Man-machine co-creation enterprise business process management system

Abstract

The invention provides a man-machine co-creation enterprise business process management system, belongs to the field of enterprise business process management, and is used for solving the problems of unreasonable task allocation, insufficient process monitoring and difficulty in knowledge reuse of a traditional system. The system comprises a task triggering and analyzing module, a disassembling and planning module, a workflow generating and optimizing module and the like. A task is obtained and analyzed through a task triggering and analyzing module, the task is disassembled by utilizing multiple strategies, workflow is determined through multi-objective optimization, man-machine execution subtasks are coordinated, the process is monitored in real time and dynamically adjusted, and meanwhile service data are collected to optimize AI capacity and accumulate enterprise knowledge. The system realizes full-life-cycle intelligent management of the business process from initiation, execution to optimization, improves the business execution efficiency and resource utilization rate of enterprises, and enhances the adaptability and core competitiveness of the enterprises to business changes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of enterprise business process management, and in particular to a human-computer co-creation enterprise business process management system. BACKGROUND

[0002] In enterprise business process management, traditional systems usually solidify business processes, lack adaptability to dynamic changes in business, and are difficult to achieve efficient collaboration between human employees and AI employees. Existing systems either rely too much on manual operations, lack intelligence in task decomposition and distribution, resulting in low execution efficiency; or only support simple AI task processing, cannot allocate tasks differently according to the capabilities of AI and humans, and cause resource waste. At the same time, traditional systems lack real-time monitoring and dynamic adjustment mechanisms during process execution, and cannot optimize the process in time when there are delays in progress or resource conflicts. In addition, enterprise business knowledge is difficult to effectively deposit and reuse, resulting in an increase in repetitive work and a lack of innovation capability. How to build a business process management system with intelligent task management, efficient human-computer collaboration, dynamic monitoring and adjustment, and continuous optimization of knowledge has become a problem to be solved. SUMMARY

[0003] The present application provides a human-computer co-creation enterprise business process management system, which can realize intelligent and collaborative management of enterprise business processes, and solve the problems of unreasonable task allocation, insufficient process monitoring and difficult knowledge reuse in traditional systems.

[0004] In a first aspect, the present application provides a human-computer co-creation enterprise business process management system. The system comprises a task triggering and analyzing module, which is configured to obtain task information manually input by human employees or task events automatically generated by AI employees monitoring enterprise information system data, and analyze the task information using natural language processing technology; a task decomposition and planning module, which is connected to the task triggering and analyzing module and configured to decompose the task according to the analyzed task information using a function decomposition method, a role decomposition method or a time sequence decomposition method, determine the execution order, dependency relationship, required resources and time of subtasks; a workflow generation and optimization module, which is connected to the task decomposition and planning module and configured to generate a candidate workflow scheme based on the subtask information, determine an optimal workflow scheme by minimizing the total number of subtasks, the number of subtasks requiring human employees to execute and the efficiency of workflow execution history using a multi-objective optimization algorithm, and combining basic recommendation degree calculation and recommendation degree influence coefficient calculation; a task execution and collaboration module, which is connected to the workflow generation and optimization module and configured to coordinate human employees and AI employees to execute subtasks according to the optimal workflow scheme, automatically trigger AI execution when the conditions for AI executable subtasks are complete, push subtasks requiring human execution, and support human-computer interaction and task result delivery; a monitoring and adjustment module, which is connected to the task execution and collaboration module and configured to monitor the progress deviation rate and resource load rate of business process execution in real time, and perform subtask redistribution, parallelism adjustment or process branch switching when the preset threshold is exceeded; a knowledge management and learning module, which is connected to the task execution and collaboration module and the monitoring and adjustment module, and configured to collect, organize and analyze business process data, improve the task processing capability of AI employees, provide knowledge retrieval for human employees, and update business process templates and rules.

[0005] By adopting the above technical solutions, a system framework covering the whole life cycle management of tasks is constructed. The task triggering and analyzing module realizes diversified initiation and accurate analysis of tasks; the task decomposition and planning module combines multiple strategies for scientific decomposition; the workflow generation and optimization module determines the optimal execution path using a multi-objective algorithm; the task execution and collaboration module ensures efficient human-computer collaboration; the monitoring and adjustment module ensures the stability of process execution; and the knowledge management and learning module promotes enterprise knowledge accumulation and capability improvement. The modules work together to realize intelligent and collaborative management of business processes.

[0006] Further, in the task triggering and analyzing module, the data monitored by the AI employees includes order quantity in the ERP system, inventory level, customer complaint rate in the CRM system and customer addition quantity.

[0007] By adopting the above technical solutions, the system can perceive real-time changes in key data in enterprise operation, timely trigger corresponding business tasks, and enhance the response capability of the system to business dynamics.

[0008] Furthermore, the functional decomposition method divides the task into multiple sub-tasks according to business functions; the role decomposition method allocates sub-tasks based on the ability advantages of human employees and AI employees; and the time-series decomposition method divides the task into pre-task, core sub-task, and post-task according to the order of task execution.

[0009] By adopting the above technical solutions, a variety of strategies are provided for task decomposition, which can flexibly and reasonably decompose tasks into sub-tasks according to different business characteristics and needs, thereby improving the scientific nature of task planning.

[0010] Furthermore, the basic recommendation degree calculation formula is as follows: ,in, These are the weighting coefficients. For workflow The number of subtasks, The minimum number of subtasks among all candidate workflows. For workflow The number of subtasks that require human execution. The minimum number of human tasks among all candidate workflows.

[0011] By adopting the above technical solutions, the number of sub-tasks and human tasks can be quantitatively calculated, providing an objective basis for the preliminary evaluation of the workflow and making the selection of workflow solutions more reasonable.

[0012] Furthermore, the recommendation influence coefficient includes a historical execution time average correction coefficient. and time series trend adjustment coefficient This represents the average historical execution time of the workflow. As the industry benchmark time, Determined based on the slope of linear regression of historical execution time series.

