Role-driven business process automatic approval method based on reinforcement learning

Abstract

The invention discloses a role-driven business process automatic approval method based on reinforcement learning, and the method comprises the following steps: S1, collecting business process data, carrying out the preprocessing, and constructing a preprocessing data set; s2, constructing a graph data structure, extracting role interaction features, and generating a role dynamic weight matrix; s3, optimizing training by adopting an improved near-end strategy, and optimizing an approval path through shearing and regularization; s4, searching an optimal hyper-parameter by using a He-Make algorithm, and optimizing a learning rate, a discount, entropy regularization and a shearing range; s5, generating an optimal approval path, and dynamically adjusting role permission and manual intervention; s6, analyzing the approval data in real time, and updating the network and the weight matrix online; and S7, deploying a strategy network, and monitoring and dynamically adjusting the approval strategy in real time. According to the method, reinforcement learning and improved near-end strategy optimization are combined with the Hereum algorithm, and dynamic intelligent optimization of the approval process and role permission is realized, so that the approval efficiency and accuracy are greatly improved, and the manual intervention rate is reduced.

Description

Technical Field

[0001] The present invention relates to the field of reinforcement learning technology, and in particular to a role-driven business process automation approval method based on reinforcement learning. Background Art

[0002] The rapid development of enterprise information management and digital transformation has led to the widespread adoption of automated business process approval systems across various industries. Traditional approval systems are primarily based on rule engines or workflow management systems, implementing approval operations through pre-defined approval processes and static role-based permission configurations. This approach offers certain advantages for simple, standardized business processes. However, with the evolving business environment and increasing complexity, traditional methods are facing numerous drawbacks, including low approval efficiency, insufficient flexibility, and an inability to adapt to dynamic business changes.

[0003] In existing technologies, the design of approval processes typically uses fixed flowcharts and static role-authority configurations. Approval decisions rely on pre-set rules or templates, lacking the ability to respond to and process various dynamic factors in the approval process in real time. Decisions at each node in the approval process struggle to fully consider real-time data, historical approval records, and the actual performance of each approval role in the business. This often results in approval paths not being optimally allocated, and even in some cases leading to approval delays, authority conflicts, or inaccurate approval results. At the same time, manual intervention in traditional approval systems is often too frequent, as fixed approval rules cannot effectively respond to rapid changes in the business environment. A large number of anomalies are easily generated during the approval process, forcing companies to continuously rely on manual review, which reduces overall approval efficiency and increases operating costs.

[0004] In recent years, with the development of machine learning and deep learning technologies, reinforcement learning, a method for obtaining optimal decision-making strategies through trial and error, has gradually gained application in various fields. Reinforcement learning methods dynamically adjust automated strategies through continuous interaction with the environment and accumulated rewards, making them better able to cope with complex and dynamic decision-making problems. However, among existing reinforcement learning applications, most methods focus on areas such as gaming and robotic control, while relatively little research has applied reinforcement learning to business process approval. Existing technologies have not yet fully utilized reinforcement learning to intelligently optimize approval paths and dynamically adjust approval role permissions.

[0005] Furthermore, while some current research attempts to incorporate reinforcement learning into business process approval systems, most often use traditional deep reinforcement learning models, such as deep Q-networks or basic policy gradient methods. These methods suffer from slow convergence and insufficient stability when faced with high-dimensional state spaces and complex decision-making processes. In approval tasks, in particular, multi-dimensional information such as approval status, approval history, and interactions between roles is involved. Efficiently integrating this data and generating accurate approval decisions poses a significant technical challenge. Furthermore, traditional reinforcement learning models are susceptible to excessively large step sizes and policy fluctuations during updates, leading to unstable approval decisions and suboptimal approval paths.

[0006] To address these issues, in recent years, academia and industry have begun exploring automated approval methods based on reinforcement learning, introducing advanced deep reinforcement learning models to achieve intelligent and dynamic optimization of approval decisions. Some studies have attempted to use proximal policy optimization models, which can improve the stability of policy updates to a certain extent by imposing clipping restrictions on policy updates. However, traditional proximal policy optimization models still have shortcomings when dealing with multi-objective optimization problems in the approval process. For example, it is difficult to achieve a balance between multiple indicators such as approval efficiency, role load balancing, approval accuracy, and approval costs. Furthermore, fixed hyperparameter settings often cannot adapt to dynamically changing business needs in actual approval scenarios, resulting in poor performance of the model in practical applications.

