource utilization, ultimately outputting an adaptive scheduling solution. The intelligent planning and scheduling module includes a multi-agent deep reinforcement learning unit, a graph neural network and variational autoencoder combined unit, a hyperparameter self-tuning unit, and a digital twin simulation unit. An execution monitoring and closed-loop feedback module is used to monitor the execution status of the adaptive scheduling solution in the actual environment online. When an anomaly is detected, the feedback information is transmitted to the intelligent planning and scheduling module to trigger re-planning. The service integration and interface management module is used to arrange and call various heterogeneous business service interfaces according to the adaptive scheduling plan output by the intelligent planning and scheduling module.

[0006] Preferably, the pretreatment process comprises the following steps: 2.1 Pull raw data from sensors, historical business logs, and third-party APIs in parallel; 2.2 Use Kalman filter to suppress noise on the original data; 2.3 Scale the cleaned data to ensure that the subsequent modules input the same dimension.

[0007] Preferably, the knowledge representation model is a hybrid representation model based on a knowledge graph, including a graph layer, an embedding layer and a semantic constraint layer. The graph layer is used to represent entities and semantic relationships between entities with a node-edge structure; the embedding layer maps the nodes and edges in the graph layer to a low-dimensional vector space for subsequent calculation and reasoning; the semantic constraint layer imposes constraints on the embedded vector based on predefined domain rules to ensure that the vector representation conforms to the consistency and integrity of the domain knowledge.

[0008] Preferably, the multi-agent deep reinforcement learning unit is used to model each execution as an agent, and the agent is collaboratively trained through a local policy network and a shared global value network to adaptively allocate tasks and resources in different business scenarios.

[0009] Preferably, the graph neural network and variational autoencoder joint unit is used to perform high-dimensional embedding representation of the dependency relationship between tasks and resources, and generate candidate scheduling solutions through variational inference.

[0010] Preferably, the hyperparameter self-tuning unit is based on a Bayesian optimization model and is used to dynamically balance and update multiple objectives online when adjusting hyperparameters such as the learning rate and exploration-utilization balance coefficient of the scheduling model. The multiple objectives include timeliness, cost, and resource utilization. The model formula of the hyperparameter self-tuning unit is: in: is the hyperparameter vector to be optimized, represents the learning rate, represents the exploration-exploitation balance coefficient; They correspond to three major optimization goals: : Timeliness, :cost, : Resource utilization; is the mean; is the standard deviation; It is The weight coefficient of the target is used to make a dynamic trade-off between timeliness, cost and resource utilization, and satisfies ; It is to explore and utilize the equilibrium constant; This is the acquisition function value used by Bayesian optimization.

[0011] Preferably, the digital twin simulation unit is used to construct a digital twin model based on the real-time physical environment information provided by the data unified modeling and storage module, and simulate the execution effect of the candidate scheduling scheme in parallel in the twin environment, and compare the simulation results with the actual monitoring data and then feed them back to the multi-agent deep reinforcement learning unit to achieve closed-loop optimization.

[0012] Preferably, the intelligent planning and scheduling module performs joint reasoning and optimization on the multiple objectives through the online training of the multi-agent deep reinforcement learning unit and the feedback linkage of the digital twin simulation unit, combined with the graph neural network and variational autoencoder joint unit and the hyperparameter self-tuning unit based on Bayesian optimization, and finally outputs an adaptive scheduling plan.

[0013] Preferably, the adaptive scheduling scheme includes a dynamic sorting scheme based on task priority, a flexible allocation scheme based on resource utilization and cost constraints, and an online adjustment scheme based on real-time execution feedback.

[0014] The present invention provides an automatic planning integrated service system based on artificial intelligence. It has the following beneficial effects: This AI-based automatic planning and integration service system achieves real-time, efficient integration and semantic unification of multi-source heterogeneous data. It uses pre-processing and modeling techniques such as Kalman filtering, normalization, and knowledge graph hybrid representation to address challenges such as data heterogeneity, noise interference, and semantic inconsistency in sensor data, historical business logs, and third-party API interfaces. This ensures the accuracy and reliability of subsequent scheduling decisions from the source. Furthermore, the modular design makes the system highly scalable and maintainable. Accessing new data sources or business scenarios requires minimal configuration in the "Data Access and Processing" and "Knowledge Representation" modules, eliminating the need for significant changes to the overall architecture and significantly reducing the cost of system deployment and upgrades.

