Multi-mode-based software architecture intelligent design and optimization system

Abstract

The invention discloses a multi-modal-based software architecture intelligent design and optimization system, which relates to the field of intelligent design and optimization, and comprises the steps of performing modal perception and preprocessing on architecture design demand information input by a user, converting the architecture design demand information into structured semantic data, and optimizing and enhancing semantic expression by introducing a reinforcement learning strategy, forming a structured user intention vector set; through a semantic mapping and reasoning processing unit, user intention vectors in the user intention vector set are constructed into a semantic-component alignment graph, and graph structure modeling and confrontation generation algorithm combination are adopted to generate a candidate structure graph set; a structure diagram generation step in the structure generation module effectively improves the rationality, diversity and adaptation capability of an automatically generated structure, and provides core support for realizing automatic construction of a target-demand-oriented architecture structure.

Description

Technical Field

[0001] The present invention relates to the field of intelligent design and optimization, and in particular to a multi-modal software architecture intelligent design and optimization system. Background Art

[0002] With the increasing functional complexity of software systems, the increasingly diverse deployment environments, and the continuous evolution of user interaction methods, traditional software architecture design methods have gradually exposed limitations such as low efficiency, weak intelligence, and poor adaptability. In recent years, the rapid development of artificial intelligence, especially multimodal perception, natural language processing, and graph neural networks, has promoted the transformation of software architecture design patterns from people-centered to intelligent collaboration. Multimodal technology enables the system to recognize and integrate inputs from different forms such as voice, text, graphic sketches, gestures, etc., so as to obtain user demand intentions more comprehensively; semantic modeling and vector expression methods realize the structured transformation of complex requirements; at the same time, semantic-driven component recommendation mechanisms and graph structure modeling algorithms also provide a feasible path for automatically generating preliminary architecture diagrams. On this basis, how to achieve accurate semantic understanding of user intentions, multimodal collaborative modeling, and structural optimization control has become an important research direction for current intelligent software architecture design.

[0003] Although existing technologies have made some progress in natural language semantic parsing, component knowledge graph construction, and preliminary structure diagram generation, there are still many problems that need to be solved. For example, the current semantic fusion accuracy between multimodal inputs is insufficient, and the semantic understanding ability is difficult to reflect the user's contextual personalized behavior; the architecture candidate structure diagram lacks reliable logic and engineering consistency verification methods, and is prone to potential conflicts or weak connection points; in addition, the existing system relies on manual intervention in the structural optimization link, lacks automatic feedback mechanism and semantic closed-loop control capabilities, making it difficult to implement structural optimization suggestions. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a multimodal-based software architecture intelligent design and optimization system that solves the problems of inaccurate semantic understanding and insufficient consistency in structure diagram generation under multimodal user input.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a multi-modal software architecture intelligent design and optimization system, which includes: The intention recognition module performs modal perception and preprocessing on the architectural design requirement information input by the user, converts it into structured semantic data, and optimizes and enhances the semantic expression by introducing reinforcement learning strategies to form a structured set of user intention vectors; the structure generation module builds a semantic-component alignment graph based on the user intention semantic vector through the semantic mapping and reasoning processing unit, and uses a combination of graph structure modeling and adversarial generation algorithm to generate a candidate structure atlas; the structure verification module calls the automatic structure inspection process to perform logical and engineering verification on the structure atlas of the candidate architecture structure scheme, and automatically visually marks potential errors and structural weaknesses to obtain a list of structural problems and an initial architecture sketch; the optimization output module makes one-click optimization suggestions based on the list of structural problems through the optimization logic unit to form the final optimized architecture scheme.

[0007] As a preferred solution of the multimodal software architecture intelligent design and optimization system described in the present invention, the following steps are performed to perform modal perception and preprocessing on the architecture design requirement information input by the user and convert it into structured semantic data: Receive various forms of user input, classify and encapsulate them according to the input modality, and form a multi-modal raw data set; Call the corresponding parsing module to perform signal recognition and conversion on the multimodal raw data set, and normalize it to obtain the semantic vector set after modality alignment; The modality-aligned semantic vector set is input into the context fusion module, and structured semantic data is generated by combining the user's historical behavior and modality credibility coefficient.

