A business data processing method and system based on visual lane modeling

Abstract

The application relates to the technical field of visual lane modeling, and discloses a business data processing method and system based on visual lane modeling, wherein a multi-task prediction model of a graph neural network and a Transformer fusion is trained based on multi-source historical business chain data, a plurality of groups of lane layout and event configuration schemes are generated in combination with real-time operations of a user captured by a design interface, the optimal scheme is selected through simulation evaluation and filtering, and visual display is performed, and finally, visual lane modeling is executed according to user feedback, through automatic data processing, high-precision model prediction, real-time interactive optimization and visual modeling, the design efficiency and quality of a business chain are improved, meanwhile, model prediction based on historical data avoids artificial experience deviation, meets the comprehensive requirements of real-time performance, reliability and resource efficiency for complex equipment development, and provides efficient technical support for the landing of model-based system engineering in the field of complex equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual lane modeling, and particularly relates to a business data processing method and system based on visual lane modeling. BACKGROUND

[0002] In the field of complex equipment development, model-based system engineering (MBSE) has become the mainstream paradigm of business chain design. The core requirement is to achieve the balance of multi-dimensional performance indicators (delay, resource, reliability) and the rapid adaptation of dynamic business requirements. The current mainstream business chain design tools still rely on the traditional mode of manual dragging and configuration (adopting a manual way to construct and drag lanes and operation units in the lanes, and adopting a manual way to configure the event types between the operation units to realize the construction of a visual business carrier canvas for different service requirements). Although this can meet the basic visualization requirements, it has significant limitations in large-scale and high-complexity business scenarios.

[0003] The existing business chain design technology mainly falls into two categories: one is a static visualization-based tool (such as IBM Rhapsody), which realizes business logic expression through UML sequence diagrams and flowcharts. However, this type of tool can only describe simple call relationships and cannot distinguish event types (such as periodic events / conditional events), and the lane position is fixed, making it difficult to dynamically reconstruct the flow. The other is a customized design method based on human experience, which relies on the understanding of the design personnel to configure the lane layout and event relationship. However, this method has three major core defects: first, the design decision is highly dependent on experience and lacks historical data support, resulting in poor consistency of schemes produced by different design personnel; second, when the complexity of the business chain increases (such as the number of lanes > 10 and the number of event relationships > 20), the efficiency of manual adjustment decreases exponentially, and it is difficult to ensure global optimality; third, it has no self-adaptive learning ability and cannot establish a prediction model based on historical design data and performance logs. The design of the lane directly relates to the performance indicators of the underlying system, such as response time, resource utilization, etc. If problems are only discovered after the design is completed through testing, the cost of later iteration is relatively high.

[0004] Although some improvement schemes currently attempt to introduce automatic configuration functions, they still cannot solve the problems of insufficient feature compactness, high real-time interaction delay, and poor multi-scenario adaptability. For example, existing schemes cannot capture user design operation features in real time and quickly generate optimization suggestions, and lack an online model updating mechanism, resulting in the need for manual re-adjustment of rules in different scenarios such as avionics systems (strong real-time) and industrial monitoring (high reliability), with a long adaptation period. Therefore, there is an urgent need in the industry for a business data processing scheme based on visual lane modeling that can ensure low delay and low memory occupation while improving the automation and accuracy of business chain design. SUMMARY

[0005] The application provides a business data processing method and system based on visual lane modeling, aiming to solve at least one of the above technical problems.

[0006] To achieve the above-mentioned purpose, the application provides a business data processing method based on visual lane modeling, which comprises the following steps:

[0007] S1: Collecting historical business chain design data from multiple source systems, performing standardization processing and time sequence alignment, and constructing a training sample set;

[0008] S2: Based on the fusion architecture of graph neural network and Transformer, a multi-task prediction model is constructed, the model parameters are trained by combining the loss function and the constructed training sample set, and a multi-task prediction model with lane layout prediction and event configuration capability is obtained;

[0009] S3: Capturing user real-time operations through the event listening module of the design interface, extracting service metadata and canvas state features in the user real-time operations, and generating a standardized input vector;

[0010] S4: Inputting the real-time standardized input vector into the trained multi-task prediction model to generate multiple lane layout and event configuration optimization schemes, and selecting the optimal scheme group through simulation evaluation;

[0011] S5: Superimposing the optimal scheme group in a visual form on the design canvas, recording user feedback data, and performing visual lane modeling according to the user feedback data.

[0012] Optionally, step S1: Collecting historical business chain design data from multiple source systems, performing standardization processing and time sequence alignment, and constructing a training sample set, specifically comprising:

[0013] S11: Parsing the version history of the business chain design file from the Git version control system, extracting lane layout data L, and organizing it into CSV format data according to version number and timestamp; wherein the lane layout data includes lane ID, service name, service type, center coordinate, width and hierarchical relationship;

[0014] S12: Parsing the event definition section in the design file, constructing an event graph, taking service operation as node and event relationship as directed edge, generating event graph data G=(V,E), and converting the event graph data G into adjacency matrix A and edge feature matrix E_feat;

[0015] S13: Collect performance indicator data P from the system test log database; wherein the performance indicator data P includes event response time, packet loss rate, CPU occupancy rate and memory usage;

[0016] S14: Time sequence alignment of lane layout data L, event graph data G and performance indicator data P based on the timestamp of the design version, forming a single sample, and using Z-score to standardize the single sample to obtain a training sample set.

