Browser-based industrial agent graphical development intent recognition method

Abstract

The application is particularly a browser-based industrial intelligent agent graphical development intention recognition method, and relates to the technical field of industrial intelligent agent graphical arrangement, comprising: carrying out denoising and normalization processing on the operation behavior data, organizing effective operation events into a structured behavior sequence in chronological order, extracting local features of the behavior sequence based on a sliding window, and generating a behavior feature vector.In the application, two types of signal sources, operation behavior data and context data, are collected, a behavior feature vector and a context feature vector are extracted respectively, and the two types of feature vectors are fused and input into an intention classification model for score calculation, so that intention judgment can comprehensively integrate information in two dimensions of operation behavior mode and arrangement environment graph structure state, the technical problem of insufficient intention recognition accuracy caused by the existing scheme only according to a single signal source is solved, and the technical effect of obtaining more comprehensive judgment basis in a variable operation mode scene is achieved.

Description

Technical Field

[0001] This invention relates to the field of graphical orchestration technology for industrial intelligent agents, and more particularly to a browser-based method for identifying the graphical development intent of industrial intelligent agents. Background Technology

[0002] In graphical orchestration scenarios for industrial intelligent agent applications, users orchestrate these applications via a browser by dragging and dropping nodes, configuring parameters, and connecting data streams. Accurately identifying the user's operational intent during the orchestration process helps the platform provide targeted assistance prompts and interactive support.

[0003] Existing intent recognition solutions typically employ rule engines or single classification models, relying solely on a single signal source to determine user intent. In graphical orchestration scenarios, user operation patterns are highly variable, and the limited basis for judgment provided by a single signal source results in insufficient intent recognition accuracy.

[0004] In addition, existing solutions mostly use batch processing, which cannot respond to incremental user input in real time and results in high processing latency.

[0005] Meanwhile, existing solutions lack the ability to perceive the orchestration stage, employing the same judgment strategy across different stages, resulting in a lack of specificity in intent judgment. Therefore, there is a need for a graphical orchestration intent recognition method for industrial intelligent agents that can integrate multi-source signals, perceive the orchestration stage, and support real-time response. Summary of the Invention

[0006] The purpose of this invention is to propose a browser-based graphical development intent recognition method for industrial intelligent agents in order to solve the above-mentioned problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: Browser-based graphical development of intent recognition methods for industrial intelligent agents includes: Collect user operation behavior data and context data in the graphical orchestration environment. The operation behavior data is obtained by listening to mouse events, keyboard events and canvas operation events, and the context data is obtained by reading the state information of the orchestration canvas. The operation behavior data is denoised and normalized, and the effective operation events are organized into a structured behavior sequence in chronological order. Local features are extracted from the behavior sequence based on a sliding window to generate a behavior feature vector. State modeling is performed on the context data, and the number of nodes, number of connections, graph topology features and error state markers are extracted from the context state to generate a context feature vector; Based on the operation type distribution in the behavior feature vector and the graph structure features in the context feature vector, the current orchestration stage is determined and a stage label is generated. The encoded vector of the stage label is concatenated with the behavior feature vector and the context feature vector to form a stage-aware fusion feature vector. This vector is then input into the intent classification model to calculate the confidence score of each intent category. The K intents ranked by confidence score are then taken as the candidate intent list and output, where K is a preset parameter.

[0008] Preferably, organizing valid operational events into a structured sequence of behaviors in chronological order includes: Each valid operation event is transformed into a standardized action representation. Each action representation is a quadruple containing the operation type, the operation target node identifier, the operation timestamp, and the operation parameters. Arrange all actions in time-stamp order to generate the structured behavior sequence; The operation type is a discrete category field, which is converted into a numerical vector using one-hot encoding; the operation timestamp is processed using Z-score normalization.

[0009] Preferably, the local feature extraction of the behavior sequence based on the sliding window includes: Set the size of the sliding window to the number of actions W contained within the window; Each time a new action is received, the sliding window moves forward one position, discarding the oldest action in the sequence, while ensuring that the window always contains the most recent W actions; Within the sliding window, the operation frequency, the number of consecutive undoes, and the hover duration are statistically analyzed to generate the behavior feature vector; Specifically, when the total number of actions in the behavior sequence is less than W, feature extraction is performed using all current actions as the window range.

