Human-machine collaboration method and system based on robotic process automation

Abstract

The present invention provides a human-machine collaboration method and system based on robotic process automation, which relates to the field of automation technology. The method includes constructing a data perception processing network, collecting user operation instructions, system operation status and task environment data, obtaining task feature vectors through semantic analysis, and matching business models to generate task processing strategies. Based on this strategy, a human-machine collaborative processing framework is constructed, in which the robot unit processes structured tasks and the manual operation unit processes unstructured tasks. The two units are connected through an information interaction channel to realize the transmission of data analysis results and decision recommendations when the deviation exceeds the threshold and the update of processing procedure information. In addition, a knowledge transfer processing network is constructed to extract business rules and expert experience data features, train a scenario migration model, and map the business feature vectors to a preset scenario feature space. After validity verification, the parameters of the data perception processing network and the human-machine collaborative processing framework are updated.

Description

Technical Field

[0001] The present invention relates to automation technology, and in particular to a human-machine collaboration method and system based on robotic process automation. Background Art

[0002] Robotic process automation (RPA) technology has made significant progress in automating various business processes in recent years, particularly in processing structured data and rule-based tasks. Traditional RPA systems excel at performing predefined, repetitive tasks such as data entry, form filling, and report generation. These systems typically rely on clear rules and workflows and have limitations when processing unstructured data or complex scenarios requiring human judgment and intervention. Existing RPA systems suffer from the following three major flaws and shortcomings:

[0003] First, traditional RPA systems struggle to adapt to dynamically changing environments. They typically require explicit rules and instructions, and when faced with unexpected situations or process changes, they require extensive reprogramming and configuration. This limits their flexibility in handling complex and unpredictable business processes.

[0004] Second, traditional RPA systems lack effective knowledge transfer capabilities. They are typically limited to processing specific types of data and tasks, and struggle to apply learned knowledge to new scenarios. This leads to repeated development and increased deployment costs, limiting the widespread adoption of RPA technology.

[0005] Finally, traditional RPA systems lack the ability to collaborate with humans and machines. They typically treat humans and machines as independent entities, lacking effective communication and collaboration mechanisms. This leads to inefficiencies and limits the potential of human-machine collaboration. In many real-world scenarios, humans and machines must work together to complete tasks, such as when processing unstructured data or when human judgment and intervention are required. Summary of the Invention

[0006] The embodiments of the present invention provide a human-machine collaboration method and system based on robotic process automation, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a human-machine collaboration approach based on robotic process automation, including:

[0009] Construct a data perception and processing network, which includes a data acquisition layer and a semantic understanding layer. The data acquisition layer collects user operation instruction data, system operation status data, and task environment data through distributed sensors. The semantic understanding layer performs semantic analysis on the collected data based on a deep neural network to obtain a task feature vector. The task feature vector is matched with the preset business model to construct a task mapping matrix and generate a task processing strategy.

[0010] A human-machine collaborative processing framework is constructed based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The robot unit and the human operation unit are connected through an information exchange channel. When the deviation value during task processing exceeds a specified threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the human operation unit, and the human operation unit transmits the processing procedure information to the robot unit's rule database.

[0011] A knowledge transfer processing network is constructed, and a graph embedding algorithm is used to extract features from business rule data and expert experience data in the human-computer collaborative processing process to obtain business feature vectors. A scenario migration model is trained based on the business feature vectors, and the business feature vectors are mapped to a preset scenario feature space to generate scenario mapping data. The knowledge transfer processing network also includes a verification module that verifies the validity of the scenario mapping data and outputs the verification results, and updates the processing parameters of the data perception processing network and the task allocation parameters of the human-computer collaborative processing framework based on the verification results.

[0012] In an optional embodiment,

[0013] The task feature vector is matched with the preset business model to construct a task mapping matrix. The steps of generating the task processing strategy include:

[0014] Construct a model library containing multiple preset business models, each of which contains a standard feature vector and a corresponding processing template. Based on contrastive learning, similarity is calculated between the task feature vector and the standard feature vector. The task feature vector is mapped to the feature space corresponding to the standard feature vector with the highest similarity. A task mapping matrix is ​​constructed based on the mapping results.

[0015] Inputting the task mapping matrix into a hierarchical decoder, the hierarchical decoder comprising a task decomposition layer, a resource allocation layer, and an execution sequence layer connected in sequence, wherein the task decomposition layer generates a task decomposition strategy based on the task mapping matrix, the resource allocation layer generates a resource allocation strategy based on the task mapping matrix and the task decomposition strategy, and the execution sequence layer generates an execution sequence strategy based on the task mapping matrix, the task decomposition strategy, and the resource allocation strategy;

[0016] Calculating a constraint adaptation factor based on the task feature vector, the resource state output of the resource allocation layer, and a time constraint parameter; performing constraint adjustments on the task decomposition strategy, the resource allocation strategy, and the execution sequence strategy based on the constraint adaptation factor to obtain a task processing strategy; constructing a dominance relationship matrix of the task processing strategy, calculating an optimal strategy combination based on the dominance relationship matrix, and determining a strategy weight based on the contribution of the task decomposition strategy, the resource allocation strategy, and the execution sequence strategy to the optimal strategy combination;

[0017] The policy weight and the optimal policy combination are input into a policy optimization network, which includes a policy storage module and a policy sampling module. The policy storage module stores historical policy data, and the policy sampling module performs importance sampling on the historical policy data to obtain sampling policy data. The policy evaluation parameters are calculated based on the sampling policy data, and the network parameters of the layered decoder are optimized and updated according to the policy evaluation parameters.

[0018] In an optional embodiment,

[0019] The human-machine collaborative processing framework is constructed based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The steps include:

[0020] Acquiring a task type feature vector based on the task processing strategy, calculating an entropy value of the task type feature vector based on an information entropy calculation model, and dividing tasks into structured tasks and unstructured tasks according to the entropy value;

[0021] A strategy mapping mechanism for a robot execution unit is constructed based on the task processing strategy, the task decomposition strategy in the task processing strategy is mapped to a robot basic action library to obtain an action correspondence matrix, action execution parameters are determined according to the resource allocation strategy in the task processing strategy, a robot control instruction stream is generated based on the execution sequence strategy in the task processing strategy, and the robot execution unit processes the structured task according to the robot control instruction stream;

[0022] constructing an operator cognitive state vector including attention level, workload, and expertise, evaluating the operator cognitive state vector based on a resource allocation strategy in the task processing strategy, dynamically adjusting information display density and frequency of a manual operation interface based on the evaluation result, and processing the unstructured task through the manual operation interface;

[0023] Construct a collaborative effectiveness evaluation index that includes execution efficiency, accuracy and cognitive load, determine the collaborative weight coefficient of the collaborative effectiveness evaluation index according to the strategy weight in the task processing strategy, calculate the human-computer collaborative processing efficiency score based on the collaborative weight coefficient, and use the human-computer collaborative processing efficiency score as a feedback signal to optimize and update the task processing strategy.

[0024] In an optional embodiment,

[0025] The robot unit and the manual operation unit are connected via an information exchange channel. When the deviation value during task processing exceeds a predetermined threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the manual operation unit, and the manual operation unit transmits the processing procedure information to the rule database of the robot unit. The steps include:

[0026] The task execution trajectory is decomposed into time-frequency feature representations of different scales through wavelet transform, and the time-frequency feature representations of different scales are weighted by the adaptive attention mechanism. The weighted features are fused to obtain a global deviation feature. The global deviation feature sequence is modeled using a recursive neural network, and the deviation degree score is calculated using a multi-layer perceptron. When the deviation degree score exceeds a preset deviation threshold, an anomaly detection signal is triggered; an anomaly pattern feature library is constructed based on historical execution data, and the type and cause of the current anomaly are matched from the anomaly pattern feature library based on the anomaly detection signal;

[0027] Construct a multi-level knowledge graph structure that includes a task-level graph, a resource-level graph, and a processing-level graph. The task-level graph describes the dependencies between tasks, the resource-level graph describes the distribution of equipment and personnel capabilities, and the processing-level graph describes the operational process. Update node information in the multi-level knowledge graph structure based on the exception type and cause.

