Query intention understanding and execution path optimization system and method based on multi-modal deep learning model

Abstract

The invention discloses a query intention understanding and execution path optimization system and method based on a multi-modal deep learning model, and the method comprises the steps: a multi-modal intention analysis module is used for receiving unstructured query input from a user, and uniformly extracting the unstructured query input as a standardized SQL; the cross-model intention recognition module is used for receiving the standardized SQL output by the multi-mode intention analysis module, recognizing a data entity, an operation type and a data model to which the data entity and the operation type relate in a query intention, and outputting an entity structure table and a query plan tree; the access mode analysis module is used for using the query plan tree, identifying a connection relationship between node objects in the tree by contrasting with an entity structure table, performing cost estimation on access of each node in the tree, and constructing a logic plan tree; and the path optimization generation module is used for executing generation, evaluation and screening of candidate execution paths on the basis of the logic plan tree output by the analysis module, and finally determining an optimal execution plan. According to the method, the problems of semantic analysis, cross-model query intention recognition and execution path intelligent selection of complex inputs such as natural languages in a multi-mode database can be solved, and the query accuracy and the execution efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of database management technology, and in particular to a system and method for query intent understanding and execution path optimization based on a multimodal deep learning model. Background Technology

[0002] With the rapid development of information technology, data types are becoming increasingly diversified, and traditional single-mode data management methods are struggling to meet the demands of modern complex applications. Multi-Model Databases have emerged to address this need, supporting multiple data models within a single database system, such as relational models, document models (e.g., JSON), key-value models, graph models, time-series models, and vector models. Through a unified query language interface or a converged query framework, multi-model databases enable users to efficiently process heterogeneous data sources within a single system, and have been widely applied in fields such as financial risk control, intelligent search, and the Industrial Internet of Things.

[0003] However, in practical applications, the query interface of multi-model databases still faces significant challenges. First, users need to understand the structure and syntax rules of different data models before querying, which presents a high learning and usage cost for non-professional users. Second, traditional database systems cannot effectively parse the underlying query intent when faced with diverse input methods such as natural language or API interactions, nor can they automatically generate corresponding standard query language. Furthermore, when handling mixed retrieval tasks such as join queries, multi-hop graph relationships, and vector approximation searches, the query optimizer of multi-model databases often fails to make efficient execution path selections based on the query intent, leading to decreased query performance.

[0004] The root of the aforementioned problems lies in the lack of a deep understanding of "query intent" in current multimodal database systems. This is particularly true when processing inputs such as natural language, semi-structured data (e.g., JSON), or nested API requests, where the system struggles to construct a unified semantic expression space. How to understand the intent implied in user input based on deep learning models and automatically parse query intent and optimize execution paths by combining heterogeneous data storage structures remains a technical bottleneck in current multimodal database research and industrial applications. This challenge not only involves feature alignment and fusion representation of multimodal data but also requires the system to possess dynamic generalization and adaptive capabilities at the levels of structure mapping, pattern recognition, and query planning. It is one of the core issues in the development of intelligent databases. Summary of the Invention

[0005] Purpose of the invention: This invention provides a query intent understanding and execution path optimization system and method based on a multimodal deep learning model, which solves the problems of semantic parsing of complex inputs such as natural language in multimodal databases, cross-model query intent recognition and intelligent selection of execution paths, thereby improving query accuracy and execution efficiency.

[0006] Technical Solution: The present invention discloses a query intent understanding and execution path optimization system based on a multimodal deep learning model, comprising: a multimodal intent parsing module, a cross-model intent recognition module, an access pattern analysis module, and a path optimization generation module; the multimodal intent parsing module receives unstructured query input from users and extracts it into standardized SQL; the cross-model intent recognition module receives the standardized SQL output by the multimodal intent parsing module, identifies the data entities, operation types, and their respective data models involved in the query intent, and outputs an entity structure table and a query plan tree; the access pattern analysis module uses the query plan tree to identify the connection relationships between node objects in the tree by referring to the entity structure table, and estimates the access cost of each node in the tree to construct a logical plan tree; the path optimization generation module generates, evaluates, and filters candidate execution paths based on the logical plan tree output by the analysis module, and finally determines the optimal execution plan.