[0013] By adopting the above technical solutions and introducing historical execution efficiency data and trend analysis, the workflow recommendation can reflect the dynamic changes in actual execution results, prioritize workflows with better efficiency performance and an improving trend, and improve the reliability of workflow solutions.

[0014] Furthermore, the conditional completeness of the AI-executable subtask means that the data input is complete, the required resources are available, and the AI ​​employee's analysis efficiency value for the subtask is higher than the task processing efficiency value of the human employee.

[0015] By adopting the above technical solutions, the specific conditions for AI to automatically execute sub-tasks are clarified, ensuring that AI executes tasks in appropriate scenarios, avoiding resource waste and execution errors, and improving the efficiency and accuracy of human-machine collaboration.

[0016] Further, in the monitoring and adjusting module, the calculation formula of the progress deviation rate is The resource load rate is determined by calculating the number of sub-tasks processed by human workers at the same time and the AI computing resource utilization.

[0017] By adopting the above technical solution, a quantitative index calculation method is provided for process monitoring, which facilitates the system to accurately judge the process execution state and provides data support for subsequent adjustment.

[0018] Further, the preset threshold is that the progress deviation rate exceeds 10%, or the number of sub-tasks processed by human workers at the same time exceeds 3, or the AI computing resource utilization exceeds 80%.

[0019] By adopting the above technical solution, the triggering condition of process adjustment is clarified, so that the system can intervene in time when the process appears abnormal, and the smooth operation of the business process is ensured.

[0020] Further, the knowledge management and learning module optimizes the data cleaning algorithm and risk assessment model of AI workers according to historical data and real-time feedback by using machine learning algorithm.

[0021] By adopting the above technical solution, the AI capability is continuously improved, so that the AI worker can better adapt to business demands and improve the quality and efficiency of task processing.

[0022] Further, the task execution and collaboration module pushes the sub-task information to human workers, including sub-task description, deadline, associated historical cases, reference scheme provided by AI workers, and data support link.

[0023] By adopting the above technical solution, rich information support is provided for human workers to execute tasks, reducing information asymmetry, helping human workers to quickly and accurately complete tasks, and improving task execution efficiency.

[0024] In summary, the present application at least contains the following beneficial effects:

[0025] 1. A full-process intelligent human-machine collaborative business process management system is provided, which improves the business execution efficiency and resource utilization rate of enterprises;

[0026] 2. Through multi-dimensional task decomposition and multi-target workflow optimization, scientific planning and efficient execution of tasks are realized;

[0027] 3. With real-time monitoring and dynamic adjustment, stable operation of business processes is ensured, and through knowledge management, the continuous improvement of enterprise capability is promoted.

[0028] It is to be understood that the description in the summary is not intended to identify key or essential features of embodiments of the application, nor is it intended to limit the scope of the application. Other aspects of the application will be evident from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0029] The above and other features, aspects, and advantages of embodiments of the present application will become more apparent from the following description in conjunction with the accompanying drawings. In the drawings, like reference numerals refer to like elements, wherein:

[0030] Figure 1 A schematic diagram of an architecture of a human-computer co-creation enterprise business process management system in embodiments of the present application is shown;

[0031] Figure 2 A schematic diagram of a human-computer co-creation enterprise business process task division method executed by a task decomposition and planning module in embodiments of the present application is shown;

[0032] Figure 3 A schematic diagram of a human-computer co-creation enterprise business process intelligent orchestration engine tuning method executed by a knowledge management and learning module in embodiments of the present application is shown. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions, and advantages of embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in embodiments of the present application with reference to the drawings in embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0034] In addition, the term "and / or" herein is merely a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects.

[0035] The present application provides a human-computer co-creation enterprise business process management system, which integrates task triggering, decomposition, optimization, execution, monitoring, and knowledge management function modules, realizes intelligent and collaborative management of enterprise business processes, and significantly improves process execution efficiency and enterprise resource utilization.

[0036] In a first aspect, embodiments of the present application disclose a human-computer co-creation enterprise business process management system. The system can be implemented as a server or included in a server.

[0037] Figure 1 A module architecture diagram of the human-machine co-creation enterprise business process management system in the embodiments of the present application is shown. The system includes a task triggering and analysis module, a task decomposition and planning module, a workflow generation and optimization module, a task execution and collaboration module, a monitoring and adjustment module, and a knowledge management and learning module.

[0038] The task triggering and analysis module serves as the entrance of the system and undertakes the dual functions of task information acquisition and analysis. At the hardware level, it is deployed in the enterprise server cluster, establishes a TCP / IP connection with ERP, CRM, OA and other information systems through the enterprise intranet, and realizes real-time data acquisition; at the software level, a microservice architecture is built using the SpringCloud framework to support distributed data processing.

[0039] There are two ways to obtain tasks, one is manual triggering, and the other is automatic triggering. Among them, manual triggering includes that human employees input the task name (such as “Double 11 promotion activity planning”), the project it belongs to, the detailed description (including activity target, budget, expected effect, etc.) and the expected completion time in the task creation window through the system interface of the Web or mobile end. The front end of the system uses the Vue.js framework to realize the interactive interface, and transmits the data asynchronously to the back-end server through AJAX technology. Automatic triggering includes that AI employees monitor enterprise information systems in real time through preset data monitoring rules. For example, a rule is set in the ERP system: when the order volume exceeds 200% of the daily order volume within 1 hour, an “order peak processing” task is automatically generated; in the CRM system, if the customer complaint rate increases by 50% within 24 hours, the “customer complaint emergency handling” task is triggered. The rule engine uses the Drools framework to realize dynamic addition and modification of rules.

[0040] The specific task analysis process includes using NLTK (Natural Language Toolkit) and a custom business dictionary to perform word segmentation, part-of-speech tagging, and named entity recognition on the task description. For example, input “formulate Q3 quarter electronic product sales promotion plan”, the system can extract “Q3 quarter” (time), “electronic product” (business object), “sales promotion plan” (task type) and other key information, and match them with the preset business template library. If the “sales promotion” template is matched, the standard process nodes and resource requirements in the template are extracted as references.