[0010] Preferably, the state modeling of the context data includes: The context state is defined as a triple containing the graph structure, the running state, and the historical operation record, wherein the graph structure is a directed graph consisting of a set of nodes and a set of connection relationships. The node in-degree distribution, node out-degree distribution, number of connected components, and longest path length are extracted from the directed graph as the graph topological features. The error status marker is a categorical field, which is converted into a numerical vector using one-hot encoding; the number of nodes, the number of connections, the number of connected components, and the longest path length are normalized using Z-score.

[0011] Preferably, determining the current arrangement stage includes: When the number of nodes is growing rapidly, the number of connections is lower than the preset ratio of the number of nodes, and the operation type is mainly drag and drop, the current stage is determined to be the building stage. When the frequency of connection operations exceeds a preset frequency threshold and the connection relationships between nodes gradually increase, the current stage is determined to be the connection stage. When the operation is concentrated on node attribute editing, and the frequency of drag and drop operations and connection operations is lower than the preset frequency threshold, the current stage is determined to be the parameter configuration stage. When the number of consecutive undoes exceeds a preset threshold, or when the error status flag is active, the current stage is determined to be the debugging and optimization stage. The stage label is a discrete category field, which is converted into the encoding vector using one-hot encoding.

[0012] Preferably, the intent classification model uses corresponding judgment weights for different stage labels: During the setup phase, the judgment weights focus on intent categories related to node type selection and layout orchestration; During the connection phase, the determination weight focuses on the intent category related to data flow association and inter-node relationships; During the parameter configuration phase, the determination weight focuses on the intent category related to parameter adjustment and attribute configuration; During the debugging and optimization phase, the judgment weights focus on intent categories related to error localization and process optimization.

[0013] Preferably, the intent classification model is implemented using a weighted scoring method, and the calculation of the confidence score for each intent category includes: For each intent category, the weight coefficients of each dimension corresponding to the intent category are multiplied one by one by the corresponding dimension components of the stage perception fusion feature vector and then summed to obtain the weighted score of the intent category. The candidate intent list is output by taking the top K intent categories from all intent categories and sorting them by weighted score.

[0014] Preferably, when semantic signal data input by the user in the search box or command input box is collected, text feature extraction is performed on the semantic signal data to generate a semantic feature vector; The behavioral feature vector, the context feature vector, and the semantic feature vector are concatenated into a three-source fusion feature vector, which is then input into the intent classification model to calculate the confidence score for each intent category. The candidate intent list and its corresponding confidence score are then output.

[0015] Preferably, the method further includes a streaming intent recognition step: Upon receiving a new operation event, it is converted into a standardized action representation and appended to the end of the behavior sequence. The sliding window moves forward one position, and statistical features are re-extracted based on the latest action sequence within the window to update the behavior feature vector. Based on the type and target of the new operation event, the context state is incrementally updated. When the new operation event is node addition, the corresponding node is added to the node set. When the new operation event is connection creation, the corresponding connection is added to the connection relationship set. When the new operation event is node deletion, the corresponding node is removed from the node set and the associated connection is removed simultaneously. When the new operation event is parameter modification, the attribute configuration of the corresponding node is updated. The context feature vector is regenerated based on the updated context state. The updated behavior feature vector and context feature vector are concatenated and input into the intent classification model to recalculate the intent score and output an updated list of candidate intents. Specifically, a minimum update interval is set. When the timestamp difference between two adjacent operation events is less than the minimum update interval, the new operation event is only appended to the behavior sequence. The context state incremental update and intent score recalculation process are not triggered temporarily. When the next operation event that meets the interval condition arrives, batch updates and intent recalculation are performed based on all the accumulated new operation events.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention collects two types of signal sources: operational behavior data and contextual data. It extracts behavioral feature vectors and contextual feature vectors respectively, and then merges the two types of feature vectors and inputs them into an intent classification model for scoring calculation. This enables intent judgment to comprehensively consider information from two dimensions: operational behavior patterns and the structure and state of the orchestration environment. This solves the technical problem of insufficient intent recognition accuracy caused by existing solutions relying on only a single signal source, and achieves the technical effect of obtaining more comprehensive judgment criteria in scenarios with changing operational patterns.

