Intelligent task processing method and system based on digital employee

Abstract

The application discloses an intelligent task processing method and system based on digital employees, and belongs to the technical field of intelligent task processing. A sequence prediction model is used to extract user historical operation behaviors to generate a preference feature vector. Based on the preference feature vector, semantic analysis is performed on a natural language instruction to determine a personalized execution configuration containing a priority and a time window. Then, a basic operation sequence is generated by combining an intention splitting task. If there is a cross-application operation, an optimal calling path is solved in an automatic integration library by using a graph association model. Finally, a task is completed by simulating manual operation, and the weight of the preference feature vector is dynamically adjusted according to a real-time execution state. The application realizes highly personalized cross-application task automatic execution, improves the accuracy of instruction analysis, and continuously enhances the adaptive ability of automatic operation through a closed-loop feedback mechanism.

Description

Technical Field

[0001] This invention relates to the field of intelligent task processing technology, and in particular discloses an intelligent task processing method and system based on digital employees. Background Technology

[0002] Intelligent task processing technology can effectively reduce the investment of office manpower and improve business operation efficiency. A search revealed that existing related patent applications (CN112783154A and CN113986561A) disclose intelligent task processing solutions that mostly rely on external network interfaces for command interaction. They only possess basic functions such as question answering and text generation, lack user behavior sequence prediction models, semantic parsing for personalized configuration, task splitting and scheduling, cross-application path optimization of graph association models, and a local closed-loop preference feature dynamic update mechanism.

[0003] Existing technologies have many shortcomings and derivative problems: First, they rely on external network resolution and scheduling, limiting local software operation permissions and making it impossible to execute automated operations continuously across applications; second, they fail to mine user operating habits, task frequency, and other preference characteristics, resulting in task execution that does not align with usage habits and has poor adaptability; third, they lack the ability to intelligently split tasks and solve for optimal call paths, leading to low efficiency in cross-software collaborative processing; and fourth, the transmission of behavioral data and operation trajectories poses a privacy risk, making it difficult to achieve local closed-loop autonomous operation of the device and failing to meet the needs of unattended, all-weather intelligent operation.

[0004] Therefore, the key issue in realizing all-weather intelligent task processing is how to enable the system to directly simulate human operation of various software on a local computer without relying on external network resolution, and to ensure that all preference memories and execution trajectories are completely closed within the user's own device. Summary of the Invention

[0005] This invention provides an intelligent task processing method and system based on digital employees, aiming to solve at least one of the defects existing in the prior art.

[0006] One aspect of the present invention relates to an intelligent task processing method based on digital employees, comprising the following steps: S100. Obtain the user's historical operation behavior sequence, use a sequence prediction model to extract features from the user's historical operation behavior sequence, and generate a user preference feature vector. The user preference feature vector includes quantitative representation data of operation habits, task frequency, and time distribution characteristics. S200. Based on the user preference feature vector, perform semantic parsing on the received natural language task instructions, identify the target task intent and determine personalized execution configuration information, including the mapping relationship between task priority, trigger time window and target operation object; S300: Combining the target task intent and personalized execution configuration information, start the timed scheduling mechanism, break down the target task, and generate a basic operation sequence that can be executed by the terminal device. The basic operation sequence uses touch trajectory, character input, and graphical interface control positioning as the basic execution unit. S400. If there are cross-application collaborative operations in the basic operation sequence, the preset automated operation unit is matched based on the graph association model to solve the optimal operation call path. The automated operation unit is stored in the pre-built automated integration library, and a single automated operation unit corresponds to an independent application execution action. S500: Following the optimal operation call path, it simulates human touch trajectory and character input behavior to complete the automated operation of terminal device tasks. During operation, it identifies the available status of graphical interface controls in real time and obtains real-time task execution status information. S600: Based on the real-time execution status information of the task, dynamically adjust the weight parameters of the task frequency feature and time distribution feature in the user preference feature vector to generate an updated user preference feature vector.

[0007] Further, step S100 includes: S110. Extract timestamp information and behavior trajectory from the interaction log to construct a sequence of user's historical operation behavior; S120. Extract task frequency features based on the user's historical operation behavior sequence, and determine time distribution features based on timestamps; S130. Use a sequence prediction model to perform feature mining on task frequency features and time distribution features to obtain user deep operation habit feature vectors. S140. By integrating task frequency features, time distribution features, and user deep operation habit feature vectors, a user preference feature vector is generated.

[0008] Further, step S200 includes: S210. Obtain natural language task instructions, and combine natural language task instructions with user preference feature vectors to determine the target task intent and instruction semantic feature representation. S220. Based on the target task intent, extract entity nouns from the semantic feature representation of the instruction to identify the target operation object; S230. Retrieve the time distribution features and task frequency features from the user preference feature vector, combine the time distribution features and the available time period of the target operation object to define the trigger time window, and determine the task priority corresponding to the target task intent based on the task frequency features. S240: Construct the mapping relationship between task priority, trigger time window and target operation object, and generate personalized execution configuration information.

[0009] Further, step S300 includes: S310. Decompose the target task intent according to the preset task splitting rules to obtain multiple sub-task nodes; S320. Determine the location of the graphical interface control based on the control identifier that matches the subtask node; S330. Extract the touch trajectory coordinates and character input sequence corresponding to the subtask node; S340. Combine the positioning of the graphical interface controls, the touch trajectory coordinates, and the character input sequence to generate a basic operation sequence that can be executed by the terminal device.

[0010] Further, step S400 includes: S410. Obtain the set of application identifiers contained in the basic operation sequence. If there are at least two different application identifiers in the set of application identifiers, extract the application state context corresponding to the cross-application collaborative operation. S420. Based on the application state context, retrieve several independent application execution actions from the pre-built automation integration library, and construct a graph association model with the independent application execution actions as nodes. S430. Determine the path weight by combining the historical execution time and historical resource consumption data of nodes in the graph association model; S440. Use the shortest path algorithm to perform path optimization on the graph association model with path weights to obtain the optimal operation call path, and match the corresponding preset automated operation unit according to the optimal operation call path.

[0011] Further, step S500 includes: S510. Analyze the optimal operation call path to obtain the initial touch coordinate sequence and character input instruction set; S520. Process the initial touch coordinate sequence to generate a simulated manual touch trajectory; S530: Based on the simulated human touch trajectory and character input instruction set, initiate the automated execution of terminal device tasks; S540: Collect the available status of controls corresponding to the graphical interface images during automated execution; S550: Generate real-time task execution status information based on the availability status of the control, complete the automated execution of tasks on the terminal device, and obtain the real-time task execution status information.

[0012] Further, step S600 includes: S610. Extract the task execution state gradient vector corresponding to the real-time execution state information of the task. S620. Determine the task frequency feature attenuation factor and time distribution feature offset based on the task execution state gradient vector. S630. Calculate the overall adjustment magnitude vector of preference feature weights using the task frequency feature attenuation factor and the time distribution feature offset. S640. Construct the preference parameter update diagonal matrix by adjusting the overall magnitude vector based on the preference feature weights; S650. The preference parameter update diagonal matrix is ​​fused and mapped with the initial feature vector. The weight parameters of task frequency features and time distribution features are dynamically adjusted in combination with the real-time execution status to generate the updated user preference feature vector.

[0013] Another aspect of the present invention relates to an intelligent task processing system based on digital employees, for implementing the above-described intelligent task processing method based on digital employees, comprising: The user preference feature vector generation module is used to obtain the user's historical operation behavior sequence, and to extract features from the user's historical operation behavior sequence using a sequence prediction model to generate user preference feature vectors. The user preference feature vectors include quantitative representation data of operation habits, task frequency, and time distribution characteristics. The personalized execution configuration information determination module is used to perform semantic parsing on the received natural language task instructions based on the user preference feature vector, identify the target task intent and determine the personalized execution configuration information. The personalized execution configuration information includes the mapping relationship between task priority, trigger time window and target operation object. The basic operation sequence generation module is used to combine the target task intent and personalized execution configuration information, start the timed scheduling mechanism, split the target task, and generate a basic operation sequence that can be executed by the terminal device. The basic operation sequence uses touch trajectory, character input, and graphical interface control positioning as the basic execution unit. The optimal operation call path solving module is used to solve the optimal operation call path by matching preset automated operation units based on the graph association model if there are cross-application collaborative operations in the basic operation sequence. The automated operation units are stored in a pre-built automated integration library, and each automated operation unit corresponds to an independent application execution action. The real-time task execution status information acquisition module is used to simulate human touch trajectory and character input behavior according to the optimal operation call path, complete the automated operation of terminal device tasks, and identify the available status of graphical interface controls in real time during operation to obtain real-time task execution status information. The updated user preference feature vector generation module is used to dynamically adjust the weight parameters of task frequency features and time distribution features within the user preference feature vector based on real-time task execution status information, thereby generating the updated user preference feature vector.

[0014] The beneficial effects achieved by this invention are as follows: This invention provides an intelligent task processing method and system based on digital employees, addressing the challenges of accurately parsing natural language instructions in automated operation scenarios of terminal devices, particularly the difficulty in combining user-specific habits with the inefficiency of cross-application collaborative operation and dynamic optimization of preference models. This invention extracts user historical operation behavior using a sequence prediction model to generate preference feature vectors. Based on these vectors, it performs semantic parsing of natural language instructions, determining personalized execution configurations including priority and time windows. Subsequently, it combines intent-based task decomposition to generate basic operation sequences. If cross-application operations exist, it uses a graph association model to solve for the optimal call path in the automation integration library. Finally, it simulates manual operation to complete the task, dynamically adjusting the weights of the preference feature vectors based on real-time execution status. This invention achieves highly personalized automated execution of cross-application tasks, improves the accuracy of instruction parsing, and continuously enhances the adaptive capability of automated operations through a closed-loop feedback mechanism. The specific beneficial effects achieved are as follows: 1. Improve the accuracy of evaluation of intelligent task parsing and automated execution. This invention collects historical user operation behavior sequences and uses a sequence prediction model to perform deep feature extraction, quantifying and generating user preference feature vectors that include operation habits, task frequency, and time distribution characteristics. This enables precise digital modeling of user behavior patterns, providing high-precision feature support for subsequent personalized task processing and avoiding the shortcomings of traditional automated scripts with fixed parameter configurations that cannot adapt to users' actual operation preferences. Relying on user preference feature vectors to assist in natural language command semantic parsing, it accurately identifies task intents and matches personalized execution configuration information, establishing a precise mapping relationship between task priority, trigger time window, and operation object, effectively solving the problems of poor scenario adaptability, misidentification of intent, and fixed parameter configurations in traditional semantic parsing. Combining task intents and personalized configurations, it accurately decomposes tasks, generating executable operation sequences based on touch trajectories, character input, and interface control positioning, ensuring the refinement and feasibility of task decomposition. By matching automated operation units and solving for the optimal call path through a graph association model, it significantly improves the accuracy of execution logic in cross-application collaborative operations, avoiding problems such as cross-scenario operation redundancy, jump errors, and execution omissions. By identifying the availability of interface controls in real time and dynamically sensing execution progress during task execution, the controllability and accuracy of the automated execution process are ensured. Based on the real-time execution status, the weights of user preference feature vectors are dynamically and iteratively updated to continuously correct model feature biases. This allows the accuracy of task identification, scheduling, and execution to continuously improve with use, significantly enhancing the overall accuracy, stability, and success rate of task processing for digital employees.

[0015] 2. Deeply explore the underlying cognitive mechanisms of user operation patterns and task execution adaptation. This invention, through full-process data modeling and intelligent iteration mechanisms, deeply mines the inherent correlation between user behavior characteristics, task attributes, and automated execution logic, overcoming the limitations of traditional automation technologies that rely solely on fixed processes and lack user-centric understanding. First, it uses sequence modeling to quantitatively represent user operating habits, task frequency, and time distribution, uncovering implicit patterns and personalized preferences in long-term user behavior, thus upgrading from surface-level operation records to understanding behavioral mechanisms. It constructs a mapping relationship between user preferences and task semantics and execution configurations, revealing the differentiated execution needs of different users under similar tasks and clarifying the matching mechanism between personalized configurations and efficient task execution. It achieves hierarchical decomposition of complex tasks into basic operation units, uncovering the logical deconstruction patterns between complex business tasks and terminal-level operations. Based on graph association models, it mines the relationships and optimal collaboration logic among various automated operation units in multiple application scenarios, clarifying the linkage mechanism across application tasks. Establish a closed-loop iterative mechanism for task execution feedback and user preference characteristics, explore the impact of task execution results on the updating of user behavior characteristics, realize the continuous evolution of digital employees' understanding of user operations, and completely change the traditional automation operation mode of "no cognition, no iteration, no adaptation".