[0007] Furthermore, unstructured query input includes natural language input (text questions, command statements), structured input (SQL fragments, form fields), semi-structured input (JSON, YAML, API parameters), and keyword / phrase input formats.

[0008] Furthermore, the multimodal intent parsing module includes an intent embedding unit, an intent alignment unit, and an intent diagnostic feedback unit;

[0009] The intent embedding unit uses BERT or Transformer models to extract 768-dimensional feature vectors of contextual intent in natural language, uses Embedding to perform structured encoding of the structured input to generate 256-dimensional feature vectors, uses serialized tree structure encoding to encode the semi-structured input to generate 512-dimensional feature vectors, and uses bag-of-words model to extract 128-dimensional feature vectors of keywords.

[0010] The intent alignment unit optimizes the representation space between different inputs, making the intent-equivalent query representations close in the vector space. It uses a multi-tower cross-modal alignment network, where each modality uses an independent encoding sub-network to project the feature vectors of different inputs onto the same 1024-dimensional intent vector space.

[0011] The intent diagnosis feedback unit checks whether the SQL structure output by the intent alignment unit meets the target syntax specification. If the detection passes, the standardized SQL is output to the next module. If the detection fails, a correction suggestion is automatically generated, and the multimodal intent parsing module is triggered to re-parse the original input based on the correction suggestion until the verification passes or the maximum number of retries is reached.

[0012] Furthermore, the cross-model intent recognition module extracts entity names, attribute fields, and relationship types from standardized SQL query statements, maps the entities to database metadata, determines their corresponding entities in relational models, graph models, or vector models, and outputs a list of linked entities. The entity structure table includes the original entity name, the mapped standard entity ID, the model type (relational / graph / vector), and key attribute fields.

[0013] Furthermore, the access pattern analysis module includes a connection relationship parsing unit and an access pattern judgment and cost estimation unit; the connection relationship parsing unit is used to identify the connection relationships between entities in the query statement based on the input structured SQL statement and entity structure table, and organize the identified relationships into a logical plan tree structure; the access pattern judgment and cost estimation unit is used to infer the access pattern and perform preliminary cost estimation for each node and its connection path in the logical plan tree.

[0014] Furthermore, access patterns include primary key query, range query, full table scan, TopK search, and vector approximation search. Figure 1 Skip queries and full-text index queries.

[0015] Furthermore, the path optimization generation module includes a path enumeration unit, a path cost evaluation and optimal path selection unit, and a plan reconstruction and optimization unit;

[0016] The path enumeration unit is used to perform full path enumeration on all possible node connection sequences based on the logical plan tree structure output by the access mode analysis module, and generate a set of candidate execution paths; each candidate path consists of a node sequence and a connection edge sequence, and is accompanied by a path cost feature vector sequence.

[0017] The path cost evaluation and optimal path selection unit is used to perform comprehensive cost evaluation and ranking of the candidate execution path set using a deep neural network (DNN) cost model; the DNN cost model takes the path cost feature vector sequence of the candidate path as input and outputs the global execution cost prediction value of the path.

[0018] The planning reconstruction and optimization unit is used to sort candidate paths and select the optimal path based on the DNN prediction values ​​and other evaluation indicators.

[0019] Accordingly, a query intent understanding and execution path optimization method based on a multimodal deep learning model includes the following steps:

[0020] Step 1: The multimodal intent parsing module receives unstructured query input, maps it to a unified intent space through the intent embedding unit and intent alignment unit, and generates initial standardized SQL; the initial standardized SQL is subjected to syntax specification detection and alignment verification with database metadata;

[0021] Step 2: Input the above standardized SQL into the cross-model intent recognition module to identify the data entities, attribute fields and operation types involved in the query intent, and further determine the data model type in the metadata to generate a structured entity structure table and query plan tree.