[0017] Optionally, in step S2, a multi-task prediction model is constructed based on a graph neural network and a Transformer fusion architecture, specifically including:

[0018] S21: Construct a double-branch neural network architecture including a local prediction branch and an event configuration branch;

[0019] The local prediction branch uses a graph attention network, and the input layer receives the lane feature matrix L_feat extracted from the lane layout data L and the spatial adjacency matrix S. The lane spatial dependency is learned through 2 layers of graph attention GAT, the number of attention heads of the graph attention GAT is 4, the dimension of the hidden layer is 64, and the expression of the attention weight is: ;

[0020] In the formula, represents the initial feature vector of the lane u, represents a linear transformation matrix, represents an attention weight vector, represents the neighborhood set of the lane u, represents a vector concatenation operation, represents the initial feature vector of the lane v, represents the neighborhood set of the lane u any one of the lanes, represents the initial feature vector of the lane ;

[0021] The event configuration branch uses a Transformer decoder, and the input layer receives the node feature matrix V_feat and the edge feature matrix E_feat extracted from the event graph data G. The event association pattern is learned through 3 layers of encoder, and the output layer outputs the event type probability distribution through the Softmax activation function.

[0022] Optionally, in step S2, the model parameters are trained by combining the loss function and the constructed training sample set, and a multi-task prediction model with lane layout prediction and event configuration capabilities is obtained, specifically including:

[0023] S22: Define a combined loss function, and the expression is:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] wherein, denotes the total loss value, denotes the loss weight, denotes the layout prediction loss, denotes the event classification loss, denotes the learning loss, denotes the number of lanes, denotes the real coordinates of lane i, denotes the predicted coordinates of lane i, denotes the number of events, denotes the real type label of event j, denotes the predicted probability that event j belongs to type k, denotes the batch size, denotes the feature representation of event b, denotes the positive sample of event b, denotes the negative sample, denotes the cosine similarity, denotes the temperature parameter;

[0029] S23: using the Adam optimizer, training the model under the conditions of preset learning rate, weight decay, batch size and training rounds, and obtaining a multi-task prediction model with lane layout prediction and event configuration capabilities after training is completed.

[0030] Optionally, step S3: capturing user real-time operations through the event listening module of the design interface, extracting service metadata and canvas state features in the user real-time operations, and generating a standardized input vector, specifically including:

[0031] S31: concentrating event listeners in the design interface to capture three types of operation events; wherein the operation events include mouse drag events, service placement events or connection operation events;

[0032] S32: calling a service registry through a backend API to extract metadata of the current operation service and convert it into a service meta feature vector; wherein the metadata includes service ID, service type, maximum processing delay, CPU resource demand, memory demand, and interface protocol;

[0033] S33: Traverse the current canvas lane, extract spatial features and density features, concatenate spatial features, density features and service element feature vectors to generate a canvas state feature vector as a standardized input vector.

[0034] Optionally, step S4: input the real-time standardized input vector into the trained multi-task prediction model to generate multiple groups of lane layout and event configuration optimization schemes, and filter the optimal scheme group through simulation evaluation, specifically including:

[0035] S41: input the standardized input vector into the layout prediction branch, output the 3 groups of lane position adjustment vectors with the highest confidence obtained by the layout prediction branch as the layout scheme, and input the standardized input vector into the event configuration branch, output the 3 groups of configuration schemes with the highest probability obtained by the event configuration branch;

[0036] S42: cross-combine the layout scheme and the configuration scheme to form 9 groups of initial candidate schemes, simulate running in the SimPy simulation environment, output several core performance indicators of each scheme, calculate the performance score using the weighted sum of the several core performance indicators, and calculate the design specification score using the sum of the mapping scores of the several design specification standards and the scheme configuration parameters of each scheme in the design constraint rule library;

[0037] S43: for the 9 groups of initial candidate schemes, calculate the comprehensive performance score according to the performance score and the design specification score, and select the 5 groups of schemes with the highest score as the optimal scheme group; wherein, the expression of the comprehensive performance score is:

[0038] ；

[0039] In the formula, S represents the comprehensive performance score, represents the performance score, represents the design specification score.

[0040] Optionally, S5: superimpose the optimal scheme group on the design canvas in a visual form, record user feedback data, and perform visual lane modeling according to the user feedback data, specifically including:

[0041] S51: superimpose each scheme in the optimal scheme group on the current design scheme in the order of the comprehensive performance score from high to low on the canvas with a preset transparency outline;

[0042] S52: label the difference between the current design scheme and the optimal scheme with a target color, and generate a detail display panel on one side of the canvas for displaying the performance score and the design specification score of the current optimal scheme;

[0043] S53: capture and record user feedback, and select the optimal scheme according to the operation instruction corresponding to the user feedback data to execute the visual swim lane modeling; wherein the operation instruction includes scheme application instruction, scheme switching instruction or scheme modification instruction corresponding to keyboard shortcut feedback data or page menu operation feedback data.

[0044] In addition, in order to achieve the above-mentioned purpose, the application also provides a business data processing system based on visual swim lane modeling, comprising:

[0045] The construction module is used for collecting historical business chain design data from multiple source systems, performing standardization processing and time sequence alignment, and constructing a training sample set.

