Marine transportation business process automation implementation method based on large language model

Abstract

The application discloses a kind of based on big language model's sea transport business process automation implementation method, belong to process automation field.This method steps include: in operation, collect multimodal data, obtain the task order list of target process;Task items in task order list are grouped;Using big language model assigns activity name for each grouping;Using reinforcement learning model assigns a callable tool for each activity name respectively;Task order list, the activity name corresponding to each task item and the tool corresponding to each activity name are input into big language model, obtain process arrangement data, then process arrangement data is converted into RPA automation script.The application can mine process without modifying original system, shorten the mining time, reduce cost, also eliminate the hidden trouble brought by modifying system, and have the advantages such as strong adaptability, with automatic iterative optimization capability.

Description

Technical Field

[0001] This invention belongs to the field of process automation, and specifically relates to a method for automating maritime business processes. Background Technology

[0002] The ocean shipping business involves a large number of repetitive operations, such as booking processing, customs document generation, and freight settlement. These processes typically require operation across multiple systems (such as ERP, email, Excel, and customs brokerage platforms). In the past, these processes relied on manual execution, which was not only inefficient but also prone to delays or losses due to operational errors.

[0003] With the development of maritime-related technologies, process automation has become a key means to improve efficiency and reduce human error. Robotic Process Automation (RPA) technology, which uses software robots to simulate manual operations and automatically complete repetitive tasks, is a research hotspot in this field. Accurate process mining is crucial for achieving process automation. Current methods involve tracking the status and events of specific data through pre-installed log points in business software systems. For example, Chinese invention patent application CN118585575A discloses a process mining method, device, electronic device, and medium based on a graph database. It extracts case identifiers and information from event logs, stores activities chronologically, mines process edge information, and ultimately stores all information in a graph database to achieve process mining. However, this approach requires secondary modifications to the software system, resulting in a long development cycle and high modification costs. Summary of the Invention

[0004] This invention proposes an automated implementation method for maritime business processes based on a large language model. Its purpose is to solve the problems of long cycle and high cost that exist in obtaining process mining by modifying existing software systems to obtain log information.

[0005] The technical solution of this invention is as follows:

[0006] A method for automating maritime business processes based on a large language model, comprising the following steps:

[0007] Step S1: Collect multimodal data while the operator is performing the target process operation, and obtain the task sequence list of the target process based on the multimodal data;

[0008] Step S2: Group the task items in the task sequence list;

[0009] Step S3: Use the large language model to assign activity names to each group;

[0010] Step S4: Use a reinforcement learning model to assign a callable tool to each activity name;

[0011] Step S5: Input the task sequence list, the activity name corresponding to each task item, and the tool corresponding to each activity name into the large language model to obtain the process orchestration data, and then convert the process orchestration data into RPA automation scripts.

[0012] As a further improvement to the method for automating maritime business processes based on a large language model, in step S1, the method for collecting multimodal data is as follows: real-time capture of multimodal data during operator operation. The multimodal data includes audio data, screen image data, and UI interaction data. When the operator triggers a recording operation, the current multimodal data is recorded and added to a temporary list.

[0013] As a further improvement to the method for automating maritime business processes based on a large language model: In step S1, obtaining the task sequence list of the target process based on multimodal data means: reading each piece of multimodal data in the temporary list, with each piece of multimodal data corresponding to one task item, and sending each piece of multimodal data to a large language model that supports multimodality to obtain the task description information corresponding to each task item.

[0014] As a further improvement to the aforementioned method for automating maritime business processes based on a large language model, the specific process of step S2 is as follows:

[0015] Step S2-1: Convert the task description information of each task item into an embedding vector;

[0016] Step S2-2: Calculate the semantic similarity between each embedding vector;

[0017] Step S2-3: Group task items by clustering based on the semantic similarity between embedded vectors and preset clustering parameters.

[0018] As a further improvement to the method for automating maritime business processes based on a large language model: In step S3, for the current group, the task description information of all task items in the group and the corresponding time are input into the large language model, so that the large language model, guided by prompt words, combines a pre-set set of activity names to assign the best activity name to the group.

[0019] As a further improvement to the aforementioned method for automating maritime business processes based on large language models: The reinforcement learning model is trained using historical data from RPA process operations.