[0008] As a preferred solution of the multimodal software architecture intelligent design and optimization system described in the present invention, the various forms of user input refer to the user's voice input, natural language text input, graphic sketches and gesture operations.

[0009] As a preferred solution of the multimodal software architecture intelligent design and optimization system of the present invention, the steps of calling the corresponding analysis module to perform signal recognition and conversion on the multimodal original data set are as follows: Convert language input into text to obtain a list of intended keywords and a functional sentence structure; Extract module nodes and connection relationships from the image sketch to obtain a set of connection paths and candidate matching semantic units of components; Identify the operation semantics from the gesture operation, and obtain the interactive operation list and the component identification and action intention correspondence table; Perform grammatical parsing, entity extraction and structure mapping on natural language text input to obtain a list of functional entities, technical attribute structures and non-functional requirement annotations.

[0010] As a preferred solution of the multimodal software architecture intelligent design and optimization system described in the present invention, wherein: the semantic expression is optimized and enhanced by introducing a reinforcement learning strategy to form a structured user intention vector set, the steps are as follows: The semantic unit codes in the structured semantic data are converted into semantic unit vectors, and the semantic unit vectors are format-aligned and semantic space-mapped to obtain a set of multimodal fusion semantic vectors; Retrieve the historical behavior data records of the current user, perform context fusion modeling with the multimodal fusion semantic vector set, and obtain the user context enhanced semantic vector set; The user context enhanced semantic vector set is input into the learning model, and the policy gradient reinforcement learning algorithm is applied to update the policy model parameters to obtain the user core intent set; The user's core intent set is mapped to a standardized terminology to obtain a structured intent-term alignment list, and the intent-term alignment list is classified according to intent type to obtain a structured user intent semantic vector.

[0011] As a preferred solution of the multimodal software architecture intelligent design and optimization system described in the present invention, the steps of combining graph structure modeling with adversarial generation algorithm to generate a candidate structure graph set are as follows: By using semantic mapping and reasoning, the structured user intent semantic vector is mapped to the target intent node in the semantic-component graph, and the candidate component set corresponding to the semantic node is screened from the architecture knowledge graph in combination with the semantic association scoring mechanism; According to the semantic mapping and reasoning processing unit and the semantic-component association information in the architecture knowledge graph library, the alignment relationship between intentions and components and the structural connection relationship between components are constructed, and the alignment relationship between intentions and components and the structural connection relationship between components are sorted through the structured graph construction engine to obtain a set of candidate component structure graphs; Using graph neural network, each candidate structure graph in the candidate component structure graph set is modeled to obtain an embedded graph structure vector set; Input a set of embedded graph structure vectors, generate multiple structural graph candidate solutions, add semantic labels to each node in each structural graph candidate solution, perform standard formatting and packaging, and obtain a set of candidate architecture structural solution structural graphs.

[0012] As a preferred solution of the multimodal software architecture intelligent design and optimization system described in the present invention, wherein: according to the automatic structure checking process, the structure atlas of the candidate architecture structure scheme is logically and engineering checked, and potential errors and structural weaknesses are automatically and visually marked to obtain a list of structural problems and an initial architecture sketch. The automatic structure checking process is called to perform logic and engineering checks on the candidate structure atlas, and potential errors and structural weaknesses are automatically and visually marked to obtain a list of structural problems. The steps are as follows: The candidate architecture structure graph set is converted into an internal graph modeling format by using a method based on semantic ontology mapping and structural specification template matching. The unified standard graph structure model is parsed and encapsulated to obtain a standardized structure graph representation set by combining the graph node type set and topology rules. Logical integration and engineering verification of the standardized structural diagram representation set to obtain a list of structural problems with the type, location and severity of each problem; Combined with the standardized structure diagram representation set and the structure problem list, the structure problems are marked in the standardized structure diagram identification set to generate annotated structure diagram sketches.

[0013] As a preferred solution of the multi-modal software architecture intelligent design and optimization system described in the present invention, wherein: the one-key optimization suggestion is made according to the structural problem list by optimizing the logic unit, the steps are as follows: Analyze the list of structural problems item by item, classify them according to the problem type, and match optimization suggestions for each problem to obtain a set of optimization suggestion solutions; Determine the applicability of each optimization suggestion in the optimization suggestion set, and automatically modify the initial structure sketch according to the optimization suggestion to obtain an updated architecture structure diagram model; The updated architecture structure diagram model is semantically expanded, technology stack completed and engineered at the deployment level to obtain the final optimized architecture solution.