[0046] The training module is used for constructing a multi-task prediction model based on a graph neural network and a Transformer fusion architecture, training model parameters through a combined loss function and the constructed training sample set, and obtaining a multi-task prediction model with swim lane layout prediction and event configuration capabilities.

[0047] The generation module is used for capturing user real-time operations through an event listening module of a design interface, extracting service metadata and canvas state features in the user real-time operations, and generating a standardized input vector.

[0048] The screening module is used for inputting the real-time standardized input vector into the trained multi-task prediction model, generating multiple groups of swim lane layout and event configuration optimization schemes, and screening the optimal scheme group through simulation evaluation.

[0049] The execution module is used for superimposing and displaying the optimal scheme group in a visual form on a design canvas, recording user feedback data, and executing visual swim lane modeling according to the user feedback data.

[0050] The application has the beneficial effects that a business data processing method and system based on visual swim lane modeling are proposed, a multi-task prediction model of a graph neural network and a Transformer fusion is trained based on multi-source historical business chain data, multiple groups of swim lane layout and event configuration schemes are generated combined with user real-time operations captured by a design interface, the optimal scheme is screened through simulation evaluation and visualized display, and finally visual swim lane modeling is executed according to user feedback, through automatic data processing, high-precision model prediction, real-time interactive optimization and visual modeling, the efficiency and quality of business chain design are improved, at the same time, model prediction based on historical data avoids artificial experience deviation, meets the comprehensive needs of real-time performance, reliability and resource efficiency for complex equipment development, and provides efficient technical support for the landing of model-based system engineering (MBSE) in the field of complex equipment. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1A flowchart of a business data processing method based on visual lane modeling of the present application. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0053] The embodiment of the present application provides a business data processing method based on visual lane modeling, referring to Figure 1 , Figure 1 A flowchart of a business data processing method based on visual lane modeling of the present application.

[0054] In the embodiment, a business data processing method based on visual lane modeling comprises the following steps:

[0055] S1: Collecting historical business chain design data from multiple source systems, performing standardized processing and time sequence alignment, and constructing a training sample set;

[0056] S2: Constructing a multi-task prediction model based on a graph neural network and a Transformer fusion architecture, training model parameters through a combined loss function and the constructed training sample set, and obtaining a multi-task prediction model with lane layout prediction and event configuration capabilities;

[0057] S3: Capturing user real-time operations through an event listening module of a design interface, extracting service metadata and canvas state features in the user real-time operations, and generating a standardized input vector;

[0058] S4: Inputting the real-time standardized input vector into the trained multi-task prediction model, generating multiple sets of lane layout and event configuration optimization schemes, and screening the optimal scheme set through simulation evaluation;

[0059] S5: Superimposing the optimal scheme set in a visual form on the design canvas, recording user feedback data, and performing visual lane modeling according to the user feedback data.

[0060] It should be noted that the existing service chain design technology mainly falls into two categories: one is a static visualization-based tool (such as IBM Rhapsody), which realizes business logic expression through UML timing diagram and flowchart, but such tool can only describe simple call relationship and cannot distinguish event types (such as periodic event / conditional event), and the lane position is fixed, making it difficult to dynamically reconstruct the flow. The other is a customized design method based on human experience, which relies on the understanding of the designer to configure the lane layout and event relationship, but this method has three major defects: first, the design decision highly depends on experience and lacks historical data support, resulting in poor consistency of the schemes produced by different designers; second, when the complexity of the service chain rises (such as the number of lanes > 10, the event relationship > 20), the efficiency of manual adjustment decreases exponentially, and it is difficult to ensure global optimality; third, there is no adaptive learning ability, and it is impossible to establish a prediction model based on historical design data and performance logs, while the lane design is directly related to the performance indicators of the underlying system, such as response time and resource utilization, through determining the data transmission path, resource scheduling logic and event interaction complexity. If problems are only discovered through testing after the design is completed, the cost of later iteration accounts for a high proportion of the total design cost.

[0061] Although some improvement schemes currently attempt to introduce automatic configuration functions, they still cannot solve the problems of insufficient feature compactness, high real-time interaction delay and poor multi-scene adaptability. For example, existing schemes cannot capture user design operation features in real time and quickly generate optimization suggestions, and lack an online model updating mechanism, resulting in the need for manual readjustment of rules in different scenarios such as avionics systems (strong real-time) and industrial monitoring (high reliability), with a long adaptation period. Therefore, there is an urgent need in the industry for a business data processing scheme based on visual lane modeling that can improve the automation and accuracy of service chain design while ensuring low delay and low memory occupation.

[0062] To solve the above problems, the present embodiment trains a multi-task prediction model based on multi-source historical service chain data, combines the user's real-time operation captured by the design interface to generate multiple lane layout and event configuration schemes, visualizes the optimal scheme after simulated evaluation, and finally performs visual lane modeling according to user feedback. Through automatic data processing, high-precision model prediction, real-time interaction optimization and visual modeling, the efficiency and quality of service chain design are improved, and the model prediction based on historical data avoids the bias of human experience, meeting the comprehensive needs of real-time, reliability and resource efficiency in complex equipment development, and providing efficient technical support for the landing of model-based system engineering (MBSE) in the field of complex equipment.