[0020] Each task item, its context information, and the activity to which it belongs in the historical data are defined as a structured state. The context information includes one or more tasks before and after the target task item. The tools used by each task item in the historical data are defined as actions. The task completion rate and execution efficiency corresponding to the implementation of each task item's action are defined as rewards, thus obtaining a training set. Then, the reinforcement learning model is trained using the training set to learn the relationship between tools and activities.

[0021] When using a trained reinforcement learning model to assign tools, the input to the reinforcement learning model is an activity name and the context corresponding to that activity name. The context includes the preceding and following tasks and the activity names corresponding to the tasks. The output is the tool recommended for that activity.

[0022] As a further improvement to the method for automating maritime business processes based on a large language model: in step S5, process orchestration data templates and examples are added to the prompts input into the large language model, as well as the usage documents of all selected tools, to drive the large language model to output process orchestration data in a structured text format.

[0023] As a further improvement to the method for automating maritime business processes based on a large language model: the process orchestration data includes process input interfaces and several sequentially arranged nodes that correspond one-to-one with the task items;

[0024] The process input interface receives externally input data when the RPA automation script is running;

[0025] Each node contains a tool call interface, a node input interface, a node processing function, and a node output interface for the corresponding task item. The node input interface is used to interface with the process input interface or the previous node output interface to obtain data and pass the obtained data to the node processing function for processing. The tool call interface is used to pass the result data processed by the node processing function to the called tool to complete the automated operation of the current task item. The node output interface is used to receive the result data output by the node processing function within this node and the data returned by the tool call interface.

[0026] As a further improvement to the aforementioned method for automating maritime business processes based on large language models, it also includes:

[0027] Step S6: Test the RPA automation script obtained in step S5, and modify the RPA automation script for any abnormal situations that occur.

[0028] As a further improvement to the aforementioned method for automating maritime business processes based on large language models, it also includes:

[0029] Step S7: Apply the modified RPA automation script to the production environment, continuously monitor the running effect, and optimize the RPA automation script generation process based on the monitoring results;

[0030] During monitoring, the following data are collected: success rate and error rate of automated tasks, percentage reduction in process automation execution time relative to the original execution time, and user satisfaction feedback. Based on the collected data, the clustering parameters used for grouping in step S2, the set of activity names used by the large language model in step S3, and the historical data used to train the reinforcement learning model are updated. The reinforcement learning model is then retrained to optimize the generation process of RPA automation scripts.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. This invention uses a multimodal large model to parse the task sequence list obtained from the actual operation process, thereby enabling process mining without modifying the original system, shortening the mining time, reducing costs, and eliminating the hidden dangers brought about by system modification.

[0033] 2. The process of generating RPA automation scripts from a task sequence list in this invention is automatically achieved by the program and a large language model. RPA developers do not need to understand the specific maritime business and software system operation process in detail. They only need to maintain the prompt words of the large model and the corresponding data to complete the script generation, thereby greatly improving work efficiency. At the same time, relying on the excellent semantic processing capabilities of the large language model, it can handle complex processes more accurately and efficiently, and has stronger adaptability.

[0034] 3. This invention introduces a reinforcement learning model to recommend tools for each activity. The advantage of reinforcement learning models lies in their greater focus on the relationship between structured states, actions, and rewards. Since structured states not only contain information about the current task item but also contextual information from historical data, the tools recommended by the reinforcement learning model are not only based on a specific activity name itself, but also on the preceding and following tasks and activities, and even more so on the completion rate in historical data. This fully evaluates the compatibility of connecting different tools, thereby recommending the optimal tool and solving the failure problem caused by the compatibility between tools when automating adjacent tasks, ensuring that the obtained automation script is optimal.

[0035] 4. The process orchestration data of this invention is in a structured text format. Each node contains a tool call interface for the current task, a node input interface, a node processing function, and a node output interface. This not only allows for the rapid conversion from process orchestration data to RPA automation scripts, but also provides good readability, which helps developers to identify and modify / optimize issues in a timely manner.

[0036] 5. After the automated script is put into use, this invention further optimizes the grouping, activity naming, and historical data of the reinforcement learning model based on the collected data, and continuously iterates to improve the automated execution effect. Detailed Implementation

[0037] The technical solution of the present invention will be described in detail below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0038] A method for automating maritime business processes based on a large language model, comprising the following steps:

[0039] Step S1: Collect multimodal data while the operator is performing the target process operation, and obtain the task sequence list of the target process based on the multimodal data.

