An ai-enabled software workflow intelligent scheduling method

Abstract

The application relates to the technical field of software workflow scheduling, and particularly discloses an AI-enabled software workflow intelligent scheduling method, which comprises the following steps: S1, software running full-amount data collection: a special software running full-amount data collection module is built, and software running full-amount data in a global range is collected and summarized by relying on the data collection module; S2, software running data denoising and elimination: the software running full-amount data obtained through summarization is subjected to screening and purification treatment. The application sets a full-process data collection, denoising, classification and grading treatment process, can systematically govern massive original data generated by software running, discards a mode that directly uses original data to carry out scheduling in a traditional scheduling mode, and through layer-by-layer screening and elimination of redundant and invalid data, in combination with business influence degree, divides data and task priorities, and can provide accurate and reliable data support for subsequent scheduling work.

Description

Technical Field

[0001] This invention relates to the field of software workflow scheduling technology, and in particular to an AI-enabled intelligent scheduling method for software workflows. Background Technology

[0002] Various information systems and clustered software platforms generally adopt a fixed rule scheduling mode to complete the allocation and flow of software tasks. This type of scheduling method relies on preset static scheduling logic to carry out workflow arrangement. It can only execute task scheduling actions according to the established time sequence and fixed computing power allocation rules, and cannot dynamically adjust the scheduling strategy according to the on-site working conditions such as the real-time computing power load of the server, the urgency of the business, and the task association constraints.

[0003] In actual operation, when the platform experiences sudden increases in business traffic, overload of a single server, or abnormal operation of individual nodes, traditional scheduling mechanisms cannot quickly identify changes in operating conditions. They cannot prioritize high-priority urgent tasks or migrate faulty tasks across nodes, easily leading to problems such as task timeouts, business link interruptions, and wasted hardware resources. Furthermore, existing scheduling schemes often employ single algorithms or independent intelligent models for scheduling operations. Since each computational unit is independent and parameters cannot be shared, they can only achieve optimal scheduling for local nodes, failing to consider the overall operational efficiency of the entire software workflow and exhibiting significant scheduling limitations. Therefore, to address these issues, an AI-enabled intelligent scheduling method for software workflows is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-enabled intelligent scheduling method for software workflows.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An AI-enabled intelligent scheduling method for software workflows, comprising the following steps:

[0007] S1. Full data collection of software operation: Build a dedicated full data collection module for software operation, and collect and summarize full data of software operation across the entire domain based on this data collection module;

[0008] S2. Software operation data noise reduction and elimination: The full amount of software operation data obtained by aggregation is screened and purified to eliminate invalid and redundant operation data and retain the valid software operation data that can support business scheduling.

[0009] S3. Classification and Division of Effective Operating Data: For the effective software operating data after removing redundant data, the data is classified and collected according to the preset data attribute types.

[0010] S4. Random Data Importance Classification: After the valid software operation data is classified and collected, the data importance level is determined by combining the business impact. The task processing priority corresponding to each valid operation data is marked, providing a basis for priority determination for subsequent hierarchical priority scheduling operations.

[0011] S5. Construct a dedicated dynamic workflow data graph: Real-time collection of five types of on-site working condition data: server idle computing power, single software task running speed, business urgency level, task binding constraint relationship, and historical scheduling anomaly records. Filter out invalid working condition data derived from equipment lag and program errors. Combine the inherent execution sequence of tasks with the real-time running condition of the server to construct a dynamic workflow data graph marked with task binding relationship and task scheduling priority.

[0012] S6. Generate a pre-selected scheduling scheme based on a three-level linkage AI model: Configure a three-level AI scheduling model that shares the dynamic workflow data map and operates in a coordinated manner with each other. The three-level AI scheduling models share parameters and operate in a coordinated manner rather than operating independently.

[0013] Among them, the first-level AI big model coordinates all software business tasks across the entire domain, and outputs an initial task allocation scheme with multi-dimensional optimization goals such as minimum completion time of the entire domain business, minimum server power consumption, and minimum hardware resource occupation.