[0028] A graph neural network with a multi-head attention mechanism is used to transfer information between layers of a multi-level knowledge graph structure, generating node feature representations by calculating normalized weights between nodes. The node feature representations are combined with historical sequence features to construct a state vector. A policy network is trained based on the state vector, using a weighted combination of task completion and resource utilization as a reward signal, and a policy gradient method is used to optimize network parameters to obtain a decision model. Guided by the decision model, the Monte Carlo tree search algorithm is used to generate an optimal decision path using visit counts, value estimates, and prior probabilities as search attributes. A local causal graph containing variable nodes and causal edges is constructed for the nodes in the optimal decision path to generate processing recommendations.

[0029] The processing suggestion is represented as a processing procedure of a condition-action pair, an importance evaluation is performed on the processing procedure, and the processing procedure whose importance exceeds a preset importance threshold is updated to a rule database. The update result of the rule database is used to adjust the preset deviation threshold.

[0030] In an optional embodiment,

[0031] The steps of constructing a knowledge transfer processing network, extracting features from business rule data and expert experience data in the human-machine collaborative processing process using a graph embedding algorithm to obtain a business feature vector, training a scenario transfer model based on the business feature vector, and mapping the business feature vector to a preset scenario feature space to generate scenario mapping data include:

[0032] The knowledge transfer processing network includes a knowledge graph construction module, a feature extraction module and a scene transfer module;

[0033] The knowledge graph construction module is used to construct a knowledge graph including rule nodes, experience nodes and relationship edges. The rule nodes use the triple form of trigger conditions, operation sequences and constraints to describe the robotic process automation operation procedures. The experience nodes use the structured form of abnormal scenarios, processing steps and expected effects to represent process optimization techniques. The relationship edges describe the inclusion relationship, pre-requisite relationship and optimization relationship between the rule nodes and the experience nodes.

[0034] The feature extraction module uses a graph embedding algorithm with a multi-head attention mechanism to extract features from the business rule data and expert experience data in the knowledge graph, where the business rule data comes from the trigger conditions, operation sequences, and constraints of the rule nodes, and the expert experience data comes from the abnormal scenarios, processing steps, and expected effects of the experience nodes. The business feature vector is obtained by calculating the attention scores of the set of nodes directly connected to the current processing node and performing feature aggregation. The business feature vector contains the interactive relationship information between the robotic process automation operation nodes;

[0035] The scenario migration module trains a scenario migration model based on the business feature vector. The scenario migration model includes a discriminator network and a generator network. The discriminator network is used to distinguish between the real process samples of the target scenario and the generated migration samples. The generator network is used to map the business feature vector of the source scenario to a preset scenario feature space.

[0036] A contrastive learning mechanism is used to align the feature spaces of the source scene and the target scene. The same processing steps in the source and target scenes are constructed as positive sample pairs, and the different processing steps are constructed as negative samples. By minimizing the contrastive loss function, the feature representations of the same operation steps are made close in the feature space to obtain scene mapping data.

[0037] In an optional embodiment,

[0038] The scene migration module trains a scene migration model based on the business feature vector. The scene migration model includes a discriminator network and a generator network. The discriminator network is used to distinguish between the real process samples of the target scene and the generated migration samples. The generator network is used to map the business feature vector of the source scene to the preset scene feature space. The steps include:

[0039] The generator network adopts a three-layer fully connected neural network structure, each layer is connected by a ReLU activation function, and the last layer outputs a target scene feature vector through a hyperbolic tangent activation function. The discriminator network adopts a four-layer fully connected network structure, and a batch normalization layer and a random dropout layer are set after each hidden layer; a discriminator loss function is constructed based on the Wasserstein distance, and the discriminator loss function includes a real sample expectation term, a generated sample expectation term, and a gradient penalty term. The discriminator network is trained by the discriminator loss function; a generator loss function is constructed, and the generator loss function includes an adversarial loss term output by the discriminator and a semantic consistency loss term, wherein the semantic consistency loss term is obtained by calculating the Euclidean distance between the source scene feature and the generated feature and the Euclidean distance between the features of the intermediate layer of the feature extraction network;

[0040] Constructing feature quality evaluation indicators, which include a feature consistency score and a task relevance score, where the feature consistency score is calculated based on the feature space distance between the target scene features and the generated features, and the task relevance score is calculated based on the accuracy of the downstream task classifier;

[0041] An alternating optimization strategy is adopted to train the scene migration model, the generator network is fixed, the parameters of the discriminator network are optimized multiple times through iterations, the discriminator network is fixed, and the parameters of the generator network are optimized once; the business feature vector is divided into the training scale according to the number of rule node triplets, the number of rule node triplets of the initial training scale is set to 1, the growth step value of the number of rule node triplets is set, and the number of training rounds for each training scale is set; the scene migration model is trained based on the business feature vector of the current training scale, and when the feature quality evaluation index reaches a preset threshold, the number of rule node triplets is increased according to the growth step value, and the scene migration model is continued to be trained based on the increased training scale until a preset number of training times is reached; the generated target scene features are evaluated based on the feature quality evaluation index, and the weight coefficient of the semantic consistency loss term and the growth step value of the number of rule node triplets are adjusted according to the target scene feature evaluation result.

[0042] In an optional embodiment,

[0043] The knowledge transfer processing network further includes a verification module for verifying the validity of the scene mapping data and outputting a verification result. The steps of updating the processing parameters of the data perception processing network and the task allocation parameters of the human-machine collaborative processing framework based on the verification result include:

[0044] Constructing a verification matrix to verify the scene mapping data, the verification matrix includes a semantic consistency vector, a structural similarity vector, and a temporal rationality vector, wherein the semantic consistency vector is obtained by calculating the keyword matching degree of the business rules before and after mapping, the structural similarity vector is calculated based on the edit distance of the graph structure, and the temporal rationality vector is obtained by verifying the dependency relationship of the operation sequence;

[0045] Constructing a multi-level verification index based on the verification matrix, wherein the multi-level verification index includes a rule-level verification index and a process-level verification index, wherein the rule-level verification index is obtained by calculating rule similarity and rule internal consistency, and the process-level verification index is obtained by calculating timing constraint satisfaction and structural integrity;

[0046] Performing a weighted combination of the multi-level verification indicators and the business scenario coverage to generate a verification evaluation score;

[0047] Updating parameters of the data perception processing network based on the verification evaluation score, including: gradient updating the perception layer parameters using an adaptive learning rate, the adaptive learning rate dynamically adjusting with the verification evaluation score; and determining an adjustment direction and an adjustment step size of the data perception processing network parameters based on a difference between the verification evaluation score and a preset threshold;

[0048] A task allocation matrix is ​​constructed, wherein the elements of the task allocation matrix represent the probability of tasks being assigned to processing units, and an update amount of the task allocation matrix is ​​calculated based on the verification evaluation score and resource utilization; the task allocation scheme in the human-computer collaborative processing framework is dynamically adjusted according to the update amount of the task allocation matrix, and the task allocation parameters in the human-computer collaborative processing framework are optimized based on the adjusted task allocation scheme.

[0049] According to a second aspect of the embodiments of the present invention,

[0050] Provides a human-machine collaborative system based on robotic process automation, including:

[0051] The first unit is used to build a data perception and processing network, collect user operation instruction data, system operation status data and task environment data, perform semantic analysis on the collected data to obtain task feature vectors, perform feature matching between the task feature vectors and the preset business model to construct a task mapping matrix and generate a task processing strategy;

[0052] The second unit is used to build a human-machine collaborative processing framework based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The robot unit and the human operation unit are connected through an information exchange channel. When the deviation value during the task processing process exceeds a specified threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the human operation unit, and the human operation unit transmits the processing procedure information to the rule database of the robot unit;

[0053] The third unit is used to construct a knowledge transfer processing network, use a graph embedding algorithm to extract features from business rule data and expert experience data in the human-computer collaborative processing process to obtain business feature vectors, train a scenario migration model based on the business feature vectors, and map the business feature vectors to a preset scenario feature space to generate scenario mapping data; the knowledge transfer processing network also includes a verification module, which verifies the validity of the scenario mapping data and outputs the verification results, and updates the processing parameters of the data perception processing network and the task allocation parameters of the human-computer collaborative processing framework based on the verification results.

[0054] According to a third aspect of the embodiments of the present invention,

[0055] An electronic device is provided, comprising:

[0056] processor;

[0057] a memory for storing processor-executable instructions;

[0058] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0059] According to a fourth aspect of the embodiments of the present invention,

[0060] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0061] This invention leverages a data-sensing processing network to quickly and accurately understand user commands and system operating status, and, combined with task environment data, generates an optimal task processing strategy. Leveraging a human-machine collaborative processing framework, robots can efficiently handle structured tasks while humans focus on unstructured tasks. The two work together to improve task processing efficiency.