[0041] Using NLTK (Natural Language Toolkit) and a custom business dictionary, the system performs word segmentation, part-of-speech tagging, and named entity recognition on task descriptions. For example, when given the input "Develop a Q3 quarterly sales and promotion plan for electronic products," the system can extract key information such as "Q3 quarter" (time), "electronic products" (business object), and "sales and promotion plan" (task type), and match it with a preset business template library. If a "sales and promotion" template is matched, the system extracts the standard process nodes and resource requirements from the template as a reference.

[0042] The task decomposition and planning module receives output information from the task triggering and parsing module, decomposes the task into executable subtasks using various strategies, and plans the execution details. Developed in Java, the module employs multithreading technology to accelerate processing and interacts with the human resources system and AI capability assessment system via a RESTful API.

[0043] In the task decomposition and planning module, there are three specific decomposition strategies: functional decomposition, role decomposition, and time-series decomposition.

[0044] The system employs several methods for task allocation. The first method, functional decomposition, breaks down tasks based on business function modules. For example, the "order processing" task is divided into sub-tasks such as "order receipt," "inventory query," "price calculation," and "logistics allocation." Each sub-task corresponds to a function; for instance, the inventory query sub-task calls the enterprise ERP system's inventory query API, obtaining real-time inventory data by passing in the order's product ID. The second method, role-based decomposition, allocates tasks based on a capability matrix of human and AI employees. The system maintains a dynamic capability matrix table, recording each employee / AI's capability score (1-10 points) in dimensions such as data processing, customer communication, and risk assessment. For example, the "contract terms review" task's "initial screening for compliance" (with clearly defined rules) is assigned to an AI employee (capability score of 9 points), while "business risk assessment" (requiring experience-based judgment) is assigned to a human employee with legal and business backgrounds (capability score of 8 points). The third method, temporal decomposition, divides tasks according to their execution time sequence. For example, the "Project R&D" task is broken down into "Requirements Research (Days 1-5)", "Prototype Design (Days 6-10)", "Code Development (Days 11-30)" and "Testing and Acceptance (Days 31-35)", with each sub-task having a clear start time, end time and milestone node.

[0045] In the task decomposition and planning module, dependency relationships and resource planning use an adjacency matrix to describe subtask dependencies, with the matrix dimension being the number of subtasks. For example, the dependency matrix for 3 subtasks is... , which means that subtask 1 is the predecessor of subtask 2, and subtask 2 is the predecessor of subtask 3. At the same time, according to historical data and resource occupation models, the resources required by each subtask, such as manpower, computing power, and storage space, are estimated, such as the code development subtask is expected to require 3 programmers and 5 server computing resources.

[0046] The workflow generation and optimization module is used to generate candidate workflow schemes and perform multi-objective optimization based on the task decomposition results. The core algorithm is written in Python, using Pandas for data processing and Scikit-learn for machine learning model training.

[0047] The candidate workflow generation process includes generating multiple workflow schemes based on the combination of decomposed subtasks. For example, the "customer order processing" task can generate scheme A: AI order cleaning Human exception handling AI delivery order generation; scheme B: AI order cleaning AI inventory verification Human secondary review AI delivery order generation. Each scheme is stored in the form of a directed acyclic graph (DAG), with subtasks as nodes and dependencies as edges.

[0048] The multi-objective optimization process includes basic recommendation degree calculation and recommendation degree influence coefficient calculation. The basic recommendation degree is calculated according to the formula . Among them, (the number of subtasks weight), (the number of human tasks weight). For example, the of scheme A is 0.5, the of scheme B is 0.3, and the of all schemes is 0.2, then . Historical execution time average correction coefficient , such as the historical average time consumption of scheme A is 10 hours, and the industry benchmark is 12 hours, then ; timing trend adjustment coefficient is determined according to the linear regression slope of the historical execution time series. If the slope is negative (efficiency improvement), then , otherwise . The final recommendation degree is calculated by the formula , and the scheme with the highest recommendation degree is selected as the optimal workflow.

[0049] The task execution and collaboration module is responsible for driving workflow execution and coordinating human-machine task processing and result interaction. It uses message queues (RabbitMQ) to achieve task distribution and real-time state updates through WebSocket technology.

[0050] Among them, the task execution logic, the AI task execution logic includes when the subtask meets the data integrity (such as field missing rate <5%), resource available (server load <80%) and AI processing efficiency is higher than human (through historical data comparison), automatically call AI service interface execution. For example, the AI order cleaning subtask calls the self-developed data cleaning algorithm to remove duplicate, correct errors and format unified processing of original order data.

[0051] Human task pushing logic includes pushing task notification to designated employees through WeChat for enterprise, email and other channels, the content includes subtask description, deadline, associated historical case link (stored in enterprise knowledge base), AI provided reference scheme (such as data analysis report, preliminary processing result) and data support link (such as related database query page).

[0052] Human-computer interaction mechanism includes that human employees can view AI execution results on system interface, select "confirm pass", "require correction" or "manual reprocessing". If you choose "require correction", the system will return the correction opinion to the AI module, and the AI will execute again after adjusting the algorithm according to the feedback; if "manual reprocessing", the processing result of human employees will be used as new data to update AI training model.

[0053] Monitoring and adjustment module is used for real-time monitoring of process execution state, and dynamic adjustment is triggered through preset threshold. It is deployed in monitoring server, and Prometheus+Grafana is used to build monitoring platform to visualize system indicators.

[0054] Monitoring indicators include progress deviation rate and resource load rate. Among them, the progress deviation rate is calculated according to the formula When , trigger warning. For example, a subtask is planned to be completed in 2 hours, and the actual time is 2.5 hours, then Dev . Resource load rate calculates the number of subtasks processed by human employees at the same time and Al computing resource utilization rate. When the number of tasks processed by human employees at the same time is , or the CPU utilization rate of AI server is , it is judged that the resource is overloaded.