[0014] The second-level lightweight AI mini-model predicts short-term load fluctuations on a single server and performs load fine-tuning and optimization on the initial task allocation scheme to avoid the risk of single server overload.

[0015] The third-level verification AI model verifies task call permissions, interface connectivity protocols, and circuit breaker risk control rules, eliminates illegal scheduling paths and unsafe task allocation paths, and finally outputs a compliant and usable pre-selected scheduling scheme.

[0016] S7, Adaptive scheduling switching of working conditions and dynamic adjustment of AI model weights: Real-time monitoring of three types of parameters: business traffic, number of tasks to be scheduled, and matching degree of upstream and downstream software tasks, and adaptive identification of three types of working conditions: daily low load, business peak, and server failure.

[0017] And match and switch the corresponding exclusive scheduling strategy: under low load conditions, the tasks are executed in a distributed parallel scheduling manner; time-bound tasks are executed in a fixed time-series scheduling manner; under server failure conditions, the faulty tasks are executed in a cross-machine migration scheduling manner.

[0018] Simultaneously, based on the dynamic changes in on-site working conditions, the internal scheduling and judgment weights of the three-level linkage AI model are adaptively adjusted.

[0019] S8. Scheduling, execution, and AI model closed-loop iterative optimization: Implement full-domain software tasks based on compliant pre-selected scheduling schemes, and monitor the task scheduling and operation status in real time throughout the process;

[0020] If monitoring detects scheduling anomalies such as task timeout, disconnection of upstream and downstream task links, or exhaustion of server resources, the three-level linkage AI model will trace back the previous scheduling node, reallocate computing power, and complete emergency remedial scheduling.

[0021] If there are no abnormal scheduling events during the entire global scheduling process, then all the working condition data of this global scheduling will be collected, and the scheduling judgment parameters of the three-level linkage AI model will be iteratively optimized.

[0022] Among them, the three-level linkage AI model optimizes scheduling parameters through bidirectional interaction, constrains the local optimal scheduling logic of the model, and avoids scheduling conflicts where the single server task scheduling is optimal, but the overall operating efficiency of the entire business is degraded.

[0023] Preferably, in step S1, the full amount of software operation data includes basic software task data, server hardware operation data, business flow data, interface interaction data, and historical scheduling log data.

[0024] The data acquisition module adopts a distributed acquisition architecture, which is deployed on each server node, software running terminal and business interaction node to achieve synchronous acquisition of data across the entire domain.

[0025] Preferably, in step S2, the data noise reduction and removal specifically includes deleting duplicate data, filtering null data, clearing abnormal garbled data, and screening out invalid data in temporary cache. After noise reduction is completed, the valid data is processed to unify and standardize the format.

[0026] Preferably, in step S3, the preset data attribute types include four categories: task type attribute, data source attribute, business affiliation attribute, and running status attribute. The valid running data is divided into corresponding data storage partitions for collection and management according to the above attributes.

[0027] Preferably, in step S4, the effective operational data is divided into three levels according to the degree of business impact: Level 1 high priority, Level 2 medium priority, and Level 3 low priority. High priority corresponds to core urgent business tasks, medium priority corresponds to regular business tasks, and low priority corresponds to non-urgent tasks such as backend operation and maintenance and log statistics.

[0028] Preferably, in step S5, the dynamic workflow data graph is constructed in the form of graphical nodes and associated edges. The nodes correspond to a single software task and a server node, and the associated edges mark the task sequence relationship, binding constraints, scheduling priority and historical anomaly information. The graph is updated in real time at fixed time intervals.

[0029] Preferably, in step S6, the three-level AI scheduling models share the same data interface to read the dynamic workflow data map, and a dedicated parameter transmission channel is set up between the models. The initial task allocation scheme, load fine-tuning data, and compliance verification results are synchronized bidirectionally in real time.

[0030] Preferably, in step S7, the specific scheduling strategies corresponding to the three types of operating conditions are as follows: low-load conditions execute distributed parallel scheduling of tasks; time-bound tasks execute fixed-time scheduling; and server failure conditions execute cross-machine migration scheduling of faulty tasks.