[0063] In a preferred embodiment, step S1: collecting historical service chain design data from multi-source systems, standardizing and time aligning, constructing a training sample set, specifically includes:

[0064] S11: parse the version history of the service chain design file from the Git version control system, extract the lane layout data L, and organize it into CSV format data according to the version number and timestamp; wherein the lane layout data includes lane ID, service name, service type, center coordinate, width and hierarchical relationship;

[0065] S12: parse the event definition section in the design file, construct an event graph, take service operations as nodes and event relationships as directed edges, generate event graph data G=(V,E), and convert the event graph data G into an adjacency matrix A and an edge feature matrix E_feat;

[0066] S13: collect performance indicator data P from the system test log database; wherein the performance indicator data P includes event response time, packet loss rate, CPU occupancy rate and memory usage;

[0067] S14: time align the lane layout data L, event graph data G and performance indicator data P based on the timestamp of the design version, form a single sample, standardize the single sample using Z-score, and obtain a training sample set.

[0068] In practical applications, first, the GitPython library is called to parse the version history of the business chain design file in the Git version control system, extract the swimlane layout data of each version, including swimlane ID, service name, service type, center coordinate, width, and hierarchical relationship, and organize the data into CSV format according to the structure of version number, timestamp, and swimlane attribute, to ensure the structured storage of swimlane information; then, the event definition section in the design file is parsed, with service operations as nodes (including operation ID, belonging swimlane ID, etc.) and event relationships as directed edges (including event type, period or timeout parameters, etc.), to build an event graph and convert it into an adjacency matrix and edge feature matrix, realizing the mathematical expression of event association. After that, the performance indicator data corresponding to the design version is collected from the system test log database (such as Elasticsearch), covering event response time, packet loss rate, CPU occupancy, and memory usage, and extreme outliers are removed, with the mean value of indicators of the same type of event being used to fill in missing values, to ensure the integrity of the performance data. Finally, the swimlane layout data, event graph data, and processed performance indicator data are time-aligned based on the timestamp of the design version submission, forming a single sample containing all three, and the continuous features in the sample are standardized by Z-score to eliminate dimensional influence, and the training set, validation set, and test set are divided in a ratio of 7:2:1, providing a standardized and highly correlated data source for subsequent model training.

[0069] In this embodiment, by collecting swimlane layout, event graph, and performance indicator data from Git, design files, and test logs, respectively, and through time alignment and Z-score standardization, a structured training sample set is obtained, successfully solving the problems of heterogeneous multi-source business chain data, low sample quality, and lack of historical data support in the background art. By unifying the text data of Git, numerical data of logs, and structured data of graphs into CSV or matrix format, the format differences of data from different sources are eliminated, and time alignment ensures the correlation of swimlane, event, and performance data, while standardization avoids the interference of dimension on model training, greatly improving the signal-to-noise ratio of training samples, effectively providing a high-quality data foundation for the multi-task prediction model, avoiding prediction bias caused by chaotic or incomplete input data, and laying a data foundation for subsequent precise prediction of swimlane layout and event configuration.

[0070] In a preferred embodiment, in step S2, a multi-task prediction model is constructed based on a graph neural network and a Transformer fusion architecture, specifically including:

[0071] S21: constructing a dual-branch neural network architecture including a local prediction branch and an event configuration branch;

[0072] The local prediction branch adopts a graph attention network, an input layer receives a lane feature matrix L_feat extracted from the lane layout data L and a spatial adjacency matrix S, a 2-layer graph attention GAT is used to learn the spatial dependence of the lane, the number of attention heads of the GAT is 4, the hidden layer dimension is 64, and the expression of the attention weight is:

[0073] ;

[0074] In the formula, denotes an initial feature vector of the lane u, denotes a linear transformation matrix, denotes an attention weight vector, denotes a neighborhood set of the lane u, denotes a vector splicing operation, denotes an initial feature vector of the lane v, denotes a neighborhood set of the lane u any lane, denotes an initial feature vector of the lane ;

[0075] The event configuration branch adopts a Transformer decoder, an input layer receives a node feature matrix V_feat and an edge feature matrix E_feat extracted from the event graph data G, a 3-layer encoder is used to learn the event association mode, and an output layer outputs an event type probability distribution through a Softmax activation function.

[0076] In practical applications, the embodiment realizes the synchronous prediction of lane layout and event configuration through a double-branch design: the layout prediction branch adopts a graph attention network (GAT), and the input layer receives a lane feature matrix and a spatial adjacency matrix extracted from the lane layout data, wherein the lane feature matrix contains features such as service type code, level, and standardized width of the lane, and the spatial adjacency matrix is calculated by the inverse of the Euclidean distance between lanes to represent the spatial similarity between lanes. The branch learns the spatial dependency between lanes through 2 layers of GAT (4 attention heads per layer, hidden layer dimension 64), dynamically allocates the importance of different lanes using the attention weight formula, and finally outputs the predicted position coordinates of each lane; the event configuration branch adopts a Transformer decoder, and the input layer receives a node feature matrix and an edge feature matrix extracted from the event graph data, wherein the node feature matrix contains attributes such as ID code and lane ID of the operation node, and the edge feature matrix records the type parameters of the event. The branch learns the event association rule between operation nodes through 3 layers of Transformer encoder (8 attention heads per layer, hidden layer dimension 128, FeedForward network dimension 512), captures the dependency in the event sequence using the self-attention mechanism, and the output layer converts the network output into an event type probability distribution through the Softmax activation function to realize the prediction of the event type.