[0007] Therefore, how to provide a role-driven business process automation approval method based on reinforcement learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One objective of the present invention is to propose a role-driven automated approval method for business processes based on reinforcement learning. This method leverages reinforcement learning techniques, an improved proximal policy optimization model, the Hippo optimization algorithm, and a graph attention network. Through deep learning and online feedback updates, it achieves intelligent optimization of approval paths and dynamic adjustment of approval role permissions in the approval process. The method first constructs a role relationship graph based on preprocessed business data and uses a graph attention network to extract interaction features between approval roles, thereby generating a dynamic weight matrix that reflects the influence of role decisions. Next, the business approval policy network is trained using an improved proximal policy optimization model. This model introduces an adaptive clipping mechanism and cross-role policy consistency regularization, effectively improving the stability and optimality of the approval decision output. Furthermore, the Hippo optimization algorithm is used to perform a global search of key hyperparameters of the network. These hyperparameters are dynamically updated online with real-time feedback data to continuously improve the model's adaptability to complex approval scenarios and its decision accuracy. Consequently, the present invention achieves accurate generation of approval paths, rational allocation of role permissions, and significant improvements in approval efficiency during the automation of the approval process. This method has the advantages of high decision-making accuracy, strong dynamic adaptability, high approval efficiency and reduced manual intervention. It provides enterprises with a safe, reliable, efficient and flexible new technical solution in digital transformation, process automation and intelligent management.

[0009] According to an embodiment of the present invention, a role-driven business process automated approval method based on reinforcement learning includes the following steps: S1. Obtain business process data, collect historical approval records, approval nodes in the approval process, approval status, and role information, construct a multidimensional dataset, and preprocess the multidimensional dataset; S2. Build a graph data structure based on the preprocessed multidimensional dataset, use a graph attention network to extract the role interaction features in the approval process, and generate a role dynamic weight matrix; S3. Build a business approval policy network, train it using an improved proximal policy optimization model, use a dynamic shearing mechanism to constrain business approval policy updates, and combine it with cross-role policy consistency regularization constraints to optimize the approval path of the business approval policy network. S4. Build a hyperparameter optimization module based on the Hippo optimization algorithm, define the optimization objectives, search for the optimal hyperparameter combination, and optimize the learning rate, discount factor, entropy regularization coefficient, and clipping range in the business approval policy network; S5. Utilize the optimized business approval strategy network to make intelligent decisions about the approval process, generate the optimal approval path, adjust the authority allocation of approval roles based on the dynamic role weight matrix, and determine whether manual intervention is required based on the current approval task status. S6. Analyze the data generated in the approval process in real time, calculate the execution results of the approval decision, update the business approval strategy network, and dynamically adjust the role dynamic weight matrix using an online collaborative update mechanism; S7. Deploy the trained business approval policy network in the business system to monitor the operation status of the approval process in real time. When the approval exception rate or role load exceeds the set threshold, dynamically adjust the business approval policy and optimize the approval path and role authority configuration.

[0010] Optionally, the S2 specifically includes: S21. Based on the pre-processed multidimensional dataset, extract the information of each approval role, approval node and approval path involved in the approval process, and build a role set With node collection , where n represents the total number of roles in the approval process, m represents the total number of nodes in the approval process, and a graph data structure G=(R,E) is constructed based on the role interactions in the approval process; S22. Based on the approval history, each edge in the graph data structure is given an initial weight, and the role interaction time attenuation factor and context correlation coefficient are introduced. To the role The interaction weight for: ; in, Representing a role To the role The number of times the approval task was transferred. is the time decay factor, is the contextual correlation coefficient, is the smoothing constant; S23. Use graph attention network to extract features from the normalized graph data structure. The input feature vector , calculation role With the role The attention coefficient between : ; Among them, W is the trainable weight matrix, a is the attention vector, represents vector concatenation, Representing a role With the role The time interval between the last interaction, Representing a role The input feature vector of Representing a role The input feature vector of is the logarithmic function, LeakyReLU is the activation function; S24. Calculate the attention coefficient Perform normalization to obtain the standardized attention coefficient : S25. The standardized attention coefficient matrix Integration generates the character dynamic weight matrix D, where each element ,The role dynamic weight matrix is used to reflect the decision-making influence of each ,approval role in the approval process.

[0011] Optionally, the S3 specifically includes: S31. Constructing the input feature vector of the business approval strategy network : ; in, represents the approval status vector, Represents the eigenvector of the corresponding approval role in the generated role dynamic weight matrix, Represents the approval history vector, Represents vector concatenation; S32. Define the business approval policy network as a parameter set The neural network model, the mapping function is in Indicates the approval decision output; S33. Improve the proximal strategy optimization model and calculate the advantage function of each sample : ; in, is the immediate reward at time i+t, is the value function, is the value function parameter, is the discount factor, t is the time step index, T is the sampling time domain, represents the state vector at time step i+t+1, represents the state vector at time step i+t; S34. Introduce an adaptive shearing mechanism to dynamically adjust the strategy probability ratio and define an adaptive shearing threshold for: ; in, is the initial clipping threshold, is the attenuation coefficient, Represents the advantage function The module length, is the standard deviation of the advantage function in the current batch, is the mean of the advantage function in the current batch, To prevent the smoothing constant from being zero in the denominator, exp is an exponential function; S35. Constructing an improved proximal strategy optimization objective function , optimize the approval path of the business approval policy network: ; in represents the strategy probability ratio, N represents the total number of training samples, Indicates that the strategy probability is Limited to the range Inside, min means the minimum value operation; S36, introduce cross-role policy consistency regularization constraints and calculate the regularization term : ; in, is the perturbation input feature vector, M is the number of similar scene samples; S37. Use the improved proximal policy optimization model to train the business approval policy network and update the network parameters : ; in, According to the objective function The calculated gradient update amount, is the adaptive clipping threshold, is the regularization coefficient, is the scaling factor, is the mean of the advantage function in the current batch, To update the network parameters.