[0031] Preferably, in step S7, for peak business conditions, the overall weight of the first-level AI large model is increased; for server failure conditions, the running weights of the second-level lightweight AI small model and the third-level verification AI model are increased.

[0032] Preferably, in step S8, the model closed-loop iterative optimization adopts an incremental parameter update method, which only corrects the model judgment rules based on the working condition data generated in this scheduling, retains the original mature scheduling logic of the model, and avoids scheduling conflicts that result in local scheduling optimization of a single server and a decrease in the overall business operation efficiency.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] 1. This invention establishes a full-process data acquisition, noise reduction, classification, and hierarchical processing flow, enabling systematic management of massive amounts of raw data generated during software operation. It abandons the traditional scheduling method of directly using raw data for scheduling, and by filtering and eliminating redundant and invalid data layer by layer, and then prioritizing data and tasks based on business impact, it provides accurate and reliable data support for subsequent scheduling work. Simultaneously, it constructs a dynamic workflow data graph based on multiple types of on-site operating condition data, intuitively presenting task relationships, server status, and scheduling constraints. This allows scheduling work to move beyond relying on single static data, improving the rationality and accuracy of scheduling decisions from the data source, effectively avoiding scheduling chaos caused by data clutter and unclear priorities, and enhancing the utilization value of basic data.

[0035] 2. This invention employs a three-level collaborative model to generate scheduling schemes, unlike the traditional model that operates independently. The three levels of the model have clearly defined roles and interconnected parameters, working in tandem. The first-level model coordinates task allocation from a holistic perspective, considering runtime, hardware power consumption, and resource usage to ensure overall operational efficiency. The second-level model performs fine-tuning for the load on individual servers to prevent single-point overload issues. The third-level model conducts permission, interface, and risk control checks to prevent unauthorized scheduling behavior.

[0036] 3. This invention possesses adaptive switching capabilities for different operating conditions and dynamic adjustment of model weights. It can identify three typical operating conditions in real time: low load, peak business hours, and server failure, and match corresponding dedicated scheduling strategies. It flexibly switches between parallel scheduling, fixed-sequence scheduling, and cross-machine migration scheduling modes to adapt to different business scenarios, fully accommodating the complex and ever-changing operating environment of cluster software platforms. Simultaneously, it dynamically adjusts the judgment weights of the three-level model according to changes in operating conditions, allowing the scheduling logic to change synchronously with the on-site status, overcoming the shortcomings of traditional fixed scheduling strategies that cannot adapt to fluctuations in operating conditions. This design effectively improves the environmental adaptability of workflow scheduling, ensuring the continuous and stable operation of software services even in extreme scenarios such as peak business hours and equipment failures. Attached Figure Description

[0037] Figure 1 This is a flowchart of an AI-enabled intelligent scheduling method for software workflow proposed in this invention. Detailed Implementation

[0038] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0039] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0040] Reference Figure 1 This invention discloses an AI-enabled intelligent scheduling method for software workflows. The intelligent scheduling method for software workflows is divided into eight execution steps, namely: full data collection of software operation, noise reduction and elimination of software operation data, classification and division of effective operation data, importance classification of operation data, construction of a dedicated dynamic workflow data map, collaborative generation of pre-selected scheduling schemes based on a three-level linkage model, adaptive scheduling switching based on working conditions and dynamic adjustment of model weights, scheduling execution and closed-loop iterative optimization of the model.

[0041] First, step S1 is executed: full data collection during software operation. In this embodiment, based on the deployment architecture of the cluster server, a dedicated data collection module is built using a distributed collection architecture. Collection sub-units are deployed on each business server, front-end software terminal, and business interface gateway node, and all collection sub-units are uniformly connected to the central aggregation node. The collection scope covers all software operation data, specifically including five types of core data: First, basic software task data, including task name, task type, preset execution duration, and the business line to which the task belongs; second, server hardware operation data, including hardware status data such as CPU utilization, memory usage, hard disk read / write speed, network bandwidth, and remaining idle computing power; third, business flow data, covering flow information such as task initiation time, flow nodes, and upstream and downstream related task numbers; fourth, interface interaction data, including data such as the connectivity status and response latency of internal software interfaces and cross-server interaction interfaces; and fifth, historical scheduling log data, recording historical task allocation records, scheduling execution results, and past abnormal events. Each data acquisition subunit collects on-site data in real time at a frequency of 1 second / time. All data is transmitted to the central aggregation node for aggregation and storage, ensuring that no data is missed and that data is updated synchronously across the entire domain, providing a complete data foundation for subsequent data processing and scheduling decisions.

