SOP flow rapid generation method based on activity features

Abstract

The invention discloses an SOP flow rapid generation method based on activity features, and particularly relates to the technical field of computer application, and the method comprises the steps of S1, activity feature dynamic extraction, S2, knowledge graph enhanced semantic understanding, S3, reinforcement learning flow optimization, and S4, real-time feedback and dynamic adjustment. Through multi-aspect technical innovation, the method has remarkable effects on flow generation, execution and knowledge updating, in the flow generation stage, multi-source heterogeneous data collection is combined with an advanced algorithm, the generation time is greatly shortened, in the execution process, knowledge maps, reinforcement learning and a real-time feedback mechanism are coordinated, the error rate is reduced, the resource utilization rate and compliance are improved, and the method is suitable for large-scale popularization and application. Meanwhile, the knowledge graph can be continuously updated according to the optimization process, the adaptability of the SOP process to service changes is enhanced, the intelligence and automation level of the service process is comprehensively improved, and an efficient, high-quality and flexible process solution is brought.

Description

Technical Field

[0001] The present invention relates to the field of computer application technology, and more particularly to a method for quickly generating an SOP process based on activity features. Background Art

[0002] In today's digitally driven business environment, Standard Operating Procedures (SOPs) are the core elements that ensure the efficient and stable operation of business processes. The effectiveness of their generation and optimization is directly related to the competitiveness of enterprises.

[0003] However, existing SOP process generation technologies expose a series of problems that need to be solved, such as:

[0004] 1. Traditional methods have serious flaws in activity feature extraction. The collection of multi-source, heterogeneous data faces numerous challenges, making it difficult to achieve real-time, comprehensive acquisition of operation logs, device sensor data, and user interaction records. Even if this data is acquired, processing methods for image and time series data are extremely limited.

[0005] 2. During the SOP process generation phase, existing technologies generally lack the ability to deeply mine and utilize business knowledge. Due to the failure to build a comprehensive knowledge graph, it is impossible to clearly define entities, attributes, and relationships based on the business domain terminology library, integrate historical SOP processes, industry standards, and expert experience, and thus fail to use graph neural networks (GNNs) for multi-hop reasoning to accurately determine the implicit relationships between process nodes. As a result, the generated SOP process logic is loose and difficult to meet the complex and ever-changing actual business needs.

[0006] 3. The process optimization link is also the weak point of traditional technology. Most existing methods rely on fixed, single optimization indicators, and lack a mechanism to dynamically adjust reward weights based on real-time feedback such as user evaluation scores and the number of abnormal information in system logs. They are difficult to adapt to the ever-changing business environment. When abnormalities occur in process execution, the detection and processing capabilities of existing technologies are insufficient, and they are unable to quickly and accurately identify abnormal nodes. They also lack effective algorithm support during local adjustments and global optimization, and are unable to generate better process paths. In addition, existing technologies have failed to form a closed-loop mechanism to feed back the optimized process to the knowledge system for continuous updating and improvement, resulting in the inability to accumulate and optimize business knowledge as practice advances, limiting the long-term development and improvement of the SOP process.

[0007] In view of the above situation, the present invention provides a method for quickly generating an SOP process based on activity features. Summary of the Invention

[0008] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for quickly generating an SOP process based on activity features to solve the problems raised in the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: a method for rapidly generating an SOP process based on activity features, specifically comprising the following steps:

[0010] S1. Dynamic extraction of activity features: This involves real-time collection of multi-source heterogeneous data, including business activity logs, device sensor data, and user interaction records. Convolutional neural networks (CNNs) are used to extract image-based activity features. Long short-term memory (LSTM) networks are used to analyze time series data. Key activity nodes are identified using an attention mechanism. Principal component analysis (PCA) or autoencoders are used to map multimodal features into a unified low-dimensional space to generate activity feature vectors.

[0011] S2. Knowledge graphs enhance semantic understanding. An ontology model is built based on the business domain terminology library to define entities, attributes, and relationships. Historical SOP processes, industry standards, and expert experience are integrated to form a knowledge graph. Activity feature vectors are semantically matched with the knowledge graph. Graph neural networks (GNNs) are used for multi-hop reasoning to generate priorities and dependencies between process nodes.

[0012] S3, reinforcement learning process optimization, uses activity feature vectors, process execution status, and environmental parameters as state inputs to define process generation actions, including node addition, sequence adjustment, and resource allocation. It also designs a reward function that integrates efficiency indicators, quality indicators, and dynamic adjustments. It uses a deep Q-network (DQN) or policy gradient algorithm, trained through historical process data and simulated environments, to generate the optimal process strategy.