[0014] As a preferred solution of the multimodal-based software architecture intelligent design and optimization system described in the present invention, the applicability judgment of each optimization suggestion in the optimization suggestion set refers to automatically judging whether the optimization suggestion is suitable for direct application in the architecture structure diagram model before applying each structural optimization suggestion.

[0015] The beneficial effects of the present invention are as follows: the user intention vector is aligned with the system component knowledge base through the semantic mapping and reasoning processing unit to establish a semantic-component graph, and the component dependency is modeled using a graph neural network, combined with an adversarial generation mechanism to generate diverse candidate structure graph solutions. The structure generation module breaks through the traditional static structure generation method based on rules or templates, realizes the dynamic conversion from semantic abstraction to structural instances, and introduces a two-way feedback mechanism of semantic constraints and structural optimization, so that the generated solution is both innovative and scalable while satisfying the expression of intent. Finally, the structural graph generation step in the structure generation module effectively improves the rationality, diversity and adaptability of the automatically generated structure, and provides core support for the automatic construction of architecture structures oriented to target requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1 Schematic diagram of the multimodal-based software architecture intelligent design and optimization system.

[0018] Figure 2 Generate a module flow chart for the structure of a multimodal-based software architecture intelligent design and optimization system.

[0019] Figure 3 This is the flow chart of the structural verification module of the multimodal-based software architecture intelligent design and optimization system.

[0020] Figure 4 This is a flow chart of the optimization output module for the multimodal-based software architecture intelligent design and optimization system. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0022] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figure 1~Figure 4 , is an embodiment of the present invention, which provides a multi-modal software architecture intelligent design and optimization system, including the following steps: The architectural design requirement information input by the user is modally perceived and preprocessed, converted into structured semantic data, and the semantic expression is optimized and enhanced by introducing reinforcement learning strategies to form a structured set of user intention vectors.

[0025] It receives input data streams in various forms through a unified interactive interface, including voice files (WAV), natural language text (TXT), sketch images (PNG / SVG) and gesture sequences (JSON), automatically identifies the modal type of each input data, and encapsulates and classifies the input data streams in various forms according to the modal type to form a multimodal raw data set in a unified format.

[0026] The modal parsing module of the object is called for various types of modal data to extract primary semantic information as follows: when the input modality is voice input, the Whisper-Large speech recognition model (large speech recognition model) is used to obtain the transcribed text and semantic labels; when the input modality is text input, the RoBERTa model (pre-trained model) and entity recognition technology are used to extract functional keywords and technical terms; when the input modality is a sketch image, the image is firstly feature extracted through a convolutional neural network, and then a graph structure reconstruction network (such as the convolutional network) is used to reconstruct the module nodes and connection paths; when the input modality is gesture input, a skeleton recognition model based on a visual transformer is used to obtain the semantics of user interaction operations, and finally each modality is parsed into: modality type, semantic structure data and modality semantic vector.

[0027] The semantic vectors extracted under different modalities are uniformly mapped into a standard semantic space to achieve normalization and alignment, and a set of semantic vectors after modality alignment is obtained. The linear mapping formula is as follows: ; in, Represents the number of the mode, Indicates The original semantic vector corresponding to the modal input, Expressed as The linear projection matrix that maps the semantic vectors of each modality to the unified semantic space, Expressed as The corresponding bias vector of each modality in the process of mapping to the unified semantic space, Represented as the mapped semantic vector, that is, the normalized semantic vector, It is represented as the common dimension of the semantic space to which all modalities are uniformly mapped after normalization.

[0028] All normalized modal semantic vector sets are input into the context fusion module, and the following two types of information are fused: the historical behavior semantic data of the current user and the credibility coefficient of each modality. Based on the weighted similarity between the semantic vector of each modality and the learning weight vector obtained by back-propagation adaptive learning, a set of weighted modal feasibility coefficients are obtained through softmax calculation (exponential normalization function calculation).

[0029] The unified semantic vector generated by the fusion result is input into the semantic decoding module, and the semantic vector is converted into structured semantic information data using the classifier model and entity mapping rule library.