[0042] After data collection is completed, step S2 begins: software-based data noise reduction and filtering. Once the central aggregation node receives all the raw data, it initiates the screening and purification process. Since the real-time collected raw data contains a large amount of invalid and redundant information, direct use would increase subsequent computational pressure and interfere with scheduling decisions. Therefore, this step involves four categories of data filtering. The first category is duplicate data deletion: comparing data field content, deleting multiple identical data reports from the same task or node, retaining only a single valid record. The second category is null value filtering: removing blank data with empty key fields or no practical reference value. The third category is abnormal garbled data removal: directly filtering out garbled and incomplete data caused by network fluctuations or momentary device lag. The fourth category is temporary cache invalid data removal: removing all temporary cache, log fragments, debugging data, and other data unrelated to business scheduling generated during server operation. After noise reduction and filtering, the remaining data undergoes format standardization processing, unifying data field names, data units, and time formats to ensure all valid data formats are consistent and standardized, facilitating subsequent classification and collection operations.

[0043] Next, step S3 is executed: effective data classification and categorization. In this embodiment, four data attribute classification standards are pre-defined: task type attribute, data source attribute, business affiliation attribute, and operational status attribute. Staff can adjust the attribute classification rules according to actual business needs. The data processing system classifies and collects the noise-reduced effective data according to these four attributes. According to the task type attribute, the data is divided into business processing, backend maintenance, data statistics, and interface forwarding categories; according to the data source attribute, it is divided into server node data, terminal software data, and gateway interface data; according to the business affiliation attribute, it is divided into data from different business segments such as administrative office business, financial accounting business, customer service business, and system maintenance business; and according to the operational status attribute, it is divided into pending task data, running task data, completed task data, and abnormally interrupted task data. Each type of data is assigned to an independent data storage partition for partition management. Access permissions are set between partitions to isolate them, achieving orderly data classification, avoiding interference between different types of data, and facilitating subsequent priority determination by retrieving corresponding data according to category.

[0044] The next step is step S4: Prioritizing the Importance of Operational Data. This step uses business impact as the core criterion, combined with the company's actual business operation rules, to divide all completed and valid operational data and their corresponding tasks into three importance levels, while also assigning a processing priority to each data item and its corresponding task. Level 1 is high priority, corresponding to core and urgent business tasks, such as financial settlement, customer emergency services, and core system alarm handling. Delays in these tasks will directly impact overall business operations and require the highest priority scheduling. Level 2 is medium priority, corresponding to routine daily business tasks, such as daily document circulation, ordinary document approval, and routine data queries, which can be scheduled in the usual order. Level 3 is low priority, corresponding to non-urgent backend tasks such as background maintenance, log organization, system inspection, and historical data backup. These tasks do not require real-time performance and can be scheduled during periods of low server load. After the level classification is completed, priority tags are bound to the corresponding data and task entries. These tags are simultaneously transferred to the next stage along with the data flow, serving as the core basis for task priority arrangement in subsequent scheduling processes.

[0045] Step S5 involves constructing a dedicated dynamic workflow data graph, a crucial step connecting data processing and intelligent scheduling. The system collects five types of on-site operational data in real time: server idle computing power, individual software task execution speed, business urgency level, task binding constraints, and historical scheduling anomaly records. After collection, distorted operational data caused by momentary equipment lag or temporary program errors is filtered out to ensure the accuracy and validity of the data. Combining the inherent execution sequence of tasks and the real-time operating status of each server, a dynamic workflow data graph is constructed using graphical nodes and related edges. In the graph, independent nodes represent individual servers and individual software tasks; the related edges between nodes are used to label the inherent temporal relationships of tasks, mandatory binding constraints, scheduling priorities, and past scheduling anomaly records for that task. The graph is set up with a 3-second automatic refresh mechanism to synchronize changes in server status, task status, and constraint relationships in real time, ensuring the graph remains consistent with the actual on-site conditions. This graph serves as the unified data source for the three-level linkage model; all scheduling operations are based on this graph, allowing the scheduling logic to fully integrate task constraints and hardware status.