[0022] Step 3: Based on the query plan tree and entity structure table, perform access mode inference and preliminary cost estimation for each access node, and infer and perform preliminary cost estimation for the connection path between each node.

[0023] Step 4: Based on the logical connection tree and path cost, combine, reorder and score possible access paths, use a DNN network to evaluate the global cost of different candidate paths, and select the optimal path to generate the final execution plan.

[0024] Furthermore, in step 1, the step of extracting the intent feature vector of the user's unstructured query input includes:

[0025] For natural language input, use BERT or Transformer models to extract 768-dimensional feature vectors;

[0026] For structured input, Embedding encoding is used to extract 256-dimensional feature vectors;

[0027] For semi-structured input, a 512-dimensional feature vector is extracted using serialized tree structure encoding;

[0028] For keyword input, a bag-of-words model is used to extract a 128-dimensional feature vector.

[0029] Furthermore, in step 1, a cross-modal alignment network with a multi-tower model structure is used to project the extracted feature vectors of the unstructured query inputs into a shared high-dimensional intent vector space; in the shared intent space, a Transformer Decoder is used to decode the aligned intent vectors and generate standardized SQL statements.

[0030] It receives standardized SQL statements generated by the multimodal intent parsing module and performs matching and verification against the database's metadata structure. The verification includes: whether the table names and field names in the SQL statement exist in the metadata; whether the operation type (such as "graph query") matches the data model (such as "relational model") of the target entity; and whether there are any undefined join relationships.

[0031] If the verification passes, the SQL structure is considered consistent and correct, and it is output to the next module.

[0032] If the validation fails (e.g., there is semantic ambiguity, SQL syntax error, or entity or attribute misalignment), a structured correction suggestion will be automatically generated.

[0033] The system feeds back the correction suggestion to the multimodal intent parsing module, which then triggers the module to re-parse the user's original unstructured input a second time, based on the correction suggestion, until the validation passes or the maximum number of retries is reached.

[0034] Furthermore, in step 2, the input SQL statement is parsed to extract the table name, field name, and join condition from the statement block;

[0035] The table names (entity names) and field names obtained from the SQL parsing are matched with the database metadata; the metadata not only includes the entity definition, but also indicates the data model type (relational model, graph model, or vector model) to which each entity belongs.

[0036] Semantic aliases and field aliases are parsed and disambiguated, and implicit connection conditions across models are identified based on the model type in the metadata.

[0037] Based on the query operation type and connection conditions obtained from the parsing, each data entity is mapped to a node in the plan tree, and its model attributes are marked. The identified connection operations are mapped to parent-child relationship edges to generate the query plan tree structure.

[0038] Furthermore, in step 4, based on the logical connection diagram, the system generates multiple candidate sets of possible access paths, extracts multimodal features for each candidate path, and maps each modal feature to the same vector space to form a multimodal path cost vector.

[0039] A self-attention mechanism model is constructed to predict the global cost of each candidate path. The network input is a multimodal path representation, and the output is the predicted global execution cost of the path. During the training phase, the network performs supervised learning based on historical execution logs and real cost labels to continuously optimize the path cost prediction accuracy.

[0040] For all candidate paths, sort them according to the global cost predicted by DNN, and select the path with the highest score as the final execution plan candidate based on the comprehensive score results;