[0030] Perform modal perception and preprocessing on the architecture design requirement information input by the user and convert it into structured semantic data. The steps are as follows: Receive various forms of user input, classify and encapsulate them according to the input modality, and form a multi-modal raw data set; The analysis module is called on the multimodal raw data set to perform signal recognition and semantic conversion, and unified normalization is performed to obtain a set of semantic vectors after modality alignment; The modality-aligned semantic vector set is input into the context fusion module, and combined with the user's historical behavior and modality credibility coefficient to generate structured semantic data.

[0031] Receive various forms of user input and mark them into four categories: text, voice, sketch, and gesture. Remove background noise from the voice and segment it into sentences, ensuring that each segment is less than 10 seconds. Then use the voice recognition engine on the terminal or cloud to input the cleaned audio segments one by one and return the corresponding recognized text. Extract the contours of the sketch image and identify the connection between the graphic elements and the sketch image. Perform action recognition on the gesture operation video frames and convert them into standard command expressions. Finally, encapsulate all input data according to the source modality to form a unified multimodal raw data set and mark their respective modality types.

[0032] Parse and extract semantic units from multimodal raw data sets: perform syntactic analysis on text and transcribed speech content to extract functional requirements, non-functional constraints, and component relationship descriptions; abstract the node and connection information of sketch content recognition to form a descriptive statement that represents the relationship between components and their structures; map gesture operation recognition results to standard design actions, and finally standardize the semantic content of all modal information and parsing rules and the domain terminology library to form a set of semantic units.

[0033] The modal type and semantic type of each semantic unit in the semantic unit set are embedded and transformed to generate a semantic representation in a unified format. The representations of the same semantic content in different modalities are aligned to remove redundant and conflicting expressions. Then, the trust coefficient is set for the modal source, and the aligned semantics in all modalities are integrated to obtain a modal aligned semantic set.

[0034] The user's recent architectural design historical behavior information, including commonly used components, design preferences and interaction modes, is combined with modal alignment semantics to enhance the contextual association ability of semantics. The fused content is then semantically grouped and classified, and the grouping results are output in the form of structured semantic data, which serves as the input basis for the subsequent construction of the architectural structure diagram, and finally a structured semantic dataset is obtained.

[0035] By introducing reinforcement learning strategies to optimize and enhance semantic expression, a structured set of user intention vectors is formed. The steps are as follows: The semantic unit codes in the structured semantic data are converted into semantic unit vectors, and the semantic unit vectors are format-aligned and semantic space-mapped to obtain a set of multimodal fusion semantic vectors; Retrieve the historical behavior data records of the current user, and perform context fusion modeling with the multimodal fusion semantic vector set to obtain the user context enhanced semantic vector set; The user context enhanced semantic vector set is used as input, and the policy is updated through the policy gradient-based reinforcement learning algorithm to obtain the user core intention set; The user's core intent set is mapped to a standardized terminology to obtain a structured intent-term alignment list, and the intent-term alignment list is classified according to intent type to obtain a structured user intent semantic vector.

[0036] A semantic encoding model (such as BERT-base) is used to encode each semantic unit of the structured semantic data and convert it into a fixed-length semantic vector. Then, all semantic unit vectors are embedded in the same semantic space through linear projection to form a preliminary set of multimodal fusion semantic vectors.

[0037] Historical behavior records including historical architecture modification logs, operation preferences, and keyword selection frequency are extracted from the user behavior data storage module, and converted into context representation vectors through behavior sequence embedding. Subsequently, the fused semantic vector is concatenated with the context vector, and the Transformer structure is introduced for fusion modeling to obtain a context-enhanced semantic vector set.

[0038] The context-enhanced semantic vector set is input into the policy learning model as the current state, and the accuracy and precision of the user's intent expression are defined as the reward function of the policy learning model. Then, the policy gradient-based method is applied to update the parameters of the policy model. After iterative training in multiple rounds, the semantic vector that can maximize the expected reward is selected as the output, which is the user's core intent set.

[0039] By referring to the standardized semantic terminology library, vector matching is performed on each core intent vector in the user core intent set, and the closest standard term is selected based on cosine similarity calculation. It is combined with the intent vector to form an intent-term alignment pair, and then the alignment pair is classified according to the intent type to form a structured user intent semantic vector set.