[0040] The target process is usually a common and highly repetitive process in the ocean shipping business, such as: generating shipping reports, invoice processing, approval workflow, extracting customer information, data entry or data cleaning, etc. It can also be a process involving cross-system operations, such as the process of extracting information from emails and updating the booking system.

[0041] The method for collecting multimodal data is as follows: real-time capture of multimodal data during operator operation. Multimodal data includes audio data, screen image data, and UI interaction data (such as web page interaction). When the operator triggers a recording operation (which can be triggered by a button or voice), the current multimodal data is recorded and added to a temporary list.

[0042] The process of obtaining the task sequence list of the target process based on multimodal data refers to: reading each piece of multimodal data from a temporary list, with each piece of multimodal data corresponding to one task item; and then feeding each piece of multimodal data into a large language model that supports multimodality to obtain the task description information corresponding to each task item. The task description information includes the task name, operation instructions, software window, etc.

[0043] Table 1: Task sequence list for a certain process.

[0044] Task Item Number Task Name Software window Operating Instructions 1 Open the approval form Approval System Open the form page of the "Approval System". 2 Fill out the approval form Approval System Enter approval information in the form 3 Submit approval form Approval System Click the submit button to complete the form submission. 4 Open email Email system Open Outlook and view the emails awaiting approval. 5 Reply to email Email system Reply to the customer's email to confirm the approval status. 6 View approval records Approval System View historical approval records in the approval system

[0045] This step uses the Tsinghua University open-source model "visualglm-6b" (multimodal large language model), combined with audio processing libraries (such as speech_recognition) and video frame extraction technology (OpenCV's cv2.VideoCapture) to process the captured data.

[0046] Step S2: Group the task items in the task sequence list.

[0047] The specific process is as follows:

[0048] Step S2-1: Convert the task description information of each task item into an embedding vector.

[0049] An embedding vector is a way to represent the semantic information of text (or other data) using a high-dimensional real-valued vector. Using a pre-trained language model (such as BERT, GPT, word2vec, or text-embedding-ada-002), the semantics of text can be encoded into a fixed-length vector.

[0050] For example, consider two task descriptions:

[0051] Task 1: "Open the approval form";

[0052] Task 2: "Fill in the form";

[0053] After being processed by an embedding model (such as GPT's embedding API), they may be transformed into the following vectors:

[0054] Task 1 Vector: [0.85, 0.27, 0.63…];

[0055] Task 2 Vector: [0.83, 0.25, 0.65…].

[0056] Step S2-2: Calculate the semantic similarity between each embedding vector.

[0057] In this embodiment, the cosine similarity between embedded vectors is used as the semantic similarity.

[0058] Each dimension of a vector is a real number, representing the projection of the text onto certain semantic features.

[0059] The cosine similarity between any two embedding vectors A and B is calculated as follows: .

[0060] Table 2: Semantic similarity among the 6 task items in Table 1.

[0061] Task Item Number 1 2 3 4 5 6 1 1 0.99 0.98 0.15 0.14 0.97 2 0.99 1 0.98 0.16 0.15 0.96 3 0.98 0.98 1 0.14 0.13 0.95 4 0.15 0.16 0.14 1 0.99 0.16 5 0.14 0.15 0.13 0.99 1 0.15 6 0.97 0.96 0.95 0.16 0.15 1

[0062] Step S2-3: Group the task items based on the semantic similarity between the embedded vectors and the preset clustering parameters.

[0063] The clustering parameters include: minimum number of groups min_samples=2, similarity threshold eps=0.05. During the clustering process, texts with similar semantics are mapped to similar vector spaces.

[0064] Clustering can be performed using methods such as k-means.

[0065] In this embodiment, tasks 1, 2, 3, and 5 are grouped into group 1, and tasks 4 and 5 are grouped into group 2.

[0066] Table 3: Grouping results.

[0067] Group numbering Task Item Number Remark 1 1, 2, 3, 6 There are tasks related to approval. 2 4, 5 Tasks related to email interaction

[0068] Step S2 has the following advantages:

[0069] 1. High flexibility: Based on semantic similarity, it adapts to different task descriptions and language expressions.

[0070] 2. High accuracy: It captures deep semantic information by embedding vectors, avoiding grouping errors caused by surface text differences.