[0041] The optimal path is mapped to an executable sequence of operations, including access order, join strategy, index selection, and subquery execution method; the final execution plan is output to drive the actual query execution, and the execution results and prediction costs can be fed back to the training module for continuous optimization of the deep learning model.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The present invention adopts a cross-modal alignment network, which can uniformly model and parse multiple modal inputs such as natural language, structured data (such as CSV tables), and semi-structured data (such as JSON and API request parameters), and intelligently optimize the query execution plan by combining the understanding of query intent, thereby significantly improving the ease of use and execution efficiency of the query; (2) The present invention proposes a global cost prediction model based on deep learning. This model models the candidate path as a sequence of feature vectors and uses the self-attention mechanism to capture the deep dependency relationship between different nodes on the path, which can more accurately predict the global total cost, thereby generating the truly optimal execution strategy for complex cross-modal queries; (3) The present invention has created a unique parsing-diagnosis The closed-loop mechanism of disconnection-feedback-reparation means that after the system decodes the intent vector into SQL, it will automatically verify it with the metadata. When ambiguity or error is found, the system will automatically generate a correction suggestion. This suggestion will be used as a new context and concatenated with the user's original input to trigger a second automatic parsing. Users do not need to master the underlying data structure and query syntax to efficiently access heterogeneous data in the multi-model database using natural language or interactive API. At the same time, the database system can automatically sense the user's intent and adjust the query execution plan, which is especially suitable for high-complexity query scenarios such as graph data and vector data. (4) It greatly reduces the threshold for using multi-model databases and improves the system's capabilities in multi-source data integration, semantic retrieval and intelligent optimization. It has broad application prospects and significant engineering value. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the system structure of the present invention.

[0044] Figure 2 This is a schematic diagram of the multimodal intent parsing module structure of the present invention.

[0045] Figure 3 This is a schematic diagram of the cross-model intent recognition module structure of the present invention.

[0046] Figure 4 This is a schematic diagram of the access pattern analysis module of the present invention.

[0047] Figure 5 This is a schematic diagram of the path optimization generation module structure of the present invention. Detailed Implementation

[0048] like Figure 1 As shown, a query intent understanding and execution path optimization system based on a multimodal deep learning model includes:

[0049] The multimodal intent parsing module is responsible for receiving and understanding various forms of user input, aligning them through a deep learning model, and converting them into unified, validated, standardized SQL.

[0050] The cross-model intent recognition module is responsible for parsing standardized SQL, identifying which data model (relational, graph, vector) each data entity involved in the query belongs to, and constructing an initial query plan tree.

[0051] The access pattern analysis module is responsible for analyzing the access patterns of each node in the query plan tree and estimating its local cost, and constructing the logical plan tree.

[0052] The path optimization generation module is responsible for enumerating all possible execution paths and using a deep neural network (DNN) to perform a global cost evaluation on paths containing multimodal features, ultimately generating the optimal execution plan.

[0053] like Figure 2 The diagram shown is a detailed structural schematic of the multimodal intent parsing module of the present invention, which is the key to realizing multimodal input understanding.

[0054] (1) Extracting multimodal inputs and processing heterogeneous inputs: Different inputs are first processed by a specific encoder to extract initial feature vectors:

[0055] For natural language input, pre-trained BERT or Transformer models are used for contextual intent encoding to extract 768-dimensional feature vectors;

[0056] For structured input (SQL fragments, forms): Learnable embedding layers are used to encode SQL keywords, table names, and field names, generating 256-dimensional feature vectors.

[0057] For semi-structured inputs, they are first serialized and then fed into the Tree-LSTM serialization tree encoder model to generate 512-dimensional feature vectors.

[0058] For keyword input, the traditional bag-of-words model is used to extract a sparse, low-dimensional 128-dimensional feature vector.

[0059] (2) In order to solve the problem of different vector dimensions and spatial heterogeneity, a cross-modal alignment network was adopted.

[0060] The cross-modal alignment network employs a multi-tower network structure. Each input modality corresponds to an independent encoding tower.

[0061] Each input modality (natural language, structured, semi-structured, keyword) has an independent encoding tower. The weights of these towers are not shared because they process completely different distributions of input features. Each tower is implemented using a three-layer fully connected network with 1024 units per layer, followed by ReLU as the activation function and a Dropout layer with a Dropout probability p=0.1. The final output of all towers is a 1024-dimensional vector, which resides in the "shared intent vector space".