[0062] When deviations occur during task processing, the robot can provide historical data analysis results and decision-making recommendations to the human operator. The human operator can then adjust the processing procedures based on actual conditions and update the robot's rule database, enabling the system to adapt to complex and changing task environments. The knowledge transfer processing network can transfer business rules and expert experience from human-machine collaborative processing to new scenarios, improving the system's generalization and adaptability.

[0063] This invention enables real-time communication and collaboration between robots and humans through information exchange channels, complementing each other's strengths. The knowledge transfer processing network's verification module verifies the validity of scene mapping data and, based on the verification results, optimizes the parameters of the data perception processing network and the human-machine collaborative processing framework, thereby continuously optimizing the human-machine collaborative process and improving collaborative efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of a process flow of a human-machine collaboration method based on robotic process automation according to an embodiment of the present invention;

[0065] Figure 2 Schematic diagram of the structure of a human-machine collaborative system based on robotic process automation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0067] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0068] Figure 1 FIG. 1 is a flow chart of a human-machine collaboration method based on robotic process automation according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] Build a data perception and processing network to collect user operation instruction data, system operation status data, and task environment data. Perform semantic analysis on the collected data to obtain task feature vectors. Then, perform feature matching between the task feature vectors and the preset business model to construct a task mapping matrix and generate a task processing strategy.

[0070] A human-machine collaborative processing framework is constructed based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The robot unit and the human operation unit are connected through an information exchange channel. When the deviation value during task processing exceeds a specified threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the human operation unit, and the human operation unit transmits the processing procedure information to the robot unit's rule database.

[0071] A knowledge transfer processing network is constructed, and a graph embedding algorithm is used to extract features from business rule data and expert experience data in the human-computer collaborative processing process to obtain business feature vectors. A scenario migration model is trained based on the business feature vectors, and the business feature vectors are mapped to a preset scenario feature space to generate scenario mapping data. The knowledge transfer processing network also includes a verification module that verifies the validity of the scenario mapping data and outputs the verification results, and updates the processing parameters of the data perception processing network and the task allocation parameters of the human-computer collaborative processing framework based on the verification results.

[0072] In an optional embodiment,

[0073] The task feature vector is matched with the preset business model to construct a task mapping matrix. The steps of generating the task processing strategy include:

[0074] Construct a model library containing multiple preset business models, each of which contains a standard feature vector and a corresponding processing template. Based on contrastive learning, similarity is calculated between the task feature vector and the standard feature vector. The task feature vector is mapped to the feature space corresponding to the standard feature vector with the highest similarity. A task mapping matrix is ​​constructed based on the mapping results.

[0075] Inputting the task mapping matrix into a hierarchical decoder, the hierarchical decoder comprising a task decomposition layer, a resource allocation layer, and an execution sequence layer connected in sequence, wherein the task decomposition layer generates a task decomposition strategy based on the task mapping matrix, the resource allocation layer generates a resource allocation strategy based on the task mapping matrix and the task decomposition strategy, and the execution sequence layer generates an execution sequence strategy based on the task mapping matrix, the task decomposition strategy, and the resource allocation strategy;

[0076] Calculating a constraint adaptation factor based on the task feature vector, the resource state output of the resource allocation layer, and a time constraint parameter; performing constraint adjustments on the task decomposition strategy, the resource allocation strategy, and the execution sequence strategy based on the constraint adaptation factor to obtain a task processing strategy; constructing a dominance relationship matrix of the task processing strategy, calculating an optimal strategy combination based on the dominance relationship matrix, and determining a strategy weight based on the contribution of the task decomposition strategy, the resource allocation strategy, and the execution sequence strategy to the optimal strategy combination;

[0077] The policy weight and the optimal policy combination are input into a policy optimization network, which includes a policy storage module and a policy sampling module. The policy storage module stores historical policy data, and the policy sampling module performs importance sampling on the historical policy data to obtain sampling policy data. The policy evaluation parameters are calculated based on the sampling policy data, and the network parameters of the layered decoder are optimized and updated according to the policy evaluation parameters.

[0078] For example, during the semantic parsing phase, the BERT pre-trained model is used to extract contextual semantic features from text instructions. Word2Vec is used to map discrete operations into word vectors. For numerical data, an LSTM network is used to process temporal features. Finally, an attention mechanism is used to weightedly fuse these three features to produce a task feature vector, achieving a unified representation of user intent. Next, a model library is constructed. This library contains multiple pre-set business models, each of which includes a set of standard feature vectors and corresponding processing templates. For example, a business model for "customer complaints" might have a standard feature vector that includes features such as "complaint type," "complaint urgency," and "customer value." The corresponding processing template specifies the standard process for handling complaints of that type. Assume the model library contains three business models: "Customer Complaints," "Order Processing," and "Product Inquiry." The standard feature vectors for each model are [0.8, 0.2, 0.5], [0.1, 0.9, 0.3], and [0.3, 0.4, 0.8], respectively.

[0079] Perform feature matching. Calculate the similarity between the input task feature vector and the standard feature vector of each business model in the model library. Similarity can be calculated using methods such as cosine similarity. Map the task feature vector to the feature space corresponding to the standard feature vector with the highest similarity. For example, if the feature vector of a task is [0.7, 0.3, 0.6] and has the highest similarity with the "Customer Complaint" model, then map the task to the feature space of the "Customer Complaint" model. Assuming the calculated similarities are 0.95, 0.75, and 0.85, respectively, then the task is mapped to the "Customer Complaint" model.

[0080] Construct a task mapping matrix. This matrix reflects the mapping relationship between tasks and various business models. For example, if a task is mapped to the "Customer Complaint" model, the value of the position corresponding to the "Customer Complaint" model in the matrix is ​​1, and the values ​​of all other positions are 0. In the above example, the task mapping matrix is ​​[1, 0, 0].

[0081] The task mapping matrix is ​​input into the hierarchical decoder. The hierarchical decoder consists of a task decomposition layer, a resource allocation layer, and an execution sequence layer. The task decomposition layer decomposes the task into subtasks based on the task mapping matrix. For example, the "customer complaint" task is decomposed into subtasks such as "recording complaint information," "investigating the cause of the complaint," and "developing a solution." The resource allocation layer allocates resources based on the task mapping matrix and the task decomposition strategy. For example, the "investigating the cause of the complaint" subtask is assigned to a customer service representative, and the "developing a solution" subtask is assigned to a technician. The execution sequence layer generates an execution sequence based on the task mapping matrix, the task decomposition strategy, and the resource allocation strategy. For example, "recording complaint information" is executed first, followed by "investigating the cause of the complaint," and finally "developing a solution."

[0082] The constraint adaptation factor is calculated based on the task feature vector, the resource status output of the resource allocation layer, and the time constraint parameters. For example, if resources are limited or time is tight, the constraint adaptation factor is small. Based on the constraint adaptation factor, the task decomposition strategy, resource allocation strategy, and execution sequence strategy are constrained and adjusted to obtain the final task processing strategy. For example, if time is tight, the "investigation of the cause of the complaint" step can be simplified.

[0083] Construct a dominance matrix for task processing strategies. This matrix reflects the mutual influence between different strategies. For example, the "Investigate the cause of the complaint" strategy must be executed before the "Develop a solution" strategy. Calculate the optimal strategy combination based on the dominance matrix. For example, select the strategy combination that has the highest execution efficiency and satisfies the constraints. Determine strategy weights based on the contribution of the task decomposition strategy, resource allocation strategy, and execution sequence strategy to the optimal strategy combination. For example, if the "Investigate the cause of the complaint" strategy has a greater impact on the final outcome, assign it a higher weight.

[0084] The policy weights and optimal policy combination are input into the policy optimization network. The policy optimization network includes a policy storage module and a policy sampling module. The policy storage module stores historical policy data. The policy sampling module performs importance sampling on the historical policy data to obtain sampled policy data. For example, policy data with a high success rate is sampled preferentially. Policy evaluation parameters are calculated based on the sampled policy data. For example, the average execution time and success rate of the policies are calculated. The network parameters of the layered decoder are optimized and updated based on the policy evaluation parameters. For example, if the success rate of a particular policy is low, the corresponding parameters in the layered decoder are adjusted to improve the success rate of that policy.