[0077] Therefore, the embodiment constructs a layout prediction branch based on a graph attention network (GAT) and an event configuration branch based on a Transformer decoder to form a double-branch neural network architecture, which respectively learns the spatial dependency of lanes and the event association mode, effectively solving the problems of insufficient feature compactness of the model in the background art, the single-task model being unable to balance the prediction of lane layout and event configuration, and poor multi-scene adaptability. Among them, the GAT focuses on the spatial dependency of key lanes through the attention mechanism, and the Transformer accurately captures the association mode of the event sequence, making the expression of the model on the spatial features of lanes and the logical features of events more compact. At the same time, the double-branch design can optimize the lane layout and event configuration tasks respectively, avoid the single model being unable to balance, significantly improve the event classification accuracy, and further support adjusting the branch loss weight according to different business scenarios (such as increasing the event configuration branch weight in the aviation electrical scene), without the need to reconstruct the model to adapt to different scene requirements, meeting the multi-scene design requirements in the complex equipment development field.

[0078] In the preferred embodiment, in step S2, the model parameters are trained by combining the loss function and the constructed training sample set to obtain a multi-task prediction model with lane layout prediction and event configuration capabilities, specifically including:

[0079] S22: define a combined loss function, the expression is:

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] wherein, denotes the total loss value, denotes the loss weight, denotes the layout prediction loss, denotes the event classification loss, denotes the representation learning loss, denotes the number of lanes, denotes the real coordinates of lane i, denotes the predicted coordinates of lane i, denotes the number of events, denotes the real type label of event j, denotes the predicted probability that event j belongs to type k, denotes the batch size, denotes the feature representation of event b, denotes the positive sample of event b, denotes the negative sample, denotes the cosine similarity, denotes the temperature parameter;

[0085] S23: using the Adam optimizer, training the model under the preset learning rate, weight decay, batch size and training round conditions, and obtaining a multi-task prediction model with lane layout prediction and event configuration capabilities after training is completed.

[0086] In actual application, the embodiment realizes efficient training of the model through loss function design and optimization strategy: first, a combined loss function is defined, and the total loss is obtained by weighted summation of the layout prediction loss, the event classification loss and the representation learning loss, wherein the layout prediction loss adopts mean square error (MSE), and the square difference average of the real coordinates and the predicted coordinates of the lane is calculated to measure the deviation of the lane layout prediction. The event classification loss adopts cross-entropy loss, and the logarithmic loss of the real type label (one-hot encoding) and the predicted probability of the event is calculated to measure the accuracy of the event type prediction. The representation learning loss adopts InfoNCE loss, and the cosine similarity of the positive sample (events of the same type) and the negative sample (events of different types) of the event feature is compared to enhance the similarity of the features of the same events and the difference of the features of different events. The loss weight can be adjusted according to the business scenario (such as higher event classification loss weight in the avionics scenario).

[0087] After that, the model is trained using the Adam optimizer, with an initial learning rate of 1e-4, weight decay of 1e-5, batch size of 32, and 100 training rounds. Gradient clipping (clip_norm=1.0) is used during training to prevent gradient explosion, and the learning rate is dynamically adjusted using the cosine annealing learning rate scheduling strategy to ensure stable convergence of the model. At the same time, an early stopping strategy is introduced, where the model performance (layout prediction MAE, event classification Acc) is evaluated every epoch using the validation set. If the validation loss does not decrease for 5 consecutive epochs, training is stopped to avoid overfitting. The model parameters with the smallest validation loss are saved to ensure good generalization ability.

[0088] In this embodiment, by defining a combined loss function containing layout prediction loss, event classification loss, and representation learning loss, the model is trained using the Adam optimizer under preset parameters to obtain a multi-task model with lane layout and event configuration prediction capabilities. This effectively solves the problems of model training overfitting, multi-task loss imbalance, and low training efficiency in the background art.

[0089] In a preferred embodiment, step S3: capturing user real-time operations through the event listener module of the design interface, extracting service metadata and canvas state features from the user real-time operations, and generating a standardized input vector, specifically including:

[0090] S31: Centralizing event listeners in the design interface to capture three types of operation events; wherein the operation events include mouse drag events, service placement events, or connection operation events;

[0091] S32: Calling the service registry through the backend API to extract the metadata of the current operation service and convert it into a service meta-feature vector; wherein the metadata includes service ID, service type, maximum processing delay, CPU resource demand, memory demand, and interface protocol;

[0092] S33: Iterating through the current canvas lane to extract spatial features and density features, and concatenating the spatial features, density features, and service meta-feature vector to generate a canvas state feature vector as a standardized input vector.

[0093] In this embodiment, first, a JavaScript event listener is integrated into the design interface developed in Vue to capture three types of core user operations in real time: mouse drag events (record drag trajectory coordinates with a sampling frequency of 50Hz), service placement events (record service initial placement coordinates and timestamps), and connection operation events (record source operation and target operation IDs). All operation data is stored in real time in the front-end localStorage in JSON format (including operation type, service ID, trajectory information, and canvas state) to ensure that operation data is not lost and can be quickly called.