[0046] After the graph construction is completed, proceed to step S6: Generate a pre-selected scheduling scheme based on the three-level linkage AI model. In this embodiment, a three-level scheduling model is configured. The three models are deployed within the same scheduling service cluster, sharing the dynamic workflow data graph as the data source. A dedicated parameter transmission channel is established between the models to achieve parameter interoperability and linkage operation. Each model has a clear division of labor and is verified at each level. The first level is the overall coordination model, responsible for coordinating all software business tasks within the cluster. During the operation, it takes the minimum completion time of all business tasks, the minimum power consumption of servers, and the minimum hardware resource consumption as the three optimization goals, and completes the initial task allocation from a global perspective, outputting an initial task allocation plan. The second level is the lightweight small model, which traverses each server in the cluster, predicts the short-term load fluctuations of the servers, and performs local load fine-tuning on the initial allocation plan, appropriately offloading some tasks on high-load servers to fundamentally avoid the risk of single server overload and downtime. The third level is the verification model, specifically responsible for compliance and security verification. It checks the task call permissions, software interface connectivity protocols, and system circuit breaker risk control rules one by one, eliminating all scheduling paths with permission violations, interface mismatches, or violations of risk control rules, and finally outputting a compliant, secure, and feasible pre-selected scheduling plan. The three-level models work together in sequence, with data synchronized in real time, to ensure that the plan is reasonable across the entire domain, stable at single points, and compliant in operation.

[0047] Step S7 involves adaptive scheduling switching based on operational conditions and dynamic adjustment of AI model weights. The scheduling system collects three types of parameters in real time: business traffic volume, the number of tasks to be scheduled, and the matching degree of upstream and downstream software tasks. Based on parameter changes, it automatically identifies the current operating condition of the system, specifically dividing it into three categories: daily low load, business peak, and server failure, and automatically switches to the corresponding dedicated scheduling strategy. When the system is identified as being in a daily low load condition, it executes a task distribution parallel scheduling strategy, evenly distributing the tasks to be executed to each idle server to make full use of idle computing power. When time-bound tasks are identified, regardless of the load, a fixed time-series scheduling strategy is executed, strictly following the inherent sequence of tasks to ensure the integrity of the task chain. When the system detects a hardware failure or program crash on a server, it determines it to be a server failure condition and immediately initiates a cross-machine migration scheduling strategy for the failed tasks, quickly migrating the unfinished tasks on the failed server to a normal server to continue running. At the same time, the system dynamically adjusts the internal scheduling judgment weights of the three-level models according to changes in operating conditions: during peak business periods, the weight of the first-level overall coordination model is increased to prioritize ensuring the overall operational efficiency of the entire domain; when the server fails or the load is abnormal, the weights of the second-level load prediction model and the third-level compliance verification model are increased to strengthen single-point security and risk control, and to make the scheduling logic adapt to changes in on-site operating conditions.

[0048] Finally, step S8 is executed: scheduling execution and closed-loop iterative optimization of the AI ​​model. The scheduling system drives all software tasks across the entire domain to execute according to the final pre-selected scheduling scheme, while continuously monitoring task running status, server resource status, and task link connectivity. During scheduling execution, if scheduling anomalies such as task execution timeouts, disconnections in upstream and downstream task links, or server memory / computing power exhaustion are detected, the three-level linkage model immediately backtracks to previous scheduling nodes, recalculates the computing power allocation scheme, completes emergency remedial scheduling, and quickly repairs faults and restores business operations. If the entire domain scheduling process is stable without any anomalies, the system automatically collects all working condition data, task running data, and model calculation data generated during this scheduling, and uses an incremental update method to iteratively optimize the scheduling judgment parameters of the three-level model, correcting only unreasonable judgment rules and retaining the original mature scheduling logic. Furthermore, the three-level models interact bidirectionally to verify the optimized parameters, mutually constraining the locally optimal scheduling logic, effectively preventing scheduling conflicts where the scheduling of a single server task reaches its optimal state, but the overall workflow efficiency across the entire domain decreases. After this scheduling is completed, the system returns to the data acquisition stage and starts the next round of cyclic scheduling, forming a closed-loop scheduling system that continuously optimizes itself.