[0016] 3. Significantly improves the practicality of the technical solution in various scenarios and its adaptability for engineering implementation. This invention constructs a fully closed-loop intelligent task processing system encompassing "behavioral modeling, semantic parsing, task decomposition, cross-domain collaboration, simulation execution, and dynamic iteration." It boasts strong engineering practicality, a wide range of scenario adaptability, and significant industrialization value. Using natural language commands as the task entry point lowers the user's operational threshold, eliminating the need for users to write complex scripts or configure cumbersome rules, thus adapting to the daily office and business automation needs of ordinary users. Through a personalized execution configuration mechanism, it can adapt to different users' operating habits, work rhythms, and task preferences, solving the problem of poor adaptability and uniformity inherent in traditional digital employees. Simultaneously, this invention adopts a standardized automated operation unit library architecture, achieving cross-application and cross-scenario task collaborative scheduling through a graph association model. It can adapt to independent tasks in a single application and complex tasks involving multiple applications, covering diverse automation scenarios such as daily office work, business process processing, and system maintenance. The overall task execution adopts a simulated human touch and input method, requiring no modification to existing terminal systems and software interfaces, exhibiting strong compatibility and rapid deployment on various terminal devices. Furthermore, this invention possesses continuous self-iterative optimization capabilities, which can continuously optimize the execution strategy as the user goes through the process, maintain a highly efficient and adaptable state in the long term, significantly reduce the costs of manual intervention, rule maintenance, and script updates, and its practicality and scalability are significantly better than traditional fixed automation solutions.

[0017] 4. Establish a self-learning closed-loop optimization mechanism to continuously improve the intelligence level of digital employees. This invention constructs a complete self-learning and self-optimizing closed-loop system through real-time execution status feedback and dynamic update mechanisms for user preference features, overcoming the shortcomings of traditional automation tools such as static configuration, fixed capabilities, and inability to evolve autonomously. This invention can dynamically adjust the weights of features such as task frequency and time distribution based on the actual execution of each task, continuously adapting to the latest user operating habits and business change trends, and achieving dynamic optimization of task scheduling strategies, intent recognition logic, and operation execution paths. In long-term operation, digital employees can autonomously accumulate user business characteristics and scenario adaptation experience, continuously reducing the probability of false triggers, incorrect executions, and missed executions, continuously improving the processing capacity and stability of complex tasks and cross-application collaborative tasks, and achieving continuous iterative upgrades of automated service quality. It can fully adapt to dynamic changes such as user business iterations, updated operating habits, and application version upgrades, possessing extremely strong long-term use value and intelligent evolution capabilities. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the intelligent task processing method based on digital employees according to the present invention. Figure 2 This is a functional block diagram of an embodiment of the intelligent task processing system based on digital employees according to the present invention.

[0019] Explanation of icon numbers: 10. User preference feature vector generation module; 20. Personalized execution configuration information determination module; 30. Basic operation sequence generation module; 40. Optimal operation call path solution module; 50. Task real-time execution status information acquisition module; 60. Updated user preference feature vector generation module. Detailed Implementation

[0020] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0021] like Figure 1As shown, the first embodiment of this invention proposes an intelligent task processing method based on digital employees, belonging to the fields of intelligent task automation, natural language semantic parsing, and temporal behavior intelligent prediction technology in digital employee RPA (Robotic Process Automation). Digital employees are the core execution entities, referring to intelligent automated programs built based on RPA robotic process automation, natural language processing, intelligent scheduling, and human-like operation simulation technology. They can simulate human intervention in repetitive office and business tasks such as terminal touch control, character input, interface operation, and cross-application collaborative business, possessing autonomous scheduling, personalized adaptation, and dynamic iterative optimization capabilities. This invention addresses industry pain points of traditional digital employee RPA, such as fixed task execution, inability to adapt to personalized user operating habits, low accuracy of natural language command recognition, chaotic cross-application task collaboration, rigid task scheduling, and inability to dynamically iteratively adapt to changes in user behavior. It proposes a full-process intelligent task processing system based on user historical behavior sequence mining, sequence prediction model feature extraction, personalized semantic parsing, task splitting and scheduling, graph association model cross-application path optimization, human-like automatic execution, and closed-loop iterative updates of user features. This invention enables personalized adaptation, intelligent scheduling, cross-application collaboration, and dynamic adaptive iterative upgrades for digital employee task execution, significantly improving the accuracy and user adaptability of unstructured natural language tasks. The invention includes the following steps: Step S100: Obtain the user's historical operation behavior sequence, use a sequence prediction model to extract features from the user's historical operation behavior sequence, and generate a user preference feature vector. The user preference feature vector includes quantitative representation data of operation habits, task frequency, and time distribution features.

[0022] This step collects all historical operation data of users on terminal devices, office systems, and business platforms, and splices them in time series to form a standardized sequence of user historical operation behavior. A lightweight sequence prediction model is used to perform deep time series feature extraction, dimension filtering, and feature normalization on the behavior sequence, quantifying three core behavioral features: user operation habits, task frequency, and time distribution. These features are then fused to generate a high-dimensional, standardized user preference feature vector, providing user behavior benchmark data for subsequent personalized semantic parsing, task configuration, and intelligent scheduling.

[0023] User historical operation behavior sequences are a dataset of all user operations arranged chronologically based on a timeline. Each sequence contains six basic fields: operation type, operation time, operation object, operation duration, application scenario, and execution result. The sequence collection period ranges from 7 to 90 days, with 30 days being preferred. The criteria for selection are: less than 7 days of sample size is insufficient, resulting in poor generalization of behavioral features; more than 90 days of data is redundant and contains outdated behavioral interference; and a 30-day period can accurately represent users' stable and normalized operation habits. Each sequence has ≥500 valid samples to ensure the effectiveness of feature extraction.

[0024] The sequence prediction model (the core algorithm and complete architecture of this step) adopts a lightweight improved BiLSTM time-series sequence prediction model with an overall five-layer architecture design, adapted for behavioral sequence feature extraction and temporal pattern mining: 1. Sequence Input Layer: Receives normalized user historical operation behavior time-series sequences, with a unified input dimension of 256 dimensions; 2. Bidirectional Temporal Encoding Layer: Mines historical preceding patterns and subsequent associations through forward and backward LSTM units, capturing long-term temporal dependencies; 3. Feature Selection Layer: Adaptively weighted using an attention mechanism to strengthen high-frequency and core operation features and weaken low-frequency invalid interference features; 4. Dimensionality Reduction Layer: Compresses high-dimensional temporal features into a fixed-dimensional feature vector; 5. Feature Output Layer: Outputs a standardized user preference feature vector. The core parameters of the model and their values ​​are determined as follows: The model iteration learning rate is set to 0.001 to 0.01, preferably 0.005; the iteration convergence threshold error is ≤0.001; the value is determined by the following: if the learning rate is too large, feature fitting will easily oscillate; if it is too small, the convergence speed will be too slow. This parameter range takes into account both fitting accuracy and computational efficiency; the model feature extraction accuracy is ≥98%.

[0025] The user preference feature vector is a standardized multidimensional vector that quantifies the patterns of personalized user operation behavior. The total dimension of the vector ranges from 64 to 256, with 128 dimensions being preferred. It includes three core quantitative features: 1. Operation habit features (20% weight): quantitative values ​​of touch preferences, input habits, interface operation order, and operation path preferences, with a value range of [0, 1]; 2. Task frequency features (40% weight): quantitative values ​​of the historical execution frequency and repetition of various business tasks, with a value range of [0, 1]; 3. Time distribution features (40% weight): quantitative values ​​of concentrated task execution periods, operation intervals, and periodic patterns, with a value range of [0, 1]. The weighting is based on the core patterns of user behavior, with task frequency and time distribution being the core basis for personalized scheduling and thus having a higher weighting. The vector dimension is moderate, balancing feature completeness and real-time computation.

[0026] Step S200: Based on the user preference feature vector, perform semantic parsing on the received natural language task instructions, identify the target task intent and determine personalized execution configuration information. The personalized execution configuration information includes the mapping relationship between task priority, trigger time window and target operation object.

[0027] The system receives unstructured natural language task instructions from the user, uses the user preference feature vector generated in step S100 as a personalized prior constraint, and performs deep natural language semantic parsing. It accurately identifies the target task category, execution requirements, and business scenario corresponding to the user input instructions, and locks in the target task intent. Combining the user's historical preference patterns, it matches and generates a mapping relationship containing task priority, trigger time window, and target operation object, forming exclusive personalized execution configuration information that adapts to the user's personalized habits.

[0028] Natural language task instructions are task execution instructions entered by users in a conversational, unstructured text format. They have no fixed syntax format and include vague instructions, abbreviated instructions, and personalized instructions. The instruction text length is 2 to 200 characters, which is suitable for daily office task input scenarios.

[0029] The semantic parsing algorithm architecture adopts a pre-trained BERT fine-tuned semantic parsing model with a four-layer architecture: semantic word segmentation layer, intent feature encoding layer, preference feature fusion layer, and configuration output layer; by fusing user preference feature vectors, it corrects general semantic parsing biases and achieves personalized intent recognition; the semantic recognition accuracy is ≥96.5%.

[0030] The target task intent is a standardized, executable task business intent corresponding to a natural language instruction. It includes subdivided intent types such as file processing, data entry, report generation, message push, and cross-application data synchronization, with intent classification accuracy ≥97%.

[0031] Personalized execution configuration information is a set of task-specific execution configuration parameters that match user behavior preferences. The quantification range and threshold of core parameters are based on the following: 1. Task Priority: Divided into 1 to 5 levels, with level 1 being the highest priority and level 5 the lowest priority. Regular high-frequency tasks are configured with levels 1 to 2 by default, while low-frequency temporary tasks are configured with levels 3 to 5. Value selection is based on matching user task frequency preferences, prioritizing the scheduling of high-frequency core tasks to ensure the efficiency of core business execution. 2. Trigger Time Window: Customizable executable time period, with a window duration range of 0.5h to 24h, preferably 1h to 12h. Value selection is based on matching the user's historical task time distribution characteristics to avoid executing tasks during periods of low user activity and adapting to user work habits. 3. Target Operation Object Mapping Relationship: Establishing a four-dimensional mapping relationship between task intent, operation application, operation control, and execution path, with a mapping matching accuracy of ≥99%.

[0032] Step S300: Combining the target task intent and personalized execution configuration information, start the timed scheduling mechanism, split the target task, and generate a basic operation sequence that can be executed by the terminal device. The basic operation sequence uses touch trajectory, character input, and graphical interface control positioning as the basic execution unit.

[0033] Based on the identified target task intent and personalized execution configuration information, the built-in intelligent timed scheduling mechanism is activated. The overall target task is hierarchically and finely broken down according to task priority and trigger time window constraints. Complex business tasks are decomposed into the smallest operation units that can be directly identified and independently executed by terminal devices. These units are then sorted and assembled according to the execution sequence to generate a standardized basic operation sequence with touch trajectory, character input, and graphical interface control positioning as the core execution units.

[0034] The timed scheduling mechanism is an intelligent hierarchical scheduling architecture based on time windows, task priorities, and task dependencies. The complete scheduling architecture consists of: a priority sorting layer, a time window verification layer, a task dependency decoupling layer, a time sequence splitting layer, and a sequence generation layer. The scheduling response latency is ≤50ms, and the scheduling accuracy is 100%. The scheduling thresholds are based on: high-priority tasks (levels 1-2) are scheduled for immediate execution, medium-priority tasks (level 3) are executed first within the time window, and low-priority tasks (levels 4-5) are executed as a fallback during idle periods.

[0035] The target task splitting rule is that the number of basic operation units after splitting a single complex task is 3 to 50, preferably 5 to 30. The basis for the value is: less than 3 units do not need to be split, and more than 50 units are prone to timing disorder and execution redundancy. This range is suitable for the splitting needs of most office business tasks.

[0036] The basic operation sequence is a standardized set of timing operations that can be directly parsed and executed by the terminal device. It is assembled based on the smallest operation unit and the execution timing error is ≤10ms.