[0015] At the scheduling execution level, this solution leverages the collaborative reasoning of multi-agent deep reinforcement learning and a graph neural network-variational autoencoder joint model, combined with Bayesian optimization hyperparameter self-tuning and closed-loop feedback from digital twin simulation. This allows for joint optimization of multiple objectives, including timeliness, cost, and resource utilization, enabling adaptive dynamic sorting, flexible allocation, and online adjustment. The system automatically monitors execution performance and triggers replanning during operation, transmitting actual feedback information back to the scheduling module to continuously refine decision-making strategies, ensuring optimal or near-optimal scheduling solutions are consistently provided in complex and changing business environments, significantly improving overall operational efficiency and resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart for implementing the invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, an embodiment of the present invention provides an automatic planning integration service system based on artificial intelligence, including a multi-source heterogeneous data access and processing module for real-time pre-processing of multi-source heterogeneous data from sensors, historical business logs and third-party APIs.

[0019] The preprocessing process includes the following steps: 2.1 Pull raw data from sensors, historical business logs, and third-party APIs in parallel; 2.2 Use Kalman filter to suppress noise on the original data; 2.3 Scale the cleaned data to ensure that the subsequent modules input the same dimension.

[0020] The unified data modeling and storage module is used to semantically model and store multi-source heterogeneous data using a knowledge representation model. The knowledge representation model is a hybrid representation model based on a knowledge graph, consisting of a graph layer, an embedding layer, and a semantic constraint layer. The graph layer represents entities and the semantic relationships between them using a node-edge structure; the embedding layer maps the nodes and edges in the graph layer into a low-dimensional vector space for subsequent computation and reasoning; and the semantic constraint layer imposes constraints on the embedded vectors based on predefined domain rules to ensure the consistency and integrity of the vector representation with domain knowledge.

[0021] The intelligent planning and scheduling module is used to infer and optimize timeliness, cost, and resource utilization, and ultimately output an adaptive scheduling plan. The intelligent planning and scheduling module includes a multi-agent deep reinforcement learning unit, a graph neural network and variational autoencoder joint unit, a hyperparameter self-tuning unit, and a digital twin simulation unit.

[0022] A multi-agent deep reinforcement learning unit is used to model each execution as an agent. The agents are trained collaboratively through a local policy network and a shared global value network to adaptively allocate tasks and resources in different business scenarios.

[0023] The graph neural network and variational autoencoder joint unit is used to perform high-dimensional embedding representation of the dependencies between tasks and resources and generate candidate scheduling solutions through variational inference.

[0024] The hyperparameter self-tuning unit, based on the Bayesian optimization model, is used to dynamically balance and update multiple objectives online when adjusting hyperparameters such as the learning rate and exploration-exploitation balance coefficient of the scheduling model. These objectives include timeliness, cost, and resource utilization. The hyperparameter self-tuning unit model formula is: in: is the hyperparameter vector to be optimized, represents the learning rate, represents the exploration-exploitation balance coefficient; They correspond to three major optimization goals: : Timeliness, :cost, : Resource utilization; is the mean; is the standard deviation; It is The weight coefficient of the target is used to make a dynamic trade-off between timeliness, cost and resource utilization, and satisfies ; It is to explore and utilize the equilibrium constant; This is the acquisition function value used by Bayesian optimization.

[0025] The digital twin simulation unit is used to build a digital twin model based on the real-time physical environment information provided by the unified data modeling and storage module, and to simulate the execution effects of candidate scheduling schemes in parallel in the twin environment. The simulation results are compared with the actual monitoring data and then fed back to the multi-agent deep reinforcement learning unit to achieve closed-loop optimization.