[0012] Optionally, the S4 specifically includes: S41. Construct a hyperparameter vector, denoted as : ; in, represents the learning rate of the business approval policy network, represents the discount factor, represents the entropy regularization coefficient, Indicates the shear range; S42. Constructing a multi-objective optimization function : ; in, represents the normalized index of approval efficiency under given hyperparameters, Indicates the role load balancing normalization index, represents the normalized index of approval accuracy, represents the normalized index of approval cost, , , and is the predetermined weight coefficient; S43. Initialize candidate hyperparameter vector population , where P is the population size; S44. Use the Hippo optimization algorithm to iteratively update the candidate hyperparameter vector. The update rule is: ; in, represents the value of the i-th candidate vector at iteration t, Indicates that the current iteration The largest best candidate vector, represents the mean of the candidate hyperparameter vector, To use the step size coefficient, To explore the step size coefficient, represents the value of the i-th candidate vector at iteration t+1; S45. Calculate the multi-objective optimization function for each candidate vector , and update the current best candidate vector to: ; in, Indicates Get the maximum value value; S46. After reaching the preset number of iterations or convergence conditions, select the final hyperparameter vector Applied to the training of the business approval policy network to update the learning rate, discount factor, entropy regularization coefficient and clipping range.

[0013] Optionally, the S5 specifically includes: S51, collect the latest status, role information, approval history and feedback data of the current approval task from the real-time approval system, and construct the input vector : ; in, Represents the latest approval status vector, Represents the eigenvector of the corresponding approval role in the generated role dynamic weight matrix, Represents the latest approval history vector, Represents vector concatenation; S52. Utilize the optimized business approval strategy network For the input vector Process and generate approval decision output , introducing the adaptive fusion function Constructing the optimal approval path : ; Among them, D represents the role dynamic weight matrix, Indicates the role in the approval history The relevant feature components, is the fusion balance coefficient, is the dynamic feedback factor, n represents the total number of roles in the approval process, In the approval task i, the role The weight of influence on the approval decision, represents the dynamic feedback factor; S53, combined with the role dynamic weight matrix D, the historical performance indicator vector of the approval role and the Role Synergy Matrix ,in Representing a role With the role The synergy between them is used to dynamically adjust the permissions of the approval roles and construct the role permission allocation vector : ; in, represents the i-th row of the character dynamic weight matrix D, Indicates the approval role Historical performance indicators, 、 and is the balance coefficient, is the activation function; S54, based on the latest approval status vector Approval decision output , calculate the manual intervention index : ; in, is the preset state threshold, Indicates that human intervention is required; S55, according to the manual intervention indicators Determine the execution mode of the approval task. The approval task will be submitted to the senior approval role or administrator for review, otherwise the business approval policy network will automatically perform the approval; S56, the optimal approval path generated , Role Permission Allocation Vector and manual intervention indicators Integrate into approval decision results and record the approval decision results in the approval result database.

[0014] Optionally, the S6 specifically includes: S61. Collect feedback data generated during the approval process from the real-time approval system and construct a feedback vector. The feedback vector includes the actual execution result and task completion time of the approval task. S62. Construct a feedback error vector to reflect the deviation between the actual approval result and the approval decision output generated by the optimized business approval strategy network; S63. Based on the feedback error and the updated input vector, including the latest approval status, the role dynamic weight, and the latest approval history, obtain the feedback update amount of the business approval policy network parameter; S64. Determine the update amount of the role dynamic weight matrix based on the feedback adjustment in combination with the approval feedback data, and update the role dynamic weight matrix D: ; Among them, D is the role dynamic weight matrix, is the updated role dynamic weight matrix, is the weight matrix obtained based on real-time feedback data, is the weight matrix update coefficient, is the feedback innovation coefficient, is the matrix variation caused by feedback; S65. Get the hyperparameter vector Perform further online optimization to generate feedback-adjusted hyperparameter vectors : ; in, represents the updated hyperparameter vector, represents the feedback-generated tuning benchmark hyperparameter vector, is the hyperparameter change amount obtained based on the approval feedback data, Update the momentum vector for the historical hyperparameters, is the hyperparameter update rate, is the feedback innovation coefficient, is the momentum update coefficient.

[0015] The beneficial effects of the present invention are: This paper combines reinforcement learning technology, an improved proximal policy optimization model, the Hippo optimization algorithm, and a graph attention network to provide a novel solution to the shortcomings of traditional automated business process approval systems in complex and dynamic business environments. By constructing a role relationship graph based on preprocessed business data and using a graph attention network to extract interaction features between approval roles, the paper generates a dynamic weight matrix that reflects the influence of each role's decision-making, enabling the precise generation of approval paths and the intelligent allocation of approval role permissions.