[0037] The basic execution unit includes three types of smallest indivisible execution units: 1. Touch trajectory unit: coordinate positioning accuracy ≤ 0.5px, touch duration 50ms~200ms, preferably 100ms; value basis: conforming to human touch operation habits, avoiding recognition failure due to excessively fast touch or inefficiency due to excessively slow touch; 2. Character input unit: single input character length 1~50 characters, input interval 30ms~100ms; 3. Graphical interface control positioning unit: control positioning matching accuracy ≥ 99.5%, accurately matching interface controls such as buttons, input boxes, pop-ups, and menus.

[0038] Step S400: If there are cross-application collaborative operations in the basic operation sequence, then rely on the graph association model to match the preset automated operation unit, solve the optimal operation call path, and store the automated operation unit in the pre-built automated integration library. Each automated operation unit corresponds to an independent application execution action.

[0039] Traverse the basic operation sequence generated in step S300, and verify whether the operation unit involves multi-application switching, cross-application data interaction, or cross-system collaborative operation. If cross-application collaborative operation is determined to exist, call the pre-built automated integration library, match various preset automated operation units based on the graph association model, quantify and calculate the execution time, error probability, and resource utilization of multiple feasible call paths, and iteratively solve the comprehensive optimal operation call path to achieve efficient and low-failure collaborative execution of cross-application tasks.

[0040] Cross-application collaborative operation refers to a coordinated operation in which the basic operation sequence requires switching between two or more applications or system platforms. The judgment threshold is: if the number of applications involved is ≥2, it is judged as a cross-application collaborative scenario, and the path optimization mechanism is activated.

[0041] The automation integration library is a pre-built database of standardized automated operation units. The library stores templates for independent application execution actions. Each automated operation unit corresponds to one independent and independent application execution action. The total number of units in the library is ≥1000, covering mainstream office, business, and terminal operation scenarios, with a unit call success rate of ≥99.8%.

[0042] An automated operation unit refers to the smallest application execution module that can be independently invoked and is standardized and encapsulated. It includes independent actions such as application startup, page navigation, data reading, content writing, file saving, and window closing. The unit execution time is ≤500ms.

[0043] The graph association model (the core algorithm and complete architecture of this step) adopts a directed weighted association graph neural network model with a five-layer architecture: 1. Node definition layer: each application and each automated operation unit is a graph node; 2. Edge weight assignment layer: execution time, error rate, and resource consumption are used as edge weights, with weight values ​​ranging from [0, 1]; 3. Path traversal layer: traverses all feasible cross-application call paths; 4. Comprehensive scoring layer: calculates the comprehensive optimal value of the path based on weights; 5. Optimal path output layer: outputs the globally optimal operation call path. The optimal path determination threshold is based on a comprehensive score ≥ 0.95, with scoring dimensions of execution efficiency (weight 0.5), stability (weight 0.3), and resource consumption (weight 0.2), ensuring that the path is efficient, stable, and has low resource consumption.

[0044] The optimal operation call path refers to the standardized operation sequence path with the shortest execution time, lowest error probability, and least system resource consumption across application scenarios, with an execution error rate of ≤0.1%.

[0045] Step S500: Following the optimal operation call path, simulate manual touch trajectory and character input behavior to complete the automated operation of terminal device tasks. During operation, identify the available status of graphical interface controls in real time and obtain real-time task execution status information.

[0046] Following the optimal operation call path obtained in step S400, the system simulates human daily touch trajectories, character input rhythms, and operating habits to drive the terminal device to complete the automated operation of the entire task process. During the continuous execution of the task, the system captures the status of graphical interface controls in real time, dynamically identifies available states such as successful loading, failure, pop-up obstruction, and loading timeout, and collects and summarizes task execution progress, execution results, and exception information in real time to generate real-time execution status information for the entire task process.

[0047] The system simulates human touch trajectories and creates an anthropomorphic touch path that matches user preferences, abandoning mechanical linear touch and adopting manually fitted curve trajectories with a trajectory similarity of ≥95%; the touch fluctuation error is ≤2px, completely simulating the randomness and habit of human operation.

[0048] The simulated character input behavior adapts to the user's input rhythm, with an input interval fluctuation range of 30ms to 100ms, simulating the fast and slow rhythm of human typing, avoiding the characteristics of mechanical uniform input, and the degree of anthropomorphism is ≥96%.

[0049] The available states of graphical interface controls include five categories: normal availability, loading, invalid, obscured, and timeout. The control loading timeout threshold is 1000ms. The criteria for determining the timeout are: if the control fails to complete loading within 1000ms, it is considered a loading error, triggering a retry or error alarm mechanism.

[0050] The real-time execution status information of the task includes execution progress (0% to 100%), single-step execution result, exception type, execution time, and control status data. The status acquisition frequency is 10fps to 30fps, preferably 20fps, to ensure that status monitoring is complete and without delay.

[0051] Step S600: Based on the real-time execution status information of the task, dynamically adjust the weight parameters of the task frequency feature and time distribution feature in the user preference feature vector to generate an updated user preference feature vector.

[0052] Using the real-time task execution status, task completion status, and user subsequent correction behavior data obtained in step S500 as iterative samples, the actual execution frequency and actual execution time distribution data of the current round of tasks are dynamically statistically analyzed. Based on the dynamic error feedback mechanism, the weight parameters of task frequency features and time distribution features in the user preference feature vector are finely adjusted in real time, while retaining stable operation habit features and updating and iterating dynamic behavior features to generate an updated user preference feature vector that adapts to the latest user operation habits, thereby achieving closed-loop continuous optimization of the intelligent adaptation capability of digital employees.

[0053] The dynamic weight parameter adjustment mechanism (complete iterative architecture) is a four-layer closed-loop iterative architecture: state sample storage layer, feature deviation calculation layer, weight fine-tuning layer, and vector reconstruction and solidification layer; 1. State sample storage layer: stores the real-time execution status and behavior data of this round; 2. Feature deviation calculation layer: calculates the feature deviation value between historical preferences and current actual operations; 3. Weight fine-tuning layer: adaptively adjusts the frequency and time distribution feature weights; 4. Vector reconstruction and solidification layer: generates updated standardized feature vectors.

[0054] Weight parameter adjustment threshold and value basis: The single weight adjustment magnitude is ≤0.05; Value basis: small and gradual iteration to avoid confusion in personalized adaptation caused by single weight mutation, and to ensure the continuity and stability of user habit iteration; When the feature deviation threshold is >0.03, the weight update is triggered. When the deviation is ≤0.03, the user habit is determined to be stable and the vector is not updated to avoid invalid iteration.

[0055] Updated user preference feature vector: Iteratively optimized standardized multidimensional feature vector, retaining the original stable operation habit features, updating the dynamically changing task frequency and time distribution features, improving feature matching accuracy by ≥5% after iteration, and continuously adapting to dynamic changes in user behavior.

[0056] The intelligent task processing method based on digital employees provided in this embodiment overcomes the shortcomings of traditional digital employee RPA, such as fixed processes, lack of personalized adaptation, inability to iterate autonomously, and inefficient cross-application collaboration. By intelligently mining user behavior sequences, personalized semantic parsing, hierarchical intelligent scheduling, graph model cross-application path optimization, human-like automatic execution, and feature closed-loop iterative updates, it constructs a full-process adaptive intelligent task processing system. This upgrades digital employees from "mechanical execution" to "intelligent adaptation and autonomous optimization," significantly improving the accuracy of natural language task recognition, cross-application collaboration efficiency, and user personalization. It is suitable for all-scenario office work and automated intelligent business processing.

[0057] Furthermore, the intelligent task processing method based on digital employees provided in this embodiment includes step S100 as follows: Step S110: Extract timestamp information and behavior trajectory from the interaction log to construct a sequence of user's historical operation behavior.

[0058] The formula for constructing a user's historical action behavior sequence is as follows: (1) In formula (1), It is a sequence of user's historical operation behavior, without units, and its value is a set containing multiple ordered triples. It is derived from the time-series reconstruction output of terminal interaction log and is used to represent the complete time-series operation behavior of the user. For the first The timestamp of each operation, measured in seconds, is derived from the terminal system clock and is used to unify the timing reference of all operations. For the first The terminal operation action code has no unit and is derived from the event classification and labeling results of the interaction log. It is used to unify the timing benchmark of all operation behaviors. For the first The hash value of the interface state at the moment of operation is unitless. It is derived from the hash calculation result of the terminal interface snapshot at the moment of operation and represents the interface context state corresponding to the operation. The total number of historical behavior samples is a unitless number derived from log statistics. The value is a positive integer ≥200, used to ensure the sufficiency and stability of the sample for preference feature statistics. The control logic of formula (1) is to extract three core data types from the terminal device interaction log in batches: operation timestamp, operation action code, and interface state hash. The time sequence is sorted in ascending order of timestamp to construct a standardized time sequence triplet behavior sequence, eliminating the problems of log data time sequence disorder and missing context, and providing a complete and well-organized basic dataset for subsequent multi-dimensional user preference feature extraction. Traditional digital employee behavior modeling only records the operation action type and completely ignores the operation time sequence correlation and interface context state, resulting in a one-sided and low-precision representation of behavior features. Formula (1) constructs a time-action-interface state triple time sequence, which fully preserves the time sequence logic of user operations and scene context correlation information, and improves the comprehensiveness and accuracy of subsequent preference feature mining from the data source level.

[0059] The intelligent task processing system obtains users' historical interaction logs in office software from the business server, extracts the timestamp information and interface click behavior trajectory from the logs, sorts them in chronological order, and constructs a standardized sequence of users' historical operation behaviors.

[0060] Step S120: Extract task frequency features based on the user's historical operation behavior sequence, and determine time distribution features based on timestamps.

[0061] The formulas for calculating task frequency characteristics and time distribution characteristics are as follows: (2) (3) In formulas (2)~(3), For the first Task frequency characteristics, unitless, derived from statistical calculations of historical behavior sequences, with a value range of [missing information]. Represents the user's attitude towards the first Frequency of operation preferences for similar tasks; This is a conditional indicator function with no unit. Its value is 1 if the condition is true and 0 if the condition is false. It is used to accurately count the number of operations of the same type of task. This is a task category index, without units, derived from a preset task category library, and its value is a positive integer, uniquely identifying a type of terminal task. For the first The time distribution feature vector of the task category is unitless and comes from the timestamp statistical calculation of the same type of task. It includes two core dimension parameters: time mean and time standard deviation. For the first The average operation time for a task category, measured in seconds, is derived from the average of all timestamps for tasks of the same category, representing the typical time points when a user operates on this type of task. For the first The standard deviation of operation time for a task category, measured in seconds, is derived from the timestamp dispersion calculation of tasks of the same category; the value is a non-negative real number, representing the degree of time fluctuation in user operation of this type of task. =The matrix transpose operator has no unit and is used to convert row vectors into column vectors to adapt to the subsequent vector concatenation and matrix operation format. The control logic of formulas (2) to (3) is based on the standardized historical behavior sequence. The operation frequency of various tasks is counted by the indicator function to quantify the user's task operation preference heat. The mean and standard deviation of the timestamp data of similar tasks are further calculated. From the two dimensions of time concentration and time fluctuation range, the time distribution law of user task operation is fully quantified to construct the basic preference feature of frequency + time. Traditional user preference modeling simply counts the number of task operations, which can only represent the operation heat and cannot depict the time law and stability of user operation. Formulas (2) to (3) construct a three-dimensional feature system of frequency-time mean-time standard deviation, which accurately covers the frequency preference, time preference and operation stability features of user task operation, and greatly improves the refinement of user preference representation.

[0062] The intelligent task processing system extracts task frequency features based on users' historical operation behavior sequences. Specifically, it counts the cumulative number of executions of various office tasks within a preset time interval. For example, it counts the frequency of report approval tasks initiated by users within a week, quantifying the high-frequency preference characteristics of user task execution. Simultaneously, the intelligent task processing system relies on the timestamp information corresponding to the behavior sequences to extract task time distribution features and analyze the fixed execution time patterns of various user tasks. For example, it identifies the temporal characteristics of users concentrating on processing approval tasks between 9:00 AM and 10:00 AM daily.

[0063] Step S130: Use a sequence prediction model to perform feature mining on task frequency features and time distribution features to obtain the user's deep operation habit feature vector.