[0094] Then the MySQL service registry is called through the backend API to extract the metadata of the current user's operation service, including service ID, service type, maximum processing delay, CPU resource requirement, memory requirement, interface protocol, the categorical type metadata (such as service type, interface protocol) is converted into numerical features by using one-hot encoding, the continuous type metadata (such as maximum processing delay, CPU resource requirement) is mapped to the [0, 1] interval by using min-max standardization, forming a service meta feature vector with a dimension of 1x12; then all the lanes in the current design canvas are traversed to extract the spatial features (current coordinates, width, adjacent lane spacing) and density features (event connection density = event edge number / lane number, layout uniformity = lane spacing variance) of the lanes, the spatial features, density features and service meta feature vector are spliced by dimension to form a canvas state feature vector (dimension 1x(12+2N+2), N is the number of current canvas lanes); finally, the feature vector is checked for integrity and reasonableness, the missing values are filled with the feature mean value of the same type lane, and the features exceeding the reasonable range are truncated according to the boundary value, and the checked feature vector is transmitted to the backend model service in real time through the WebSocket protocol to meet the real-time interaction requirements.

[0095] Therefore, the embodiment captures user operations through the event listener of the design interface, calls the service registry to extract service metadata and encode, combines the spatial and density features of the canvas lanes, and generates a standardized input vector, effectively solving the problems of high real-time interaction delay, incomplete feature extraction and inconsistent feature format in the background technology.

[0096] In a preferred embodiment, step S4: input the real-time standardized input vector into the trained multi-task prediction model to generate multiple groups of lane layout and event configuration optimization schemes, and filter the optimal scheme group through simulation evaluation, specifically including:

[0097] S41: input the standardized input vector into the layout prediction branch, output the 3 groups of lane position adjustment vectors with the highest confidence obtained by the layout prediction branch as the layout scheme, input the standardized input vector into the event configuration branch, and output the 3 groups of configuration schemes with the highest probability obtained by the event configuration branch;

[0098] S42: cross-combine the layout scheme and the configuration scheme to form 9 groups of initial candidate schemes, simulate running in the SimPy simulation environment, output several core performance indicators of each scheme, calculate the performance score by using the weighted sum of the several core performance indicators, and calculate the design specification score by using the sum of the mapping scores of the several preset design specification standards and each scheme with respect to the scheme configuration parameters in the design constraint rule library.

[0099] S43: For the 9 groups of initial candidate schemes, the comprehensive performance score is calculated according to the performance score and the design specification score, and the top 5 groups of schemes with the highest scores are selected as the optimal scheme group; wherein, the expression of the comprehensive performance score is:

[0100] ；

[0101] In the formula, S represents the comprehensive performance score, represents the performance score, represents the design specification score.

[0102] In this embodiment, first, the standardized input vectors extracted in real time are respectively input into the two branches of the multi-task prediction model, the layout prediction branch outputs the 3 groups of lane position adjustment vectors with the highest confidence, forming 3 groups of layout schemes (completely adjusted according to the predicted position, only adjusting the non-core lane while keeping the core lane position, and fine-tuning based on the predicted position), and the adjustment process ensures that the lanes are not overlapped; the event configuration branch outputs the 3 groups of event configuration schemes with the highest probability distribution of event types (such as periodic events and response events) and corresponding parameters (such as periodic values and timeout times).

[0103] Then, the 3 groups of layout schemes and the 3 groups of event schemes are cross combined to form 9 groups of initial candidate schemes, each scheme is simulated to run for 1 hour in the SimPy simulation environment, and 4 core performance indicators including event response time, CPU utilization, memory utilization and logical clarity are output. After standardizing the indicators, the performance score is obtained by weighted summation according to the preset weights (event response time 0.4, CPU utilization 0.25, memory utilization 0.2, and logical clarity 0.15); at the same time, according to the design constraint rule library (including service level compliance, lane layout uniformity, event type matching degree, ICD protocol version, and core lane position), each scheme is checked item by item, each specification is scored 20 points, the total score is normalized to obtain the design specification score; finally, the comprehensive score of each scheme is calculated according to the comprehensive score formula, and the schemes with a layout difference of less than 5 pixels and consistent event configuration are removed, and the top 5 groups of schemes are reserved as the optimal scheme group, ensuring that the schemes have high performance and compliance.

[0104] Therefore, by inputting the standardized input vector into the model to generate multiple groups of layout and event schemes, and cross combining them to evaluate the performance score through SimPy simulation and calculate the specification score according to the design constraint rule library, the optimal scheme group is selected according to the comprehensive score, effectively solving the problems of low efficiency of manual scheme selection, difficulty in balancing performance and compliance, and insufficient diversity of schemes in the background art. The combination, evaluation and selection of schemes are automatically completed, providing high-quality scheme selection for subsequent visual modeling.

[0105] In a preferred embodiment, S5: superimpose the optimal scheme group in a visual form on the design canvas, record user feedback data, and perform visual swim lane modeling according to the user feedback data, specifically comprising: S51: sequentially superimpose each scheme in the optimal scheme group on the canvas in a preset transparency profile according to the comprehensive performance score from high to low, and the current design scheme;

[0106] S52: label the difference between the current design scheme and the optimal scheme with a target color, and generate a detail display panel on one side of the canvas for displaying the performance score and design specification score of the current optimal scheme;

[0107] S53: capture and record user feedback, and perform visual swim lane modeling according to the operation instruction corresponding to the user feedback data; wherein the operation instruction includes scheme application instruction, scheme switching instruction or scheme modification instruction corresponding to keyboard shortcut feedback data or page menu operation feedback data.