[0016] Furthermore, this invention improves upon traditional reinforcement learning methods by employing an adaptive shearing mechanism and cross-role policy consistency regularization, effectively balancing exploration and stability. This allows the business approval policy network to maintain a high degree of stability and adaptability during updates. Furthermore, by integrating the Hippo optimization algorithm to globally optimize key network hyperparameters and enabling online updates using real-time feedback data, this invention can continuously adapt to changes in complex approval scenarios, significantly improving the accuracy and efficiency of approval decisions.

[0017] As a result, this invention not only improves the flexibility and intelligence of the approval process automation system in handling multidimensional data and dynamic tasks, but also significantly reduces reliance on human intervention, shortens the approval cycle, and optimizes resource allocation. Overall, this invention offers benefits such as high decision-making accuracy, strong dynamic adaptability, high approval efficiency, rational role and authority allocation, and excellent system operational stability, providing secure, reliable, and efficient technical support for enterprise digital transformation and intelligent management. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flowchart of the role-driven business process automation approval method based on reinforcement learning proposed by the present invention; Figure 2 This is a schematic diagram of constructing a role dynamic weight matrix for the role-driven business process automation approval method based on reinforcement learning proposed in the present invention. DETAILED DESCRIPTION

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0020] refer to Figure 1 and Figure 2 The role-driven business process automation approval method based on reinforcement learning includes the following steps: S1. Obtain business process data, collect historical approval records, approval nodes in the approval process, approval status, and role information, construct a multidimensional dataset, and preprocess the multidimensional dataset; S2. Build a graph data structure based on the preprocessed multidimensional dataset, use a graph attention network to extract the role interaction features in the approval process, and generate a role dynamic weight matrix; S3. Build a business approval policy network, train it using an improved proximal policy optimization model, use a dynamic shearing mechanism to constrain business approval policy updates, and combine it with cross-role policy consistency regularization constraints to optimize the approval path of the business approval policy network. S4. Build a hyperparameter optimization module based on the Hippo optimization algorithm, define the optimization objectives, search for the optimal hyperparameter combination, and optimize the learning rate, discount factor, entropy regularization coefficient, and clipping range in the business approval policy network; S5. Utilize the optimized business approval strategy network to make intelligent decisions about the approval process, generate the optimal approval path, adjust the authority allocation of approval roles based on the dynamic role weight matrix, and determine whether manual intervention is required based on the current approval task status. S6. Analyze the data generated in the approval process in real time, calculate the execution results of the approval decision, update the business approval strategy network, and dynamically adjust the role dynamic weight matrix using an online collaborative update mechanism; S7. Deploy the trained business approval policy network in the business system to monitor the operation status of the approval process in real time. When the approval exception rate or role load exceeds the set threshold, dynamically adjust the business approval policy and optimize the approval path and role authority configuration.

[0021] In this embodiment, S2 specifically includes: S21. Based on the pre-processed multidimensional dataset, extract the information of each approval role, approval node and approval path involved in the approval process, and build a role set With node collection , where n represents the total number of roles in the approval process, m represents the total number of nodes in the approval process, and a graph data structure G=(R,E) is constructed based on the role interactions in the approval process; S22. Based on the approval history, each edge in the graph data structure is given an initial weight, and the role interaction time attenuation factor and context correlation coefficient are introduced. To the role The interaction weight for: ; in, Representing a role To the role The number of times the approval task was transferred. is the time decay factor, is the contextual correlation coefficient, is the smoothing constant; S23. Use graph attention network to extract features from the normalized graph data structure. The input feature vector , calculation role With the role The attention coefficient between : ; Among them, W is the trainable weight matrix, a is the attention vector, represents vector concatenation, Representing a role With the role The time interval between the last interaction, Representing a role The input feature vector of Representing a role The input feature vector of is the logarithmic function, LeakyReLU is the activation function; S24. Calculate the attention coefficient Perform normalization to obtain the standardized attention coefficient : S25. The standardized attention coefficient matrix Integration generates the character dynamic weight matrix D, where each element ,The role dynamic weight matrix is used to reflect the decision-making influence of each ,approval role in the approval process.

[0022] In this embodiment, S3 specifically includes: S31. Constructing the input feature vector of the business approval strategy network : ; in, represents the approval status vector, Represents the eigenvector of the corresponding approval role in the generated role dynamic weight matrix, Represents the approval history vector, Represents vector concatenation; S32. Define the business approval policy network as a parameter set The neural network model, the mapping function is in Indicates the approval decision output; S33. Improve the proximal strategy optimization model and calculate the advantage function of each sample : ; in, is the immediate reward at time i+t, is the value function, is the value function parameter, is the discount factor, t is the time step index, T is the sampling time domain, represents the state vector at time step i+t+1, represents the state vector at time step i+t; S34. Introduce an adaptive shearing mechanism to dynamically adjust the strategy probability ratio and define an adaptive shearing threshold for: ; in, is the initial clipping threshold, is the attenuation coefficient, Represents the advantage function The module length, is the standard deviation of the advantage function in the current batch, is the mean of the advantage function in the current batch, To prevent the smoothing constant from being zero in the denominator, exp is an exponential function; S35. Constructing an improved proximal strategy optimization objective function , optimize the approval path of the business approval policy network: ; in represents the strategy probability ratio, N represents the total number of training samples, Indicates that the strategy probability is Limited to the range Inside, min means the minimum value operation; S36, introduce cross-role policy consistency regularization constraints and calculate the regularization term : ; in, is the perturbation input feature vector, M is the number of similar scene samples; S37. Use the improved proximal policy optimization model to train the business approval policy network and update the network parameters : ; in, According to the objective function The calculated gradient update amount, is the adaptive clipping threshold, is the regularization coefficient, is the scaling factor, is the mean of the advantage function in the current batch, To update the network parameters.