[0013] S4, real-time feedback and dynamic adjustment, collect process execution data in real time, identify abnormal nodes through Isolation Forest or LSTM autoencoder, perform local adjustment and global optimization, feed back the optimized process to the knowledge graph, and update domain knowledge.

[0014] Preferably, in step S1, the use of a convolutional neural network (CNN) to extract image-type activity features is to input image data of a screenshot of the device operation interface into the convolutional neural network (CNN), and the convolutional neural network (CNN) extracts key features in the image through multiple convolutional layers and pooling layers. The size and number of convolution kernels of the convolutional layer and the pooling method of the pooling layer are adjusted according to the resolution and feature complexity of the image.

[0015] Preferably, in step S1, the use of the long short-term memory network LSTM to analyze the time series data is to input the time series data of the operation steps into the long short-term memory network LSTM according to time steps. The long short-term memory network LSTM controls the transmission and update of information through the forget gate, input gate and output gate to capture the long-term dependencies in the time series data. The number of hidden layers and the number of neurons of the long short-term memory network LSTM are set according to the length and complexity of the time series data.

[0016] Preferably, in step S2, when the ontology model is constructed based on the business domain terminology library, the entities include business-related objects of "order" and "payment";

[0017] The attributes include entity-related properties of "amount" and "time";

[0018] The relationships include logical connections between "contains" and "trigger" entities.

[0019] Preferably, in step S2, the multi-hop reasoning performed using the graph neural network GNN takes the information of nodes and edges in the knowledge graph as input. The graph neural network GNN performs multi-hop reasoning by aggregating and propagating information of neighboring nodes to determine the implicit relationship between process nodes. The number of layers of the graph neural network GNN and the number of neurons in each layer are adjusted according to the scale and complexity of the knowledge graph.

[0020] Preferably, in step S3, when designing the reward function, the efficiency indicators include process execution time T and resource utilization R, wherein the process execution time T is the total time spent from the start to the end of the process, and the resource utilization R is the ratio of the actual amount of resources used to the total amount of resources;

[0021] The quality indicators include error rate E and compliance C, where error rate E is the ratio of the number of errors that occur during process execution to the total number of executions, and compliance C determines whether the process meets the requirements based on preset compliance rules, with a value of 0 indicating non-compliance and 1 indicating compliance.

[0022] The dynamic adjustment is to dynamically adjust the reward weight according to the real-time feedback of the user evaluation score S and the number of abnormal information N in the system log;

[0023] The reward function is F = w1×(1 / T)+w2×R+w3×(1-E)+w4×C+w5×S+w6×(1 / N), where w1, w2, w3, w4, w5 and w6 are dynamically adjusted weight coefficients.

[0024] Preferably, in step S3, when the deep Q network DQN is used for training, a Q network is constructed, the input of which is the state, and the output is the Q value corresponding to each action. The network parameters are updated by minimizing the mean square error between the Q value estimate and the target Q value. The target Q value is calculated by the maximum Q value of the next state and the current reward, that is,

[0025] Among them, s is the current state, a is the current action, r is the current reward, γ is the discount factor, s′ is the next state, and a′ is the action in the next state.

[0026] Preferably, in step S4, the identification of abnormal nodes by Isolation Forest is to use the node-related features in the process execution data as input to construct an isolation forest model, wherein the isolation forest model constructs multiple isolation trees by randomly selecting features and split points, calculates the path length from each node to the root node, and determines whether the node is an abnormal node based on the abnormality of the path length. The abnormality score Siso is calculated based on the degree of deviation between the average path length of the node and the average path length of the normal nodes.

[0027] Preferably, in step S4, when the global optimization uses a genetic algorithm to regenerate the process path, the process nodes are encoded as chromosomes, and a new process path is generated through selection, crossover and mutation operations. The selection operation adopts a roulette wheel selection method, and the probability of being selected is determined according to the process execution effect corresponding to each chromosome, such as the reward function value. The crossover operation is performed according to a certain crossover probability P. c Exchange some genes of the two chromosomes, and the mutation operation is performed according to a certain mutation probability P m Randomly changing genes in chromosomes.

[0028] Preferably, in step S4, the optimized process is fed back to the knowledge graph to update the domain knowledge, which is to add new process nodes, relationships between nodes and relevant information during process execution, such as node time consumption and error types, to the knowledge graph, and update and expand the entities, attributes and relationships in the knowledge graph to improve the domain knowledge.