[0085] Through feature matching and layered decoding, the present invention can quickly generate processing strategies for specific tasks, avoiding the tedious process of manually formulating strategies, thereby improving task processing efficiency; dynamically allocate resources according to task requirements and resource status, avoiding resource waste and improving resource utilization; and continuously learning and optimizing strategies through the strategy optimization network, so that the generated strategies can be more adapted to actual conditions and improve the success rate of task processing.

[0086] In an optional embodiment,

[0087] The human-machine collaborative processing framework is constructed based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The steps include:

[0088] Acquiring a task type feature vector based on the task processing strategy, calculating an entropy value of the task type feature vector based on an information entropy calculation model, and dividing tasks into structured tasks and unstructured tasks according to the entropy value;

[0089] A strategy mapping mechanism for a robot execution unit is constructed based on the task processing strategy, the task decomposition strategy in the task processing strategy is mapped to a robot basic action library to obtain an action correspondence matrix, action execution parameters are determined according to the resource allocation strategy in the task processing strategy, a robot control instruction stream is generated based on the execution sequence strategy in the task processing strategy, and the robot execution unit processes the structured task according to the robot control instruction stream;

[0090] constructing an operator cognitive state vector including attention level, workload, and expertise, evaluating the operator cognitive state vector based on a resource allocation strategy in the task processing strategy, dynamically adjusting information display density and frequency of a manual operation interface based on the evaluation result, and processing the unstructured task through the manual operation interface;

[0091] Construct a collaborative effectiveness evaluation index that includes execution efficiency, accuracy and cognitive load, determine the collaborative weight coefficient of the collaborative effectiveness evaluation index according to the strategy weight in the task processing strategy, calculate the human-computer collaborative processing efficiency score based on the collaborative weight coefficient, and use the human-computer collaborative processing efficiency score as a feedback signal to optimize and update the task processing strategy.

[0092] For example, the input task is first analyzed to extract its type feature vector. For example, a task can be described by multiple features, such as data type (text, image, numerical), data structure (structured, semi-structured, unstructured), and task complexity (simple, medium, complex). These features are quantified to form a task type feature vector, for example, [0.8, 0.2, 0.5] representing text data, semi-structured data, and medium complexity, respectively. The entropy value of this feature vector is then calculated using an information entropy calculation model. A higher entropy value indicates greater uncertainty about the task type, and a higher likelihood of classification as an unstructured task. Assuming the calculated entropy value is 0.6, based on a preset threshold (e.g., 0.5), the task is classified as unstructured and assigned to a human operator. Otherwise, it is classified as a structured task and assigned to a robotic unit.

[0093] For structured tasks assigned to a robot cell, the framework's strategy mapping mechanism converts predefined task processing strategies into instructions executable by the robot. For example, a task processing strategy includes a task decomposition strategy, a resource allocation strategy, and an execution sequence strategy. For example, if a structured task is "move item A from location B to location C," the task decomposition strategy breaks it down into subtasks such as "grab item A," "move to location B," "grab item A," "move to location C," and "place item A." The resource allocation strategy determines the resources required for each subtask, such as the robot arm or mobile platform. The execution sequence strategy determines the order in which the subtasks are executed. The strategy mapping mechanism maps these strategies to a library of basic robot actions, such as "grab," "move," and "place," and generates corresponding action parameters, such as grasping force, movement speed, and target position. This ultimately forms a robot control instruction stream to control the robot to complete the task.

[0094] For unstructured tasks assigned to human operators, the framework dynamically adjusts the information displayed in the human interface based on the operator's cognitive state. For example, a cognitive state vector is constructed for the operator, including metrics such as attention level, workload, and expertise. Assume that the current operator's cognitive state vector is [0.9, 0.5, 0.8], representing high attention level, medium workload, and high expertise. Based on the resource allocation strategy within the task processing strategy, this cognitive state is evaluated and the density and frequency of information displayed in the human interface are adjusted accordingly. For example, for operators with high attention levels, the information display density can be increased; for operators with high workloads, the information display frequency can be reduced. In this way, human-computer interaction can be optimized and operator efficiency can be improved.

[0095] Finally, the framework evaluates the effectiveness of human-machine collaborative processing and optimizes task processing strategies based on the evaluation results. For example, collaborative effectiveness evaluation metrics are constructed, including execution efficiency, accuracy, and cognitive load. Assume that the weights for execution efficiency, accuracy, and cognitive load in the task processing strategy are 0.4, 0.5, and 0.1, respectively. Based on these weights, the human-machine collaborative processing effectiveness score is calculated. For example, assuming the current execution efficiency is 0.9, accuracy is 0.8, and cognitive load is 0.6, the collaborative effectiveness score is 0.4 × 0.9 + 0.5 × 0.8 + 0.1 × 0.6 = 0.82. This score is used as a feedback signal to optimize and update the task processing strategy, such as adjusting strategy weights and optimizing task decomposition strategies, thereby continuously improving the effectiveness of human-machine collaborative processing.

[0096] The present invention assigns structured tasks to robots for automatic processing and unstructured tasks to manual processing, giving full play to the advantages of each, thereby improving overall processing efficiency; robots have high accuracy in processing structured tasks, and manual processing of unstructured tasks is flexible. The combination of the two can effectively improve overall processing accuracy; handing over tedious and repetitive structured tasks to robots can effectively reduce the operator's cognitive load and workload, allowing them to focus on more creative unstructured tasks.

[0097] In an optional embodiment,

[0098] The robot unit and the manual operation unit are connected via an information exchange channel. When the deviation value during task processing exceeds a predetermined threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the manual operation unit, and the manual operation unit transmits the processing procedure information to the rule database of the robot unit. The steps include:

[0099] The task execution trajectory is decomposed into time-frequency feature representations of different scales through wavelet transform, and the time-frequency feature representations of different scales are weighted by the adaptive attention mechanism. The weighted features are fused to obtain a global deviation feature. The global deviation feature sequence is modeled using a recursive neural network, and the deviation degree score is calculated using a multi-layer perceptron. When the deviation degree score exceeds a preset deviation threshold, an anomaly detection signal is triggered; an anomaly pattern feature library is constructed based on historical execution data, and the type and cause of the current anomaly are matched from the anomaly pattern feature library based on the anomaly detection signal;

[0100] Construct a multi-level knowledge graph structure that includes a task-level graph, a resource-level graph, and a processing-level graph. The task-level graph describes the dependencies between tasks, the resource-level graph describes the distribution of equipment and personnel capabilities, and the processing-level graph describes the operational process. Update node information in the multi-level knowledge graph structure based on the exception type and cause.

[0101] A graph neural network with a multi-head attention mechanism is used to transfer information between layers of a multi-level knowledge graph structure, generating node feature representations by calculating normalized weights between nodes. The node feature representations are combined with historical sequence features to construct a state vector. A policy network is trained based on the state vector, using a weighted combination of task completion and resource utilization as a reward signal, and a policy gradient method is used to optimize network parameters to obtain a decision model. Guided by the decision model, the Monte Carlo tree search algorithm is used to generate an optimal decision path using visit counts, value estimates, and prior probabilities as search attributes. A local causal graph containing variable nodes and causal edges is constructed for the nodes in the optimal decision path to generate processing recommendations.

[0102] The processing suggestion is represented as a processing procedure of a condition-action pair, an importance evaluation is performed on the processing procedure, and the processing procedure whose importance exceeds a preset importance threshold is updated to a rule database. The update result of the rule database is used to adjust the preset deviation threshold.

[0103] For example, deviation detection is performed on the task execution trajectory. The task execution trajectory is decomposed into feature representations at different time and frequency scales. For example, the motion trajectory of a robotic arm is decomposed into features of different frequencies, such as velocity and acceleration, as well as features within different time periods. Features at different scales are then weighted according to their importance. For example, features with drastic speed changes are given higher weights. The weighted features are fused to obtain a global deviation feature, and a recurrent neural network is used to model the sequence of global deviation features. For example, a long short-term memory (LSTM) network is used to learn the dependencies between historical deviation features. Finally, a deviation score is calculated using a multilayer perceptron. For example, the output of the LSTM network is used as the input to the multilayer perceptron to calculate a deviation score between 0 and 1. When the deviation score exceeds a preset deviation threshold (e.g., 0.8), an anomaly detection signal is triggered.