[0040] Through the semantic mapping and reasoning processing unit, the user intention semantic vector is matched with the component semantics in the architecture component library and semantic reasoning is performed to build a semantic-component alignment graph containing nodes and edges. Combined with the knowledge rule set in the optimization suggestion module, a candidate structure graph set is generated through graph traversal and reconstruction operations.

[0041] Each semantic unit in the structured semantic data is input into the semantic encoding model (which can be a pre-trained language model) to generate an intermediate semantic vector. Then, through linear transformation, all semantic vectors are mapped into a unified semantic representation space to form a set of multimodal fused semantic vectors.

[0042] The user's historical interaction semantic records are retrieved from the user portrait or behavior database and input into the same semantic encoding model. The historical interaction semantic records are encoded into a set of historical behavior semantic vectors. For each current semantic vector, the semantic correlation between the current semantic vector and the historical behavior semantics is calculated, the attention weight is constructed, and the historical behaviors are weightedly fused to form a context-enhanced semantic vector.

[0043] The set of context-enhanced semantic vectors is used as the current state information and input into the reinforcement learning strategy model to learn the optimal intent selection strategy. The reinforcement learning configuration is as follows: the state is each context-enhanced semantic vector, the action is whether to select the semantic vector as the core design intent, the reward signal is generated based on the success rate score generated by the subsequent structure diagram or the user interaction feedback evaluation function, and the strategy function models the selection decision in a probabilistic form.

[0044] The intent selection model is trained and updated through the strategy optimization formula to generate a core design intent set.

[0045] The terms in the core intent set output by the reinforcement learning phase are matched and mapped with the standard component names and domain terms in the knowledge base through the semantic term standardization module, and are classified according to the semantic type of the intent to obtain a structured intent semantic vector data structure.

[0046] The candidate structure graph set is generated by combining graph structure modeling and adversarial generation algorithm. The steps are as follows: Based on the structured user intention semantic vector, the intention nodes of the semantic-component graph are obtained through semantic mapping reasoning, and the candidate component set corresponding to the semantic node is extracted through the semantic association scoring mechanism; According to the semantic mapping and reasoning processing unit and the architecture knowledge graph library, the intent-component alignment relationship and component structure relationship are constructed, and the structured graph construction engine is used to organize them to obtain a set of candidate component structure graphs; Using graph neural network, we extract node and edge features from each candidate structure graph in the candidate component structure graph set, and combine it with adversarial generation algorithm to obtain the graph structure vector embedding set. The semantic vector and graph structure vector embedding set are taken as input to generate multiple candidate structural graph solutions, and semantic labels are added to each node in each structural graph, and standard formatting and packaging are performed to obtain a set of candidate architectural structural solution structural graphs.

[0047] Taking structured user intent semantic information as input, each piece of intent information is associated with the corresponding intent node in the semantic-component knowledge graph through the semantic mapping reasoning process. Then, the weighted semantic similarity scoring mechanism is used to calculate the semantic relevance between each piece of intent information and the candidate component nodes in the graph, and several of the most matching component nodes are screened out according to the scores to obtain a list of candidate components matched in the preliminary semantic component graph.

[0048] The predefined component connection rules in the structured knowledge graph are called to legally combine the candidate components in the candidate component list, and a candidate structure graph with a topological structure between multiple components is constructed. Each candidate structure graph is then passed through the structure graph construction engine to complete the standardized connection structure organization, and the graph neural network algorithm is used for modeling to extract the overall structural features corresponding to each structure graph, forming multiple candidate structure graphs and their structural embedding representation results.

[0049] An adversarial algorithm for graph structure generation is used to input semantic information and structural graph representation results to construct a variety of possible structural graph candidate schemes. Then, a structural rationality judgment module is introduced to evaluate each structural graph scheme. Finally, each node in the structural graph is marked with standardized semantic labels and component identities to complete the unified encapsulation of the structural graph format, forming an architecture and candidate scheme set that can be used for subsequent verification and optimization.

[0050] The automatic structural inspection process is called to perform logic and engineering verification on the structural atlas of the candidate architectural schemes, and potential errors and structural weaknesses are marked through automatic visual inspection to obtain a list of structural problems and an initial architectural sketch.