[0055] The dynamic adjustment strategy includes subtask redistribution, parallelism adjustment, and flow branch switching. Subtask redistribution includes redistributing overloaded tasks to individuals with higher scores in the same ability dimension based on the employee / AI capability matrix. For example, a certain employee's "customer complaint handling" task is assigned to other idle employees with higher communication skills scores. Parallelism adjustment includes converting serial subtasks to parallel when the dependency relationship allows. For example, "data collection" and "literature research" subtasks have no data dependency, so they can be started simultaneously, reducing overall time consumption. Flow branch switching includes automatically switching to a backup workflow template when a node has been manually reprocessed for three consecutive times and marking the original process for optimization.

[0056] The knowledge management and learning module is used to build an enterprise business knowledge base and achieve knowledge sedimentation and AI capability improvement. Elasticsearch is used to build a knowledge retrieval engine, and TensorFlow is used to implement machine learning model training. It has functions such as knowledge collection and organization, AI capability optimization, knowledge sharing and application, etc.

[0057] The knowledge collection and organization function is used to automatically collect data during task execution, including successful cases, failure lessons, optimization suggestions, etc. For example, the experience of "coupon distribution strategy adjustment to increase sales by 20%" in a certain promotion activity is organized into a knowledge card, with keywords such as "promotion activity", "coupon strategy", and "sales increase".

[0058] The AI capability optimization function is used to optimize AI algorithm models using historical data and real-time feedback. For example, by analyzing a large number of contract review cases, the rules weight of the natural language processing model is adjusted to improve the accuracy of clause compliance judgment; according to the correction data of human employees on risk assessment results, the parameters of the risk assessment algorithm are optimized.

[0059] The knowledge sharing and application function is used to provide knowledge retrieval functions for human employees, supporting multi-dimensional searches such as keywords, tags, and case types. At the same time, new knowledge is integrated into business process templates, such as updating the optimized order processing process to the system template library for subsequent task calls.

[0060] To better understand the system, the system collaborative workflow is illustrated with an example. Specifically, taking the "618 big promotion activity preparation" task of an e-commerce enterprise as an example,

[0061] Task triggering and analysis: the operation department manually inputs the task through the system interface, describing it as "plan and execute the 618 big promotion activity, target sales growth 50%, budget 2 million yuan, start on June 1st". The task triggering and analysis module parses the key information, matches the "promotion activity" template, and obtains the standard process nodes and resource requirements.

[0062] Task decomposition and planning: Adopt role decomposition method to decompose tasks into "activity plan planning (human)", "advertisement material production (human + AI)", "inventory preparation (AI)", "traffic estimation and distribution (AI)", etc. Subtasks, and determine the dependency relationship and time plan.

[0063] Workflow generation and optimization: Generate three workflow schemes, select the scheme with the least total number of subtasks, reasonable proportion of human tasks, and high historical execution efficiency through multi-objective optimization calculation: AI inventory preparation → human activity plan planning → AI traffic estimation and distribution → human + AI advertisement material production.

[0064] Task execution and collaboration: The system automatically starts the AI inventory preparation task, calls the ERP system data to calculate the required commodity inventory, and at the same time pushes the activity plan planning task to the planner. After the planner completes the plan, the AI performs traffic estimation and outputs an analysis report for human review and confirmation.

[0065] Monitoring and adjustment: During the execution process, the monitoring module finds that the advertisement material production progress is lagging behind, with a progress deviation rate of 15%, immediately redistributes part of the design task to other designers, and adjusts the serial "material review" and "modification" subtasks to parallel, shortening the overall cycle.

[0066] Knowledge management and learning: After the activity is completed, the knowledge management and learning module collects activity data, organizes the experience of "precise traffic distribution strategy to improve conversion rate by 30%" into the database, and optimizes the traffic prediction algorithm model according to the feedback of humans on the AI traffic estimation results.

[0067] Additionally, the task decomposition and planning module can also mark whether a subtask is more suitable for human employees or AI employees. The following will explain in detail with a specific task division process.

[0068] Figure 2 A flowchart of a human-machine co-creation-oriented enterprise business process task division method executed by the task decomposition and planning module in the embodiments of the present application is shown. The method specifically includes the following steps:

[0069] S21: Receive enterprise business process arrangement tasks and decompose them into multiple subtasks.

[0070] In the enterprise business process management system, the task receiving module listens to business process arrangement requests from various departments of the enterprise in real time, such as market research tasks, order processing tasks, project planning tasks, etc. When a specific task is received, first analyze the task description through natural language processing (NLP) technology, extract task keywords, targets, constraint conditions, etc. Information, build a task initial information table.

[0071] Subsequently, the task decomposition module decomposes the overall task into multiple subtasks according to the task complexity using a recursive decomposition algorithm. For example, the "annual marketing activity planning" task can be decomposed into "market research analysis", "activity scheme design", "resource allocation management", "effectiveness evaluation feedback", and other subtasks. During the decomposition process, the principle of "functional independence and clear interface" is followed to ensure that each subtask has clear input and output and executability, and the subtasks are connected through data interaction interfaces to realize information transmission.

[0072] S22: For each subtask, the AI employee's AI analysis efficiency value and the human employee's task processing efficiency value facing the subtask are calculated respectively.

[0073] The method of this step specifically includes AI employee's AI analysis efficiency value calculation and human employee's task processing efficiency value calculation facing the subtask.

[0074] The calculation of the AI employee's AI analysis efficiency value facing the subtask includes two core links: basic analysis efficiency value determination and analysis capability adjustment coefficient determination.

[0075] The basic analysis efficiency value determination can use the historical efficiency method and / or the capability description method.

[0076] The historical efficiency method includes collecting efficiency data of the AI employee in completing the subtask in the past, including task execution time, data processing amount, result accuracy, etc. Taking the task execution time as an example, assuming that the historical data is , first calculate the mean and the standard deviation , and apply the 3 times standard deviation principle to eliminate abnormal values outside the range , to obtain the effective data , and the basic analysis efficiency value is .

[0077] The capability description method includes constructing an AI employee capability label system, including "data cleaning", "rule reasoning", "text generation", and other labels, each label corresponding to a basic efficiency value of 0-100 points. For the required capabilities of the subtask, the corresponding labels are matched from the AI employee capability label library, for example, the "market research analysis" subtask requires "data cleaning" and "statistical analysis" capabilities, and if the AI employee has these two labels and the assigned values are 90 points and 85 points respectively, and the weights are 0.6 and 0.4 respectively, then points.