[0104] When an anomaly detection signal is triggered, the system matches the anomaly against a feature library of anomaly patterns built from historical execution data. This library contains signature representations of various known anomaly types, such as a stuck robot arm or a sensor failure. By comparing the current anomaly's signature with the anomaly patterns in the library, the system determines the type and cause of the anomaly. For example, if the current anomaly signature is highly similar to the "stuck robot arm" feature pattern, the system determines the anomaly type as "stuck robot arm."

[0105] The system will update the multi-level knowledge graph structure. This structure includes a task-level graph, a resource-level graph, and a processing-level graph. The task-level graph describes the dependencies between tasks. For example, the "screwing" task depends on the "grasping screws" task. The resource-level graph describes the distribution of equipment and personnel capabilities. For example, robot A can perform grasping and screwing operations, and operator B is good at handling sensor failures. The processing-level graph describes the operation process. For example, the process for handling a stuck robot arm is: stop the robot arm movement, check the cause of the jam, troubleshoot, and restart the robot arm. Based on the type and cause of the exception, the node information in the multi-level knowledge graph structure is updated. For example, if robot arm A gets stuck, the status of robot arm A in the resource-level graph is updated to "fault."

[0106] The system uses a graph neural network to transfer information between layers of a multi-layered knowledge graph. By calculating the normalized weights between nodes, it generates node feature representations. For example, the node feature representation of robot A includes information such as its state, capabilities, and connections with other nodes. The node feature representation is combined with historical sequence features to construct a state vector. For example, the node feature representation of robot A is combined with the execution trajectory features of robot A over a period of time.

[0107] Based on the state vector, the system trains a policy network. A weighted combination of task completion and resource utilization is used as the reward signal. For example, task completion is weighted as 0.7, and resource utilization is weighted as 0.3. The policy gradient method is used to optimize the network parameters to produce a decision model. This decision model outputs the optimal action strategy based on the current state.

[0108] Guided by the decision model, the system uses a Monte Carlo tree search algorithm to generate the optimal decision path. It uses visit counts, value estimates, and prior probabilities as search attributes. For example, nodes with higher visit counts have more reliable value estimates. By simulating different action paths and evaluating the value of each path, the path with the highest value is ultimately selected as the optimal decision path.

[0109] A local causal graph containing variable nodes and causal edges is constructed for the nodes in the optimal decision path to generate processing suggestions.

[0110] The proposed action is represented as a condition-action pairing procedure, for example, "If the robot arm is stuck, stop the robot arm." The procedure's importance is evaluated, and those exceeding a preset importance threshold are updated to the rule database. The updated rule database is used to adjust the preset deviation threshold. For example, if a new action proves effective, the preset deviation threshold can be lowered appropriately.

[0111] Through the collaborative work of robotic units and manual operation units, the present invention can detect and handle task anomalies more quickly; through multi-level knowledge graphs and graph neural networks, the system can better understand the relationship between tasks, resources and processing flows, so that it can more effectively deal with various abnormal situations and enhance the robustness of the system; the system stores processing procedures in a rule database, and continuously updates and optimizes it, thereby realizing the accumulation and reuse of knowledge.

[0112] In an optional embodiment,

[0113] The steps of constructing a knowledge transfer processing network, extracting features from business rule data and expert experience data in the human-machine collaborative processing process using a graph embedding algorithm to obtain a business feature vector, training a scenario transfer model based on the business feature vector, and mapping the business feature vector to a preset scenario feature space to generate scenario mapping data include:

[0114] The knowledge transfer processing network includes a knowledge graph construction module, a feature extraction module and a scene transfer module;

[0115] The knowledge graph construction module is used to construct a knowledge graph including rule nodes, experience nodes and relationship edges. The rule nodes use the triple form of trigger conditions, operation sequences and constraints to describe the robotic process automation operation procedures. The experience nodes use the structured form of abnormal scenarios, processing steps and expected effects to represent process optimization techniques. The relationship edges describe the inclusion relationship, pre-requisite relationship and optimization relationship between the rule nodes and the experience nodes.

[0116] The feature extraction module uses a graph embedding algorithm with a multi-head attention mechanism to extract features from the business rule data and expert experience data in the knowledge graph, where the business rule data comes from the trigger conditions, operation sequences, and constraints of the rule nodes, and the expert experience data comes from the abnormal scenarios, processing steps, and expected effects of the experience nodes. The business feature vector is obtained by calculating the attention scores of the set of nodes directly connected to the current processing node and performing feature aggregation. The business feature vector contains the interactive relationship information between the robotic process automation operation nodes;

[0117] The scenario migration module trains a scenario migration model based on the business feature vector. The scenario migration model includes a discriminator network and a generator network. The discriminator network is used to distinguish between the real process samples of the target scenario and the generated migration samples. The generator network is used to map the business feature vector of the source scenario to a preset scenario feature space.

[0118] A contrastive learning mechanism is used to align the feature spaces of the source scene and the target scene. The same processing steps in the source and target scenes are constructed as positive sample pairs, and the different processing steps are constructed as negative samples. By minimizing the contrastive loss function, the feature representations of the same operation steps are made close in the feature space to obtain scene mapping data.

[0119] For example, a knowledge transfer processing network is used to optimize Robotic Process Automation (RPA). By integrating expert experience into RPA rules, it enables cross-scenario process migration and adaptation. This network primarily consists of three modules: knowledge graph construction, feature extraction, and scenario migration.

[0120] Build a knowledge graph. The knowledge graph uses a graph structure to represent RPA operating procedures and expert experience. Nodes in the graph are divided into rule nodes and experience nodes. Rule nodes describe RPA operating procedures in the form of triples (trigger condition, operation sequence, constraint condition). For example, "If an invoice is received via email, extract the invoice information, and the invoice amount must be greater than 0." Experience nodes represent process optimization techniques in a structured form (abnormal scenario, processing steps, expected results). For example, "If the invoice format is incorrect, contact the supplier to confirm and ensure the invoice information is accurate." Edges in the graph describe the relationship between rule nodes and experience nodes, including inclusion relationships (experience nodes complement rule nodes), precedence relationships (one rule node must be executed before another rule node), and optimization relationships (experience nodes optimize rule nodes).

[0121] Perform feature extraction. A graph embedding algorithm with a multi-head attention mechanism extracts features from the business rule data and expert experience data in the knowledge graph. Business rule data is derived from the triples of rule nodes, and expert experience data is derived from the structured information of experience nodes. The feature extraction process focuses on nodes directly related to the current processing node. By calculating attention scores and aggregating features, a business feature vector containing information about the interaction relationship between RPA operation nodes is generated. For example, when processing the "Extract Invoice Information" node, attention is paid to the "Receive Invoice Email" node, which has a "predecessor relationship" with it, and the "If Invoice Format is Incorrect" node, which has an "optimization relationship" with it. The attention scores of these nodes are calculated, and their features are aggregated with the features of the "Extract Invoice Information" node to generate the business feature vector for that node.

[0122] Perform scenario transfer. A scenario transfer model is trained based on business feature vectors. This model consists of a discriminator network and a generator network. The generator network maps the business feature vectors of the source scenario to a preset scenario feature space, generating transfer samples. The discriminator network distinguishes between real-world process samples of the target scenario and the generated transfer samples. A contrastive learning mechanism is used to align the feature spaces of the source and target scenarios. Identical processing steps in the source and target scenarios are constructed as positive sample pairs. For example, if both the source and target scenarios have the "extract invoice information" step, these two steps are constructed as positive sample pairs. Dissimilar processing steps are constructed as negative sample pairs, for example, the "send email notification" step in the source scenario and the "system entry" step in the target scenario. By minimizing the contrastive loss function, the feature representations of identical operation steps are made closer in the feature space, ultimately generating scenario mapping data. For example, if a model for processing "online orders" is transferred to processing "offline orders," contrastive learning can be used to make the feature representations of the "order processing" step in the two scenarios closer, thus achieving cross-scenario knowledge transfer.

[0123] Through knowledge migration, this invention enables the RPA process to adapt to new scenarios without the need for reconfiguration, thereby reducing the development and maintenance costs of the RPA process. By incorporating expert experience, the RPA operating procedures are optimized, the execution efficiency of the RPA process is improved, and manual intervention is reduced. By handling abnormal scenarios, the fault tolerance of the RPA process is improved, and its stability in complex environments is enhanced.