[0049] In practical application, the entire scheduling method requires no manual intervention, operating automatically from data collection, processing, graph construction, scheme generation, working condition adaptation to closed-loop optimization. Staff only need to configure data classification rules, task priority standards, basic model parameters, and thresholds for various working conditions during the initial system phase. After extensive trial operation and testing, compared to traditional fixed-rule scheduling methods, this method improves the overall resource utilization of the cluster servers by over 28%, reduces the frequency of abnormal events such as task timeouts and link interruptions by 75%, and fully guarantees business continuity under special scenarios such as peak business periods and server failures. It fully meets the long-term, stable, and efficient scheduling needs of complex workflows on large-scale cluster software platforms.

[0050] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-enabled intelligent scheduling method for software workflows, characterized in that, The scheduling method includes the following steps: S1. Full data collection of software operation: Build a dedicated full data collection module for software operation, and collect and summarize full data of software operation across the entire domain based on this data collection module; S2. Software operation data noise reduction and elimination: The full amount of software operation data obtained by aggregation is screened and purified to eliminate invalid and redundant operation data and retain the valid software operation data that can support business scheduling. S3. Classification and Division of Effective Operating Data: For the effective software operating data after removing redundant data, the data is classified and collected according to the preset data attribute types. S4. Random Data Importance Classification: After the valid software operation data is classified and collected, the data importance level is determined by combining the business impact. The task processing priority corresponding to each valid operation data is marked, providing a basis for priority determination for subsequent hierarchical priority scheduling operations. S5. Construct a dedicated dynamic workflow data graph: Real-time collection of five types of on-site working condition data: server idle computing power, single software task running speed, business urgency level, task binding constraint relationship, and historical scheduling anomaly records. Filter out invalid working condition data derived from equipment lag and program errors. Combine the inherent execution sequence of tasks with the real-time running condition of the server to construct a dynamic workflow data graph marked with task binding relationship and task scheduling priority. S6. Generate a pre-selected scheduling scheme based on a three-level linkage AI model: Configure a three-level AI scheduling model that shares the dynamic workflow data map and operates in a coordinated manner. The three-level AI scheduling models share parameters and operate in a coordinated manner rather than independently. Among them, the first-level AI big model coordinates all software business tasks across the entire domain, and outputs an initial task allocation scheme with multi-dimensional optimization goals such as minimum completion time of the entire domain business, minimum server power consumption, and minimum hardware resource occupation. The second-level lightweight AI mini-model predicts short-term load fluctuations on a single server and performs load fine-tuning and optimization on the initial task allocation scheme to avoid the risk of single server overload. The third-level verification AI model verifies task call permissions, interface connectivity protocols, and circuit breaker risk control rules, eliminates illegal scheduling paths and unsafe task allocation paths, and finally outputs a compliant and usable pre-selected scheduling scheme. S7, Adaptive scheduling switching of working conditions and dynamic adjustment of AI model weights: Real-time monitoring of three types of parameters: business traffic, number of tasks to be scheduled, and matching degree of upstream and downstream software tasks, and adaptive identification of three types of operating conditions: daily low load, business peak, and server failure. And match and switch the corresponding exclusive scheduling strategy: under low load conditions, the tasks are executed in a distributed parallel scheduling manner; time-bound tasks are executed in a fixed time-sequence scheduling manner; under server failure conditions, the faulty tasks are executed in a cross-machine migration scheduling manner. Simultaneously, based on the dynamic changes in on-site working conditions, the internal scheduling and judgment weights of the three-level linkage AI model are adaptively adjusted. S8. Scheduling, execution, and AI model closed-loop iterative optimization: Implement full-domain software tasks based on compliant pre-selected scheduling schemes, and monitor the task scheduling and operation status in real time throughout the process; If monitoring detects scheduling anomalies such as task timeout, disconnection of upstream and downstream task links, or exhaustion of server resources, the three-level linkage AI model will trace back the previous scheduling node, reallocate computing power, and complete emergency remedial scheduling. If there are no abnormal scheduling events during the entire global scheduling process, then all the working condition data of this global scheduling will be collected, and the scheduling judgment parameters of the three-level linkage AI model will be iteratively optimized. Among them, the three-level linkage AI model optimizes scheduling parameters through bidirectional interaction, constrains the local optimal scheduling logic of the model, and avoids scheduling conflicts where the single server task scheduling is optimal, but the overall operating efficiency of the entire business is degraded.