[0051] Receive a set of structured user intent semantic vectors, each of which represents a user core intent, and analyze the set of structured user intent semantic vectors through the semantic mapping and reasoning processing unit, identify the semantic nodes corresponding to the intent, extract candidate components according to the intent-component association rules, and obtain the "intent-component candidate set" structure, where the semantic association scoring formula is as follows: ; in, Indicates Intent vectors, Expressed as component semantic vectors, Express intention With components The semantic matching degree of Represented as the user intention vector currently being processed, Represents the semantic vector of the candidate component currently being processed.

[0052] The semantic mapping and reasoning processing unit and the architectural knowledge graph library are called to extract the alignment relationship between each pair of intent and components and the prior structural relationship between candidate components, and then sorted and reconstructed through the structured graph construction engine to form an initial connection relationship graph between candidate components.

[0053] The candidate component structure graph set is input into the graph neural network model, the overall representation of the graph is extracted, and the embedded vector representation set of the candidate structure graph set is generated. The node information update formula is as follows: ; in, In the graph neural network, the node In the The feature representation of the layer, Represented as a node exist The feature representation of the layer, Representation Node The set of adjacent nodes of Representation Node The number of adjacent edges of Representation Node The number of adjacent edges of Expressed as The weight matrix of the layer, Represents the current layer number of the graph neural network, Indicates the target node currently being updated. Representation and Node All neighboring nodes.

[0054] The semantic vector set and the graph structure are embedded in the set, all the sets are input into the structure graph generator module, and the combination mechanism of the structure generator and the discriminator is used to generate multiple structure graph candidate solutions. The generator constructs an architecture diagram with preliminary engineering logic, structural rationality and semantic consistency based on the semantic-component-structure graph embedding, adds semantic labels to the component nodes of each structure graph, and encapsulates them in a standard format, and finally generates a set of candidate architecture structure scheme structure graphs.

[0055] According to the automatic structural inspection process, the candidate structural scheme structural atlas is logically and engineering checked, and potential errors and structural weaknesses are automatically visually marked to obtain a list of structural problems and an initial structural sketch. The automatic structural inspection process is called to perform logical and engineering checks on the candidate structural atlas, and potential errors and structural weaknesses are automatically visually marked to obtain a list of structural problems. The steps are as follows: Parse the structure graph data of the candidate architecture structure graph set, convert it into the internal graph modeling format, establish a unified standard graph structure model, and parse and encapsulate the unified standard graph structure model to obtain a standardized structure graph representation set; Logical integration and engineering verification of the standardized structural diagram representation set to obtain a list of structural problems with the type, location and severity of each problem; Combined with the standardized structure diagram representation set and the structure problem list, problems are marked in the structure diagram and annotated structure diagram sketches are generated.

[0056] The structure diagram files in the candidate architecture structure diagram set are parsed through semantic rules and structural grammatical constraints to obtain component nodes, connection edges, dependency types and interface description information in each diagram. The extracted graph metadata is then input into the structure parsing engine and uniformly converted into a system-internal defined graph modeling format that includes unified node numbers, connection semantic types and deployment attribute fields. Finally, a set of standard format structure diagram models is obtained, which can ensure that each sub-graph must contain a deployment entry node, there must be a reasonable dependency path between all service components, and there must be no isolated nodes or dead loop paths.

[0057] A structural dependency matrix is ​​constructed for each structural diagram model, and a consistency analysis is performed in combination with the predefined engineering constraint rule library, which mainly includes: logical consistency analysis and engineering constraint verification. Each type of problem is then assigned a unique problem type code, and the severity of the problem is quantitatively rated according to the following rules: Level 1 problems (critical blocking types): such as backbone service disconnection and illegal circular dependencies; Level 2 problems (structural irregularities): such as incorrect module reuse and repeated interfaces; Level 3 problems (optimizable types): such as unused components and repeated logical paths. Finally, a list of structural problems is obtained, including problem type, location information and severity.

[0058] The list of structural problems is mapped to the original structural diagram, corresponding to the location node or connection edge of each problem, and a color and image annotation strategy is adopted according to the severity of the problem. The annotations in the diagram are then grouped into a unified layer to form an interactive structural sketch. Each diagram is bound to the corresponding list of problems, and finally a set of annotated initial structural sketches is generated.