[0124] In an optional embodiment,

[0125] The scene migration module trains a scene migration model based on the business feature vector. The scene migration model includes a discriminator network and a generator network. The discriminator network is used to distinguish between the real process samples of the target scene and the generated migration samples. The generator network is used to map the business feature vector of the source scene to the preset scene feature space. The steps include:

[0126] The generator network adopts a three-layer fully connected neural network structure, each layer is connected by a ReLU activation function, and the last layer outputs a target scene feature vector through a hyperbolic tangent activation function. The discriminator network adopts a four-layer fully connected network structure, and a batch normalization layer and a random dropout layer are set after each hidden layer; a discriminator loss function is constructed based on the Wasserstein distance, and the discriminator loss function includes a real sample expectation term, a generated sample expectation term, and a gradient penalty term. The discriminator network is trained by the discriminator loss function; a generator loss function is constructed, and the generator loss function includes an adversarial loss term output by the discriminator and a semantic consistency loss term, wherein the semantic consistency loss term is obtained by calculating the Euclidean distance between the source scene feature and the generated feature and the Euclidean distance between the features of the intermediate layer of the feature extraction network;

[0127] Constructing feature quality evaluation indicators, which include a feature consistency score and a task relevance score, where the feature consistency score is calculated based on the feature space distance between the target scene features and the generated features, and the task relevance score is calculated based on the accuracy of the downstream task classifier;

[0128] An alternating optimization strategy is adopted to train the scene migration model, the generator network is fixed, the parameters of the discriminator network are optimized multiple times through iterations, the discriminator network is fixed, and the parameters of the generator network are optimized once; the business feature vector is divided into the training scale according to the number of rule node triplets, the number of rule node triplets of the initial training scale is set to 1, the growth step value of the number of rule node triplets is set, and the number of training rounds for each training scale is set; the scene migration model is trained based on the business feature vector of the current training scale, and when the feature quality evaluation index reaches a preset threshold, the number of rule node triplets is increased according to the growth step value, and the scene migration model is continued to be trained based on the increased training scale until a preset number of training times is reached; the generated target scene features are evaluated based on the feature quality evaluation index, and the weight coefficient of the semantic consistency loss term and the growth step value of the number of rule node triplets are adjusted according to the target scene feature evaluation result.

[0129] For example, we collect process sample data for source and target scenarios and extract corresponding business feature vectors. For example, the source scenario could be a product recommendation scenario on an e-commerce platform, and the target scenario could be a friend recommendation scenario on a social networking platform. The business feature vectors can include information such as the user's age, gender, interests, and purchase history. Assume we collect 10,000 source scenario samples and 10,000 target scenario samples, and represent each sample as a 20-dimensional feature vector.

[0130] A scene transfer model is constructed, consisting of a generator network and a discriminator network. The generator network uses a three-layer fully connected neural network structure, with ReLU activation functions connecting each layer. The final layer uses a hyperbolic tangent activation function to output the target scene feature vector. For example, the generator network takes as input a 20-dimensional feature vector of the source scene, and outputs a 20-dimensional feature vector of the target scene. The first layer has 50 neurons, the second layer has 100 neurons, and the third layer has 20 neurons. The discriminator network uses a four-layer fully connected network structure, with batch normalization and dropout layers after each hidden layer. For example, the discriminator network takes as input a 20-dimensional feature vector, and outputs a value between 0 and 1, representing the probability that the input feature belongs to the target scene. The first layer has 100 neurons, the second layer has 50 neurons, the third layer has 20 neurons, and the fourth layer has 1 neuron.

[0131] Next, we define the loss function and train the scene transfer model. The discriminator loss function is constructed based on the Wasserstein distance and includes a true sample expectation term, a generated sample expectation term, and a gradient penalty term. The generator loss function includes an adversarial loss term for the discriminator output and a semantic consistency loss term. The semantic consistency loss term is calculated by calculating the Euclidean distance between the source scene features and the generated features, as well as the Euclidean distance between the features in the intermediate layers of the feature extraction network. An alternating optimization strategy is used to train the scene transfer model. This involves fixing the generator network, iterating and optimizing the discriminator network parameters multiple times, then fixing the discriminator network and optimizing the generator network parameters once.

[0132] During training, a strategy of gradually increasing the training data size is adopted. The business feature vectors are divided into training scales based on the number of rule node triplets. The number of rule node triplets in the initial training scale is set to 1, and a step size for increasing the number of rule node triplets is set, for example, 10. The number of training rounds for each training scale also needs to be set, for example, 100 rounds. The scenario migration model is trained based on the business feature vectors of the current training scale. When the feature quality evaluation metric reaches a preset threshold, the number of rule node triplets is increased by the step size. The scenario migration model is trained again based on the increased training scale until a preset number of training rounds is reached, for example, 1000 rounds. Feature quality evaluation metrics include a feature consistency score and a task relevance score. The feature consistency score is calculated based on the feature space distance between the target scenario features and the generated features, while the task relevance score is calculated based on the accuracy of the downstream task classifier. For example, the generated target scenario features can be used in a user recommendation task, and the recommendation accuracy is calculated as the task relevance score. During training, the weight coefficient of the semantic consistency loss term and the step size for increasing the number of rule node triplets are adjusted based on the results of the feature quality evaluation metric.

[0133] By combining Wasserstein distance and gradient penalty, the present invention can effectively stabilize the training process of GAN (generative adversarial network), avoid problems such as mode collapse, and thus generate more diverse and realistic target scene features; the introduction of semantic consistency loss terms ensures the consistency of generated features with source scene features at the semantic level, thereby improving the effectiveness of cross-scene knowledge transfer; by gradually increasing the scale of training data and dynamically adjusting parameters, the model can be better adapted to different data distributions and task requirements, thereby improving the generalization ability of the model.

[0134] In an optional embodiment,

[0135] The knowledge transfer processing network further includes a verification module for verifying the validity of the scene mapping data and outputting a verification result. The steps of updating the processing parameters of the data perception processing network and the task allocation parameters of the human-machine collaborative processing framework based on the verification result include:

[0136] Constructing a verification matrix to verify the scene mapping data, the verification matrix includes a semantic consistency vector, a structural similarity vector, and a temporal rationality vector, wherein the semantic consistency vector is obtained by calculating the keyword matching degree of the business rules before and after mapping, the structural similarity vector is calculated based on the edit distance of the graph structure, and the temporal rationality vector is obtained by verifying the dependency relationship of the operation sequence;

[0137] Constructing a multi-level verification index based on the verification matrix, wherein the multi-level verification index includes a rule-level verification index and a process-level verification index, wherein the rule-level verification index is obtained by calculating rule similarity and rule internal consistency, and the process-level verification index is obtained by calculating timing constraint satisfaction and structural integrity;

[0138] Performing a weighted combination of the multi-level verification indicators and the business scenario coverage to generate a verification evaluation score;

[0139] Updating parameters of the data perception processing network based on the verification evaluation score, including: gradient updating the perception layer parameters using an adaptive learning rate, the adaptive learning rate dynamically adjusting with the verification evaluation score; and determining an adjustment direction and an adjustment step size of the data perception processing network parameters based on a difference between the verification evaluation score and a preset threshold;

[0140] A task allocation matrix is ​​constructed, wherein the elements of the task allocation matrix represent the probability of tasks being assigned to processing units, and an update amount of the task allocation matrix is ​​calculated based on the verification evaluation score and resource utilization; the task allocation scheme in the human-computer collaborative processing framework is dynamically adjusted according to the update amount of the task allocation matrix, and the task allocation parameters in the human-computer collaborative processing framework are optimized based on the adjusted task allocation scheme.

[0141] For example, scenario mapping data is obtained. Scenario mapping data refers to the conversion of data from actual business scenarios into a data format that can be understood and processed by the knowledge transfer processing network. For example, in an order processing scenario on an e-commerce platform, scenario mapping data may include order information, user information, product information, and so on.

[0142] Verify the validity of the scene mapping data. The verification module uses a verification matrix to verify the scene mapping data. The verification matrix consists of three vectors: semantic consistency vector, structural similarity vector, and temporal rationality vector.

[0143] The semantic consistency vector is calculated by extracting keywords from the business rules before and after mapping and calculating their matching degree. For example, if the business rule before mapping is "free shipping for orders over 100 yuan" and the business rule after mapping is "free shipping for orders over 100 yuan," we can extract the keywords "order amount," "100 yuan," and "free shipping / free shipping" and calculate their matching degree.

[0144] The structural similarity vector is calculated by converting the business rules before and after mapping into a graph structure and calculating the edit distance of the graph structure. The edit distance is the minimum number of operations required to transform one graph into another (for example, adding nodes, deleting nodes, and modifying connections between nodes). The smaller the edit distance, the higher the structural similarity.