[0023] In this embodiment, the S4 specifically includes: S41. Construct a hyperparameter vector, denoted as : ; in, represents the learning rate of the business approval policy network, represents the discount factor, represents the entropy regularization coefficient, Indicates the shear range; S42. Constructing a multi-objective optimization function : ; in, represents the normalized index of approval efficiency under given hyperparameters, Indicates the role load balancing normalization index, represents the normalized index of approval accuracy, represents the normalized index of approval cost, , , and is the predetermined weight coefficient; S43. Initialize candidate hyperparameter vector population , where P is the population size; S44. Use the Hippo optimization algorithm to iteratively update the candidate hyperparameter vector. The update rule is: ; in, represents the value of the i-th candidate vector at iteration t, Indicates that the current iteration The largest best candidate vector, represents the mean of the candidate hyperparameter vector, To use the step size coefficient, To explore the step size coefficient, represents the value of the i-th candidate vector at iteration t+1; S45. Calculate the multi-objective optimization function for each candidate vector , and update the current best candidate vector to: ; in, Indicates Get the maximum value value; S46. After reaching the preset number of iterations or convergence conditions, select the final hyperparameter vector Applied to the training of the business approval policy network to update the learning rate, discount factor, entropy regularization coefficient and clipping range.

[0024] In this embodiment, the S5 specifically includes: S51, collect the latest status, role information, approval history and feedback data of the current approval task from the real-time approval system, and construct the input vector : ; in, Represents the latest approval status vector, Represents the eigenvector of the corresponding approval role in the generated role dynamic weight matrix, Represents the latest approval history vector, Represents vector concatenation; S52. Utilize the optimized business approval strategy network For the input vector Process and generate approval decision output , introducing the adaptive fusion function Constructing the optimal approval path : ; Among them, D represents the role dynamic weight matrix, Indicates the role in the approval history The relevant feature components, is the fusion balance coefficient, is the dynamic feedback factor, n represents the total number of roles in the approval process, In the approval task i, the role The weight of influence on the approval decision, represents the dynamic feedback factor; S53, combined with the role dynamic weight matrix D, the historical performance indicator vector of the approval role and the Role Synergy Matrix ,in Representing a role With the role The synergy between them is used to dynamically adjust the permissions of the approval roles and construct the role permission allocation vector : ; in, represents the i-th row of the character dynamic weight matrix D, Indicates the approval role Historical performance indicators, 、 and is the balance coefficient, is the activation function; S54, based on the latest approval status vector Approval decision output , calculate the manual intervention index : ; in, is the preset state threshold, Indicates that human intervention is required; S55, according to the manual intervention indicators Determine the execution mode of the approval task. The approval task will be submitted to the senior approval role or administrator for review, otherwise the business approval policy network will automatically perform the approval; S56, the optimal approval path generated , Role Permission Allocation Vector and manual intervention indicators Integrate into approval decision results and record the approval decision results in the approval result database.

[0025] In this embodiment, S6 specifically includes: S61. Collect feedback data generated during the approval process from the real-time approval system and construct a feedback vector. The feedback vector includes the actual execution result and task completion time of the approval task. S62. Construct a feedback error vector to reflect the deviation between the actual approval result and the approval decision output generated by the optimized business approval strategy network; S63. Based on the feedback error and the updated input vector, including the latest approval status, the role dynamic weight, and the latest approval history, obtain the feedback update amount of the business approval policy network parameter; S64. Determine the update amount of the role dynamic weight matrix based on the feedback adjustment in combination with the approval feedback data, and update the role dynamic weight matrix D: ; Among them, D is the role dynamic weight matrix, is the updated role dynamic weight matrix, is the weight matrix obtained based on real-time feedback data, is the weight matrix update coefficient, is the feedback innovation coefficient, is the matrix variation caused by feedback; S65. Get the hyperparameter vector Perform further online optimization to generate feedback-adjusted hyperparameter vectors : ; in, represents the updated hyperparameter vector, represents the feedback-generated tuning benchmark hyperparameter vector, is the hyperparameter change amount obtained based on the approval feedback data, Update the momentum vector for the historical hyperparameters, is the hyperparameter update rate, is the feedback innovation coefficient, is the momentum update coefficient.

[0026] Example 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to the procurement contract approval process of a large enterprise. The traditional approval method relies on a fixed rule engine, the approval process is rigid, and the role authority allocation is unreasonable, resulting in low approval efficiency, frequent manual intervention, and a long approval cycle, which affects the company's operational efficiency and cost management. For example, the company's contract approval involves multiple departments, including the procurement department, the finance department, the legal department and senior management. Under the traditional model, all contract approvals need to be carried out according to a fixed process. Regardless of the size of the contract amount or the credit status of the supplier, they must go through multiple approval levels. Since the process cannot be adjusted dynamically, low-risk contracts require approval from multiple management levels, while high-risk contracts may not trigger a review in time. The approval resources are unevenly distributed and the approval process is seriously delayed.