[0017] 2. This invention divides the orchestration process into four stages—building, connecting, parameter configuration, and debugging / optimization—based on operation type distribution and graph structure characteristics. It employs corresponding judgment weights for each stage, addressing the technical problem of existing solutions lacking orchestration stage awareness, leading to a lack of targeted intent judgment. This achieves the technical effect of intent scoring calculation adapting to the differences in user behavior across different stages. Furthermore, by incrementally updating the behavior sequence and context state and calculating the intent score on a rolling basis, the intent recognition result is updated each time a new operation event is received. This solves the technical problem of high processing latency in batch processing methods, achieving the technical effect of real-time response in streaming intent recognition. Attached Figure Description

[0018] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0020] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0021] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0022] In this embodiment, it includes: An industrial automation company is using a browser-based graphical orchestration platform for industrial agents to build a quality inspection agent application for a battery production line. Operators need to orchestrate visual inspection nodes, data preprocessing nodes, anomaly detection nodes, and alarm push nodes into a complete quality inspection data stream. The platform's front-end browser layer monitors the operator's mouse, keyboard, and canvas operation events in real time and continuously reads the graph structure state of the orchestration canvas.

[0023] Step S100: Acquire multi-source signal data The operator drags a visual inspection node (node ​​001), a data preprocessing node (node ​​002), an anomaly detection node (node ​​003), and an alarm push node (node ​​004) sequentially onto the arrangement canvas. Then, the operator attempts to create a connection between node 001 and node 002, and enters a threshold of 0.85 in the parameter input box of node 003. The front-end browser layer collects the following raw operation behavior event stream and context data, as shown in Table 1 below.

[0024] Table 1. Raw Operation Event Flow Data Collection E001 Mouse movement Blank area of ​​the canvas March 11, 20XX, 09:14:02.103 Coordinates (312, 178) E002 Node drag and drop Node 001 March 11, 20XX, 09:14:05.441 Landing point coordinates (320, 180) E003 Mouse movement Blank area of ​​the canvas March 11, 20XX, 09:14:06.012 Coordinates (315, 181) E004 Node drag and drop Node 002 March 11, 20XX, 09:14:09.887 Landing point coordinates (480, 180) E005 Node drag and drop Node 003 March 11, 20XX, 09:14:14.223 Landing point coordinates (640, 180) E006 Node drag and drop Node 004 March 11, 20XX, 09:14:18.556 Landing point coordinates (800, 180) E007 Connection creation Node 001 → Node 002 March 11, 20XX, 09:14:23.114 none E008 Text input Node 003 parameter box March 11, 20XX, 09:14:31.667 The threshold is set to 0.85. Meanwhile, the context data read by the front end is: there are 4 nodes and 1 connection on the canvas, and node 003 has a warning state that the parameters are not fully configured.

[0025] Step S201: Denoising and Normalization Processing The original event stream is denoised to filter out mouse movement events (E001, E003) that have no actual orchestration significance, retain valid operation events, and generate a set of valid operation events.

[0026] Table 2. Set of effective operational events after denoising V001 Node drag and drop Node 001 March 11, 20XX, 09:14:05.441 Landing point coordinates (320, 180) V002 Node drag and drop Node 002 March 11, 20XX, 09:14:09.887 Landing point coordinates (480, 180) V003 Node drag and drop Node 003 March 11, 20XX, 09:14:14.223 Landing point coordinates (640, 180) V004 Node drag and drop Node 004 March 11, 20XX, 09:14:18.556 Landing point coordinates (800, 180) V005 Connection creation Node 001 → Node 002 March 11, 20XX, 09:14:23.114 none V006 Text input Node 003 parameter box March 11, 20XX, 09:14:31.667 The threshold is set to 0.85. Of the original 8 events, E001 and E003 were invalid mouse movement events. After filtering, 6 valid events were retained, with a noise filtering rate of 25%.

[0027] Step S202: Generate a structured behavior sequence Each event in the set of valid action events is transformed into a standardized four-tuple action representation, arranged by timestamp, to generate a behavior sequence. ,common One action.

[0028] right The fields use one-hot encoding, and the operation type set includes 5 categories: {node drag and drop, connection creation, text input, parameter modification, and undo / redo}. The node drag and drop encoding is as follows: The connection creation code is The text input code is .

[0029] right The fields are Z-score normalized. The original timestamps of the 6 actions are converted to relative seconds as follows: 5.441, 9.887, 14.223, 18.556, 23.114, and 31.667, in seconds (based on 09:14:00.000).

[0030] mean and standard deviation The calculation is as follows: ; ; With action Taking timestamp standardization as an example, the second is: ; in Second.