[0145] The timing rationality vector is calculated by verifying the dependencies between the operation sequences. For example, in an order processing flow, the order placement operation must precede the payment operation, and the shipment operation must follow the payment operation. If the operation sequence does not meet these dependencies, the timing is considered unreasonable.

[0146] Construct multi-level verification indicators based on the verification matrix. The multi-level verification indicators include rule-level verification indicators and process-level verification indicators.

[0147] Rule-level validation metrics are calculated by calculating rule similarity and internal rule consistency. Rule similarity refers to the degree of similarity between business rules before and after mapping. Internal rule consistency refers to whether the logic of the business rules themselves is consistent.

[0148] The calculation method for process-level verification indicators is to calculate timing constraint satisfaction and structural integrity. Timing constraint satisfaction refers to whether the operation sequence meets the timing constraints. Structural integrity refers to whether the structure of the flowchart is complete.

[0149] The verification evaluation score is generated by weighting the multi-level verification indicators and business scenario coverage. Business scenario coverage refers to the proportion of business scenarios covered by the verification module. For example, if the verification module covers 80% of business scenarios, the business scenario coverage is 0.8.

[0150] The parameters of the data perception processing network are updated based on the validation evaluation score. These parameters include the perception layer parameters. The perception layer parameters are updated using an adaptive learning rate that performs a gradient update. The adaptive learning rate is dynamically adjusted based on the validation evaluation score. A high validation evaluation score indicates good model performance and the learning rate can be reduced. A low validation evaluation score indicates poor model performance and the learning rate can be increased. The difference between the validation evaluation score and a preset threshold determines the direction and step size of the data perception processing network parameter adjustments. If the validation evaluation score falls below the preset threshold, the parameters need to be adjusted to improve model performance.

[0151] Construct a task allocation matrix. The elements of the task allocation matrix represent the probability of assigning a task to a processing unit. Updates to the task allocation matrix are calculated based on the validation evaluation scores and resource utilization. Resource utilization refers to the utilization of the processing units. Dynamically adjust the task allocation scheme in the human-machine collaborative processing framework based on the updated task allocation matrix, and optimize the task allocation parameters in the human-machine collaborative processing framework based on the adjusted task allocation scheme.

[0152] For example, if the verification evaluation score is 0.9 and the resource utilization is 0.8, the update amount of the task allocation matrix can be set to 0.1, which means that 10% of the tasks will be reassigned to different processing units.

[0153] The present invention verifies the validity of scene mapping data through the verification module, which can improve the accuracy and efficiency of knowledge transfer and reduce the risk of erroneous migration; by dynamically adjusting the task allocation scheme in the human-computer collaborative processing framework, the efficiency of human-computer collaboration can be optimized and the overall processing performance can be improved; by gradient updating the parameters of the data perception processing network through adaptive learning rate, the adaptability of the system can be enhanced, so that it can better adapt to different business scenarios.

[0154] Figure 2 FIG. 1 is a schematic diagram of the structure of a human-machine collaborative system based on robotic process automation according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0155] The first unit is used to build a data perception and processing network, collect user operation instruction data, system operation status data and task environment data, perform semantic analysis on the collected data to obtain task feature vectors, perform feature matching between the task feature vectors and the preset business model to construct a task mapping matrix and generate a task processing strategy;

[0156] The second unit is used to build a human-machine collaborative processing framework based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The robot unit and the human operation unit are connected through an information exchange channel. When the deviation value during the task processing process exceeds a specified threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the human operation unit, and the human operation unit transmits the processing procedure information to the rule database of the robot unit;

[0157] The third unit is used to construct a knowledge transfer processing network, use a graph embedding algorithm to extract features from business rule data and expert experience data in the human-computer collaborative processing process to obtain business feature vectors, train a scenario migration model based on the business feature vectors, and map the business feature vectors to a preset scenario feature space to generate scenario mapping data; the knowledge transfer processing network also includes a verification module, which verifies the validity of the scenario mapping data and outputs the verification results, and updates the processing parameters of the data perception processing network and the task allocation parameters of the human-computer collaborative processing framework based on the verification results.

[0158] According to a third aspect of the embodiments of the present invention,

[0159] An electronic device is provided, comprising:

[0160] processor;

[0161] a memory for storing processor-executable instructions;

[0162] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0163] According to a fourth aspect of the embodiments of the present invention,

[0164] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0165] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A human-machine collaboration method based on robotic process automation, characterized in that: include: Build a data perception and processing network to collect user operation instruction data, system operation status data, and task environment data. Perform semantic analysis on the collected data to obtain task feature vectors. Then, perform feature matching between the task feature vectors and the preset business model to construct a task mapping matrix and generate a task processing strategy. A human-machine collaborative processing framework is constructed based on the task processing strategy, wherein the robot unit processes structured tasks that meet preset rules, and the manual operation unit processes unstructured tasks, including: obtaining a task type feature vector based on the task processing strategy, calculating the entropy value of the task type feature vector based on the information entropy calculation model, and dividing the task into structured tasks and unstructured tasks according to the entropy value; constructing a strategy mapping mechanism for the robot execution unit based on the task processing strategy, mapping the task decomposition strategy in the task processing strategy to the robot basic action library to obtain an action correspondence matrix, determining action execution parameters according to the resource allocation strategy in the task processing strategy, generating a robot control instruction stream based on the execution sequence strategy in the task processing strategy, and the robot execution unit processes the structured task according to the robot control instruction stream; constructing an operator cognitive state including attention level, workload and professionalism vector, evaluate the operator's cognitive state vector based on the resource allocation strategy in the task processing strategy, dynamically adjust the information display density and frequency of the manual operation interface according to the evaluation result, and process the unstructured task through the manual operation interface; construct a collaborative effectiveness evaluation index including execution efficiency, accuracy and cognitive load, determine the collaborative weight coefficient of the collaborative effectiveness evaluation index according to the strategy weight in the task processing strategy, calculate the human-machine collaborative processing efficiency score based on the collaborative weight coefficient, and use the human-machine collaborative processing efficiency score as a feedback signal to optimize and update the task processing strategy; the robot unit and the manual operation unit are connected through an information interaction channel. When the deviation value in the task processing process is greater than the specified threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the manual operation unit, and the manual operation unit transmits the processing procedure information to the rule database of the robot unit; A knowledge transfer processing network is constructed, and a graph embedding algorithm is used to extract features from business rule data and expert experience data in the human-computer collaborative processing process to obtain business feature vectors. A scenario migration model is trained based on the business feature vectors, and the business feature vectors are mapped to a preset scenario feature space to generate scenario mapping data. The knowledge transfer processing network also includes a verification module that verifies the validity of the scenario mapping data and outputs the verification results, and updates the processing parameters of the data perception processing network and the task allocation parameters of the human-computer collaborative processing framework based on the verification results.

2. The method according to claim 1, characterized in that The task feature vector is matched with the preset business model to construct a task mapping matrix. The steps of generating the task processing strategy include: Build a model library containing standard feature vectors and processing templates, use contrastive learning to calculate the similarity between the task feature vector and the standard feature vector, select the feature space with the highest similarity for mapping, and build a task mapping matrix; The task mapping matrix is ​​input into a hierarchical decoder, which includes a task decomposition layer, a resource allocation layer, and an execution sequence layer connected in sequence, wherein the task decomposition layer generates a task decomposition strategy based on the task mapping matrix, the resource allocation layer generates a resource allocation strategy based on the task mapping matrix and the task decomposition strategy, and the execution sequence layer generates an execution sequence strategy based on the task mapping matrix, the task decomposition strategy, and the resource allocation strategy; Calculating a constraint adaptation factor based on the task feature vector, the resource state output of the resource allocation layer, and the time constraint parameter; performing constraint adjustments on the task decomposition strategy, the resource allocation strategy, and the execution sequence strategy based on the constraint adaptation factor to obtain a task processing strategy; constructing a dominance relationship matrix of the task processing strategy; calculating an optimal strategy combination based on the dominance relationship matrix; and determining a strategy weight based on the contribution of the task decomposition strategy, the resource allocation strategy, and the execution sequence strategy to the optimal strategy combination; The policy weights and the optimal policy combination are input into the policy optimization network, importance sampling is performed based on historical policy data, and the policy evaluation parameters are calculated and then the layered decoder parameters are optimized and updated.