[0027] This invention, applied to the enterprise's procurement contract approval system, aims to optimize the approval process, reduce approval time, improve approval efficiency, and rationally allocate permissions for each approval role. First, the system extracts multidimensional data from historical contract approval data, including approval roles, contract amounts, supplier reputation, and historical approval paths, to construct a role relationship graph. It then uses a graph attention network to extract interaction features between approval roles and generate a dynamic role weight matrix. Then, based on the optimized business approval strategy network, the system intelligently determines the optimal approval path based on the contract amount, risk level, historical approval records, and the behavior patterns of the approval roles, and dynamically adjusts the permissions of the approval roles. For example, procurement contracts with contract amounts less than 500,000 yuan and reputable suppliers can have a simplified approval process, with direct review by the procurement manager. For contracts with larger contract amounts and lower supplier credit scores, the system automatically adds financial and legal approval steps and, when necessary, triggers manual intervention to ensure the security of high-risk contracts.

[0028] After the system of the present invention was officially put into operation in the enterprise, the optimization effects of the approval process before and after the launch were compared. Before the launch, the average contract approval time of the enterprise was 21.4 hours, of which the approval process for contracts with a value of less than 500,000 yuan accounted for 65%, but still needed to be reviewed by senior management, resulting in a long approval time for low-risk contracts. The manual intervention rate is as high as 50%, that is, one out of every two contract approvals requires manual adjustment of the approval process or role authority, resulting in a large amount of human resources being occupied. In addition, the role authority distribution of contract approval is seriously unbalanced, the approval task load of the purchasing manager is too high, the approval tasks of the finance and legal departments are unevenly distributed, and the approval tasks of some senior management personnel are too concentrated, affecting the overall approval efficiency.

[0029] Four months after the system of the present invention was put into operation, the approval efficiency was significantly improved. The average time for contract approval was reduced to 6.8 hours, a reduction of 68.2% compared with the traditional process; among them, 80% of contracts below 500,000 yuan were directly approved by the purchasing manager without the intervention of senior management, making the approval process more efficient. The manual intervention rate dropped to 12%, significantly reducing the reliance on manual adjustments in the approval process. The allocation of roles and permissions has become more reasonable, the approval tasks of the purchasing manager have been reduced by 40%, and the approval task load of the finance and legal departments has become more balanced. In addition, after the contract approval process was optimized, the company's operating costs for procurement contract approval were reduced by 27%, the approval error rate dropped by 18%, and the accuracy and efficiency of the approval process were comprehensively improved.

[0030] Overall, the application of this system in the enterprise procurement approval process significantly optimizes the approval path, improves the accuracy of approval decisions, reduces manual intervention, and increases the degree of automation. It also achieves a rational allocation of approval roles, reduces operating costs, and improves approval efficiency and contract execution accuracy. This optimization achievement is not only applicable to procurement contract approval scenarios but can also be extended to multiple fields such as finance, healthcare, and government affairs, providing an efficient, flexible, and intelligent solution for enterprise digital transformation and intelligent approval.

[0031] Table 1 Comparison of contract approval optimization effects ; From the above contract approval optimization effect comparison table, it can be seen that the role-driven business process automation approval method based on reinforcement learning proposed in the present invention has significantly improved the approval efficiency in actual application and effectively reduced manual intervention. Under the traditional approval model, the average time for enterprise contract approval is as long as 21.4 hours. After the application of the present invention, the approval time is shortened to 6.8 hours, a reduction of 68.2%. In the specific approval amount classification, the approval time for contracts below 500,000 yuan is reduced from 16.2 hours to 5.2 hours, a reduction of 67.9%; and the approval time for contracts above 1 million yuan is reduced from 32.8 hours to 11.5 hours, a reduction of 64.9%. These data fully prove that the present invention can achieve significant acceleration of the approval process in contract approval scenarios of different scales, thereby effectively improving work efficiency.

[0032] Furthermore, the manual intervention rate in traditional approval systems was as high as 50%. However, after applying this invention, this rate was reduced to 12%, a 76.0% decrease. For high-risk contracts, the manual intervention rate dropped from 75% to 30%, while for low-risk contracts, the manual intervention rate dropped from 38% to 9%. This demonstrates that the automated approval system constructed through this invention not only achieves intelligent approval decision-making, but also significantly reduces human intervention, lowers the risk of approval errors, and improves the reliability of the entire approval process.

[0033] In terms of role authority allocation, under the traditional approval process, purchasing managers shouldered as much as 68% of approval tasks. However, after optimization using this invention, this proportion was reduced to 41%, a 40% reduction in burden, allowing purchasing managers to focus on more critical tasks. Simultaneously, the proportion of approval tasks handled by the finance and legal departments increased from 20% and 12% to 30% and 29%, respectively, representing increases of 50% and 141.7%, respectively. This ensures a rational allocation of approval tasks and avoids resource waste. As a result, the entire approval system achieves a more balanced role configuration, ensuring that each department can play its due role, thereby improving overall approval efficiency and fairness.