[0026] The intelligent planning and scheduling module conducts joint reasoning and optimization of multiple objectives through the online training of the multi-agent deep reinforcement learning unit and the feedback linkage of the digital twin simulation unit, combined with the graph neural network and variational autoencoder joint unit and the hyperparameter self-tuning unit based on Bayesian optimization, and finally outputs an adaptive scheduling plan.

[0027] The specific implementation is as follows: Application scenarios: Time-sensitive urban terminal delivery scheduling, urban "last mile" fast delivery, and order on-time rate requirement of ≥95%.

[0028] Module collaboration details: Multi-agent Deep Reinforcement Learning Unit: Network structure: Each agent contains two layers of fully connected policy networks, and the value network consists of three layers of convolution plus two layers of full connection.

[0029] Priority experience replay: Historical samples are sorted according to the temporal difference error, and important samples have a higher sampling probability.

[0030] Parameter synchronization: Every 10 minutes, synchronize the global value network parameters to all agents.

[0031] Graph Neural Networks and Variational Autoencoder Units: Graph construction: Nodes represent "distribution points" or "order clusters", and edges represent road connections and adjacent relationships.

[0032] Message passing: Multi-hop graph convolution aggregates neighbor information and updates edge weights based on the real-time road congestion index.

[0033] Candidate path generation: The decoder outputs multiple delivery paths and selects the top 5 optimal paths based on the reconstruction error.

[0034] Hyperparameter self-tuning unit: Bayesian optimization: A Gaussian process model with an auto-correlated variance-squared exponential kernel is used to perform online tuning of the learning rate and exploration coefficient.

[0035] Tuning cycle: Run the collection function every 12 hours to select new hyperparameters.

[0036] Digital Twin Simulation Unit: Simulation step size: 1 minute as the step size to simulate the vehicle driving and parking process.

[0037] Physical mapping: Mapping real-time GPS, road speed, and traffic light data onto a virtual road network.

[0038] Deviation feedback: When the average deviation between the simulation time and the actual time exceeds 5%, the model fine-tuning is immediately triggered and the deviation log is recorded.

[0039] Monitoring and Alerting: Indicator collection: Capture punctuality rate, average delivery time, number of re-planning, etc. through the monitoring system.

[0040] Visualization: Key indicators are presented in real time on the visualization platform. When the number of replanning exceeds 2 per hour, an alarm is sent through the messaging platform.

[0041] Performance indicators: Average punctuality rate increased from 82% to 97%; the number of route adjustments during peak hours decreased by 30%, and the idle vehicle rate decreased by 12%.

[0042] The execution monitoring and closed-loop feedback module is used to monitor the execution status of the adaptive scheduling scheme in the actual environment online. When an anomaly is detected, the feedback information is passed to the intelligent planning and scheduling module to trigger re-planning.

[0043] The service integration and interface management module orchestrates and calls heterogeneous business service interfaces based on the adaptive scheduling solutions output by the intelligent planning and scheduling module. These adaptive scheduling solutions include dynamic sorting based on task priority, flexible allocation based on resource utilization and cost constraints, and online adjustment based on real-time execution feedback. Example 1

[0044] This embodiment is different from the first embodiment in that the application scenario of this embodiment is cost-sensitive small-batch multi-variety manufacturing scheduling.

[0045] Application scenario: machining workshops with frequent switching between multiple varieties, focusing on reducing conversion downtime and tool change costs.

[0046] Module collaboration details: Multi-agent Deep Reinforcement Learning Unit: Agent division: Each production line is an agent.

[0047] Local strategy: Input the current queue length, equipment occupancy status and order deadline, and output the decision on the next production batch.

[0048] Shared Value Network: Evaluates the total switching cost of the entire workshop and integrates the states of all agents through the Critic network.

[0049] Graph Neural Networks and Variational Autoencoder Units: Graph construction: Nodes include "workpiece batch", "equipment" and "process", with side labels of "processing time" and "conversion time".

[0050] Deep embedding: After multiple rounds of neighbor aggregation, the node vector is concatenated with the historical switching records.

[0051] Production schedule generation: The variational autoencoder outputs the top 10 production schedules and uses heuristic cost estimation to quickly select the top 3.