[0071] 3. Wide range of applicable scenarios: It can be used in complex business processes, especially in scenarios with diverse task descriptions.

[0072] This step achieves efficient and accurate task grouping by employing a task semantic similarity-based grouping strategy combined with text embedding and clustering algorithms. This method is particularly suitable for business processes with complex and diverse task descriptions, such as approval processes and customer service processes. The grouping results can be further optimized by combining actual business needs with the capabilities of pre-trained language models.

[0073] Step S3: Use the large language model to assign activity names to each group.

[0074] For the current group, the task description information and corresponding time of all tasks in the group are input into the large language model. Under the guidance of prompt words, the large language model combines the pre-set set of activity names to assign the best activity name to the group.

[0075] The prompts in this step should clearly define the input and the expected output, and reference the set of activity names as a knowledge base to provide a reference for the large language model.

[0076] In this embodiment, the activity name of group 1 is "Approval Operation" and the activity name of group 2 is "Email Communication".

[0077] Step S4: Use a reinforcement learning model to assign a callable tool to each activity name.

[0078] The reinforcement learning model is a Deep Q-Network, which uses RPA to process historical data and trains the reinforcement learning model through experience replay and target network.

[0079] Each task item, its context information, and the activity to which it belongs in the historical data are defined as a structured state. The context information includes one or more preceding and following tasks. The tools used by each task item in the historical data are defined as actions, and the task completion rate and execution efficiency corresponding to the actions of each task item are defined as rewards, thus obtaining a training set. The reinforcement learning model is then trained using the training set to learn the correlation between tools and activities.

[0080] The tools mentioned refer to the connectors available to the RPA system, which need to be prepared in advance, for example:

[0081] - Email processing: Outlook, Sendmail.

[0082] - Data entry / management: SharePoint, Google Forms.

[0083] - Approval process: Microsoft Forms, Planner.

[0084] - Notification services: RSS, Teams.

[0085] When using a trained reinforcement learning model to assign tools, the input to the reinforcement learning model is an activity name and the context corresponding to that activity name. The context includes the preceding and following tasks and the activity names corresponding to the tasks. The output is the tool recommended for that activity.

[0086] In this embodiment, the tool assigned to "Approval Operations" is Microsoft Forms, and the tool assigned to "Email Communication" is Outlook.

[0087] Step S5: Input the task sequence list, the activity name corresponding to each task item, and the tool corresponding to each activity name into the large language model to obtain the process orchestration data, and then convert the process orchestration data into RPA automation scripts.

[0088] In this step, process orchestration data templates and examples (in XML format) are added to the prompts input into the large language model, along with the usage documents for all selected tools, driving the large language model to output process orchestration data in a structured text format.

[0089] The process orchestration data includes a process input interface and several sequentially arranged nodes that correspond one-to-one with each task item. The process input interface receives externally input data during RPA automation script runtime. Each node contains a tool call interface, a node input interface, a node processing function, and a node output interface for the corresponding task item. The node input interface interfaces with the process input interface or a previous node output interface to obtain data and passes the obtained data to the node processing function for processing. The tool call interface passes the result data processed by the node processing function to the called tool to complete the automation operation of the current task item. The node output interface receives the result data output by the node processing function within this node and the data returned by the tool call interface.

[0090] If no data processing is required within a node, the node processing function can directly pass the received data to the tool call interface and the node output interface.

[0091] When converting workflow orchestration data into RPA automation scripts, programmatic conversion should be prioritized. Since the workflow orchestration data is in structured text format, the conversion to RPA automation scripts can be completed through string concatenation and template application.

[0092] Step S6: Test the RPA automation script obtained in step S5, and modify the RPA automation script for any abnormal situations that occur.

[0093] During testing, ensure that:

[0094] - All tasks were triggered and executed correctly.

[0095] - Connectors integrate seamlessly with business systems.

[0096] - The output results are as expected.

[0097] Step S7: Apply the modified RPA automation script to the production environment, continuously monitor the running effect, and optimize the RPA automation script generation process based on the monitoring results.

[0098] During monitoring, the following data is collected: success rate and error rate of automated tasks, the percentage reduction in process automation execution time compared to the original execution time, and user satisfaction feedback. Based on the collected data, the clustering parameters in step S2, the activity name set in step S3, and the historical data in step S4 are updated to retrain the reinforcement learning model and optimize the RPA automation script generation process.