[0078] In the method of this step, the analysis capability adjustment coefficient can be determined based on one or more of the independent completion frequency ratio, similar task association, and time sequence factor.

[0079] The independent completion frequency ratio The specific determination method includes: the statistics module recording the frequency of AI employees independently completing the sub-task. and the frequency of assignment ,when hour, For example, if assigned 100 attempts, and completes 80 independently, then... .

[0080] Similar task association ( The specific method for determining the similarity between subtasks includes: The subtask similarity calculation model is based on dimensions such as task description, required skills, and processing logic, and uses a cosine similarity algorithm to calculate the similarity between subtasks. A similarity threshold of 0.7 is set. For similar subtasks with similarities higher than the threshold, their independent completion frequency ratios are collected and weighted averaged, with the weights being the normalized results of the similarity values. For example, subtask A and subtask B... The similarities were 0.8 and 0.75, respectively. If the values ​​are 0.9 and 0.85, then... .

[0081] Time-series factors can specifically include one or more of the following: trend coefficient, volatility coefficient, and extreme value coefficient. Among these, the trend coefficient ( ) The determination method includes arranging the analytical capability adjustment coefficients by time series and obtaining the trend equation through linear regression fitting. The trend coefficient is adjusted according to the sign and magnitude of the slope b, for example... This indicates an improvement in capability, and is assigned a correction factor of 1.05; b A correction factor of 0.95 is assigned at this time.

[0082] Volatility coefficient ( ) The determination method includes calculating the standard deviation of the adjustment factor within a preset time period (e.g., 3 months). ,like If the fluctuation threshold is met, a stability correction coefficient of 1.02 is assigned; otherwise, it is 1.

[0083] Extreme value coefficient ( ) The determination method includes recording historical maximum values. and minimum value Current value near Multiplying by 0.98, it approaches... Multiply by 1.02.

[0084] The comprehensive adjustment coefficient for time series factors, determined by comprehensively considering the trend coefficient, volatility coefficient, and extreme value coefficient of time series factors ( Specifically, various factors are integrated through a correlation algorithm.

[0085]

[0086] Therefore, the efficiency value of AI analysis ( It can be calculated that,

[0087] The specific calculation process of the task processing efficiency value of human employees facing sub-tasks includes two parts: the calculation of the basic processing efficiency value and the calculation of the processing capacity adjustment coefficient.

[0088] The calculation of the basic processing efficiency value involves establishing a human employee skill matrix, covering dimensions such as "professional skills," "work experience," and "communication skills." Each dimension is assigned a matching score (0-100 points) to the skills required for the sub-task. For example, the sub-task "customer communication" requires the skills of "communication skills" and "industry knowledge." If an employee's scores for these two skills are 95 and 80 points respectively, with weights of 0.7 and 0.3, then the basic processing efficiency value is...

[0089]

[0090] The calculation of the processing capacity adjustment factor includes collecting historical efficiency data of human employees in completing the subtask and calculating the average completion time. Compared with industry benchmark time The ratio of these values ​​yields the processing capacity adjustment coefficient. (when If the average employee time is 2 hours and the industry benchmark is 3 hours, then .

[0091] Therefore, the task processing efficiency value can be calculated. .

[0092] In step S22, for the calculation of AI analysis efficiency values, the historical efficiency method and the capability description method can be used in parallel. A weighted fusion (e.g., historical efficiency method with 60% weight, capability description method with 40% weight) is used to obtain the final basic analysis efficiency value, avoiding the limitations of a single method. For example, when an AI employee first handles the "data cleaning" subtask, there is no historical data, so the capability description method can be used entirely for assignment; as historical data accumulates, the weight of the historical efficiency method is gradually increased.

[0093] In the analysis capability adjustment coefficient, the similarity calculation for the task association component requires the establishment of a multi-dimensional feature vector, including the task's input and output formats, processing logic steps, and required algorithm types. For example, the tasks "user data cleaning" and "order data cleaning" have similar input data structures and both involve deduplication and format conversion in their processing logic, resulting in a similarity score of 0.85, which is higher than the threshold of 0.7, thus allowing them to be included in the analysis of similar tasks.

[0094] For the processing capacity adjustment coefficient of human employees, industry benchmark efficiency data can be determined through industry reports, public databases or enterprise historical average data to ensure the objectivity of the evaluation. For example, the industry benchmark time for the "contract review" subtask is 4 hours, and the average completion time of an employee is 3 hours, so the adjustment coefficient is 1.33, reflecting that the processing efficiency is higher than the industry average level.

[0095] S23: Compare the AI analysis efficiency value and the task processing efficiency value, and assign the subtask to an AI employee or a human employee.

[0096] The comparison module built in the server can first determine the size relationship between the AI analysis efficiency value and the task processing efficiency value:

[0097] If , the subtask is assigned to an AI employee, who completes the task using the high efficiency of data processing and rule execution;

[0098] If , it is assigned to a human employee to take advantage of their flexible decision-making and complex problem-solving skills.

[0099] When the difference between the two is (preset efficiency difference threshold, such as 5 points), a subtask complexity factor is introduced for secondary judgment. The complexity factor is evaluated through three dimensions: processing logic complexity, required skill diversity, and task goal ambiguity, each dimension is rated on a 5-point scale (1-5 points), and the weighted average complexity value is obtained. If (complexity threshold, such as 3 points), it is preferentially assigned to a human employee; otherwise, it is assigned to an Al employee.

[0100] To achieve automated decision-making, the system builds a subtask assignment decision tree model, with inputs including subtask characteristics (complexity, required skills), AI employee historical data (independent completion frequency, processing time), and human employee historical data (completion quality, efficiency), and the output is the assignment result. Further, the random forest algorithm is used to optimize the model, and the voting mechanism of multiple decision trees is used to improve accuracy and reduce the risk of overfitting of a single model.