[0108] In this embodiment, the optimal scheme group selected is first sorted by comprehensive score from high to low, and the swim lanes of each scheme are sequentially superimposed on the design canvas with a 60% transparent blue profile, wherein the core swim lanes (such as "flight control calculation" and "sensor fusion") use a 2px wide profile line, and the non-core swim lanes use a 1px wide profile line, facilitating the user to quickly identify the priority of the scheme.

[0109] Then the difference between the optimal scheme and the current design is marked with a red dashed line box, including the area where the swim lane needs to be moved, the newly added or modified event edge, when the user mouse hovers over the difference area, a prompt box is popped up to display the difference description (such as "swim lane right shift 150 pixels, can reduce event response time 12 ms"), and a detail panel is popped up on the right side of the canvas to display the performance score, design specification score, core parameters (such as event period, ICD protocol version) and recommended reasons (such as "based on 50 historical avionics project data, this layout CPU utilization is reduced by 10%") of the scheme, supporting the user to click the "split screen comparison" button to display the current design scheme on the left and the optimal scheme on the right, and intuitively view the difference between the two.

[0110] Finally, the user feedback operation is monitored, including keyboard shortcut (Alt+Enter one-key application of all adjustments, Alt+L only apply layout adjustment, Alt+E only apply event configuration adjustment) and right-click menu operation (apply the current scheme, ignore the current scheme, customize the adjustment parameter), after receiving the user feedback instruction, the system automatically updates the swimlane position and event configuration in the canvas, displays the progress bar prompt processing progress during the application process (such as "application: 3 / 5 swimlane complete adjustment"), and pops up a "scheme application successful, modeling completed" prompt box after the adjustment is completed, and at the same time, the user feedback data (accept = 1, modify = 0.5, reject = 0) is stored in the MongoDB database according to the structure of "user ID-scheme ID-feedback type-operation timestamp", which provides user preference data for subsequent online model learning.

[0111] Therefore, the embodiment can effectively solve the problems of poor user interaction experience, modeling results not meeting user needs and feedback data not being reusable in the background technology by superimposing the optimal scheme group on the current canvas in turn with a translucent outline according to the comprehensive score, marking the scheme difference and generating a detail panel to display the score, capturing user feedback (shortcut / menu operation), and executing visual swimlane modeling according to the feedback instruction.

[0112] In the description of the embodiments of the application, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0113] In the description of the embodiments of the application, it should be understood that "-" and "~" represent a range between two values, and the range includes the endpoints. For example: "A-B" represents a range greater than or equal to A and less than or equal to B. "A~B" represents a range greater than or equal to A and less than or equal to B.

[0114] In the description of the embodiments of the application, the term "and / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects.

[0115] Although the embodiments of the application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirit of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for business data processing based on visualized lane modeling, characterized in that, The method comprises the following steps: S1: Collecting historical service chain design data from a multi-source system, performing standardization processing and time sequence alignment, and constructing a training sample set; specifically comprising: S11: Parsing the version history of the service chain design file from the Git version control system, extracting the lane layout data L, and organizing it into CSV format data according to the version number and timestamp; wherein the lane layout data includes lane ID, service name, service type, center coordinate, width and hierarchical relationship; S12: Parsing the event definition section in the design file, constructing an event graph, taking service operations as nodes and event relationships as directed edges, generating event graph data G=(V,E), and converting the event graph data G into an adjacency matrix A and an edge feature matrix E_feat; S13: Collecting performance indicator data P from a system test log database; wherein the performance indicator data P includes event response time, packet loss rate, CPU occupancy and memory usage; S14: Aligning the lane layout data L, event graph data G and performance indicator data P in time sequence based on the timestamp of the design version, forming a single sample, standardizing the single sample using Z-score, and obtaining a training sample set; S2: Based on the graph neural network and the Transformer fusion architecture, a multi-task prediction model is constructed, the model parameters are trained through the combination of the loss function and the constructed training sample set, and a multi-task prediction model with lane layout prediction and event configuration capabilities is obtained; specifically comprising: S21: Constructing a double-branch neural network architecture including a local prediction branch and an event configuration branch; Wherein the local prediction branch adopts a graph attention network, the input layer receives the lane feature matrix L_feat extracted from the lane layout data L and the spatial adjacency matrix S, learns the lane spatial dependency relationship through 2 layers of graph attention GAT, the number of attention heads of the graph attention GAT is 4, the hidden layer dimension is 64, and the expression of the attention weight is: wherein, denotes an initial feature vector of a lane u, denotes a linear transformation matrix, denotes an attention weight vector, denotes a neighborhood set of a lane u, denotes a vector concatenation operation, denotes an initial feature vector of a lane v, denotes a neighborhood set of a lane u any one of the lanes, denotes an initial feature vector of a lane . Wherein the event configuration branch adopts a Transformer decoder, the input layer receives the node feature matrix V_feat and the edge feature matrix E_feat extracted from the event graph data G, learns the event correlation pattern through 3 layers of encoder, and the output layer outputs the event type probability distribution through the Softmax activation function; S3: Capturing user real-time operations through the event listening module of the design interface, extracting service metadata and canvas state features in the user real-time operations, and generating a standardized input vector; S4: Inputting the real-time standardized input vector into the trained multi-task prediction model to generate multiple sets of lane layout and event configuration optimization schemes, and selecting the optimal scheme set through simulation evaluation; S5: Superimposing the optimal scheme set in a visual form on the design canvas, recording user feedback data, and performing visual lane modeling according to the user feedback data.