[0062] This invention employs triplet loss as the contrastive learning loss. The training objective is to minimize the geometric distance between semantically equivalent queries in the 1024-dimensional space, while maximizing the distance between semantically irrelevant queries. Specifically, in each step of training, a triplet is sampled from the training data:

[0063] Anchor point: A query input, such as a natural language question, is output as a 1024-dimensional vector after passing through its encoding tower;

[0064] Positive samples: Inputs that are semantically equivalent to anchors but come from different structures, such as the 1024-dimensional vector output by the standard SQL corresponding to the question after passing through its encoding tower.

[0065] Negative samples: Inputs that are not semantically related to the anchor, such as 1024-dimensional vectors of other SQL queries in the same batch.

[0066] Triple loss function The definition is as follows:

[0067]

[0068] in This represents the Euclidean distance (L2 distance) between two vectors. It is a preset boundary hyperparameter. Finally, a 1024-dimensional intent feature vector is output.

[0069] (3) Decoding and Feedback (Intent Diagnosis Feedback Unit)

[0070] The 1024-dimensional intent vector is fed into a Transformer Decoder, which consists of a stack of six decoder layers. Each layer contains masked multi-head self-attention and multi-head cross-attention. Masked multi-head self-attention ensures that when predicting the i-th word, it can only rely on the output of the (i-1)-th word and above. For multi-head cross-attention, and All are composed of this 1024-dimensional Generated through linear projection. This comes from the output of the previous masked self-attention step. The output of the top layer of the decoder is fed into a final linear layer, the output dimension of which is equal to the size of the SQL vocabulary. Finally, a softmax function is used to generate a probability distribution for each word in the vocabulary, and the model selects the word with the highest probability as the output of this step. This process is repeated until the generated sequence ends, resulting in a standardized SQL statement. This SQL statement is not executed immediately but is sent to the intent diagnostic feedback unit for verification.

[0071] The system uses a lightweight SQL parser to perform lexical and syntactic analysis on the generated SQL statements, constructing an abstract syntax tree (AST). All identifiers, namely table names and field names, are extracted from the AST.

[0072] The system will match the extracted list of identifiers with the database metadata.

[0073] If the test passes: the SQL is valid, and the result is output to the next module.

[0074] If the detection fails: This unit will automatically generate a correction suggestion. It uses edit distance as the string similarity algorithm for fuzzy matching and generates a suggestion. This correction suggestion, along with the user's original input, is concatenated using text and separated by a special delimiter [SEP], and then sent back to step 1 of this module for a second parsing.

[0075] like Figure 3 The diagram shown is a detailed structural schematic of the cross-model intent recognition module of the present invention. This module receives validated standardized SQL output by the escaping module.

[0076] The SQL is parsed using a syntax parser, extracting all entities (table names), attributes (field names), and join conditions from clauses such as SELECT, FROM, and WHERE. The system maintains a global metadata directory that stores table structures and identifies the data model to which each entity belongs. The entity identification and linking unit in this module matches the entities parsed from the SQL with the metadata directory to accurately identify whether the query is a cross-model query. If all entities belong to the same model, the single-model optimization process begins; otherwise, cross-model optimization is performed.

[0077] This module outputs an entity structure table that indicates all entities involved in this query and their data model types, as well as an initial query plan tree labeled with model types.

[0078] like Figure 4The diagram shown is a detailed structural schematic of the access pattern analysis module of the present invention. This module receives... Figure 3 The output includes annotated query plan tree and entity structure table.

[0079] Based on the tree structure, the logical plan tree is constructed by identifying the connection relationships between nodes. The access mode judgment and cost estimation unit of this module traverses each node of the logical plan tree, infers all possible access modes based on its data model type, and estimates the local cost.

[0080] For relational model nodes: infer whether it is a primary key query, an index query, or a full table scan.

[0081] For graph model nodes: the inference is Figure 1 Skip query or subgraph matching.