[0059] By optimizing logical units, one-click optimization suggestions are made based on the list of structural problems to form the final optimized architecture solution.

[0060] Receive a list of structural problems, each of which includes: problem type, location of the components involved, and problem severity level. Enter the problem list into the optimization suggestion check engine, call the embedded optimization knowledge base, use semantic matching and rule mapping to correspond to the optimization suggestions for each problem, and obtain the optimization suggestion solution set. The optimization suggestion matching scoring formula is as follows: ; in, Indicates Structural problem items, Indicates Optimization suggestions, Indicates The semantic vector representation of the question is Expressed as The proposed semantic vector representation is Expressed as The first question Semantic similarity scores for the proposed

[0061] The suitability of each optimization suggestion is analyzed through the suitability scoring function to screen out the optimization suggestions that are safe and reasonable for application to the initial structural sketch, and the correction operation is triggered in the automatic structural diagram to obtain the updated architectural model.

[0062] The updated structure diagram model is input into the deployment packaging engine, and deployment-related metadata (such as computing resource tags and security policy levels) is added to each component. At the same time, historical projects and contextual preferences are combined to infer and generate the component runtime environment. Through the engineering packaging in the deployment packaging engine, a deployment template is automatically generated to finally obtain an optimized architecture structure diagram.

[0063] In summary, the present invention achieves this by: aligning the user intention vector with the system component knowledge base through a semantic mapping and reasoning processing unit, establishing a semantic-component graph, and using a graph neural network to model component dependencies, combined with an adversarial generation mechanism to generate diverse candidate structure graph solutions. The structure generation module breaks through the traditional static structure generation method based on rules or templates, realizes the dynamic conversion from semantic abstraction to structural instances, and introduces a two-way feedback mechanism of semantic constraints and structural optimization, so that the generated solution is both innovative and scalable while satisfying the expression of intent. Finally, the structure graph generation step in the structure generation module effectively improves the rationality, diversity and adaptability of the automatically generated structure, and provides core support for the automatic construction of architecture structures that meet target requirements.

[0064] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-modal software architecture intelligent design and optimization system, characterized by: include, The intention recognition module performs modal perception and preprocessing on the architectural design requirement information input by the user, converts it into structured semantic data, and optimizes and enhances the semantic expression by introducing a reinforcement learning strategy to form a structured user intention vector set; The structure generation module constructs the user intention vectors in the user intention vector set into a semantic-component alignment graph through the semantic mapping and reasoning processing unit, and combines graph structure modeling with adversarial generation algorithm to generate a candidate structure graph set; The structural verification module calls the automatic structural inspection process to perform logic and engineering verification on the candidate structural atlas, and obtains a list of structural problems by automatically visually marking potential errors and structural weaknesses; The output module is optimized by optimizing the logical units and providing one-click optimization suggestions based on the list of structural problems to form the final optimized architecture solution.

2. The multimodal software architecture intelligent design and optimization system according to claim 1, characterized in that: The steps of performing modal perception and preprocessing on the architecture design requirement information input by the user and converting it into structured semantic data are as follows: Receive various forms of user input, classify and encapsulate them according to the input modality, and form a multi-modal raw data set; Call the corresponding parsing module to perform signal recognition and conversion on the multimodal raw data set, and normalize it to obtain the semantic vector set after modality alignment; The modality-aligned semantic vector set is input into the context fusion module, and structured semantic data is generated by combining the user's historical behavior and modality credibility coefficient.

3. The multimodal software architecture intelligent design and optimization system according to claim 2, characterized in that: The various forms of user input refer to the user's voice input, natural language text input, graphic sketches, and gesture operations.

4. The multimodal software architecture intelligent design and optimization system according to claim 2, characterized in that: The corresponding analysis module is called to perform signal recognition and conversion on the multimodal raw data set. The steps are as follows: Convert language input into text to obtain a list of intended keywords and a functional sentence structure; Extract module nodes and connection relationships from the image sketch to obtain a set of connection paths and candidate matching semantic units of components; Identify the operation semantics from the gesture operation, and obtain the interactive operation list and the component identification and action intention correspondence table; Perform grammatical parsing, entity extraction and structure mapping on natural language text input to obtain a list of functional entities, technical attribute structures and non-functional requirement annotations.