2. The business data processing method based on visualization lane modeling of claim 1, wherein, In step S2, the model parameters are trained through the combination of the loss function and the constructed training sample set, and a multi-task prediction model with lane layout prediction and event configuration capabilities is obtained, specifically comprising: S22: define a combined loss function, the expression is: wherein, denotes the total loss value, denotes the loss weight, denotes the layout prediction loss, denotes the event classification loss, denotes the learning loss, denotes the number of lanes, denotes the real coordinates of lane i, denotes the predicted coordinates of lane i, denotes the number of events, denotes the real type label of event j, denotes the predicted probability that event j belongs to type k, denotes the batch size, denotes the feature representation of event b, denotes the positive sample of event b, denotes the negative sample, denotes the cosine similarity, denotes the temperature parameter; S23: adopt the Adam optimizer, train the model under the condition of preset learning rate, weight decay, batch size and training round, and obtain a multi-task prediction model with lane layout prediction and event configuration capability after training is completed.

3. The business data processing method based on visualization lane modeling of claim 1, wherein, Step S3: capture the user's real-time operation through the event listening module of the design interface, extract the service metadata and canvas state features in the user's real-time operation, and generate a standardized input vector, specifically including: S31: concentrate event listeners in the design interface to capture three types of operation events; wherein the operation events include mouse drag events, service placement events or connection operation events; S32: call the service registry through the backend API, extract the metadata of the current operation service, and convert it into a service meta feature vector; wherein the metadata includes service ID, service type, maximum processing delay, CPU resource demand, memory demand, interface protocol; S33: traverse the current canvas lane, extract spatial features and density features, concatenate the spatial features, density features and service meta feature vector to generate a canvas state feature vector as a standardized input vector.

4. The business data processing method based on visualization lane modeling of claim 1, wherein, Step S4: input the real-time standardized input vector into the trained multi-task prediction model to generate multiple lane layout and event configuration optimization schemes, and select the optimal scheme group through simulation evaluation, specifically including: S41: input the standardized input vector into the layout prediction branch, output the three groups of lane position adjustment vectors with the highest confidence obtained by the layout prediction branch as the layout scheme, and input the standardized input vector into the event configuration branch, output the three groups of configuration schemes with the highest probability obtained by the event configuration branch; S42: cross combine the layout scheme and the configuration scheme to form 9 groups of initial candidate schemes, simulate running in the SimPy simulation environment, output several core performance indicators of each scheme, calculate the performance score using the weighted sum of several core performance indicators, and calculate the design specification score using the preset several design specification standards and the mapping score sum of each scheme for scheme configuration parameters in the design constraint rule library; S43: for the 9 groups of initial candidate schemes, calculate the comprehensive performance score according to the performance score and the design specification score, and select the top 5 groups of schemes with the highest score as the optimal scheme group; wherein the expression of the comprehensive performance score is: In the formula, S represents a comprehensive performance score, represents a performance score, represents a design specification score.

5. The business data processing method based on visualization lane modeling of claim 1, wherein, S5: superimpose the optimal scheme group in a visual form on the design canvas, record user feedback data, and perform visual lane modeling according to the user feedback data, specifically including: S51: superimpose each scheme in the optimal scheme group on the current design scheme in the order of the comprehensive performance score in a preset transparency outline on the canvas; S52: label the difference between the current design scheme and the optimal scheme with a target color, and generate a detail display panel on one side of the canvas to display the performance score and the design specification score of the current optimal scheme. S53: Capture and record user feedback, and select the optimal scheme for visual swim lane modeling according to the operation instruction corresponding to the user feedback data; wherein the operation instruction includes scheme application instruction, scheme switching instruction or scheme modification instruction corresponding to keyboard shortcut feedback data or page menu operation feedback data.

6. A business data processing system based on visualized lane modeling, characterized by, For executing the method of any one of claims 1-5, comprising: a construction module for collecting historical business chain design data from multi-source systems, performing standardization processing and time sequence alignment, and constructing a training sample set; a training module for constructing a multi-task prediction model based on a graph neural network and a Transformer fusion architecture, training model parameters by combining a loss function and the constructed training sample set, and obtaining a multi-task prediction model with swim lane layout prediction and event configuration capabilities; a generation module for capturing user real-time operations through an event listening module of a design interface, extracting service metadata and canvas state features in the user real-time operations, and generating a standardized input vector; a screening module for inputting the real-time standardized input vector into the trained multi-task prediction model, generating multiple groups of swim lane layout and event configuration optimization schemes, and screening the optimal scheme group through simulation evaluation; an execution module for superimposing and displaying the optimal scheme group in a visual form on the design canvas, recording user feedback data, and executing visual swim lane modeling according to the user feedback data.

Images