[0064] The formula for calculating the feature vector of deep user operation habits is: (4) In formula (4), This is a feature vector representing the user's deep operational habits. It is unitless, derived from feature extraction by a sequence encoder, and has a value dimension of 64~256. It is used to represent the user's implicit and long-term operational inertia features. It is a time sequence encoder, without units, derived from a pre-trained LSTM / Transformer time sequence model; used to mine implicit temporal correlations and operational inertia in behavioral sequences; The set of trainable parameters for the sequence encoder is unitless and comes from model pre-training and iterative optimization. It is used to adjust the accuracy of model feature extraction. The control logic of formula (4) is to input the constructed triplet historical behavior sequence into the pre-trained temporal sequence encoder. Relying on the model's temporal feature mining capability, it automatically learns the implicit correlation rules of continuous user operations, mines deep operation habits that cannot be identified by manual rules, and outputs a high-dimensional implicit operation habit feature vector. Traditional operation habit recognition relies on manual rule definition and can only identify explicit and simple operation rules. It cannot adapt to the dynamic and complex changes in user operation inertia. Formula (4) adopts a data-driven temporal coding method to automatically mine the implicit operation habits of users. It does not require manual rule intervention and can adaptively adapt to the dynamic changes in users' long-term operation preferences.

[0065] The intelligent task processing system employs a sequence prediction model to perform deep feature mining on task frequency and temporal distribution characteristics. This sequence prediction model is a deep learning temporal network with temporal memory and gating adjustment capabilities, incorporating a built-in memory unit and a three-gating mechanism. During feature mining, frequency and temporal distribution characteristics are used as model inputs. A forgetting gate filters out noise data such as accidental and invalid user operations. An input gate updates the user's high-frequency, regular, and valid operation states, and an output gate outputs a temporal hidden layer state vector. This hidden layer state vector accurately represents the dynamic evolution of user behavior across multiple time points, revealing deep-level user operation habit characteristics. For example, it identifies a fixed operation chain where a user reviews financial statements and then proceeds with budget approval.

[0066] Step S140: Integrate task frequency features, time distribution features, and user deep operation habit feature vectors to generate user preference feature vectors.

[0067] The formula for generating user preference feature vectors is: (5) In formula (5), The user preference feature vector is unitless and comes from the concatenation and fusion of multi-dimensional features. The value dimension is 320~512, providing a unified feature input for subsequent semantic parsing and personalized configuration. This is a set of frequency features for all task categories, without units, and derived from all task categories. Summary set; This is a set of time distribution features for all task categories, without units, and derived from all task categories. The control logic of formula (5) is to dimensionally concatenate and fuse the quantified task frequency feature set, time distribution feature set, and deep operation habit feature vector extracted by the model, integrating the user's explicit operation preferences and implicit operation inertia to construct a fully covered, high-precision standardized user preference feature vector. Traditional preference features only integrate explicit frequency or time features, resulting in a single feature dimension and a one-sided representation. Formula (5) achieves three-dimensional fusion of explicit frequency, time features, and implicit habit features, comprehensively covering the user's operation preference characteristics, and providing accurate user feature support for subsequent fuzzy natural language instruction parsing and personalized task configuration.

[0068] The intelligent task processing system combines and reduces the dimensions of task frequency characteristics, time distribution characteristics, and operation habit characteristics to generate a standardized user preference feature vector, thereby achieving a precise quantitative characterization of users' personalized work habits and task requirements.

[0069] Preferably, the intelligent task processing method based on digital employees provided in this embodiment includes step S200 as follows: Step S210: Obtain natural language task instructions, and combine natural language task instructions with user preference feature vectors to determine the target task intent and instruction semantic feature representation.

[0070] The formula for determining the target mission intent and instruction semantic features is as follows: (6) In formula (6), This is a preference-enhanced semantic feature vector, which is unitless, derived from the output of the semantic parsing model, and has a value dimension of 128. It represents the precise semantics and task intent of natural language instructions. The input of natural language task instructions to the user is a string type, which comes from the real-time input of the end user. It is the original task instruction text to be parsed and is the original input carrier for semantic parsing. A preference-guided semantic parser, unitless, derived from a fine-tuned BERT pre-trained model, used to enhance instruction semantic understanding by combining user preference features, eliminate semantic ambiguity of natural language instructions, and accurately pinpoint the core task intent; The user preference feature vector is unitless and originates from step S140. Its value dimension is 320~512. It provides prior information on the user's historical operation preferences for the semantic parsing process, thereby enhancing the semantic feature preferences. The control logic of formula (6) is to input the natural language command input by the user and the pre-constructed user preference feature vector into the fine-tuning semantic parsing model. Based on the user's historical operation preference features, the semantics of the command are enhanced in context and ambiguous, accurately extracting the core semantic features of the command, eliminating invalid semantic interference, identifying the user's true task intent, and finally outputting a standardized preference-enhanced semantic feature vector with fixed dimensions and accurate semantics, providing accurate semantic basis for subsequent task splitting and automated execution. Traditional semantic parsing only relies on the features of the command text itself to carry out static semantic recognition, without associating it with the user's personalized operation preferences. When faced with non-standardized commands such as colloquial, ambiguous, and abbreviated commands, it is very easy to have semantic misjudgment and intent recognition deviation problems. Formula (6) introduces user historical preference features into the semantic parsing process. It uses personalized preference prior information to assist in semantic judgment and eliminate instruction ambiguity, effectively adapting to various non-standard natural language input scenarios and greatly improving the accuracy of intent recognition and semantic parsing robustness of complex spoken instructions and fuzzy instructions.

[0071] The intelligent task processing system receives natural language task instructions input by the user, such as "arrange departmental weekly report summary". Combining the pre-constructed user preference feature vector, it completes task intent recognition and semantic representation, determines that the target task intent is a document summary task, and outputs the corresponding instruction semantic features.

[0072] Step S220: Based on the target task intent, extract entity nouns from the semantic feature representation of the instruction to identify the target operation object.

[0073] The formula for identifying the target operation object is: (7) In formula (7), The final target operation object is a string type, derived from the entity database similarity matching results, representing the precise target operation object of the terminal corresponding to the current task; The preset operation object entity library has no unit and is derived from the summary collection of all operable objects in the terminal application, providing a standardized candidate object pool for entity matching. The entity embedding vector is a unitless vector derived from the output of the entity pre-trained embedding model. It uniquely represents the inherent semantic features of each entity in the entity library and is used for vector similarity matching operations. This is a cosine similarity calculation, unitless, with a range of values. , is used to quantify the semantic matching degree between the instruction semantic feature vector and the entity object embedding vector. The larger the value, the higher the semantic similarity. The control logic of formula (7) is to use the preference-enhanced instruction semantic feature vector output in step S210 as the matching benchmark, and perform cosine similarity operation on it one by one with the embedding vectors of all entities in the preset operation object entity library, quantify the semantic fit between each candidate entity and the task instruction, and select the entity object with the highest similarity as the accurate target operation object of the current task, so as to realize accurate matching and positioning of entities under fuzzy instructions and non-standard expressions. Traditional automated entity recognition relies heavily on fixed keyword rule matching, which can only adapt to standard sentence instructions and cannot be compatible with complex scenarios such as colloquial expressions, synonym replacement, and content omission. The entity matching fault tolerance rate is low and the matching failure and object recognition error are very easy to occur. Formula (7) adopts a semantic vector-level fuzzy similarity matching mechanism, completely gets rid of the constraints of fixed keywords, and completes entity matching based on global semantic features, which greatly improves the fault tolerance and accuracy of entity recognition under non-standard natural language instructions and adapts to more complex task scenarios.

[0074] The intelligent task processing system extracts the core entity noun "weekly report" from the semantic features of the instructions, and accurately locates the target operation object as the department's weekly report data file.

[0075] Step S230: Retrieve the time distribution features and task frequency features from the user preference feature vector, combine the time distribution features with the available time period of the target operation object to define the trigger time window, and determine the task priority corresponding to the target task intent based on the task frequency features.

[0076] The formula for determining task priority and trigger time window is as follows: (8) (9) In formulas (8)~(9), This represents the dynamic priority of the task, which is unitless and derived from the weighted activation calculation result of preference features. Its value range is... The larger the value, the higher the execution priority of the corresponding target task; The sigmoid activation function is unitless and derived from standard neural network activation operations. It is used to normalize the result of priority-weighted operations to a specific value. Standardized intervals enable unified prioritization. This is a frequency weighting coefficient, without units, derived from system preset control parameters, with a value range of 0.2 to 0.8, used to control the degree of influence of historical frequency characteristics of tasks on task priority; The time distribution weight coefficient is a unitless parameter derived from the system's preset control parameters, with a value range of 0.2 to 0.8. It is used to control the degree of influence of the stability of the task's time distribution on the task priority. The target task frequency feature is unitless and comes from the preference statistics parameters extracted in step S120. It is the historical operation frequency parameter corresponding to the current target task. The target task time distribution feature vector is unitless and comes from the preference statistics parameters extracted in step S120. It is the time distribution parameter corresponding to the current target task. The modulus of the time distribution feature vector is dimensionless and derived from the result of vector modulo operation. It is used to quantify the overall stability of the target task's operation time. The personalized trigger time window for the task, in seconds, is derived from the user's time preference adaptation calculation and represents the effective time interval during which the target task is allowed to be automatically executed. The historical average operation time of the target task, in seconds, is derived from the timestamp statistics of similar tasks and serves as the baseline center value for the personalized trigger time window. The time tolerance coefficient, in units of seconds, is derived from the preset adaptation parameters and ranges from 30 to 300 seconds. It is used to adapt to the reasonable fluctuation deviation of the user's daily operation time and ensure the fault tolerance of the time window. The control logic of formulas (8) to (9) is to retrieve the historical task frequency characteristics and time distribution characteristics corresponding to the current target task, perform a two-dimensional weighted operation through the frequency weight coefficient and the time distribution weight coefficient, and normalize it through the Sigmoid activation function to generate a dynamic task priority that fits the user's historical operation habits, so that the user's high-frequency operation and time-stable tasks can obtain higher execution priority; at the same time, based on the historical average operation time of the target task as the central benchmark, the preset time tolerance coefficient is superimposed, and combined with the available time period constraint of the target operation object, a special trigger time window that adapts to the user's personalized operation schedule is defined to realize the personalized configuration of task execution timing and priority. The task execution priority and trigger time window of traditional digital employees are all uniformly fixed configurations of the system. Standardized parameters are used to adapt to all users, which cannot adapt to the differences in operation frequency, work and rest habits and operation time fluctuation characteristics of different users, and the personalized adaptation capability is extremely poor. Formulas (8) to (9) dynamically and adaptively generate task priorities and personalized trigger time windows based on users' historical real preference data. They rely on users' own operational behavior characteristics to complete task scheduling configuration, realize refined and personalized intelligent task scheduling for each user, and greatly improve the fit of automated task execution and user experience.

[0077] The system retrieves the time distribution features and task frequency features from the user preference feature vector to determine the task trigger time window and task priority, respectively. The trigger time window is determined by performing an interval intersection operation between the time distribution features of similar historical tasks performed by the user and the available time period of the target file to determine the optimal automatic execution time period. For example, if the user's historical habit is to process document summary tasks every Friday afternoon from 3 PM to 5 PM, and the available submission time for departmental weekly reports is after 2 PM on Friday afternoon, then by comparing the time interval intersection, Friday afternoon from 3 PM to 5 PM is designated as the dedicated trigger time window for the task. The task priority is determined by reading the historical execution frequency of the target task based on the task frequency features. If the task is a high-frequency task with a fixed cycle, the intelligent task processing system's priority determination module assigns it the highest execution weight and marks it as a high-priority task.

[0078] Step S240: Construct the mapping relationship between task priority, trigger time window and target operation object, and generate personalized execution configuration information.