[0082] For vector model nodes: Infer whether to use vector approximation search or TopK retrieval.

[0083] An edge represents the flow of data from one operation node (source node) to the next operation node (target node). The cost estimation unit in this module calculates the combined transmission and transformation cost for each edge through the following steps:

[0084] First, the system obtains statistical information such as the histogram of the source nodes, and calculates the selection factor based on the query conditions of the source nodes. Calculate the estimated number of rows This value represents the number of data objects that flow through this edge.

[0085] The system checks the data model types of the source node and the target node to determine the transmission type.

[0086] The main costs of same-mode transmission are memory copying or local I / O.

[0087] Cross-modal transmission costs include data serialization, memory copying, and object format conversion.

[0088] The system uses the following formula to calculate the total cost of the edges:

[0089] in The memory overhead for transferring unit data between different execution engines. To determine the unit transformation cost, the system maintains a transformation cost matrix, based on... and From the table:

[0090] If it is a relational model Graphical model: Includes hash lookup and object reconstruction overhead for mapping relational IDs to internal node IDs of the graph engine.

[0091] If it is a relationship Vector model: Includes the computational overhead of converting relations into vectors.

[0092] If it is a graph / vector Relational model: It includes the overhead of expanding complex graph paths or high-dimensional vector results into standard two-dimensional relational tables.

[0093] The system will calculate Mark it on the corresponding connection edge in the logical plan tree.

[0094] The system attaches a preliminary cost estimate to each node and each edge (data transmission) of the logical plan tree.

[0095] like Figure 5 The diagram shown is a detailed structural schematic of the path optimization generation module of the present invention.

[0096] Path enumeration unit based on Figure 4 The logical plan tree is generated using a dynamic programming algorithm to produce all possible execution paths.

[0097] For each candidate execution path, the system extracts a multimodal feature vector to describe the path.

[0098] For relational features: statistical information on the relational operations involved in the path;

[0099] For graph features: statistical information on the graph operations involved;

[0100] For vector features: statistical information on the vector operations involved;

[0101] For edge features: edge cost between models;

[0102] By splicing Figure 4 The local cost, edge cost, and statistical information are calculated and zeroed up to a unified 128 dimensions;

[0103] Input multimodal feature vectors into a deep neural network based on a self-attention mechanism to predict the global cost of a path.

[0104] The input vector first passes through an embedding layer, then is fed into a 6-layer Transformer encoder block. Each block contains an 8-head self-attention layer and a feedforward network. ReLU is used as the activation function. The final output vector sequence of the encoder stack is first average pooled, then fed into a fully connected regression layer, ultimately outputting a prediction of the global total cost of the candidate path.

[0105] The path cost evaluation and optimal path selection unit scores and ranks all candidate paths and selects the path with the lowest predicted cost.

[0106] The planning, refactoring, and optimization unit transforms this optimal path into a final, executable physical execution plan. This unit can then feed back the actual cost of this execution to the training module for continuous model optimization.

[0107] After the query execution engine actually executes the optimal plan selected by this module, the monitoring component collects the actual cost of that plan. The actual cost includes, but is not limited to, total query time (ms), actual I / O read / write volume, and CPU consumption. The path feature sequence used for prediction in this query, the cost initially predicted by the DNN model, and the actual cost are sent to the historical query log. The model is then retrained periodically in batches to balance performance and cost. For example, retraining is triggered whenever 1000 new query records are collected, or at fixed intervals.

[0108] The training module extracts all historical data from the data pool as the training set. The DNN model is then retrained on the training set, using MSLE as the loss function. The weights of the DNN model are updated through training. Through optimization units, the DNN model can automatically learn dynamic changes in the database.