[0031] Step S203: Extract behavioral feature vectors Configure sliding window size Total number of actions currently performed. The window covers the 5 most recent actions. Within this window, statistical characteristics such as operation frequency, number of consecutive undoes, and hover duration are collected to generate a behavioral feature vector. .

[0032] In this embodiment, the preset frequency threshold is exemplarily set to 0.3 (i.e., when the proportion of this type of operation in the window exceeds 30%, it is considered to exceed the threshold), and the preset number threshold is exemplarily set to 3 times. Those skilled in the art can adjust the above thresholds according to actual application scenarios and operating habits, and the above specific values ​​do not constitute a limitation on the present invention; Table 3. Statistical characteristics of behavior within the sliding window. Node dragging frequency Number of drag events within the window / Total number of window actions 3 / 5=0.600 Connection creation frequency Number of connection events within the window / Total number of window actions 1 / 5=0.200 Text input frequency Number of text input events in the window / Total number of window actions 1 / 5=0.200 Number of consecutive cancellations Continuous undo event count within the window 0 Average operating interval Mean time difference between adjacent operations within the window 5.445 seconds Step S300: Context state modeling and feature extraction Based on the collected context data, construct the context state. .in It contains 4 nodes and 1 directed connection (node ​​001 → node 002), forming a directed graph structure. From The extracted graph topology features are as follows: Node 001 has an out-degree of 1 and an in-degree of 0; Node 002 has an out-degree of 0 and an in-degree of 1; Nodes 003 and 004 are both isolated nodes with an out-degree and an in-degree of 0. The number of connected components is 3 (Nodes 001 and 002 form one connected component, while Nodes 003 and 004 are independent), and the longest path length is 1. The error status is marked as active (node ​​003 has a parameter warning).

[0033] One-hot encoding is used for error status marking: the error status activation encoding is... Error state not activated code is , currently take Numerical features such as the number of nodes, the number of connections, the number of connected components, and the longest path length are Z-score standardized to generate context feature vectors. .

[0034] Table 4 Summary of Context State Features Number of nodes 4 Z-score standardization Number of connections 1 Z-score standardization Number of connected components 3 Z-score standardization Longest path length 1 Z-score standardization Average in-degree of nodes 0.250 Z-score standardization Average out-degree of nodes 0.250 Z-score standardization Error status flags activation One-hot encoding [1,0] Step S350: Arrangement Stage Judgment based on and The characteristics in the text are used to determine the current arrangement stage.

[0035] The current number of nodes is 4, which is rapidly increasing; the number of connections is 1, which is lower than the preset ratio for the number of nodes (the preset ratio is 0.5, meaning the number of connections should not be less than...). (Currently there is only 1 item, which is below the threshold); the node dragging frequency within the window is 0.600, which is the most frequent operation type. All three conditions meet the judgment rules for the construction phase, therefore the generated phase label is the construction phase label.

[0036] It should be noted that although the error status flag is active, the number of consecutive undoes is 0, which does not meet the judgment criteria for the debugging and optimization phase. The judgment criteria for the setup phase take precedence.

[0037] The setup stage tags use one-hot encoding. The stage category set is {Setup Stage, Connection Stage, Parameter Configuration Stage, Debugging and Optimization Stage}. The setup stage is coded as follows: .

[0038] Step S400: Feature fusion and intent score calculation One-hot encoded vector of stage label behavioral feature vector and context feature vectors Concatenate into a stage-aware fusion feature vector Input intent classification model.

[0039] Currently in the development phase, the intent classification model's judgment weights focus on intent categories related to node type selection and layout orchestration. Taking a weighted scoring method as an example, the following settings are used: Calculate the weighted score for each intent category: ; in for Total number of dimensions For the next stage of construction The intent category corresponds to the first Weight coefficients of dimensional features.

[0040] Table 5. Candidate Intent List (Top-3) 1 Add data processing node 0.873 High frequency of drag-and-drop, few connections, and dominant weight during the setup phase. 2 Complete the connections between nodes 0.761 There are isolated nodes 003 and 004, and the number of connected components is 3. 3 Adjust node layout position 0.634 Drag and drop is the primary operation type. Meanwhile, since the threshold for semantic signal data collected in step S100 is set to 0.85, feature extraction is performed on the text to generate a semantic feature vector. to , After being concatenated into a three-source fusion feature vector and recalculated, the confidence of the threshold values ​​for intent configuration node parameters is improved, and the intent is added to the candidate intent list.