[0079] The formula for generating personalized execution configuration information is: (10) In formula (10), It is a set of personalized execution configurations, without units, derived from the results of multi-parameter structured integration calculations. It is used to integrate three core configuration parameters: task priority, trigger time window, and target operation object, to form an integrated and standardized task execution configuration. This is an integrated function for personalized configuration. It has no unit and is derived from a preset structured mapping logic. It is used to establish binding relationships between multi-dimensional task parameters and output standardized configuration information that can be directly scheduled. The final target operation object is a string type, derived from the entity matching result in step S220, and is the operation object parameter configured for task execution. The personalized trigger time window for the task, in seconds, is derived from the time window definition result in step S230 and is a time constraint parameter configured for task execution. The task dynamic priority has no unit and originates from the priority determination result in step S230. It is the scheduling priority parameter configured for task execution. The control logic of formula (10) is to structurally integrate the target operation object identified in step S220, the personalized trigger time window dynamically generated in step S230, and the dynamic task priority. Through the preset configuration integration function, a one-to-one fixed mapping relationship is constructed for the three core parameters. The configuration is uniformly encapsulated into a set of personalized task execution configurations with standard format, complete parameters, and direct scheduling and calling capabilities. This provides unified and accurate pre-configuration support for subsequent task splitting, control positioning, and automated execution. In the traditional digital employee task execution configuration, the three parameters of operation object, execution time, and task priority are independent and fixed in configuration. There is no associated mapping relationship, the parameter scheduling is scattered, the adaptability is poor, and it is impossible to achieve integrated collaborative scheduling based on user preferences. Formula (10) constructs a multi-parameter integrated binding mapping configuration structure to achieve unified management of task object, execution time, and scheduling priority. This completely solves the defects of traditional configuration fragmentation, fixedness, and insufficient scheduling flexibility, and provides standardized and personalized configuration support for subsequent full-process automated intelligent task processing.

[0080] The intelligent task processing system ultimately establishes a one-to-one mapping relationship between task priorities, trigger time windows, and target data files, generating personalized task execution configuration information that adapts to the user's individual habits.

[0081] Furthermore, in the intelligent task processing method based on digital employees provided in this embodiment, step S300 includes: Step S310: Decompose the target task intent according to the preset task splitting rules to obtain multiple sub-task nodes.

[0082] The formula for calculating subtask nodes is: (11) In formula (11), It is a set of subtask sequences, without units, derived from the output of task splitting operations, and its value is a set containing multiple ordered subtask nodes, representing a standardized sequence of executable units after the splitting of a complex task; Encodes the target task intent, which is unitless, derived from the task intent recognition result, and takes discrete numerical values, uniquely distinguishing different types of user target tasks; This is a task splitting function with no unit. It is derived from a pre-defined rule-based operation logic and is used to input intent encoding and rule base, and output a standardized subtask sequence. The task splitting rule base is unitless and derived from manual preset configurations, containing standardized splitting logic and node rules for various tasks. The control logic of formula (11) is based on the identified target task intent code, matching the corresponding standardized splitting logic within the task splitting rule base, and calling the task splitting function to decompose the complete and complex user task into multiple ordered, independent, and executable sub-task nodes, forming a structured sub-task execution sequence. Traditional task splitting uses fixed hard-coded logic, which only adapts to a single task type, resulting in extremely poor versatility and scalability. Formula (11) adopts a rule-based intent matching splitting mechanism, which can be compatible with the accurate splitting of multiple complex tasks, greatly improving the versatility and scenario adaptability of task splitting.

[0083] When the intelligent task processing system recognizes the target task intent of "automatically filling out travel expense reports," it decomposes the macro-level task intent hierarchically according to preset task decomposition rules, resulting in multiple standardized sub-task nodes. The preset task decomposition rules are a parsing strategy based on business process graphs, mapping high-level business intents into standardized micro-operational steps. The intelligent task processing system, through its built-in process parsing engine, breaks down "filling out travel expense reports" into ordered sub-task nodes such as opening the expense report system, creating a new form page, entering the travel city, and entering the expense amount.

[0084] S320. Determine the location of the graphical interface controls based on the control identifier that matches the subtask node.

[0085] The formula for determining the location of graphical interface controls is: (12) In formula (12), For the first The coordinates of each subtask corresponding to the control, in pixels (px), are derived from the output of the visual positioning algorithm. The values ​​are two-dimensional coordinate vectors within the screen range, representing the precise screen position of the target control. For the first Each subtask node has no unit and is derived from the split output result of formula (11). The value is a standardized subtask unit, corresponding to a single terminal operation task. This is a visual positioning algorithm for UI controls. It is unitless, derived from a preset visual detection model, and is used to output precise coordinates of controls based on subtasks and UI images. The real-time interface image matrix of the terminal is unitless and is derived from real-time screenshots of the terminal. Its dimensions are consistent with the device screen resolution, representing the panoramic information of the current terminal interface. The control logic of formula (12) is to traverse all the sub-task nodes that have been split, combine the interface image matrix collected in real time by the terminal, and rely on the visual positioning algorithm to dynamically identify the target operation control corresponding to each sub-task, output accurate control coordinates in real time, and adapt to dynamic scenarios such as interface layout changes and device resolution switching. Traditional automated operations rely on fixed coordinate positioning, and positioning failure and task error problems are very likely to occur after interface iteration updates and device replacement. Formula (12) adopts a real-time visual dynamic positioning mode, gets rid of the fixed coordinate constraints, and significantly improves the stability of control positioning and cross-scene adaptation capabilities.

[0086] The intelligent task processing system retrieves the unique control identifiers corresponding to each sub-task node to accurately locate the graphical user interface (GUI) controls. The control identifier is a unique set of identification parameters for interface interactive elements, including resource name, hierarchical path, text attributes, control type, and other dimensional information. The control location process specifically involves the system real-time parsing of the current terminal screen's view tree structure and traversing and matching the control identifier features carried by the sub-task nodes with the attributes of the view tree nodes. For example, for the "Create Form" sub-task, the intelligent task processing system searches the view tree for the target node with the text attribute "Create Reimbursement" and the control type "Button." Upon successful matching, it reads the absolute screen boundary coordinates of the control to complete the GUI control location.

[0087] Step S330: Extract the touch trajectory coordinates and character input sequence corresponding to the subtask node.

[0088] The formula for extracting touch trajectory coordinates and character input sequence is as follows: (13) (14) In formulas (13)~(14), For the first Each subtask touch trajectory sequence is unitless and originates from the trajectory generation module output. It takes the value of multiple sets of temporal touch coordinates, representing the continuous touch operation trajectory that simulates human operation. For the first Each touch point operation moment, in seconds, originates from timing logic generated data, takes the value of a positive real number, and is used to construct the timing relationship of continuous touch operations; For the first The horizontal coordinate of each touch point, in pixels (px), is derived from the trajectory generation result and its value is within the screen resolution range, representing the horizontal position of the touch point. For the first The vertical coordinate of each touch point, in pixels (px), is derived from the trajectory generation result and its value is within the screen resolution range, representing the vertical position of the touch point. For the first The total number of touch points for each subtask, without unit, is derived from the task requirement statistics and takes the value of a positive integer, limiting the number of trajectory points for a single subtask; For the first Each subtask character input sequence has no unit, originates from the character encoding conversion result, and takes the value of an ordered set of character encodings, representing the text input content corresponding to the subtask. No. Each character encoding has no unit, originates from character mapping data, takes discrete values, and each encoding uniquely corresponds to one input character. For the first The total number of characters input for each subtask is unitless, derived from the text requirement statistics, and is a positive integer, limiting the length of input characters for a single subtask. The control logic of formulas (13) to (14) is to generate a continuous touch coordinate trajectory carrying precise timing information based on the operational requirements of each subtask, simulating continuous operations such as manual clicking, sliding, and long pressing; and simultaneously generating an ordered character encoding sequence to match the task text input requirements, forming a complete standardized simulation operation unit. Traditional automated operations only use discrete single-point clicks, and the operation trajectory is mechanical and rigid, making it very easy for the platform's risk control system to identify and intercept. Formulas (13) to (14) construct a temporally continuous simulated touch trajectory, highly restoring the continuity and randomness of manual operations, effectively avoiding risk control detection, and significantly improving the task execution pass rate.

[0089] The intelligent task processing system extracts the standard touch trajectory coordinates and character input sequences corresponding to each subtask node.

[0090] Step S340: Combine the graphical interface control positioning, touch trajectory coordinates, and character input sequence to generate a basic operation sequence that can be executed by the terminal device.

[0091] The formula for generating the basic operation sequence is: (15) In formula (15), It is a basic sequence of operations that can be executed by the terminal. It has no unit and comes from the integration result of multiple sub-task operation units. Its value is a standardized set of device execution instructions, representing a complete set of task operation instructions. The ordered union operator has no unit and is derived from standard set operations. It is used to merge all subtask operation units in the order of execution. For the first The coordinates of the control corresponding to each subtask, in pixels (px), are derived from the output of formula (12) and represent the control positioning information of the subtask operation. For the first Each subtask touch trajectory sequence is unitless and comes from the output of formula (13), representing the touch operation trajectory information of the subtask; For the first Each subtask character input sequence, without units, originates from the output of formula (14) and represents the text input information of the subtask. The control logic of formula (15) integrates the control positioning coordinates, continuous touch trajectory, and character input sequence of all subtasks according to the execution order of the subtasks, and encapsulates them into a standardized operation instruction sequence that can be recognized and executed by the terminal, thus completing the accurate conversion of the task semantic intent to the terminal operation instruction. Traditional automation instruction formats are messy and lack a unified standard, resulting in poor cross-terminal and cross-application adaptability. Formula (15) builds an atomic operation unit integration system, standardizes and encapsulates various operation instructions, and greatly improves the universality and cross-scenario adaptability of automation instructions.

[0092] The positioning information of GUI controls, touch trajectory coordinates, and character input sequences are structurally combined to generate standardized basic operation sequences that can be directly parsed and executed by the terminal.

[0093] Preferably, the intelligent task processing method based on digital employees provided in this embodiment includes step S400 as follows: Step S410: Obtain the set of application identifiers contained in the basic operation sequence. If there are at least two different application identifiers in the set of application identifiers, extract the application state context corresponding to the cross-application collaborative operation.

[0094] The formula for calculating the application state context is: (16) (17) In formulas (16)~(17), This is the set of application IDs contained in the operation sequence. It has no unit and comes from the parsing result of the basic operation sequence. Its value is a set of multiple application identifiers, representing all terminal applications involved in the task. It serves as a unique identifier for a single application, has no unit, originates from the system's preset application code, and takes the value of discrete characters or numbers, uniquely distinguishing different terminal applications. The total number of applications involved in the task, without units, is derived from the statistical results of the application ID set, and takes the value of a positive integer, representing the number of applications participating in cross-application tasks; This is a cross-application operation marker with no unit. It is derived from the scenario determination result and takes the value of 0 or 1, where 1 represents a cross-application scenario and 0 represents a single application scenario. It is a set of multi-application state contexts, without units, derived from the collection and summary results of multi-application state, and its value is a set of multi-application state parameters, representing the real-time running status of all related applications; A single application state context, without units, originates from the real-time state collection results of a single application and represents the real-time parameters of the corresponding application, such as process, interface, and operating status. The control logic of formula (17) is to parse the basic operation sequence to obtain all associated application identifiers, count the number of applications, and determine the cross-application collaborative operation scenario when the number of applications is greater than or equal to 2; and synchronously collect the real-time operating state context of all associated applications to provide complete scenario data support for subsequent cross-application path optimization. Traditional digital employees only support independent operation of a single application and lack the ability to identify cross-application scenarios and context awareness, making it impossible to handle complex tasks involving multi-application collaboration. Formula (17) can accurately identify cross-application scenarios and collect application status globally, greatly expanding the scenario coverage of intelligent task processing.

[0095] The intelligent task processing system parses the set of application identifiers corresponding to the task to be executed. If multiple applications, such as customer management systems and enterprise communication software, are involved, it determines that the current task has a cross-application collaborative operation requirement. The intelligent task processing system collects the running parameters of each application in real time and obtains the application state context corresponding to cross-application collaboration. The application state context is a set of parameters characterizing the real-time running environment, page status, and data flow progress of each application, including the order page display status of the customer management system and the session connection status of the enterprise communication software.

[0096] Step S420: Based on the application state context, retrieve several independent application execution actions from the pre-built automation integration library, and construct a graph association model with the independent application execution actions as nodes.

[0097] The formula for constructing a graph association model is: (18) In formula (18), It is a cross-application operation association graph model, which has no unit and is derived from the output of the model construction, representing the topological dependency relationship between cross-application operation units. This is a set of graph model nodes, without units, derived from the results of the automation integration library. Its value is a set of multiple operation nodes, with each node corresponding to an independent automation operation unit. Each is a single automated operation unit node, without a unit, derived from an entry stored in the integration library, and uniquely corresponds to an independent application execution action; This is a set of edges in the graph model, without units, derived from the modeling results of operation dependencies. Its value is a set of multiple associated edges, representing the jump and dependency logic between operation units; The operation unit is associated with an edge, which has no unit and originates from the logical relation mapping result, representing the first... The first operation unit and the first The execution dependencies of each operation unit. The control logic of formula (18) is based on the collected multi-application state context, and retrieves all the execution actions required for cross-application tasks from the automation integration library. It uses independent execution actions as model nodes and the jump dependencies between actions as connecting edges to build a complete cross-application operation topology graph model. Traditional cross-application operations use fixed chain execution logic, which cannot adapt to complex tasks with multiple branches and multiple jumps, and has extremely low fault tolerance and flexibility. Formula (18) relies on the graph model to depict the operation topology structure, supports the processing of complex tasks with multiple paths, and provides core model support for global path optimization.