Claims

1. A query intent understanding and execution path optimization system based on a multi-modal deep learning model, characterized in that, Comprise: A multi-modal intent parsing module, a cross-model intent recognition module, an access pattern analysis module and a path optimization generation module; The multi-modal intent parsing module is used for receiving unstructured query input from a user and uniformly extracting it into standardized SQL; the cross-model intent recognition module is used for receiving the standardized SQL output by the multi-modal intent parsing module, identifying data entities involved in the query intent, operation types and data models to which they belong, and outputting an entity structure table and a query plan tree; the access pattern analysis module is used for using the query plan tree, identifying connection relationships between node objects in the tree against the entity structure table, and cost estimating access to each node in the tree, to construct a logical plan tree; the path optimization generation module is used for generating, evaluating and screening candidate execution paths on the basis of the logical plan tree output by the analysis module, and finally determining an optimal execution plan. 2.The multi-modal deep learning model based query intent understanding and execution path optimization system of claim 1, wherein, The multi-modal intent parsing module comprises an intent embedding unit, an intent alignment unit and an intent diagnosis feedback unit; The intent embedding unit uses a BERT or Transformer model to extract a 768-dimensional feature vector of the context intent in natural language, uses Embedding to generate a 256-dimensional feature vector by structurally encoding structured data, uses a serialized tree structure encoding to generate a 512-dimensional feature vector by encoding semi-structured input, and uses a bag-of-words model to extract a 128-dimensional feature vector of keywords; The intent alignment unit optimizes the representation space between different inputs so that queries expressing the same intent are close in the vector space, uses a multi-tower structure cross-modal alignment network, each modality uses an independent encoding subnetwork, and projects the feature vectors of different inputs into the same 1024-dimensional intent vector space; The intent diagnosis feedback unit detects whether the SQL structure output by the intent alignment unit meets the target syntax specification, and if the detection passes, the standardized SQL is output to the next module; If the detection fails, a correction suggestion is automatically generated, and the multi-modal intent parsing module is triggered to reparse the original input in combination with the correction suggestion until the verification passes or the maximum number of retries is reached. 3.The multi-modal deep learning model based query intent understanding and execution path optimization system of claim 1, wherein, The cross-model intent recognition module extracts entity names, attribute fields and relationship types in the query statement from the standardized SQL, maps the entities to database metadata, determines the corresponding entities in the relational model, graph model or vector model, and outputs a linked entity list; the entity structure table includes the original entity name, the mapped standard entity ID, the model type to which it belongs and the key attribute field. 4.The multi-modal deep learning model based query intent understanding and execution path optimization system as claimed in claim 1, wherein, The access pattern analysis module comprises a connection relationship analysis unit and an access pattern judgment and cost estimation unit; The connection relationship analysis unit is used for identifying the connection relationships between entities in the query statement based on the input structured SQL statement and the entity structure table, and organizing the identified relationships into a logical plan tree structure; the access pattern judgment and cost estimation unit is used for inferring the access pattern and preliminarily estimating the cost of each node and its connection path in the logical plan tree. 5.The multi-modal deep learning model based query intent understanding and execution path optimization system as claimed in claim 1, wherein, The path optimization generation module comprises a path enumeration unit, a path cost evaluation and optimal path selection unit and a plan reconstruction and optimization unit; a path enumeration unit configured to perform full path enumeration on all possible node connection sequences based on a logical plan tree structure output by the access pattern analysis module, to generate a candidate execution path set; each candidate path is composed of a node sequence and a connection edge sequence, and is attached with a path cost feature vector sequence; a path cost evaluation and optimal path selection unit configured to use a deep neural network (DNN) cost model to comprehensively evaluate and sort the candidate execution path set; the DNN cost model takes the path cost feature vector sequence of the candidate path as input and outputs a global execution cost prediction value of the path; a plan reconstruction and optimization unit configured to sort the candidate paths based on the DNN prediction value and other evaluation indicators, and select an optimal path. 6.The method based on the multi-modal deep learning model-based query intent understanding and execution path optimization system according to claim 1, characterized in that, The method comprises the following steps: Step 1: The multi-modal intent analysis module receives an unstructured query input, maps it to a unified intent space through an intent embedding unit and an intent alignment unit, and generates an initial standardized SQL; performs syntax specification detection on the initial standardized SQL and performs alignment verification with the database metadata; Step 2: Input the above standardized SQL into the cross-model intent recognition module to identify the data entities, attribute fields and operation types involved in the query intent, and further determine the data model type thereof in the metadata to generate a structured entity structure table and a query plan tree; Step 3: Based on the query plan tree and the entity structure table, the access mode of each access node is inferred and a preliminary cost estimate is made, and the connection path between nodes is inferred and a preliminary cost estimate is made; Step 4: Based on the logical connection tree and the path cost, the possible access paths are combined, reordered and scored, the global cost of different candidate paths is evaluated using a DNN network, and the optimal path is selected to generate a final execution plan. 7.The query intent understanding and execution path optimization method based on the multi-modal deep learning model according to claim 1, as claimed in claim 6, wherein, In step 1, the intent feature vector extraction step of the user's unstructured query input includes: For natural language input, use BERT or Transformer model to extract 768-dimensional feature vector; For structured input, use Embedding encoding to extract 256-dimensional feature vector; For semi-structured input, use serialized tree structure encoding to extract 512-dimensional feature vector; For keyword input, use bag-of-words model to extract 128-dimensional feature vector. 8.The query intent understanding and execution path optimization method based on the multi-modal deep learning model according to claim 1, as claimed in claim 6, wherein, In step 1, use a cross-modal alignment network with a multi-tower model structure to project the extracted feature vectors of the unstructured query input into a shared high-dimensional intent vector space; in the shared intent space, use a Transformer Decoder decoder to decode the aligned intent vectors to generate a standardized SQL statement; Receive the standardized SQL statement generated by the multi-modal intent analysis module and match it with the metadata structure of the database; The verification content includes: whether the table name and field name in the SQL statement exist in the metadata; whether the operation type matches the data model of the target entity; whether there is an undefined connection relationship; If the verification is passed, it is determined that the SQL structure is consistent and correct, and it is output to the next module; If the check fails, automatically generate a structured correction suggestion; The system feeds the correction suggestion back to the multi-modal intent analysis module and triggers the module to combine the correction suggestion to re-analyze the user's original unstructured input for a second time until the check passes or the maximum number of retries is reached. 9.The query intent understanding and execution path optimization method based on the multi-modal deep learning model according to claim 1, as claimed in claim 6, wherein, In step 2, the input SQL statement is parsed to extract table names, field names, and connection conditions in the statement block; The table names and field names obtained by parsing the SQL are matched with the database metadata; the metadata not only contains entity definitions, but also specifies the data model type to which each entity belongs; The semantic alias and field alias are parsed and disambiguated, and the implicit connection conditions across models are identified according to the model type in the metadata; According to the query operation type and connection conditions obtained by parsing, each data entity is mapped to a node in the plan tree and its model attribute is marked, and the identified connection operation is mapped to a parent-child relationship edge to generate a query plan tree structure. 10.The query intent understanding and execution path optimization method based on the multi-modal deep learning model according to claim 1, as claimed in claim 6, wherein, In step 4, according to the logical connection graph, the system generates a set of multiple possible access path candidates, extracts multi-modal features for each candidate path, and maps each modal feature to the same vector space to form a multi-modal path cost vector; An auto-attention mechanism model is constructed to predict the global cost of each candidate path, with the network input being the multi-modal path representation and the output being the predicted global execution cost of the path; in the training phase, the network is supervised by historical execution logs and real cost labels to continuously optimize the path cost prediction accuracy; All candidate paths are sorted based on the global cost predicted by the DNN, and the path with the highest score is selected as the final execution plan candidate based on the comprehensive score results; The optimal path is mapped to an executable operation sequence, including access order, connection strategy, index selection, and subquery execution mode; The final execution plan is output to drive the actual query execution, and the execution result and predicted cost are fed back to the training module for continuous optimization of the deep learning model.

Images