[0041] Step S500: Incremental Update of Streaming Intent Recognition After completing the above operations, the operator continues to create the connection between node 002 and node 003, triggering a new operation event E009.

[0042] Step S501 converts E009 (connection creation, node 002→node 003, timestamp 09:14:38.221) into a standardized action. Appended to the end of the behavior sequence, the sequence is updated. , The sliding window moves forward, covering the previous window. Re-extract statistical features and update The connection creation frequency increased from 0.200 to 0.400, while the node dragging frequency decreased from 0.600 to 0.400.

[0043] Step S502: The new operation event type is connection creation, in the connection relationship set. A connection is added between node 002 and node 003. After the context state is incrementally updated, the number of connections increases from 1 to 2, the number of connected components decreases from 3 to 2 (nodes 001, 002, and 003 are now connected), and the longest path length increases from 1 to 2. Regenerate. .

[0044] Step S503 is based on the updated and Reassessment of the orchestration phase: The frequency of connection operations increases to 0.400, the number of connections between nodes gradually increases, the conditions for determining the connection phase are met, and the phase label is updated to the connection phase label. Recalculate the intent score and output the updated list of candidate intents.

[0045] Table 6. Candidate Intent List (Top-3) after Streaming Update 1 Complete the connections between nodes 0.891 The ranking rose from 2nd to 1st. 2 Configure connection data mapping 0.802 Add to candidate list 3 Add data processing node 0.598 The ranking dropped from 1st to 3rd. In addition, the minimum update interval in step S500 Set to 500 milliseconds. If the operator triggers the operation event again within 300 milliseconds after 09:14:38.221 (e.g., 09:14:38.521), the event will only be appended to the behavior sequence and the recalculation process of steps S502 and S503 will not be triggered for the time being. Batch updates will be performed when the next event that meets the interval condition arrives.

[0046] The logical chain of the entire data flow is as follows: The 8 raw events collected in step S100 are filtered into 6 valid events in step S201. In step S202, they are transformed into a structured behavior sequence and encoded and standardized. In step S203, statistical features such as drag frequency and connection frequency are extracted within the sliding window to generate... Step S300 extracts features such as the number of nodes, the number of connected components, and error states from the canvas graph structure to generate... Step S350 combines the two types of feature vectors to determine that the current stage is being built, and generates stage labels with one-hot encoding; Step S400 concatenates the three types of vectors into The input intent classification model is then fed back, and the output is a candidate intent list with the addition of data processing nodes as the top priority. In step S500, after a new connection event arrives, the feature vector is incrementally updated, the stage label is switched to the connection stage, the candidate intent list is updated accordingly, and the connection between the completed nodes is moved to the top priority. The output data of each step is used as the direct input of the next step, ensuring the continuity and logical consistency of the data flow.

[0047] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0048] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0049] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A browser-based graphical development intent recognition method for industrial intelligent agents, characterized in that, include: Collect user operation behavior data and context data in the graphical orchestration environment. The operation behavior data is obtained by listening to mouse events, keyboard events and canvas operation events, and the context data is obtained by reading the state information of the orchestration canvas. The operation behavior data is denoised and normalized, and the effective operation events are organized into a structured behavior sequence in chronological order. Local features are extracted from the behavior sequence based on a sliding window to generate a behavior feature vector. State modeling is performed on the context data, and the number of nodes, number of connections, graph topology features and error state markers are extracted from the context state to generate a context feature vector; Based on the operation type distribution in the behavior feature vector and the graph structure features in the context feature vector, the current orchestration stage is determined and a stage label is generated. The encoded vector of the stage label is concatenated with the behavior feature vector and the context feature vector to form a stage-aware fusion feature vector. This vector is then input into the intent classification model to calculate the confidence score of each intent category. The K intents ranked by confidence score are then taken as the candidate intent list and output, where K is a preset parameter.

2. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, Organizing effective operational events into a structured sequence of behaviors in chronological order includes: Each valid operation event is transformed into a standardized action representation. Each action representation is a quadruple containing the operation type, the operation target node identifier, the operation timestamp, and the operation parameters. Arrange all actions in time-stamp order to generate the structured behavior sequence; The operation type is a discrete category field, which is converted into a numerical vector using one-hot encoding; the operation timestamp is processed using Z-score normalization.

3. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, Local feature extraction of the behavioral sequence based on a sliding window includes: Set the size of the sliding window to the number of actions W contained within the window; Each time a new action is received, the sliding window moves forward one position, discarding the oldest action in the sequence, while ensuring that the window always contains the most recent W actions; Within the sliding window, the operation frequency, the number of consecutive undoes, and the hover duration are statistically analyzed to generate the behavior feature vector; Specifically, when the total number of actions in the behavior sequence is less than W, feature extraction is performed using all current actions as the window range.

4. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, State modeling of the context data includes: The context state is defined as a triple containing the graph structure, the running state, and the historical operation record, wherein the graph structure is a directed graph consisting of a set of nodes and a set of connection relationships. The node in-degree distribution, node out-degree distribution, number of connected components, and longest path length are extracted from the directed graph as the graph topological features. The error status marker is a categorical field, which is converted into a numerical vector using one-hot encoding; the number of nodes, the number of connections, the number of connected components, and the longest path length are normalized using Z-score.

5. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, Determining the current arrangement stage includes: When the number of nodes is growing rapidly, the number of connections is lower than the preset ratio of the number of nodes, and the operation type is mainly drag and drop, the current stage is determined to be the building stage. When the frequency of connection operations exceeds a preset frequency threshold and the connection relationships between nodes gradually increase, the current stage is determined to be the connection stage. When the operation is concentrated on node attribute editing, and the frequency of drag and drop operations and connection operations is lower than the preset frequency threshold, the current stage is determined to be the parameter configuration stage. The operation is concentrated on node attribute editing, which means that within the sliding window, the proportion of attribute editing operations (including text input and parameter modification) to the total number of operations in the window exceeds a first proportion threshold; the first proportion threshold is greater than the sum of the frequency of drag operations and the frequency of connection operations. When the number of consecutive undoes exceeds a preset threshold, or when the error status flag is active, the current stage is determined to be the debugging and optimization stage. The stage label is a discrete category field, which is converted into the encoding vector using one-hot encoding.

6. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, The intent classification model uses corresponding decision weights for labels at different stages: During the setup phase, the judgment weights focus on intent categories related to node type selection and layout orchestration; During the connection phase, the determination weight focuses on the intent category related to data flow association and inter-node relationships; During the parameter configuration phase, the determination weight focuses on the intent category related to parameter adjustment and attribute configuration; During the debugging and optimization phase, the judgment weights focus on intent categories related to error localization and process optimization.

7. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, The intent classification model uses a weighted scoring method, and the calculation of the confidence score for each intent category includes: For each intent category, the weight coefficients of each dimension corresponding to the intent category are multiplied one by one by the corresponding dimension components of the stage perception fusion feature vector and then summed to obtain the weighted score of the intent category. The candidate intent list is output by taking the top K intent categories from all intent categories and sorting them by weighted score.

8. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, When semantic signal data is collected from the user's input in the search box or command input box, text feature extraction is performed on the semantic signal data to generate a semantic feature vector; The behavioral feature vector, the context feature vector, and the semantic feature vector are concatenated into a three-source fusion feature vector, which is then input into the intent classification model to calculate the confidence score for each intent category. The candidate intent list and its corresponding confidence score are then output.

9. The browser-based graphical development intent recognition method for industrial intelligent agents according to claim 1, characterized in that, It also includes a streaming intent recognition step: Upon receiving a new operation event, it is converted into a standardized action representation and appended to the end of the behavior sequence. The sliding window moves forward one position, and statistical features are re-extracted based on the latest action sequence within the window to update the behavior feature vector. Based on the type and target of the new operation event, the context state is incrementally updated. When the new operation event is node addition, the corresponding node is added to the node set. When the new operation event is connection creation, the corresponding connection is added to the connection relationship set. When the new operation event is node deletion, the corresponding node is removed from the node set and the associated connection is removed simultaneously. When the new operation event is parameter modification, the attribute configuration of the corresponding node is updated. The context feature vector is regenerated based on the updated context state. The updated behavior feature vector and context feature vector are concatenated and input into the intent classification model to recalculate the intent score and output an updated list of candidate intents. Specifically, a minimum update interval is set. When the timestamp difference between two adjacent operation events is less than the minimum update interval, the new operation event is only appended to the behavior sequence. The context state incremental update and intent score recalculation process are not triggered temporarily. When the next operation event that meets the interval condition arrives, batch updates and intent recalculation are performed based on all the accumulated new operation events.

Images