[0098] The intelligent task processing system, based on the application state context, retrieves standardized, independent execution actions from a pre-defined automated integration library. This library is a resource repository encapsulating common interfaces, driver instructions, and standardized operation units for various applications. The system extracts atomic execution actions such as reading order data, launching application windows, and sending session messages. Using each atomic action as a node and action dependencies as edges, it constructs a directed acyclic graph (DAG) model to represent the execution order and data dependencies of cross-application actions.

[0099] Step S430: Determine the path weight by combining the historical execution time and historical resource usage data of the nodes in the graph association model.

[0100] The formula for determining path weight is: (19) In formula (19), For nodes To node The weight of the associated edge is unitless and comes from the result of the weighted calculation of the two objectives. It takes the value of a non-negative real number. The smaller the value, the better the quality of the corresponding execution path. This is the time consumption weight coefficient, which has no unit and is derived from preset control parameters. Its value range is 0.4 to 0.6, and it is used to control the weight of the impact of execution time on path quality. The average execution time of unit jump in history, in seconds, is derived from the historical execution log statistics. It takes a positive real number and represents the average time of jump between two operation units. This is a resource weighting coefficient, dimensionless, derived from preset control parameters, with a value range of 0.4 to 0.6, satisfying... , used to adjust the weight of the impact of resource consumption on path quality; The average resource consumption of unit jumps is a unitless value derived from historical resource monitoring statistics. It is a positive real number representing the memory and computing power consumption level of the unit jump process. The control logic of formula (19) is to retrieve the historical execution time and resource consumption data of each operation unit jump in the graph model. Through the weighted calculation of time consumption and resource consumption as dual objectives, the overall quality of each jump path is quantified, and the weight of all related edges is assigned. Traditional path selection only considers the execution time, which easily selects inferior paths with low execution time, high lag, and high resource consumption. Formula (19) adopts dual objective weighted modeling to balance execution efficiency and equipment stability, making the path selection results more consistent with the actual equipment operating conditions.

[0101] The intelligent task processing system quantifies the path weights between nodes based on historical execution time and memory resource usage data of each node in the graph association model. These path weights are comprehensive quantitative indicators representing the time and resource costs of task flow. For example, the intelligent task processing system statistically analyzes the average response time of the "read order" action and the peak memory usage of the "launch window" action, and uses a weighted fusion algorithm to calculate the comprehensive cost, which is then used as the edge weights of the graph model.

[0102] Step S440: Use the shortest path algorithm to perform path optimization on the graph association model with path weights to obtain the optimal operation call path, and match the corresponding preset automated operation unit according to the optimal operation call path.

[0103] The formula for calculating the optimal operation call path is: (20) In formula (20), This represents the globally optimal operation call path, which has no unit. It is derived from the shortest path algorithm optimization result and takes the value of the ordered operation sequence with the smallest weight accumulation value, representing the optimal task execution path. It is a set of all feasible execution paths, without units, derived from the graph model traversal results, and includes all executable paths from the task start point to the end point. These are the associated edges contained in the path, without units, derived from the path decomposition results, and represent the unit jump links contained in the feasible path; The weight of the edge between nodes is unitless and derived from the calculation result of formula (19), representing the comprehensive quality weight of a single jump link. The control logic of formula (20) is to traverse all feasible execution paths in the graph model using the shortest path algorithm, calculate the total weight accumulation value of each path, and select the path with the smallest total weight as the globally optimal operation call path, thus completing the optimal scheduling and matching of the automated operation unit. Traditional cross-application scheduling adopts a local greedy optimization strategy, focusing only on the single-step optimality, which easily leads to inefficient global execution and task failure. Formula (20) realizes global-dimensional path optimization, which greatly improves the overall running efficiency and execution stability of cross-application tasks.

[0104] The intelligent task processing system employs a shortest path algorithm to traverse all branches of the graph association model. By accumulating the weights of each path segment, it selects the execution sequence with the minimum global total weight. For example, the intelligent task processing system compares the combined costs of two types of execution branches: direct data push and clipboard relay. It then selects the optimal operation call path with the shortest time and lowest resource consumption, and matches it with the corresponding automated operation unit to achieve efficient collaborative connection between cross-application tasks.

[0105] Furthermore, in the intelligent task processing method based on digital employees provided in this embodiment, step S500 includes: Step S510: Analyze the optimal operation call path to obtain the initial touch coordinate sequence and character input instruction set.

[0106] By combining the initial touch coordinate sequence with the character input instruction set, a standardized terminal execution instruction set is generated. The formula for generating the standardized terminal execution instruction set is as follows: (twenty one) In formula (21), This is a global sequence of control coordinates in pixels (px), derived from the path parsing output, summarizing the control positioning coordinates of all operations in the optimal path; The original touch trajectory sequence is unitless and comes from the path parsing output, summarizing all continuous touch operation trajectories in the optimal path. It is a global character input instruction set with no unit, derived from the path parsing output, and summarizes all text input sequences in the optimal path. This is a path parsing function with no unit. It is derived from the preset decoding operation logic and is used to decode the topology path into terminal-recognizable practical instructions. The globally optimal operation call path is unitless and originates from the optimization result of formula (20), serving as the core input object for path decoding. The control logic of formula (21) is to call a preset path parsing function to decode the optimal topology execution path into three types of standardized terminal instructions: control coordinates, touch trajectory, and character input, thereby achieving seamless conversion from the optimal path to actual operation instructions. Traditional optimal path results lack a standardized conversion mechanism and cannot be directly adapted to terminal device execution. Formula (21) establishes a unified decoding conversion rule, achieving precise docking between path optimization results and terminal actual operation instructions, and improving the adaptability of automated execution.

[0107] The intelligent task processing system analyzes the optimal operation call path and separates the initial touch coordinate sequence and character input instruction set.

[0108] Step S520: Process the initial touch coordinate sequence to generate a simulated manual touch trajectory.

[0109] The formula for generating simulated human touch trajectories is: (twenty two) In formula (22), To smoothly optimize the simulated touch trajectory, it is unitless and derived from the trajectory optimization output, representing a highly realistic and smooth touch trajectory that closely replicates human operation. This is a Gaussian smoothing function with no unit. It is derived from a preset trajectory optimization algorithm and is used to eliminate mechanical inflection points and sudden velocity changes in the original trajectory. The original touch trajectory sequence is unitless and comes from the analysis result of formula (21). It is the initial mechanical trajectory data to be optimized. The trajectory smoothing coefficient is dimensionless and derived from preset adjustment parameters, with a value range of 2 to 5. A larger value indicates a higher degree of trajectory smoothing optimization. The control logic of formula (22) is to use a Gaussian smoothing algorithm to optimize and correct the original rigid touch trajectory, eliminate machine characteristics such as mechanical inflection points and speed abrupt changes, and generate a simulated touch trajectory with uniform speed, natural shape, and conformity to human operating habits. Traditional automated touch trajectories have obvious mechanical characteristics and are easily identified and intercepted by risk control systems. Formula (22) restores the randomness and smoothness of human operation through trajectory smoothing optimization, effectively avoids risk control detection, and significantly improves the task execution pass rate.

[0110] The intelligent task processing system reconstructs the initial touch coordinate sequence in a human-like manner, generating simulated human touch trajectories. This eliminates the mechanical, linear, automated operation characteristics, making the device's operation behavior conform to the real-world operation patterns of humans. Specifically, the intelligent task processing system extracts the starting and ending coordinates of the touch trajectory, adds dynamic control points to construct a third-order Bézier curve model, divides the trajectory interval according to preset motion parameters, and calculates the uniform movement step size. Simultaneously, it superimposes pixel-level random jitter deviations conforming to a Gaussian distribution on the curve path to simulate the acceleration changes and minute hand deviations of human operation, transforming the mechanical jump coordinate sequence into a continuous, non-linear, human-like touch coordinate point set.

[0111] Step S530: Based on the simulated human touch trajectory and character input instruction set, start the automated execution of terminal device tasks.

[0112] The formula for automated task execution on terminal devices is: (twenty three) In formula (23), The terminal executes the driver function, which has no unit and comes from the device's built-in execution engine. It is used to drive the terminal device to complete a full set of automated operations. The coordinate sequence of the global control, in pixels (px), is derived from the output of formula (21) and provides a positioning basis for automated operations. To smooth and optimize the simulated touch trajectory, the unitless formula is derived from the optimization result of formula (22), providing trajectory support for automated touch operation; The global character input instruction set is unitless and originates from the output of formula (21), providing instruction support for automated text input operations. The control logic of formula (23) relies on the optimized control positioning coordinates, simulated touch trajectory, and standardized character input instructions to call the terminal execution drive function, completely replicating the human operation logic and operation form, and completing the entire process of terminal automation task execution. Formula (23) integrates multi-dimensional standardized instructions and simulated operation trajectory to achieve full-process anthropomorphic automatic execution, solving the defects of traditional automated operations being rigid, easily intercepted, and poorly adaptable.

[0113] The intelligent task processing system relies on generated simulated human touch trajectories and character input command sets to drive terminal devices to execute automated tasks. During task execution, the intelligent task processing system captures screen interface images in real time, identifies the bounding boxes of key controls such as input boxes and function buttons through an object detection model, and extracts the pixel grayscale features and color saturation features within the bounding boxes.

[0114] Step S540: Collect the available status of controls corresponding to the graphical interface images during the automated execution process.

[0115] The formula for collecting the available state of a control is: (twenty four) In formula (24), For the first The status of an operation control is unavailable, without unit, and is derived from the status detection output. The value is 0 or 1, where 1 indicates that the control is available and 0 indicates that the control is inactive. This is a UI control state detection algorithm. It has no unit and is derived from a visual recognition computing model. It is used to determine the real-time running status of controls based on UI images. For the first The real-time interface frame image of the terminal corresponding to the step operation has no unit. It is obtained from real-time screenshots during the execution process and represents the panoramic state of the terminal interface before the single-step operation. For the first The coordinates of the step operation control, in pixels (px), are derived from the analysis result of formula (21) and provide a basis for the positioning of the control status. The control logic of formula (24) is to collect the terminal interface image in real time before each automated operation is executed, combine it with the control positioning coordinates, and use a visual detection algorithm to determine the availability status of the target control, and promptly capture fault conditions such as page loading abnormalities, control failures, and page jump errors. Traditional automated execution has no real-time status feedback, and the task fails directly when an abnormality occurs, with extremely poor fault tolerance. Formula (24) realizes real-time visual status verification throughout the entire process, which can accurately capture execution abnormalities and provide data support for subsequent error correction and retry, parameter iteration and optimization.

[0116] The intelligent task processing system combines preset grayscale and saturation thresholds to determine whether controls are in non-interactive states such as loading, obscuring, or grayed-out and disabled. It identifies and structurally encapsulates the real-time interactive attributes and coordinate information of each control to form a standardized control usability state.

[0117] Step S550: Generate real-time task execution status information based on the availability status of the control, complete the automated execution of the terminal device task and obtain the real-time task execution status information.

[0118] Integrate real-time task execution status information to generate a real-time task execution status sequence. The formula for generating the real-time task execution status sequence is as follows: (25) In formula (25), It is a real-time execution state sequence of a task, without units, derived from the integration of multi-dimensional state information, and represents the time-series running state data of the entire task process; For the first The available status of each operation control is unitless and comes from the detection result of formula (24), representing the effective status of the control in a single-step operation; For the first The step operation execution timestamp, in seconds, is derived from the system clock record and is used to mark the specific timing node of abnormal operating conditions; The total number of execution steps is a unitless number derived from the statistical results of the operation process. It is a positive integer representing the total number of operation steps in the complete task. The control logic of formula (25) is to integrate the available status of the controls and the execution timestamp of each step according to the task execution sequence, construct a time-sequential and structured task execution state sequence, and fully record the normal operation and abnormal working condition data of the task to support the subsequent iterative optimization of the preference model. Formula (25) constructs a full-process time-sequential state dataset, realizes the quantification and traceability of task execution conditions, and provides a core data foundation for the closed-loop self-optimization of the system.