[0052] Hyperparameter self-tuning unit: New objective: Add "switching frequency" as an additional optimization objective to Bayesian optimization.

[0053] Dynamic weight adjustment: If the average switching downtime exceeds 8 hours for 3 consecutive days, the cost weight will be automatically increased.

[0054] Collection cycle: The collection function is executed every 6 hours within a 48-hour rolling window.

[0055] Digital Twin Simulation Unit: Simulation model: discrete event simulation, each device switching is treated as a simulation event.

[0056] Equipment parameter injection: read actual processing speed and failure rate from the programmable logic controller.

[0057] Error feedback: When the error between the simulation cost and the cost recorded in the production execution system exceeds 7%, an error analysis report is generated and the operation and maintenance department is notified.

[0058] Quality and Audit: Logging: Each time a scheduling plan is generated, all input features and output sequences are stored in the ELK log system.

[0059] KPI report: Generate monthly manufacturing cost savings and production efficiency improvement reports for management decision-making.

[0060] Effect indicators: average switching downtime decreased by 18%, comprehensive utilization rate of production lines increased by 10%; unit product processing cost decreased by 12%, and early order delivery rate increased by 7%. Example 2

[0061] This embodiment is different from the first embodiment in that the application scenario of this embodiment is the elastic scheduling of cloud computing resources with severe fluctuations.

[0062] Application scenario: In a public cloud multi-tenant environment, load and instance prices fluctuate dramatically over time and bidding periods.

[0063] Module collaboration details: Multi-agent Deep Reinforcement Learning Unit: Agent definition: Each virtual machine cluster or container group is modeled as an agent.

[0064] Policy network: uses a long short-term memory network plus a fully connected layer to capture load timing characteristics.

[0065] Global Value Network: Integrates resource utilization and service level indicators of all tenants to guide joint optimization.

[0066] Graph Neural Networks and Variational Autoencoder Units: Graph construction: Nodes include "tenants," "application services," and "virtual machine instances," with sidebands including "call frequency" and "network latency."

[0067] Embedding process: After multi-hop message transmission, the node vector is spliced ​​with the price curve within the time window.

[0068] Scaling plan generation: The variational autoencoder outputs multiple scaling strategies in parallel and quickly scores them based on the SLA achievement rate and cost model.

[0069] Hyperparameter self-tuning unit: Newly included metric: The preemptible instance success rate is used as an additional optimization target.

[0070] Online adjustment: Evaluate the probability of sudden load increases every hour and dynamically increase the exploration coefficient.

[0071] Weight adjustment: If the SLA default rate exceeds 1% for two consecutive hours, the timeliness weight will be automatically increased.

[0072] Digital Twin Simulation Unit: Simulation scenario: replays the load, instance price, and network jitter data for the past 48 hours.

[0073] Parallel testing: Verify the resource utilization of candidate scaling curves in parallel in the digital twin environment.

[0074] Comparative feedback: When the deviation between the simulated resource utilization and the actual monitoring data exceeds 5%, model retraining is triggered.

[0075] Operation and maintenance and alarm: Multi-dimensional monitoring: Use the monitoring system to capture CPU, memory, network bandwidth, and cost curves.

[0076] Intelligent Alerts: Detect abnormal scaling behavior based on rules and push alerts through the messaging platform.

[0077] Visual Report: Combined with the elastic search engine on the business intelligence platform, the monthly cost savings and SLA achievement rate report is displayed.

[0078] Performance indicators: Average resource utilization increased by 10%; elastic scaling costs were reduced by approximately 18%; and the service level agreement breach rate was kept below 0.5%.