[0099] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A method for automating maritime business processes based on a large language model, characterized in that: The steps include: Step S1: Collect multimodal data while the operator is performing the target process operation, and obtain the task sequence list of the target process based on the multimodal data; In step S1, obtaining the task sequence list of the target process based on multimodal data means: reading each piece of multimodal data in the temporary list, with each piece of multimodal data corresponding to one task item, and sending each piece of multimodal data to a large language model that supports multimodality to obtain the task description information corresponding to each task item; Step S2: Group the task items in the task sequence list; The specific process of step S2 is as follows: Step S2-1: Convert the task description information of each task item into an embedding vector; Step S2-2: Calculate the semantic similarity between each embedding vector; Step S2-3: Group task items by clustering based on the semantic similarity between embedded vectors and preset clustering parameters; Step S3: Use the large language model to assign activity names to each group; Step S4: Use a reinforcement learning model to assign a callable tool to each activity name; Step S5: Input the task sequence list, the activity name corresponding to each task item, and the tool corresponding to each activity name into the large language model to obtain the process orchestration data, and then convert the process orchestration data into RPA automation scripts.

2. The method for automating maritime business processes based on a large language model as described in claim 1, characterized in that, In step S1, the method of collecting multimodal data is as follows: real-time capture of multimodal data during operator operation. Multimodal data includes audio data, screen image data, and UI interaction data. When the operator triggers a recording operation, the current multimodal data is recorded and added to a temporary list.

3. The method for automating maritime business processes based on a large language model as described in claim 1, characterized in that: In step S3, for the current group, the task description information of all task items in the group and the corresponding time are input into the large language model. Under the guidance of prompt words, the large language model combines the pre-set set of activity names to assign the best activity name to the group.

4. The method for automating maritime business processes based on a large language model as described in claim 1, characterized in that: Use RPA processes to manipulate historical data to train reinforcement learning models: Each task item, its context information, and the activity to which it belongs in the historical data are defined as a structured state. The context information includes one or more tasks before and after the target task item. The tools used by each task item in the historical data are defined as actions. The task completion rate and execution efficiency corresponding to the implementation of each task item's action are defined as rewards, thus obtaining a training set. Then, the reinforcement learning model is trained using the training set to learn the relationship between tools and activities. When using a trained reinforcement learning model to assign tools, the input to the reinforcement learning model is an activity name and the context corresponding to that activity name. The context includes the preceding and following tasks and the activity names corresponding to the tasks. The output is the tool recommended for that activity.

5. The method for automating maritime business processes based on a large language model as described in claim 1, characterized in that: In step S5, process orchestration data templates and examples are added to the prompts input into the large language model, along with the usage documents for all selected tools, driving the large language model to output process orchestration data in a structured text format.

6. The method for automating maritime business processes based on a large language model as described in claim 1, characterized in that: The process orchestration data includes process input interfaces and several sequentially arranged nodes that correspond one-to-one with the task items; The process input interface receives externally input data when the RPA automation script is running; Each node contains a tool call interface, a node input interface, a node processing function, and a node output interface for the corresponding task item. The node input interface is used to connect with the process input interface or the previous node output interface to obtain data, and then pass the obtained data to the node processing function for processing. The tool call interface is used to pass the result data processed by the node processing function to the called tool to complete the automated operation of the current task item; The node output interface is used to receive the result data output by the node processing function within this node and the data returned by the tool call interface.

7. The method for automating maritime business processes based on a large language model as described in claim 1, characterized in that, Also includes: Step S6: Test the RPA automation script obtained in step S5, and modify the RPA automation script for any abnormal situations that occur.

8. The method for automating maritime business processes based on a large language model as described in claim 7, characterized in that, Also includes: Step S7: Apply the modified RPA automation script to the production environment, continuously monitor the running effect, and optimize the RPA automation script generation process based on the monitoring results; During monitoring, the following data are collected: success rate and error rate of automated tasks, percentage reduction in process automation execution time relative to the original execution time, and user satisfaction feedback. Based on the collected data, the clustering parameters used for grouping in step S2, the set of activity names used by the large language model in step S3, and the historical data used to train the reinforcement learning model are updated. The reinforcement learning model is then retrained to optimize the generation process of RPA automation scripts.