5. The multimodal software architecture intelligent design and optimization system according to claim 1, characterized in that: The steps of optimizing and enhancing the semantic expression by introducing the reinforcement learning strategy to form a structured user intention vector set are as follows: The semantic unit codes in the structured semantic data are converted into semantic unit vectors, and the semantic unit vectors are format-aligned and semantic space-mapped to obtain a set of multimodal fusion semantic vectors; Retrieve the historical behavior data records of the current user, perform context fusion modeling with the multimodal fusion semantic vector set, and obtain the user context enhanced semantic vector set; The user context enhanced semantic vector set is input into the learning model, and the policy gradient reinforcement learning algorithm is applied to update the policy model parameters to obtain the user core intent set; The user's core intent set is mapped to a standardized terminology to obtain a structured intent-term alignment list, and the intent-term alignment list is classified according to intent type to obtain a structured user intent semantic vector.

6. The multimodal software architecture intelligent design and optimization system according to claim 1, characterized in that: The steps of combining graph structure modeling with adversarial generation algorithm to generate candidate structure graph sets are as follows: By using semantic mapping and reasoning, the structured user intent semantic vector is mapped to the target intent node in the semantic-component graph, and the candidate component set corresponding to the semantic node is screened from the architecture knowledge graph in combination with the semantic association scoring mechanism; According to the semantic mapping and reasoning processing unit and the semantic-component association information in the architecture knowledge graph library, the alignment relationship between intentions and components and the structural connection relationship between components are constructed, and the alignment relationship between intentions and components and the structural connection relationship between components are sorted through the structured graph construction engine to obtain a set of candidate component structure graphs; Using graph neural network, each candidate structure graph in the candidate component structure graph set is modeled to obtain an embedded graph structure vector set; Input a set of embedded graph structure vectors, generate multiple structural graph candidate solutions, add semantic labels to each node in each structural graph candidate solution, perform standard formatting and packaging, and obtain a set of candidate architecture structural solution structural graphs.

7. The multimodal software architecture intelligent design and optimization system according to claim 1, characterized in that: The automatic structural inspection process is called to perform logic and engineering verification on the candidate structural atlas, and potential errors and structural weaknesses are automatically visually marked to obtain a list of structural problems. According to the automatic structural inspection process, the structural atlas of the candidate architecture structure scheme is logically and engineering verified, and potential errors and structural weaknesses are automatically visually marked to obtain a list of structural problems and an initial architecture sketch. The steps are as follows: The candidate architecture structure graph set is converted into an internal graph modeling format by using a method based on semantic ontology mapping and structural specification template matching. The unified standard graph structure model is parsed and encapsulated to obtain a standardized structure graph representation set by combining the graph node type set and topology rules. Logical integration and engineering verification of the standardized structural diagram representation set to obtain a list of structural problems with the type, location and severity of each problem; Combined with the standardized structure diagram representation set and the structure problem list, the structure problems are marked in the standardized structure diagram identification set to generate annotated structure diagram sketches.

8. The multimodal software architecture intelligent design and optimization system according to claim 7, characterized in that: The method based on semantic ontology mapping and structural specification template matching automatically unifies the node semantics and topological structure in the original structural diagram into a standard modeling format by establishing semantic concept mapping rules and a structural diagram pattern template library.

9. The multimodal software architecture intelligent design and optimization system according to claim 1, characterized in that: The steps of optimizing the logic unit and making one-click optimization suggestions based on the list of structural problems are as follows: Analyze the list of structural problems item by item, classify them according to the problem type, and match optimization suggestions for each problem to obtain a set of optimization suggestion solutions; Determine the applicability of each optimization suggestion in the optimization suggestion set, and automatically modify the initial structure sketch according to the optimization suggestion to obtain an updated architecture structure diagram model; The updated architecture structure diagram model is semantically expanded, technology stack completed and engineered at the deployment level to obtain the final optimized architecture solution.

10. The multimodal software architecture intelligent design and optimization system according to claim 9, characterized in that: The applicability determination for each optimization suggestion in the optimization suggestion set refers to automatically determining whether the optimization suggestion is suitable for direct application in the architecture structure diagram model before applying each structural optimization suggestion.

Images