[0029] The technical effects and advantages of the present invention are as follows:

[0030] 1. By collecting multi-source heterogeneous data in real time and using advanced algorithms such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) to quickly extract activity features, and then using principal component analysis (PCA) or autoencoders to achieve multimodal feature fusion and generate activity feature vectors, compared with traditional methods, this method eliminates the need for extensive manual data combing and analysis, and can quickly complete feature extraction and fusion, greatly shortening the initial preparation time for SOP process generation and significantly reducing process generation time from hours or even days to minutes.

[0031] 2. In the process generation stage, the present invention uses knowledge graphs to enhance semantic understanding and uses graph neural networks (GNNs) to perform multi-hop reasoning to determine the relationship between process nodes, ensuring that the generated process logic is rigorous, scientific, and reasonable. In addition, during the process execution, the reinforcement learning process optimization step designs a reward function that includes multi-dimensional indicators such as efficiency and quality, and generates the optimal process strategy through algorithm training such as the deep Q network (DQN). At the same time, the real-time feedback and dynamic adjustment mechanism can detect and process abnormal nodes in a timely manner, effectively reducing the error rate of process execution, improving resource utilization and compliance, and comprehensively improving the execution quality of the SOP process.

[0032] 3. The knowledge graph constructed by the present invention is not static, but can be updated according to the optimized process in real-time feedback and dynamic adjustment. When new situations or changes in demand arise in business activities, the system can regenerate or adjust the SOP process based on the updated knowledge graph. For example, when business rules are updated, new equipment or operating links are introduced, the system can adapt quickly, flexibly adjust the process, and maintain efficient operation, which greatly enhances the adaptability of the SOP process to the ever-changing business environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 are within the scope of protection of the present invention.

[0035] The present invention provides a method for quickly generating an SOP process based on activity features, which specifically includes the following steps:

[0036] S1. Dynamic extraction of activity features

[0037] By adapting to different data interfaces, it enables real-time acquisition of multi-source heterogeneous data, including business activity operation logs, device sensor data, and user interaction records. For example, log collection tools can be used to connect to the operation log system, sensor drivers can be used to collect device sensor data, and front-end interaction monitoring technology can be used to collect user interaction records.

[0038] The image data is input into the convolutional neural network (CNN). For images of different resolutions, if the resolution is high, a 3x3 convolution kernel is used to accurately capture subtle features. If the resolution is low, a 5x5 convolution kernel is used to improve computational efficiency. Depending on the complexity of the interface elements in the image, when the elements are rich, the number of convolution kernels is increased from 64 to 128 to better extract features. The pooling layer uses the maximum pooling method. For small-size images, a 2x2 pooling window is used, and for large-size images, a 3x3 pooling window is used.

[0039] The time series data of the operation steps are input into the long short-term memory network (LSTM) by time step. If the time series data is short and the changes are simple, set one hidden layer with 64 neurons. If the data is long and the dependencies are complex, increase the hidden layer to three layers and set the number of neurons to 128. Use the forget gate, input gate, and output gate to control information transmission and update, so as to capture the long-term dependencies in the time series data.

[0040] Integrate the image features extracted by CNN, the time series features extracted by LSTM, and other related features. If the data dimension is high and the correlation is strong, use principal component analysis (PCA) to reduce the dimension, determine the number of principal components, and ensure that the data after dimensionality reduction retains more than 95% of the information. If the data distribution is complex, use an autoencoder to map the multimodal features to a unified low-dimensional space after training to generate an activity feature vector.

[0041] S2. Knowledge Graph Enhances Semantic Understanding

[0042] Build an ontology model based on the business domain terminology library, define entities such as "order" and "payment" when involved in order processing, attributes such as "amount" and "time" that are related to entities, and logical connections between entities such as "include" and "trigger" relationships. Collect historical SOP processes, industry standards, and expert experience to build a knowledge graph;

[0043] Input the node and edge information in the knowledge graph into the graph neural network GNN. If the knowledge graph is small and the number of nodes and edges is small, set up a 2-layer network with 32 neurons in each layer. If the graph is large and complex, increase it to 5 layers with 64 neurons in each layer. Through the aggregation and propagation of neighbor node information, multi-hop reasoning is carried out to determine the implicit relationship between process nodes.

[0044] S3. Reinforcement Learning Process Optimization

[0045] Taking the activity feature vector, process execution status and environment parameters as state input, the process generation actions include node addition, such as deciding whether to add a manual review node based on specific conditions, sequence adjustment, adjusting the sequence of process steps according to relevant rules, and resource allocation, allocating server resources according to resource demand.