[0034] In terms of economic benefits, the application of this invention has reduced procurement approval operating costs from 1.1 million yuan to 800,000 yuan, a 27% reduction. The approval error rate has also dropped from 22% to 4%, significantly reducing contract execution issues caused by approval errors. This significant reduction in the approval error rate has further improved the accuracy of contract execution and the security of corporate operations, reducing the economic losses caused by approval errors.

[0035] In summary, the role-driven business process automation approval method of the present invention realizes the intelligent optimization of the approval process by utilizing advanced technologies such as reinforcement learning, improved proximal strategy optimization model, Hippo optimization algorithm and graph attention network. The system can extract the key features of the approval role from historical data, construct a dynamic weight matrix, and realize the accurate generation of the approval path and the reasonable allocation of role permissions through continuous online updates. The actual application results show that this method not only greatly shortens the approval time and reduces the manual intervention rate, but also optimizes the role division, improves the approval accuracy and enterprise operation efficiency, has significant economic benefits and management advantages, and provides reliable and efficient technical support for enterprise digital transformation and intelligent approval.

[0036] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A role-driven business process automation approval method based on reinforcement learning, characterized by: The steps include: S1. Obtain business process data, collect historical approval records, approval nodes in the approval process, approval status, and role information, construct a multidimensional dataset, and preprocess the multidimensional dataset; S2. Build a graph data structure based on the preprocessed multidimensional dataset, use a graph attention network to extract the role interaction features in the approval process, and generate a role dynamic weight matrix; S3. Build a business approval policy network, train it using an improved proximal policy optimization model, use a dynamic shearing mechanism to constrain business approval policy updates, and combine it with cross-role policy consistency regularization constraints to optimize the approval path of the business approval policy network. S4. Build a hyperparameter optimization module based on the Hippo optimization algorithm, define the optimization objectives, search for the optimal hyperparameter combination, and optimize the learning rate, discount factor, entropy regularization coefficient, and clipping range in the business approval policy network; S5. Utilize the optimized business approval strategy network to make intelligent decisions about the approval process, generate the optimal approval path, adjust the authority allocation of approval roles based on the dynamic role weight matrix, and determine whether manual intervention is required based on the current approval task status. S6. Analyze the data generated in the approval process in real time, calculate the execution results of the approval decision, update the business approval strategy network, and dynamically adjust the role dynamic weight matrix using an online collaborative update mechanism; S7. Deploy the trained business approval policy network in the business system to monitor the operation status of the approval process in real time. When the approval exception rate or role load exceeds the set threshold, dynamically adjust the business approval policy and optimize the approval path and role authority configuration.

2. The role-driven business process automation approval method based on reinforcement learning according to claim 1 is characterized in that: The S2 specifically includes: S21. Based on the pre-processed multidimensional dataset, extract the information of each approval role, approval node and approval path involved in the approval process, and build a role set With node collection , where n represents the total number of roles in the approval process, m represents the total number of nodes in the approval process, and a graph data structure G=(R,E) is constructed based on the role interactions in the approval process; S22. Based on the approval history, each edge in the graph data structure is given an initial weight, and the role interaction time attenuation factor and context correlation coefficient are introduced. To the role The interaction weight for: ； in, Representing a role To the role The number of times the approval task was transferred. is the time decay factor, is the contextual correlation coefficient, is the smoothing constant; S23. Use graph attention network to extract features from the normalized graph data structure. The input feature vector , calculation role With the role The attention coefficient between : ； Among them, W is the trainable weight matrix, a is the attention vector, represents vector concatenation, Representing a role With the role The time interval between the last interaction, Representing a role The input feature vector of Representing a role The input feature vector of is the logarithmic function, LeakyReLU is the activation function; S24. Calculate the attention coefficient Perform normalization to obtain the standardized attention coefficient : S25. The standardized attention coefficient matrix Integration generates the character dynamic weight matrix D, where each element ,The role dynamic weight matrix is used to reflect the decision-making influence of each ,approval role in the approval process.

3. The role-driven business process automation approval method based on reinforcement learning according to claim 1 is characterized in that: The S3 specifically includes: S31. Constructing the input feature vector of the business approval strategy network : ; in, represents the approval status vector, Represents the eigenvector of the corresponding approval role in the generated role dynamic weight matrix, Represents the approval history vector, Represents vector concatenation; S32. Define the business approval policy network as a parameter set The neural network model, the mapping function is in Indicates the approval decision output; S33. Improve the proximal strategy optimization model and calculate the advantage function of each sample : ； in, is the immediate reward at time i+t, is the value function, is the value function parameter, is the discount factor, t is the time step index, T is the sampling time domain, represents the state vector at time step i+t+1, represents the state vector at time step i+t; S34. Introduce an adaptive shearing mechanism to dynamically adjust the strategy probability ratio and define an adaptive shearing threshold for: ； in, is the initial clipping threshold, is the attenuation coefficient, Represents the advantage function The module length, is the standard deviation of the advantage function in the current batch, is the mean of the advantage function in the current batch, To prevent the smoothing constant from being zero in the denominator, exp is an exponential function; S35. Constructing an improved proximal strategy optimization objective function , optimize the approval path of the business approval policy network: ; in represents the strategy probability ratio, N represents the total number of training samples, Indicates that the strategy probability is Limited to the range Inside, min means the minimum value operation; S36, introduce cross-role policy consistency regularization constraints and calculate the regularization term : ; in, is the perturbation input feature vector, M is the number of similar scene samples; S37. Use the improved proximal policy optimization model to train the business approval policy network and update the network parameters : ; in, According to the objective function The calculated gradient update amount, is the adaptive clipping threshold, is the regularization coefficient, is the scaling factor, is the mean of the advantage function in the current batch, To update the network parameters.