[0119] The intelligent task processing system updates task execution progress and exception indicators in real time based on the availability status of controls, generates real-time task execution status information, and fully records the running status of automated tasks.

[0120] Preferably, the intelligent task processing method based on digital employees provided in this embodiment includes step S600 as follows: Step S610: Extract the task execution state gradient vector corresponding to the real-time execution state information of the task.

[0121] The formula for calculating the gradient vector of the task execution state is: (26) In formula (26), The gradient vector of the task execution state is dimensionless and comes from the result of the temporal state difference operation. It represents the overall deviation of the current task execution state from the historical baseline state. This is the real-time execution status sequence of this task, without units, derived from the integrated result of formula (25), and serves as the real-time data benchmark for deviation calculation; The historical standard benchmark execution state sequence is unitless and derived from the summary of the average of the best historical states of similar tasks, serving as the standard reference data for deviation calculation. The control logic of formula (26) is to perform a difference operation between the current task's real-time execution state sequence and the historical standard benchmark state sequence to quantify the overall operational deviation and degree of abnormality of the task, and accurately locate the fluctuation of the task's execution effect. Traditional preference updates lack quantitative deviation basis, and iterative optimization is highly blind. This formula (26) quantifies the operational deviation through state gradient, accurately locks the optimization direction, and realizes targeted iterative updates of the preference model.

[0122] The intelligent task processing system extracts real-time task execution status information and calculates the task execution status gradient vector within a continuous time window. This gradient vector is used to quantitatively characterize the operational fluctuation trend of automated tasks and the changing patterns of system resource load. Based on the response latency and task success indicators in the real-time status, the intelligent task processing system calculates the latency difference ratio between adjacent time windows, using this ratio as the quantified value of the task execution status gradient vector.

[0123] Step S620: Determine the task frequency feature attenuation factor and time distribution feature offset based on the task execution state gradient vector.

[0124] The formulas for determining the task frequency characteristic attenuation factor and time distribution characteristic offset are as follows: (27) (28) In formulas (27)~(28), This is a task frequency characteristic attenuation factor, dimensionless, derived from exponential calculation results, with a value range of [value range missing]. The larger the deviation, the smaller the attenuation factor, and the greater the adjustment of the historical frequency preference weight. It is a natural exponential function, without units, derived from standard mathematical operations, and used to construct a nonlinear mapping relationship between state deviation and decay factor; This is the gradient decay coefficient, which is dimensionless and derived from preset control parameters. Its value range is 0.1 to 1.0, and it is used to control the rate of decay of frequency characteristics. The magnitude of the gradient vector of the task execution state is a unitless value derived from the result of the gradient vector modulus operation, representing the overall deviation of the task execution. This is the time distribution feature offset, in seconds, derived from the deviation offset calculation result, and used to correct the deviation error of the user's operation time preference. This is the time offset coefficient, which has no unit and is derived from preset control parameters. Its value range is 10~60, and it is used to control the adjustment range of time preference. The symbolic function is a unitless function derived from standard mathematical functions, taking values ​​of -1, 0, and 1. It is used to determine the positive and negative directions of the deviation and to determine the correction direction of the time offset. The control logic of formulas (27) to (28) is to adaptively calculate the task frequency characteristic attenuation factor based on the state gradient magnitude. The larger the task operation deviation, the lower the credibility of the historical frequency preference, and the stronger the attenuation control effect. At the same time, the time distribution characteristic offset is generated by combining the deviation direction and amplitude to accurately correct the deviation of the user's time operation preference. The traditional user preference model has fixed weights and cannot adapt to the dynamic changes of user operation habits and equipment conditions. Formulas (27) to (28) rely on the real-time operation deviation to adaptively generate adjustment parameters, realize the dynamic correction of preference weights, and adapt to the dynamically changing user operation habits.

[0125] When the gradient vector of the task execution state indicates a sharp increase in task response latency and a surge in system load, the intelligent task processing system calculates the task frequency characteristic decay factor by combining a preset negative exponential function with the gradient magnitude. This factor is used to quantify and reduce the automatic triggering frequency of the corresponding task. At the same time, based on the direction and magnitude of the gradient change, the system solves for the task time distribution characteristic offset, which is used to characterize the precise offset duration of shifting the task execution period to the idle interval of system resources.

[0126] Step S630: Calculate the overall adjustment magnitude vector of preference feature weights using the task frequency feature attenuation factor and time distribution feature offset.

[0127] The formula for calculating the overall adjustment magnitude vector of preference feature weights is: (29) In formula (29), This is a vector representing the overall adjustment magnitude of the preference feature weights. It is dimensionless and comes from the result of the fusion calculation of the two-dimensional parameters, representing the overall correction magnitude of the preference weights in the frequency and time dimensions. The task frequency characteristic attenuation factor is a unitless factor derived from the calculation result of formula (27) and is used to adjust the frequency weight adjustment ratio. This is the basic weight vector for task frequency features. It is unitless, derived from the system's preset initial parameters, and represents the baseline weights that characterize the preference for the frequency dimension. The time distribution feature offset, in seconds, is derived from the calculation result of formula (28) and is used to adjust the adjustment range of the time weight. The time distribution feature is a basic weight vector, which is unitless and comes from the system's preset initial parameters, representing the baseline weight of the time dimension preference. The control logic of formula (29) is to integrate the task frequency feature decay factor, the time distribution feature offset and the corresponding basic weight parameters, calculate the weight adjustment amount of the frequency feature and the time feature respectively, and integrate them to generate an overall weight adjustment magnitude vector, so as to simultaneously complete the accurate correction of the two-dimensional preference weights. This formula (29) realizes the coordinated dynamic adjustment of multi-dimensional preference weights, which solves the problems of single-dimensional update, low adjustment accuracy and poor adaptability of traditional models.

[0128] The intelligent task processing system uses the task frequency feature attenuation factor as the frequency dimension adjustment coefficient and the time distribution feature offset as the time dimension penalty term. It then uses linear combination operations to solve for the overall adjustment magnitude vector of the task parameter preference feature weights.

[0129] Step S640: Construct the preference parameter update diagonal matrix by adjusting the overall magnitude vector based on the preference feature weights.

[0130] The formula for constructing the diagonal matrix for preference parameter updates is: (30) In formula (30), The diagonal matrix is ​​updated for preference parameters. It is unitless and derived from the result of vector-to-matrix conversion. It is used to achieve differentiated and independent updates of multi-dimensional preference weights. This is a diagonal matrix generation function, which is unitless and derived from standard matrix operations. It is used to convert a one-dimensional weight adjustment vector into a diagonal update matrix. The overall adjustment magnitude vector of the preference feature weights is dimensionless and derived from the calculation result of formula (29), serving as the core input data for matrix generation. The control logic of formula (30) is to convert the one-dimensional weight adjustment magnitude vector into a diagonal update matrix, thereby achieving independent correction of the preference parameters in each dimension and avoiding mutual interference during the update process of different feature dimensions. Formula (30) achieves decoupled update of multi-dimensional preference parameters through matrix operations, significantly improving the accuracy and stability of the iterative update of the preference model.

[0131] The intelligent task processing system constructs a two-dimensional preference parameter update diagonal matrix with task type as the row dimension and time period as the column dimension. The overall adjustment magnitude vector of preference feature weights is filled into the corresponding matrix unit to form a feature update matrix with diagonal optimization as the main feature.

[0132] S650. The preference parameter update diagonal matrix is ​​fused and mapped with the initial feature vector. The weight parameters of task frequency features and time distribution features are dynamically adjusted in combination with the real-time execution status to generate the updated user preference feature vector.

[0133] The updated formula for generating user preference feature vectors is: (31) In formula (31), The updated user preference feature vector is unitless and comes from the result of the preference iteration calculation. It represents the latest user operation preference after the update and serves as the optimization benchmark for the next round of task processing. This is the user preference feature vector for the current round, which is unitless and comes from the retention parameters of the previous iteration. It serves as the base vector for this update. The diagonal matrix is ​​updated for preference parameters. It has no unit and is derived from the calculation result of formula (30). It is the core control matrix for preference weight update. The initial baseline preference feature vector for the user is unitless and originates from the modeling results in step S130, representing the user's original inherent preference baseline. The control logic of formula (31) relies on the diagonal update matrix to correct the deviation between the current preference vector and the initial baseline vector, dynamically adjust the frequency and time-dimension preference weight parameters, generate a preference feature vector that adapts to the user's latest operating habits and equipment conditions, and complete the closed-loop self-optimization of task execution-state feedback-preference iteration. The traditional digital employee preference model is a static fixed model that cannot adapt to user habit drift and equipment condition changes. Formula (31) builds a complete closed-loop self-optimization system, realizes online dynamic iteration of the user preference model, and continuously improves long-term operating accuracy, scenario adaptability and intelligence.

[0134] The intelligent task processing system performs matrix-dot multiplication mapping between the preference parameter update diagonal matrix and the historical user preference feature vector, dynamically corrects the task frequency feature weight and time distribution feature weight, and iteratively generates an updated user preference feature vector. This enables adaptive optimization of the digital employee task execution strategy and continuously improves the adaptability and stability of intelligent task processing.

[0135] Please see Figure 2This embodiment provides an intelligent task processing system based on digital employees, used to implement the aforementioned intelligent task processing method based on digital employees. It includes a user preference feature vector generation module 10, a personalized execution configuration information determination module 20, a basic operation sequence generation module 30, an optimal operation call path solving module 40, a task real-time execution status information acquisition module 50, and an updated user preference feature vector generation module 60. The user preference feature vector generation module 10 acquires the user's historical operation behavior sequence, uses a sequence prediction model to extract features from the historical operation behavior sequence, and generates a user preference feature vector. The user preference feature vector includes quantitative representation data of operation habits, task frequency, and time distribution characteristics. The personalized execution configuration information determination module 20 performs semantic parsing on the received natural language task instructions based on the user preference feature vector, identifies the target task intent, and determines personalized execution configuration information. The personalized execution configuration information includes the mapping relationship between task priority, trigger time window, and target operation object. The basic operation sequence generation module 30... The system combines the target task intent and personalized execution configuration information to initiate a timed scheduling mechanism, break down the target task, and generate a basic operation sequence that can be executed by the terminal device. The basic operation sequence uses touch trajectory, character input, and graphical interface control positioning as the basic execution units. The optimal operation call path solving module 40 is used to solve the optimal operation call path by matching preset automated operation units based on the graph association model if there are cross-application collaborative operations in the basic operation sequence. The automated operation units are stored in a pre-built automated integration library, and each automated operation unit corresponds to an independent application execution action. The task real-time execution status information acquisition module 50 is used to simulate human touch trajectory and character input behavior according to the optimal operation call path to complete the automated operation of the terminal device task. During the operation, the availability status of graphical interface controls is identified in real time to obtain the task real-time execution status information. The updated user preference feature vector generation module 60 is used to dynamically adjust the weight parameters of task frequency features and time distribution features in the user preference feature vector based on the task real-time execution status information to generate an updated user preference feature vector.

[0136] The intelligent task processing method and system based on digital employees provided in this embodiment have the following beneficial effects compared with the prior art: 1. Improve the accuracy of evaluation of intelligent task parsing and automated execution. This embodiment collects historical user operation behavior sequences and uses a sequence prediction model to perform deep feature extraction, quantifying and generating user preference feature vectors containing operation habits, task frequency, and time distribution features. This achieves accurate digital modeling of user behavior patterns, providing high-precision feature support for subsequent personalized task processing and avoiding the shortcomings of traditional automated scripts with fixed parameter configurations that cannot adapt to users' real operation preferences. Relying on user preference feature vectors to assist in natural language instruction semantic parsing, it accurately identifies task intents and matches personalized execution configuration information, establishing a precise mapping relationship between task priority, trigger time window, and operation object, effectively solving the problems of poor scenario adaptation, misidentification of intent, and fixed parameter configurations in traditional semantic parsing. Combining task intents and personalized configurations, it accurately decomposes tasks, generating executable operation sequences based on touch trajectories, character input, and interface control positioning, ensuring the refinement and feasibility of task decomposition. By matching automated operation units and solving for the optimal call path through a graph association model, it significantly improves the accuracy of execution logic in cross-application collaborative operations, avoiding problems such as cross-scenario operation redundancy, jump errors, and execution omissions. By identifying the availability of interface controls in real time and dynamically sensing execution progress during task execution, the controllability and accuracy of the automated execution process are ensured. Based on the real-time execution status, the weights of user preference feature vectors are dynamically and iteratively updated to continuously correct model feature biases. This allows the accuracy of task identification, scheduling, and execution to continuously improve with use, significantly enhancing the overall accuracy, stability, and success rate of task processing for digital employees.