[0046] Efficiency indicators include process execution time T, which is the total duration from the start to the end of the process; resource utilization R, which is the ratio of actual resource usage to total resource usage; quality indicators include error rate E, which is the ratio of the number of errors in process execution to the total number of executions; and compliance C, which determines whether the process is compliant based on preset compliance rules, with compliance being 1 and non-compliance being 0. Reward weights are dynamically adjusted based on real-time user feedback scores S and the number of exception messages N in the system log.

[0047] The reward function is F = w1×(1 / T)+w2×R+w3×(1-E)+w4×C+w5×S+w6×(1 / N). In actual operation, the weight coefficient can be adjusted according to different situations.

[0048] Construct a Q network whose input is the state and output is the Q value corresponding to each action. In the simulation environment, the network parameters are updated by minimizing the mean square error between the Q value estimate and the target Q value. The target Q value is obtained by the maximum Q value of the next state and the current reward, that is, Where s is the current state, a is the current action, r is the current reward, γ is the discount factor, which ranges from 0.9 to 0.99, such as 0.95, s′ is the next state, and a′ is the action in the next state. The optimal process strategy is generated by continuous training in historical process data and simulation environments.

[0049] S4. Real-time feedback and dynamic adjustment

[0050] Real-time collection of process execution data, including node processing time, error information, and other information. Node-related features in the process execution data are input into the Isolation Forest model. The model randomly selects features and cut points to construct multiple isolation trees, calculates the path length from each node to the root node, and determines whether the node is abnormal based on the degree of abnormality of the path length. When the average path length of a node deviates significantly from the average path length of normal nodes and the calculated anomaly score Siso exceeds the preset threshold, the node is determined to be an abnormal node.

[0051] When an abnormal node is detected, local adjustments are made, such as reallocating auditors. If multiple processes have abnormalities and involve overall path problems, a genetic algorithm is used for global optimization. The process nodes are encoded as chromosomes, and new paths are generated through selection, crossover, and mutation operations. The selection operation uses a roulette wheel selection method to determine the selection probability based on the process execution effect (such as the reward function value) corresponding to each chromosome. The crossover operation is based on the crossover probability P. c (such as 0.8) Exchange some genes of two chromosomes, and the mutation operation is based on the mutation probability P m (e.g. 0.01) Randomly change chromosome genes to generate a better process path;

[0052] Feed the optimized process back to the knowledge graph, update the domain knowledge, add new process nodes, relationships between nodes, and process execution related information, such as node time consumption, error types, etc., to the knowledge graph, update and expand the entities, attributes, and relationships therein, improve the domain knowledge system, and provide knowledge support for subsequent SOP process generation.

[0053] In summary, the method for rapidly generating an SOP process based on activity features of the present invention first collects multi-source heterogeneous data such as operation logs, device sensor data, and user interaction records in business activities in real time by adapting multiple data interfaces, and uses technologies such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) to extract image and time series activity features respectively, and then uses principal component analysis (PCA) or autoencoders to achieve multimodal feature fusion to generate activity feature vectors.

[0054] Then, we build an ontology model based on the business domain terminology library, build a knowledge graph based on historical SOP processes, and use graph neural networks (GNNs) for multi-hop reasoning to determine the relationships between process nodes.

[0055] Then, the activity feature vector, process execution status, and environmental parameters are used as state inputs to define process generation actions. A reward function is designed that includes efficiency, quality, and other indicators and dynamically adjusts weights. The optimal process strategy is generated through training with algorithms such as the Deep Q-Network (DQN). During process execution, data is collected in real time, and abnormal nodes are detected using the Isolation Forest model. Local adjustments are then made, or global optimization is performed using a genetic algorithm.

[0056] Finally, the optimized process is fed back to the knowledge graph to update the domain knowledge, so as to achieve rapid generation and continuous optimization of the SOP process.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for quickly generating SOP processes based on activity features, characterized by: The specific steps include: S1. Dynamic extraction of activity features: This involves real-time collection of multi-source heterogeneous data, including business activity logs, device sensor data, and user interaction records. Convolutional neural networks (CNNs) are used to extract image-based activity features. Long short-term memory (LSTM) networks are used to analyze time series data. Key activity nodes are identified using an attention mechanism. Principal component analysis (PCA) or autoencoders are used to map multimodal features into a unified low-dimensional space to generate activity feature vectors. S2. Knowledge graphs enhance semantic understanding. An ontology model is built based on the business domain terminology library to define entities, attributes, and relationships. Historical SOP processes, industry standards, and expert experience are integrated to form a knowledge graph. Activity feature vectors are semantically matched with the knowledge graph. Graph neural networks (GNNs) are used for multi-hop reasoning to generate priorities and dependencies between process nodes. S3, reinforcement learning process optimization, uses activity feature vectors, process execution status, and environmental parameters as state inputs to define process generation actions, including node addition, sequence adjustment, and resource allocation. It also designs a reward function that integrates efficiency indicators, quality indicators, and dynamic adjustments. It uses a deep Q-network (DQN) or policy gradient algorithm, trained through historical process data and simulated environments, to generate the optimal process strategy. S4, real-time feedback and dynamic adjustment, collect process execution data in real time, identify abnormal nodes through Isolation Forest or LSTM autoencoder, perform local adjustment and global optimization, feed back the optimized process to the knowledge graph, and update domain knowledge.

2. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S1, the use of the convolutional neural network (CNN) to extract image-related activity features is to input image data of a screenshot of the device operation interface into the convolutional neural network (CNN). The convolutional neural network (CNN) extracts key features in the image through multiple convolutional layers and pooling layers. The size and number of convolution kernels of the convolutional layer and the pooling method of the pooling layer are adjusted according to the resolution and feature complexity of the image.

3. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S1, the use of the long short-term memory network LSTM to analyze the time series data is to input the time series data of the operation steps into the long short-term memory network LSTM according to time steps. The long short-term memory network LSTM controls the transmission and update of information through the forget gate, input gate and output gate to capture the long-term dependencies in the time series data. The number of hidden layers and the number of neurons of the long short-term memory network LSTM are set according to the length and complexity of the time series data.

4. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S2, when the ontology model is constructed based on the business domain terminology library, the entities include business-related objects such as "order" and "payment"; The attributes include entity-related properties of "amount" and "time"; The relationships include logical connections between "include" and "trigger" entities.

5. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S2, the multi-hop reasoning using the graph neural network GNN takes the information of nodes and edges in the knowledge graph as input. The graph neural network GNN performs multi-hop reasoning by aggregating and propagating information of neighbor nodes to determine the implicit relationship between process nodes. The number of layers of the graph neural network GNN and the number of neurons in each layer are adjusted according to the scale and complexity of the knowledge graph.

6. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S3, when designing the reward function, the efficiency indicators include process execution time T and resource utilization R, where process execution time T is the total time spent from the start to the end of the process, and resource utilization R is the ratio of the actual amount of resources used to the total amount of resources; The quality indicators include error rate E and compliance C, where error rate E is the ratio of the number of errors that occur during process execution to the total number of executions, and compliance C determines whether the process meets the requirements based on preset compliance rules, with a value of 0 indicating non-compliance and 1 indicating compliance. The dynamic adjustment is to dynamically adjust the reward weight according to the real-time feedback of the user evaluation score S and the number of abnormal information N in the system log; The reward function is F = w1×(1 / T)+w2×R+w3×(1-E)+w4×C+w5×S+w6×(1 / N), where w1, w2, w3, w4, w5 and w6 are dynamically adjusted weight coefficients.

7. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S3, when the deep Q network DQN is used for training, a Q network is constructed. The input of the network is the state, and the output is the Q value corresponding to each action. The network parameters are updated by minimizing the mean square error between the Q value estimate and the target Q value. The target Q value is calculated by the maximum Q value of the next state and the current reward, that is, Among them, s is the current state, a is the current action, r is the current reward, γ is the discount factor, s′ is the next state, and a′ is the action in the next state.

8. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S4, the identification of abnormal nodes through Isolation Forest is to use the node-related features in the process execution data as input to construct an isolation forest model. The isolation forest model constructs multiple isolation trees by randomly selecting features and split points, calculates the path length from each node to the root node, and determines whether the node is an abnormal node based on the abnormality of the path length. The anomaly score Siso is calculated based on the degree of deviation between the average path length of the node and the average path length of normal nodes.

9. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S4, when the global optimization uses the genetic algorithm to regenerate the process path, the process nodes are encoded as chromosomes, and a new process path is generated through selection, crossover and mutation operations. The selection operation adopts the roulette wheel selection method, and the probability of being selected is determined according to the process execution effect corresponding to each chromosome, such as the reward function value. The crossover operation is carried out according to a certain crossover probability P. c Exchange some genes of the two chromosomes, and the mutation operation is performed according to a certain mutation probability P m Randomly changing genes in chromosomes.

10. The method for rapidly generating an SOP process based on activity features according to claim 1, wherein: In step S4, the optimized process is fed back to the knowledge graph to update the domain knowledge, which means adding new process nodes, relationships between nodes, and relevant information during process execution, such as node time consumption and error types, to the knowledge graph, and updating and expanding the entities, attributes, and relationships in the knowledge graph to improve the domain knowledge.

Images