4. The role-driven business process automation approval method based on reinforcement learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Construct a hyperparameter vector, denoted as : ; in, represents the learning rate of the business approval policy network, represents the discount factor, represents the entropy regularization coefficient, Indicates the shear range; S42. Constructing a multi-objective optimization function : ; in, represents the normalized index of approval efficiency under given hyperparameters, Indicates the role load balancing normalization index, represents the normalized index of approval accuracy, represents the normalized index of approval cost, , , and is the predetermined weight coefficient; S43. Initialize candidate hyperparameter vector population , where P is the population size; S44. Use the Hippo optimization algorithm to iteratively update the candidate hyperparameter vector. The update rule is: ; in, represents the value of the i-th candidate vector at iteration t, Indicates that the current iteration The largest best candidate vector, represents the mean of the candidate hyperparameter vector, To use the step size coefficient, To explore the step size coefficient, represents the value of the i-th candidate vector at iteration t+1; S45. Calculate the multi-objective optimization function for each candidate vector , and update the current best candidate vector to: ; in, Indicates Get the maximum value value; S46. After reaching the preset number of iterations or convergence conditions, select the final hyperparameter vector Applied to the training of the business approval policy network to update the learning rate, discount factor, entropy regularization coefficient and clipping range.

5. The role-driven business process automation approval method based on reinforcement learning according to claim 1 is characterized in that: The S5 specifically includes: S51, collect the latest status, role information, approval history and feedback data of the current approval task from the real-time approval system, and construct the input vector : ; in, Represents the latest approval status vector, Represents the eigenvector of the corresponding approval role in the generated role dynamic weight matrix, Represents the latest approval history vector, Represents vector concatenation; S52. Utilize the optimized business approval strategy network For the input vector Process and generate approval decision output , introducing the adaptive fusion function Constructing the optimal approval path : ; Among them, D represents the role dynamic weight matrix, Indicates the role in the approval history The relevant feature components, is the fusion balance coefficient, is the dynamic feedback factor, n represents the total number of roles in the approval process, In the approval task i, the role The weight of influence on the approval decision, represents the dynamic feedback factor; S53, combined with the role dynamic weight matrix D, the historical performance indicator vector of the approval role and the Role Synergy Matrix ,in Representing a role With the role The synergy between them is used to dynamically adjust the permissions of the approval roles and construct the role permission allocation vector : ; in, represents the i-th row of the character dynamic weight matrix D, Indicates the approval role Historical performance indicators, 、 and is the balance coefficient, is the activation function; S54, based on the latest approval status vector Approval decision output , calculate the manual intervention index : ; in, is the preset state threshold, Indicates that human intervention is required; S55, according to the manual intervention indicators Determine the execution mode of the approval task. The approval task will be submitted to the senior approval role or administrator for review, otherwise the business approval policy network will automatically perform the approval; S56, the optimal approval path generated , Role Permission Allocation Vector and manual intervention indicators Integrate into approval decision results and record the approval decision results in the approval result database.

6. The role-driven business process automation approval method based on reinforcement learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Collect feedback data generated during the approval process from the real-time approval system and construct a feedback vector. The feedback vector includes the actual execution result and task completion time of the approval task. S62. Construct a feedback error vector to reflect the deviation between the actual approval result and the approval decision output generated by the optimized business approval strategy network; S63. Based on the feedback error and the updated input vector, including the latest approval status, the role dynamic weight, and the latest approval history, obtain the feedback update amount of the business approval policy network parameter; S64. Determine the update amount of the role dynamic weight matrix based on the feedback adjustment in combination with the approval feedback data, and update the role dynamic weight matrix D: ; Among them, D is the role dynamic weight matrix, is the updated role dynamic weight matrix, is the weight matrix obtained based on real-time feedback data, is the weight matrix update coefficient, is the feedback innovation coefficient, is the matrix variation caused by feedback; S65. Get the hyperparameter vector Perform further online optimization to generate feedback-adjusted hyperparameter vectors : ; in, represents the updated hyperparameter vector, represents the feedback-generated tuning benchmark hyperparameter vector, is the hyperparameter change amount obtained based on the approval feedback data, Update the momentum vector for the historical hyperparameters, is the hyperparameter update rate, is the feedback innovation coefficient, is the momentum update coefficient.

Images