[0137] 2. Deeply explore the underlying cognitive mechanisms of user operation patterns and task execution adaptation. This embodiment utilizes full-process data modeling and intelligent iteration mechanisms to deeply explore the inherent correlations between user behavior characteristics, task attributes, and automated execution logic, overcoming the limitations of traditional automation technologies that rely solely on fixed processes and lack user-centric understanding. First, it uses sequence modeling to quantitatively represent user operating habits, task frequency, and time distribution, uncovering implicit patterns and personalized preferences in long-term user behavior, thus upgrading from superficial operation records to a deeper understanding of behavioral mechanisms. It constructs a mapping relationship between user preferences and task semantics and execution configurations, revealing the differentiated execution needs of different users under similar tasks and clarifying the matching mechanism between personalized configurations and efficient task execution. It achieves hierarchical decomposition of complex tasks into basic operation units, uncovering the logical deconstruction patterns between complex business tasks and terminal-level operations. Finally, it leverages graph association models to explore the relationships and optimal collaboration logic among automated operation units across multiple application scenarios, clarifying the linkage mechanism across application tasks. Establish a closed-loop iterative mechanism for task execution feedback and user preference characteristics, explore the impact of task execution results on the updating of user behavior characteristics, realize the continuous evolution of digital employees' understanding of user operations, and completely change the traditional automation operation mode of "no cognition, no iteration, no adaptation".

[0138] 3. Significantly improves the practicality of the technical solution in various scenarios and its adaptability to engineering implementation. This embodiment constructs a fully closed-loop intelligent task processing system encompassing "behavioral modeling, semantic parsing, task decomposition, cross-domain collaboration, simulation execution, and dynamic iteration." It boasts strong engineering practicality, a wide range of scenario adaptability, and significant industrialization value. This invention uses natural language commands as the task entry point, lowering the user's operational threshold and eliminating the need for users to write complex scripts or configure cumbersome rules, thus adapting to the daily office and business automation needs of ordinary users. Through a personalized execution configuration mechanism, it can adapt to different users' operating habits, work rhythms, and task preferences, solving the problem of poor adaptability and uniformity inherent in traditional digital employees. Simultaneously, this invention adopts a standardized automated operation unit library architecture, achieving cross-application and cross-scenario task collaborative scheduling through a graph association model. It can adapt to independent tasks in a single application and complex tasks involving multiple applications, covering diverse automation scenarios such as daily office work, business process processing, and system maintenance. The overall task execution adopts a simulated human touch and input method, requiring no modification to existing terminal systems and software interfaces, exhibiting strong compatibility and rapid deployment on various terminal devices. Furthermore, this invention possesses continuous self-iterative optimization capabilities, which can continuously optimize the execution strategy as the user goes through the process, maintain a highly efficient and adaptable state in the long term, significantly reduce the costs of manual intervention, rule maintenance, and script updates, and its practicality and scalability are significantly better than traditional fixed automation solutions.

[0139] 4. Establish a self-learning closed-loop optimization mechanism to continuously improve the intelligence level of digital employees. This embodiment constructs a complete self-learning and self-optimization closed-loop system through real-time execution status feedback and dynamic update mechanism of user preference features, overcoming the shortcomings of traditional automation tools such as static configuration, fixed capabilities, and inability to evolve autonomously. This invention can dynamically adjust the feature weights such as task frequency and time distribution based on the actual execution of each task, continuously adapting to the latest user operating habits and business change trends, and achieving dynamic optimization of task scheduling strategies, intent recognition logic, and operation execution paths. During long-term operation, digital employees can autonomously accumulate user business characteristics and scenario adaptation experience, continuously reducing the probability of false triggers, incorrect executions, and missed executions, continuously improving the processing capacity and stability of complex tasks and cross-application collaborative tasks, and achieving continuous iterative upgrades of automated service quality. It can fully adapt to dynamic changing scenarios such as user business iterations, updated operating habits, and application version upgrades, possessing extremely strong long-term use value and intelligent evolution capabilities.

[0140] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A smart task processing method based on digital employees, characterized in that, Includes the following steps: S100. Obtain the user's historical operation behavior sequence, use a sequence prediction model to extract features from the user's historical operation behavior sequence, and generate a user preference feature vector. The user preference feature vector includes quantitative representation data of operation habits, task frequency, and time distribution features. S200. Based on the user preference feature vector, perform semantic parsing on the received natural language task instructions, identify the target task intent and determine personalized execution configuration information, wherein the personalized execution configuration information includes the mapping relationship between task priority, trigger time window and target operation object; S300. Combining the target task intent and personalized execution configuration information, start the timed scheduling mechanism, split the target task, and generate a basic operation sequence that can be executed by the terminal device. The basic operation sequence uses touch trajectory, character input, and graphical interface control positioning as the basic execution unit. S400. If the basic operation sequence has cross-application collaborative operations, then the preset automated operation unit is matched based on the graph association model to solve the optimal operation call path. The automated operation unit is stored in a pre-built automated integration library, and a single automated operation unit corresponds to an independent application execution action. S500: According to the optimal operation call path, simulate the manual touch trajectory and character input behavior to complete the automated operation of the terminal device task. During the operation, identify the available status of the graphical interface control in real time and obtain the real-time execution status information of the task. S600. Based on the real-time execution status information of the task, dynamically adjust the weight parameters of the task frequency feature and time distribution feature in the user preference feature vector to generate an updated user preference feature vector.

2. The intelligent task processing method based on digital employees according to claim 1, characterized in that, Step S100 includes: S110. Extract timestamp information and behavior trajectory from the interaction log to construct a sequence of user's historical operation behavior; S120. Extract task frequency features based on the user's historical operation behavior sequence, and determine time distribution features based on timestamps; S130. Use a sequence prediction model to perform feature mining on task frequency features and time distribution features to obtain user deep operation habit feature vectors. S140. By integrating task frequency features, time distribution features, and user deep operation habit feature vectors, a user preference feature vector is generated.

3. The intelligent task processing method based on digital employees according to claim 2, characterized in that, Step S200 includes: S210. Obtain natural language task instructions, and combine the natural language task instructions with the user preference feature vector to determine the target task intent and instruction semantic feature representation. S220. Based on the target task intent, extract entity nouns from the instruction semantic feature representation to identify the target operation object; S230. Retrieve the time distribution feature and task frequency feature data from the user preference feature vector, combine the time distribution feature and the available time period of the target operation object to define the trigger time window, and determine the task priority corresponding to the target task intent based on the task frequency feature. S240. Construct the mapping relationship between the task priority, the trigger time window and the target operation object, and generate personalized execution configuration information.

4. The intelligent task processing method based on digital employees according to claim 1, characterized in that, Step S300 includes: S310. Decompose the target task intent according to the preset task splitting rules to obtain multiple sub-task nodes; S320. Determine the location of the graphical interface control based on the control identifier that matches the subtask node; S330. Extract the touch trajectory coordinates and character input sequence corresponding to the subtask node; S340. The positioning of the graphical interface control, the coordinates of the touch trajectory, and the character input sequence are combined to generate a basic operation sequence that can be executed by the terminal device.

5. The intelligent task processing method based on digital employees according to claim 4, characterized in that, Step S400 includes: S410. Obtain the set of application identifiers contained in the basic operation sequence. If there are at least two different application identifiers in the set of application identifiers, extract the application state context corresponding to the cross-application collaborative operation. S420. Based on the application state context, retrieve several independent application execution actions from the pre-built automated integration library, and construct a graph association model with the independent application execution actions as nodes. S430. Determine the path weight by combining the historical execution time and historical resource consumption data of the nodes in the graph association model; S440. The shortest path algorithm is used to perform path optimization on the graph association model carrying the path weight to obtain the optimal operation call path, and the corresponding preset automated operation unit is matched according to the optimal operation call path.

6. The intelligent task processing method based on digital employees according to claim 1, characterized in that, Step S500 includes: S510. Analyze the optimal operation call path to obtain the initial touch coordinate sequence and character input instruction set; S520. Process the initial touch coordinate sequence to generate a simulated manual touch trajectory; S530. Based on the simulated manual touch trajectory and the character input instruction set, initiate automated task execution of the terminal device; S540. Collect the available status of controls corresponding to the graphical interface images during the automated execution process; S550. Generate real-time task execution status information based on the availability status of the control, complete the automated execution of the terminal device task, and obtain the real-time task execution status information.

7. The intelligent task processing method based on digital employees according to claim 6, characterized in that, Step S600 includes: S610. Extract the task execution state gradient vector corresponding to the real-time execution state information of the task. The formula for calculating the gradient vector of the task execution state is: ； in, The gradient vector of the task execution state. This is the real-time execution status sequence for this task. This is a historical standard benchmark execution state sequence; S620. Determine the task frequency feature attenuation factor and time distribution feature offset based on the task execution state gradient vector. The formulas for determining the task frequency characteristic attenuation factor and time distribution characteristic offset are as follows: ； ； in, This is a task frequency characteristic attenuation factor. It is a natural exponential function. The gradient decay coefficient is... The magnitude of the gradient vector of the task execution state. This represents the offset of the time distribution characteristics. This is the time offset coefficient. It is a symbolic function; S630. Calculate the overall adjustment magnitude vector of the preference feature weight using the task frequency feature attenuation factor and the time distribution feature offset. The formula for calculating the overall adjustment magnitude vector of preference feature weights is: ； in, Adjust the overall magnitude vector of the preference feature weights. The basic weight vector for task frequency features. The basic weight vector for time distribution characteristics; S640. Construct a preference parameter update diagonal matrix based on the overall adjustment magnitude vector of the preference feature weights; S650. The preference parameter update diagonal matrix is ​​fused and mapped with the initial feature vector, and the weight parameters of task frequency feature and time distribution feature are dynamically adjusted in combination with the real-time execution status to generate an updated user preference feature vector.

8. The intelligent task processing method based on digital employees according to claim 7, characterized in that, In step S640, the formula for constructing the diagonal matrix for updating the preference parameters is: ； in, Update the diagonal matrix for preference parameters. This is a function for generating diagonal matrices.

9. The intelligent task processing method based on digital employees according to claim 8, characterized in that, In step S650, the updated formula for generating the user preference feature vector is: ； in, This is the updated user preference feature vector. This is the user preference feature vector for the current round. This is the user's initial baseline preference feature vector.

10. A digital employee-based intelligent task processing system, used to implement the digital employee-based intelligent task processing method as described in any one of claims 1 to 9, characterized in that, include: The user preference feature vector generation module is used to obtain the user's historical operation behavior sequence, use a sequence prediction model to extract features from the user's historical operation behavior sequence, and generate a user preference feature vector. The user preference feature vector includes quantitative representation data of operation habits, task frequency, and time distribution features. The personalized execution configuration information determination module is used to perform semantic parsing on the received natural language task instructions based on the user preference feature vector, identify the target task intent and determine personalized execution configuration information, wherein the personalized execution configuration information includes the mapping relationship between task priority, trigger time window and target operation object; The basic operation sequence generation module is used to combine the target task intent and personalized execution configuration information, start the timed scheduling mechanism, split the target task, and generate a basic operation sequence that can be executed by the terminal device. The basic operation sequence uses touch trajectory, character input, and graphical interface control positioning as the basic execution unit. The optimal operation call path solving module is used to solve the optimal operation call path by matching preset automated operation units based on the graph association model if there are cross-application collaborative operations in the basic operation sequence. The automated operation units are stored in a pre-built automated integration library, and each automated operation unit corresponds to an independent application execution action. The task real-time execution status information acquisition module is used to simulate human touch trajectory and character input behavior according to the optimal operation call path, complete the automated operation of terminal device tasks, and identify the available status of graphical interface controls in real time during operation to obtain task real-time execution status information. The updated user preference feature vector generation module is used to dynamically adjust the weight parameters of task frequency features and time distribution features within the user preference feature vector based on the real-time execution status information of the task, and generate the updated user preference feature vector.

Images