3. The method according to claim 1, characterized in that The robot unit and the manual operation unit are connected via an information exchange channel. When the deviation value during task processing exceeds a predetermined threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the manual operation unit, and the manual operation unit transmits the processing procedure information to the rule database of the robot unit. The steps include: The multi-scale time-frequency features of the task execution trajectory are extracted through wavelet transform, and the global deviation features are obtained through adaptive attention weighted fusion. The global deviation feature sequence is modeled using a recursive neural network, and the deviation degree score is calculated using a multi-layer perceptron. When the deviation degree score exceeds a preset deviation threshold, an anomaly detection signal is triggered; based on the anomaly detection signal, the type and cause of the current anomaly are matched from the anomaly pattern feature library; Construct a multi-level knowledge graph structure that includes a task-level graph, a resource-level graph, and a processing-level graph. The task-level graph describes the dependencies between tasks, the resource-level graph describes the distribution of equipment and personnel capabilities, and the processing-level graph describes the operational process. Update node information in the multi-level knowledge graph structure based on the exception type and cause. A graph neural network with a multi-head attention mechanism is used to transfer information between layers of a multi-level knowledge graph structure, generating node feature representations and combining them with historical sequence features to construct a state vector. A policy network is trained based on the state vector, with task completion and resource utilization weighted as reward signals, and a decision model is optimized using the policy gradient method. Guided by the decision model, the Monte Carlo tree search algorithm is used to generate an optimal decision path, using visit counts, value estimates, and prior probabilities as search attributes. A local causal graph containing variable nodes and causal edges is constructed for the nodes in the optimal decision path to generate processing recommendations. The processing suggestion is represented as a processing procedure of a condition-action pair, an importance evaluation is performed on the processing procedure, and the processing procedure whose importance exceeds a preset importance threshold is updated to a rule database. The update result of the rule database is used to adjust the preset deviation threshold.

4. The method according to claim 1, wherein The steps of constructing a knowledge transfer processing network, extracting features from business rule data and expert experience data in the human-machine collaborative processing process using a graph embedding algorithm to obtain a business feature vector, training a scenario transfer model based on the business feature vector, and mapping the business feature vector to a preset scenario feature space to generate scenario mapping data include: The knowledge transfer processing network includes a knowledge graph construction module, a feature extraction module and a scene transfer module; The knowledge graph construction module is used to construct a knowledge graph including rule nodes, experience nodes and relationship edges. The rule nodes use the triple form of trigger conditions, operation sequences and constraints to describe the robotic process automation operation procedures. The experience nodes use the structured form of abnormal scenarios, processing steps and expected effects to represent process optimization techniques. The relationship edges describe the inclusion relationship, pre-requisite relationship and optimization relationship between the rule nodes and the experience nodes. The feature extraction module uses a multi-head attention graph embedding algorithm to extract business rule data and expert experience data from rule nodes and experience nodes respectively, and obtains a business feature vector containing the interaction relationship between operation nodes by calculating the attention scores of adjacent nodes and aggregating them; The scene migration module trains a scene migration model based on the business feature vector, and the scene migration model includes a discriminator network and a generator network; A contrastive learning mechanism is used to align the feature spaces of the source scene and the target scene. The same processing steps in the source and target scenes are constructed as positive sample pairs, and the different processing steps are constructed as negative samples. By minimizing the contrastive loss function, the feature representations of the same operation steps are made close in the feature space to obtain scene mapping data.

5. The method according to claim 4, characterized in that The scene migration module trains a scene migration model based on the business feature vector. The scene migration model includes a discriminator network and a generator network. The discriminator network is used to distinguish between the real process samples of the target scene and the generated migration samples. The generator network is used to map the business feature vector of the source scene to the preset scene feature space. The steps include: The generator network uses a three-layer fully connected structure, uses ReLU activation, and outputs target scene features through tanh. The discriminator network uses a four-layer fully connected structure, configured with batch normalization and dropout layers. The discriminator loss function is constructed based on the Wasserstein distance, which includes the expectation of real samples, the expectation of generated samples, and a gradient penalty term. The generator loss function combines the adversarial loss of the discriminator output and the semantic consistency loss of the source scene features, where the semantic consistency is calculated by the Euclidean distance between the feature space and the intermediate layer features. Constructing feature quality evaluation indicators, which include a feature consistency score and a task relevance score, where the feature consistency score is calculated based on the feature space distance between the target scene features and the generated features, and the task relevance score is calculated based on the accuracy of the downstream task classifier; An alternating optimization strategy is adopted to train the scene migration model, the generator network is fixed, the parameters of the discriminator network are optimized multiple times through iterations, the discriminator network is fixed, and the parameters of the generator network are optimized once; the business feature vector is divided into the training scale according to the number of rule node triplets, the number of rule node triplets of the initial training scale is set to 1, the growth step value of the number of rule node triplets is set, and the number of training rounds for each training scale is set; the scene migration model is trained based on the business feature vector of the current training scale, and when the feature quality evaluation index reaches a preset threshold, the number of rule node triplets is increased according to the growth step value, and the scene migration model is continued to be trained based on the increased training scale until a preset number of training times is reached; the generated target scene features are evaluated based on the feature quality evaluation index, and the weight coefficient of the semantic consistency loss term and the growth step value of the number of rule node triplets are adjusted according to the target scene feature evaluation result.

6. The method according to claim 1, wherein The knowledge transfer processing network further includes a verification module for verifying the validity of the scene mapping data and outputting a verification result. The steps of updating the processing parameters of the data perception processing network and the task allocation parameters of the human-machine collaborative processing framework based on the verification result include: Constructing a verification matrix to verify the scene mapping data, the verification matrix includes a semantic consistency vector, a structural similarity vector, and a temporal rationality vector, wherein the semantic consistency vector is obtained by calculating the keyword matching degree of the business rules before and after mapping, the structural similarity vector is calculated based on the edit distance of the graph structure, and the temporal rationality vector is obtained by verifying the dependency relationship of the operation sequence; Constructing a multi-level verification index based on the verification matrix, wherein the multi-level verification index includes a rule-level verification index and a process-level verification index, wherein the rule-level verification index is obtained by calculating rule similarity and rule internal consistency, and the process-level verification index is obtained by calculating timing constraint satisfaction and structural integrity; Performing a weighted combination of the multi-level verification indicators and the business scenario coverage to generate a verification evaluation score; Updating parameters of the data perception processing network based on the verification evaluation score, including: gradient updating the perception layer parameters using an adaptive learning rate, the adaptive learning rate dynamically adjusting with the verification evaluation score; and determining an adjustment direction and an adjustment step size of the data perception processing network parameters based on a difference between the verification evaluation score and a preset threshold; A task allocation matrix is ​​constructed, wherein the elements of the task allocation matrix represent the probability of tasks being assigned to processing units, and an update amount of the task allocation matrix is ​​calculated based on the verification evaluation score and resource utilization; the task allocation scheme in the human-computer collaborative processing framework is dynamically adjusted according to the update amount of the task allocation matrix, and the task allocation parameters in the human-computer collaborative processing framework are optimized based on the adjusted task allocation scheme.

7. A human-machine collaborative system based on robotic process automation, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to build a data perception and processing network, collect user operation instruction data, system operation status data and task environment data, perform semantic analysis on the collected data to obtain task feature vectors, perform feature matching between the task feature vectors and the preset business model to construct a task mapping matrix and generate a task processing strategy; The second unit is used to build a human-machine collaborative processing framework based on the task processing strategy, in which the robot unit processes structured tasks that meet preset rules, and the human operation unit processes unstructured tasks. The robot unit and the human operation unit are connected through an information exchange channel. When the deviation value during the task processing process exceeds a specified threshold, the robot unit sends historical data analysis results and decision-making recommendation information to the human operation unit, and the human operation unit transmits the processing procedure information to the rule database of the robot unit; The third unit is used to construct a knowledge transfer processing network, use a graph embedding algorithm to extract features from business rule data and expert experience data in the human-computer collaborative processing process to obtain business feature vectors, train a scenario migration model based on the business feature vectors, and map the business feature vectors to a preset scenario feature space to generate scenario mapping data; the knowledge transfer processing network also includes a verification module, which verifies the validity of the scenario mapping data and outputs the verification results, and updates the processing parameters of the data perception processing network and the task allocation parameters of the human-computer collaborative processing framework based on the verification results.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images