2. The AI-enabled intelligent scheduling method for software workflows according to claim 1, characterized in that, In step S1, the full amount of software operation data includes basic software task data, server hardware operation data, business flow data, interface interaction data, and historical scheduling log data. The data acquisition module adopts a distributed acquisition architecture, which is deployed on each server node, software running terminal and business interaction node to achieve synchronous acquisition of data across the entire domain.

3. The AI-enabled intelligent scheduling method for software workflows according to claim 1, characterized in that, In step S2, the data noise reduction and removal specifically includes deleting duplicate data, filtering null data, clearing abnormal garbled data, and filtering out invalid data in temporary cache. After noise reduction is completed, the valid data is processed to unify and standardize the format.

4. The AI-enabled intelligent scheduling method for software workflows according to claim 1, characterized in that, In step S3, the preset data attribute types include four categories: task type attribute, data source attribute, business affiliation attribute, and running status attribute. The valid running data is divided into corresponding data storage partitions for collection and management according to the above attributes.

5. The AI-enabled intelligent scheduling method for software workflows according to claim 1, characterized in that, In step S4, the effective operational data is divided into three levels according to the degree of business impact: Level 1 high priority, Level 2 medium priority, and Level 3 low priority. High priority corresponds to core urgent business tasks, medium priority corresponds to regular business tasks, and low priority corresponds to non-urgent tasks such as backend operation and maintenance and log statistics.

6. The AI-enabled intelligent scheduling method for software workflow according to claim 1, characterized in that, In step S5, the dynamic workflow data graph is constructed in the form of graphical nodes and associated edges. Nodes correspond to a single software task and server node, and associated edges are labeled with task sequence relationships, binding constraints, scheduling priorities and historical anomaly information. The graph is updated in real time at fixed time intervals.

7. The AI-enabled intelligent scheduling method for software workflows according to claim 1, characterized in that, In step S6, the three-level AI scheduling models share the same data interface to read the dynamic workflow data map, and a dedicated parameter transmission channel is set up between the models. The initial task allocation scheme, load fine-tuning data, and compliance verification results are synchronized bidirectionally in real time.

8. The AI-enabled intelligent scheduling method for software workflow according to claim 1, characterized in that, In step S7, the specific scheduling strategies corresponding to the three types of operating conditions are as follows: low load conditions execute distributed parallel scheduling of tasks; time-bound tasks execute fixed time-series scheduling; and server failure conditions execute cross-machine migration scheduling of faulty tasks.

9. The AI-enabled intelligent scheduling method for software workflow according to claim 1, characterized in that, In step S7, for peak business conditions, the overall weight of the first-level AI large model is increased; for server failure conditions, the running weights of the second-level lightweight AI small model and the third-level verification AI model are increased.

10. The AI-enabled intelligent scheduling method for software workflows according to claim 1, characterized in that, In step S8, the model closed-loop iterative optimization adopts an incremental parameter update method, which only corrects the model judgment rules based on the working condition data generated in this scheduling, retains the original mature scheduling logic of the model, and avoids scheduling conflicts that result in local scheduling optimization of a single server and a decrease in the overall business operation efficiency.

Images