[0101] In the complexity factor evaluation, the processing logic complexity mainly examines whether the task has branching judgments, exception handling and other unstructured processes; the required skill diversity refers to the number of skills needed to mobilize to complete the task; and the task goal ambiguity refers to whether the goal is clear in the task description and whether there is ambiguity. For example, the "new product creative planning" task has a complexity score of 4.5 points due to ambiguous goals and the need for multi-disciplinary skills, which is higher than the threshold of 3 points, so even if the efficiency value is close, it is preferentially assigned to a human employee.​​

[0102] The training data of the decision tree model needs to include at least 3 years of historical task allocation and execution result data, and the model generalization ability is ensured through cross-validation. In the random forest algorithm, bootstrap sampling is used for training samples of each decision tree, and a random subspace method is used for feature selection to reduce the correlation between trees and improve the robustness of the model.

[0103] In combination with the above, the task decomposition and planning module establishes a task division mechanism based on human-machine efficiency comparison, quantitatively evaluates the task processing ability of AI employees and human employees through multiple dimensions, realizes scientific allocation of sub-tasks, and improves the efficiency of human-machine collaboration; Introducing historical data, ability labels, similar task associations, and timing factors, etc. Multi-dimensional parameters make the analysis ability adjustment coefficient more accurately reflect the actual performance of employees, solve the subjectivity problem of traditional allocation methods; Through the decision tree model and the random forest algorithm, automatic allocation decisions are realized, combined with the complexity factor to handle the fuzzy scene with close processing efficiency values, improve the reliability and adaptability of task allocation, and provide effective support for intelligent management of enterprise business processes.

[0104] Additionally, the knowledge management and learning module can also filter enterprise business process orchestration records that benefit the training of the system to facilitate better self-learning and optimization of the system.

[0105] Figure 3 A flowchart of a human-machine co-creation enterprise business process intelligent orchestration engine optimization method executed by the knowledge management and learning module in the embodiments of the present application is shown. The method specifically includes the following steps:

[0106] S31: Evaluate the ability level of AI employees in the AI employee library

[0107] The AI employee library stores the ability dimension data of all AI employees in the enterprise, and each AI employee has multiple ability dimensions, such as "data cleaning", "process planning", "exception handling", etc. For the first ability dimension of the first AI employee , the ability level is determined as follows: Specifically, the ability influence coefficient and the time decay factor are used to determine.

[0108] The determination method of the ability influence coefficient includes: calculating the AI analysis efficiency value, and solving the ratio, that is, . This coefficient reflects the performance of the AI employee in this ability dimension relative to the industry average level, and the value range is , the higher the value, the stronger the ability.

[0109] ​The calculation of the AI analysis efficiency value includes two core links: basic analysis efficiency value determination and analysis capability adjustment coefficient determination.

[0110] The basic analysis efficiency value can be determined by historical efficiency method and / or capability description method. The historical efficiency method includes collecting efficiency data of AI employees in completing the subtask in the past, including task execution time, data processing amount, result accuracy, etc. Taking the task execution time as an example, assuming that the historical data is , first calculate the mean and standard deviation , remove abnormal values beyond the range of , get valid data , then the basic analysis efficiency value is . The capability description method includes constructing an AI employee capability label system, including labels such as "data cleaning", "rule reasoning", "text generation", etc. Each label corresponds to a basic efficiency value of 0-100 points. For the required capabilities of the subtask, match the corresponding labels from the AI employee capability label library, for example, the "market research analysis" subtask requires "data cleaning" and "statistical analysis" capabilities, if the AI employee has these two labels and the assigned values are 90 points and 85 points respectively, the weights are 0.6 and 0.4 respectively, then points.

[0111] The analysis capability adjustment coefficient can be determined based on one or more of the independent completion frequency ratio, similar task association, and timing factor. The calculation method of the independent completion frequency ratio includes that the statistical module records the frequency of the AI employee independently completing the subtask and the assigned frequency , when , . For example, if assigned 100 times, independently completed 80 times, then . The calculation method of the similar task association ( ) includes that the subtask similarity calculation model calculates the similarity between subtasks based on task description, required skills, processing logic, etc. dimensions, using the cosine similarity algorithm. Set the similarity threshold to 0.7, for similar subtasks with a similarity higher than the threshold, collect their independent completion frequency ratios and take a weighted average, with the weight being the normalized result of the similarity value. For example, the similarity of subtask A and subtask is 0.8 and 0.75 respectively, and is 0.9 and 0.85 respectively, then . The timing factor can be calculated based on one or more of the trend coefficient, the fluctuation coefficient, and the extreme value coefficient. The trend coefficient ( ) is calculated by arranging the analysis capability adjustment coefficient in chronological order and fitting a trend equation through linear regression , Trend coefficient is adjusted according to the sign and magnitude of slope b, for example , when it indicates the ability improvement, a correction factor of 1.05 is given; b , a correction factor of 0.95 is given. Volatility coefficient ( ) The calculation method includes calculating the standard deviation of adjustment coefficient in a preset period (such as 3 months) , if (volatility threshold), a stable correction factor of 1.02 is given, otherwise 1. Extreme coefficient ( ) The calculation method includes recording the historical maximum value and minimum value , the current value is multiplied by 0.98 when it is close to , and multiplied by 1.02 when it is close to .

[0112] Based on the above, the comprehensive adjustment coefficient ( ) can be calculated by integrating various factors through association algorithm,

[0113]

[0114] Therefore, the AI analysis efficiency value can be calculated,

[0115] For the calculation of AI analysis efficiency value, historical efficiency method and ability description method can be used in parallel, and the final basic analysis efficiency value is obtained by weighted fusion (such as historical efficiency method accounts for 60% weight and ability description method accounts for 40% weight), avoiding the limitations of single method. For example, when a certain AI employee first processes the "data cleaning" subtask, there is no historical data, and the ability description method can be completely relied on for assignment; with the accumulation of historical data, the weight of historical efficiency method is gradually increased.