[0079] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An automatic planning integrated service system based on artificial intelligence, characterized in that: include: Multi-source heterogeneous data access and processing module, used for real-time pre-processing of multi-source heterogeneous data from sensors, historical business logs, and third-party APIs; A unified data modeling and storage module is used to perform semantic modeling and storage on the multi-source heterogeneous data through a knowledge representation model; An intelligent planning and scheduling module is used to reason and optimize timeliness, cost, and resource utilization, ultimately outputting an adaptive scheduling solution. The intelligent planning and scheduling module includes a multi-agent deep reinforcement learning unit, a graph neural network and variational autoencoder combined unit, a hyperparameter self-tuning unit, and a digital twin simulation unit. An execution monitoring and closed-loop feedback module is used to monitor the execution status of the adaptive scheduling solution in the actual environment online. When an anomaly is detected, the feedback information is transmitted to the intelligent planning and scheduling module to trigger re-planning. The service integration and interface management module is used to arrange and call various heterogeneous business service interfaces according to the adaptive scheduling plan output by the intelligent planning and scheduling module.

2. The artificial intelligence-based automatic planning integrated service system according to claim 1, characterized in that: The pre-treatment process comprises the following steps: 2.1 Pull raw data from sensors, historical business logs, and third-party APIs in parallel; 2.2 Use Kalman filter to suppress noise on the original data; 2.3 Scale the cleaned data to ensure that the subsequent modules input the same dimension.

3. The automatic planning integrated service system based on artificial intelligence according to claim 1, characterized in that: The knowledge representation model is a hybrid representation model based on the knowledge graph, including a graph layer, an embedding layer and a semantic constraint layer. The graph layer is used to represent entities and semantic relationships between entities using a node-edge structure; the embedding layer maps the nodes and edges in the graph layer to a low-dimensional vector space to facilitate subsequent calculation and reasoning; the semantic constraint layer imposes constraints on the embedded vectors based on predefined domain rules to ensure that the vector representation conforms to the consistency and integrity of the domain knowledge.

4. The artificial intelligence-based automatic planning integrated service system according to claim 1, characterized in that: The multi-agent deep reinforcement learning unit is used to model each execution as an agent, and the agent is collaboratively trained through a local policy network and a shared global value network to adaptively allocate tasks and resources in different business scenarios.

5. The automatic planning integrated service system based on artificial intelligence according to claim 1, characterized in that: The graph neural network and variational autoencoder joint unit is used to perform high-dimensional embedding representation of the dependency relationship between tasks and resources, and generate candidate scheduling solutions through variational inference.

6. The automatic planning integrated service system based on artificial intelligence according to claim 1, characterized in that: The hyperparameter self-tuning unit is based on a Bayesian optimization model and is used to dynamically balance and update multiple objectives online when adjusting hyperparameters such as the learning rate and exploration-utilization balance coefficient of the scheduling model. The multiple objectives include timeliness, cost, and resource utilization. The model formula of the hyperparameter self-tuning unit is: in: is the hyperparameter vector to be optimized, represents the learning rate, represents the exploration-exploitation balance coefficient; They correspond to three major optimization goals: Timeliness, cost, Resource utilization; is the mean; is the standard deviation; It is The weight coefficient of the target is used to make a dynamic trade-off between timeliness, cost and resource utilization, and satisfies ; It is to explore and utilize the equilibrium constant; This is the acquisition function value used by Bayesian optimization.

7. The automatic planning integrated service system based on artificial intelligence according to claim 1, characterized in that: The digital twin simulation unit is used to build a digital twin model based on the real-time physical environment information provided by the data unified modeling and storage module, and to simulate the execution effect of the adaptive scheduling scheme in parallel in the twin environment, and to compare the simulation results with the actual monitoring data and then feed them back to the multi-agent deep reinforcement learning unit.

8. The artificial intelligence-based automatic planning integrated service system according to claim 1 or 6, characterized in that: The intelligent planning and scheduling module, through the online training of the multi-agent deep reinforcement learning unit and the feedback linkage of the digital twin simulation unit, combines the graph neural network and variational autoencoder joint unit and the hyperparameter self-tuning unit based on Bayesian optimization, to jointly reason and optimize the multiple objectives and finally output an adaptive scheduling plan.

9. The automatic planning integrated service system based on artificial intelligence according to claim 1, characterized in that: The adaptive scheduling scheme includes a dynamic sorting scheme based on task priority, a flexible allocation scheme based on resource utilization and cost constraints, and an online adjustment scheme based on real-time execution feedback.

Images