[0116] In the analysis of the same task correlation part in the ability adjustment coefficient, the similarity calculation needs to establish a multi-dimensional feature vector, including task input and output format, processing logic steps, required algorithm type, etc. For example, "user data cleaning" and "order data cleaning" tasks, because the input data structure is similar, the processing logic includes deduplication and format conversion, the similarity can be determined as 0.85, which is higher than the threshold value 0.7, so it is included in the same task analysis.

[0117] The calculation of time decay factor considers the generation time and the last tuning interval of the ability dimension, and adopts a logistic function combined with time normalization processing:

[0118]

[0119] wherein, is the decay rate parameter (default 0.3), is the preset decay threshold time (default 6 months). This factor embodies the decay effect of the capability over time, and the degree of historical accumulation of the capability dimension.

[0120] Comprehensive adjustment coefficient The determination of is manually adjusted by the enterprise according to the business strategy, for example, a positive adjustment (such as 1.2) is given to the capability dimension related to the core business, and a negative adjustment (such as 0.8) is given to the marginal capability, and the value range is .

[0121] The final capability level formula is .

[0122] The system automatically updates the AI employee's capability level once a month, and generates a capability matrix table, for example:

[0123] AI Employee ID Ability Dimension Ability Influence Coefficient Time Decay Factor Comprehensive Adjustment Coefficient Ability Level Al－001 Data Cleaning 1.8 0.92 1.0 1.66 Al－001 Exception Handling 0.7 0.85 1.5 0.89

[0124] In step S31, considering the dynamic adjustment of the capability dimension evaluation, specifically, the "enterprise historical data period" in the time decay factor is usually set as the complete accumulation period of the enterprise business process data (such as 3 years), to ensure that the decay calculation of new capability dimensions is based on a reasonable time range. For example, the time normalization factor will be reduced accordingly when the newly deployed "intelligent prediction" capability dimension has been generated for less than 1 year, to avoid the virtual high capability level caused by insufficient data accumulation.

[0125] S32: Determine the optimization value degree of the subtask.

[0126] The optimization value degree of the subtask is determined by the basic optimization value degree and the correction coefficient.

[0127] wherein, the calculation process of the basic optimization value degree includes: for each subtask in the enterprise business process arrangement record , first analyze the set of capability dimensions involved , and calculate the basic optimization value degree by the following steps : correlation complexity evaluation, scoring from three dimensions of processing logic complexity (1-5 points, the more complex the logic, the higher the score), required skill diversity (number of skill types, plus 1 point for each additional type), and task goal ambiguity (1-5 points, the more ambiguous the goal, the higher the score), and taking the average value as TaskComplexity . Capability dimension weight, according to the business focus of the enterprise, assign a weight (totaling 1) to each capability dimension, for example, "data cleaning" weight 0.3, "process planning" weight 0.4. Finally determine the basic optimization value degree

[0128] The generation process of the correction coefficient includes calculating the average level of each ability dimension in the Al employee library , and the target ability level . Comparison:

[0129]

[0130] If the average level of a certain ability dimension is lower than the target , the correction coefficient is greater than 1, which strengthens the tuning value of this dimension; otherwise, it remains 1.

[0131] The tuning value degree is corrected by using the correction coefficient to obtain the tuning value degree , for example, a sub-task involves the "exception handling" ability ( ), and the basic value degree is 10, then the corrected value degree is 12.

[0132] S33: Screen the enterprise business process orchestration record set.

[0133] The method of this step uses a multi-objective optimization function to screen the enterprise business process orchestration record set.

[0134] Balancing the number of records, negative tuning value, and ability dimension satisfaction is required to construct the objective function, as follows,

[0135]

[0136] Among them, is the number of records, is the control of the negative value weight, is the control of the ability satisfaction weight; is the negative tuning value, and the record with the proportion of abnormal process exceeding is marked as negative, and the value degree is set to -5; is the ability dimension tuning value degree target, which is determined according to the enterprise strategy (such as the core ability target is 1000, and the ordinary ability is 500).

[0137] Use the greedy algorithm to iterate and select, initialize, candidate record set , and the current tuning value of the ability dimension ; iterate and select, each time select the record that makes decrease the fastest from the remaining records, and update ; the termination condition is that all ability dimension tuning values reach the target , or the record set size reaches the preset upper limit (such as 500 records). During the screening process, the records with are automatically removed, and the records are sorted according to Sort in descending order, prioritizing records with higher value.

[0138] In steps S32 and S33, the multi-dimensional correlation of value can be optimized. For example, in the evaluation of the correlation complexity between subtasks and ability dimensions, the processing logic complexity is determined by the number of condition judgments and abnormal branches in the parsing process nodes; the diversity of required skills is directly counted by the number of skill tags mentioned in the subtask description; the ambiguity of the task objective is analyzed by natural language processing technology to analyze the ambiguity of the task description, such as using the BERT model to calculate the text entropy value, the higher the entropy value, the higher the ambiguity.

[0139] In step S33, an optimized implementation of the screening algorithm can be considered. For example, in the implementation of the greedy algorithm, a priority queue is used to store the "capacity gap reduction contribution value" of the candidate records, and the time complexity of each iteration is O(n log n). (M is the number of remaining records). To avoid getting trapped in local optima, a global scan is performed every 50 iterations to ensure that high-value records are not missed.

[0140] S34: Optimize using a filtered record set.

[0141] By using the filtered record set, i.e. the enterprise business process orchestration record set obtained through filtering, the enterprise business process intelligent orchestration engine can be optimized. The optimization method can adopt existing methods, or the following optimizations can be performed.

[0142] Consider introducing an incremental tuning mechanism, with the initial tuning using a full screening of the entire record set. The AI ​​employee is optimized using supervised learning algorithms (such as gradient boosting trees). The input consists of sub-task features and capability dimension labels, and the output is process orchestration decisions. Continuous optimization is performed, with adjustments made monthly based on the aforementioned filtering algorithm after each new business record is added, while retaining historically valid records. For the current mean portion, avoid repeatedly optimizing low-value data.

[0143] For evaluating the optimization effect, after optimization, the improvement effect is verified by indicators such as the compliance rate of the following capability dimensions, optimization efficiency, and the proportion of negative value. Among them, the compliance rate of capability dimensions is... ; The capability improvement value brought by optimizing the number of records per unit of efficiency Negative value percentage .

[0144] For the whole tuning method, consider the version control of incremental tuning, for example, the system generates a unique identifier for each tuning version, records the tuning time, uses the record set, and changes the ability dimension information. When an abnormal decrease in a certain ability dimension is found after tuning, it can quickly roll back to the previous version and analyze the negative data impact of the new records to achieve traceability of the tuning process.

[0145] Based on the above, the knowledge management and learning module establishes a targeted tuning mechanism based on AI employee ability gap, accurately identifies tuning needs through multi-dimensional ability assessment and dynamic correction coefficient, and improves the support ability of AI employees for the core business of the enterprise; An intelligent screening algorithm that takes into account the number of records, negative value and ability satisfaction is proposed to avoid the resource waste of traditional full tuning, reducing the tuning data size by 30%-50% while maintaining the tuning effect; Design of incremental tuning and version control mechanism to support continuous optimization and traceability of AI employee ability, ensure long-term adaptation to dynamic changes in business processes, and significantly improve the industry adaptability and decision reliability of the intelligent orchestration engine.

[0146] It should be noted that for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0147] In summary, the present application at least contains the following beneficial effects:

[0148] 1. Through multi-module cooperation, the whole life cycle intelligent management of enterprise business processes is realized, and compared with traditional systems, the process execution efficiency is improved by more than 30% on average;

[0149] 2. Based on multi-objective optimization and human-machine ability matrix task allocation strategy, reduce 20%-40% of human resource waste and task processing time;

[0150] 3. Dynamic monitoring and knowledge accumulation mechanism enables the system to quickly adapt to business changes while continuously accumulating the core competitiveness of the enterprise, providing data-driven support for process optimization.

[0151] The above description is merely exemplary of the application and of the application of the principles thereof and the application is not limited to the disclosed technical features or combinations thereof. It is intended to be apparent to one skilled in the art that the scope of the disclosure is not limited to the technical solutions formed by the specific combinations of the technical features disclosed above, and also includes other technical solutions formed by the combinations of the technical features disclosed above or their equivalent features without departing from the above disclosed concept. For example, technical solutions formed by the mutual replacement of the above features and technical features with similar functions disclosed in the application (but not limited to) are also included.

Claims

1. A human-machine co-creation enterprise business process management system, characterized in that, include: The task triggering and parsing module is used to obtain task information manually entered by human employees or task events automatically generated by the AI ​​employee monitoring enterprise information system, and to parse the task information using natural language processing technology. The task decomposition and planning module, connected to the task triggering and parsing module, is used to decompose the task according to the parsed task information, using the functional decomposition method, role decomposition method, or time sequence decomposition method, and determine the execution order, dependencies, and required resources and time of the sub-tasks. The workflow generation and optimization module, connected to the task decomposition and planning module, is used to generate candidate workflow schemes based on subtask information. Through a multi-objective optimization algorithm, with the objectives of minimizing the total number of subtasks, minimizing the number of subtasks that need to be executed by human employees, and maximizing the historical efficiency of workflow execution, the optimal workflow scheme is determined by combining basic recommendation degree calculation and recommendation degree influence coefficient calculation. The task execution and collaboration module is connected to the workflow generation and optimization module. It is used to coordinate human employees and AI employees to execute sub-tasks according to the optimal workflow scheme, automatically trigger AI execution when the conditions for AI to execute sub-tasks are complete, push sub-tasks that need to be executed by humans, and support human-computer interaction and task result transmission. The monitoring and adjustment module is connected to the task execution and collaboration module and is used to monitor the progress deviation rate and resource load rate of the business process in real time. When the preset threshold is exceeded, the sub-task is reassigned, the parallelism is adjusted, or the process branch is switched. The knowledge management and learning module, connected to the task execution and collaboration module and the monitoring and adjustment module, is used to collect, organize and analyze business process data, improve the task processing capabilities of AI employees, provide knowledge retrieval for human employees, and update business process templates and rules.

2. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, In the task triggering and parsing module, the data monitored by the AI ​​employee includes order volume and inventory level in the ERP system, and customer complaint rate and new customer number in the CRM system.

3. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, The functional decomposition method divides tasks into multiple sub-tasks according to business functions; the role decomposition method allocates sub-tasks based on the ability advantages of human employees and AI employees; the time-series decomposition method divides tasks into pre-tasks, core tasks, and post-tasks according to the order of task execution.

4. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, The basic recommendation degree calculation formula is as follows: ,in, These are the weighting coefficients. For workflow The number of subtasks, The minimum number of subtasks among all candidate workflows. For workflow The number of subtasks that require human execution. The minimum number of human tasks among all candidate workflows.

5. The human-machine co-creation enterprise business process management system according to claim 4, characterized in that, The recommendation influence coefficient includes a historical execution time average correction coefficient. and time series trend adjustment coefficient This represents the average historical execution time of the workflow. As the industry benchmark time, Determined based on the slope of linear regression of historical execution time series.

6. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, The conditions for AI to execute sub-tasks are complete, meaning that the data input is complete, the required resources are available, and the AI ​​employee's analysis efficiency value for the sub-task is higher than the task processing efficiency value of the human employee.

7. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, In the monitoring and adjustment module, the formula for calculating the progress deviation rate is: Resource load rate is determined by calculating the number of subtasks processed simultaneously by human employees and the AI ​​calculation resource utilization rate.

8. The human-machine co-creation enterprise business process management system according to claim 7, characterized in that, The preset thresholds are: a progress deviation rate exceeding 10%, or a human employee handling more than 3 sub-tasks simultaneously, or an AI computing resource utilization rate exceeding 80%.

9. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, The knowledge management and learning module utilizes machine learning algorithms to optimize the data cleaning algorithm and risk assessment model for AI employees based on historical data and real-time feedback.

10. The human-machine co-creation enterprise business process management system according to claim 1, characterized in that, The sub-task information pushed to human employees by the task execution and collaboration module includes sub-task description, deadline, associated historical cases, reference solutions provided by AI employees, and data support links.

Images