Building electromechanical BIM model information rapid retrieval method and system

Abstract

The invention discloses a building electromechanical BIM model information rapid retrieval method and system, and the method comprises the steps: generating composite retrieval parameters fusing semantic keywords and three-dimensional coordinate constraints according to a multi-mode retrieval instruction inputted by a user; on the basis of the composite retrieval parameters, constructing a dynamic search space by utilizing a hierarchical graph convolutional network, and generating a candidate model index structure of multi-dimensional feature coding; inputting the candidate model index into a multi-objective optimization engine, performing real-time optimization on a search path by adopting a dynamic pruning algorithm driven by reinforcement learning, and outputting a candidate model set of which the confidence coefficient is higher than a preset confidence threshold after pruning; and on the basis of the candidate model set, associated equipment nodes are expanded through a knowledge graph embedding and complementing technology, and an enhanced retrieval result set containing the hidden associated equipment is generated. By utilizing the embodiment of the invention, efficient, multi-dimensional and multi-modal information accurate positioning and quick retrieval can be realized in a large-scale complex BIM model.

Description

Technical Field

[0001] The present invention belongs to the technical field of data retrieval, and particularly relates to a method and system for quickly retrieving building mechanical and electrical BIM model information. Background Art

[0002] With the deep development of the construction industry towards digitization and informatization, building information modeling (BIM), as the core technology for realizing the whole life cycle management of buildings, has been widely applied in stages such as design, construction, operation and maintenance. The BIM model integrates rich multi-disciplinary information such as space, structure, equipment, pipelines, etc., providing strong support for intelligent building management. However, with the continuous expansion of project scale and the improvement of model complexity, how to quickly and accurately retrieve the required mechanical and electrical information has become an important technical problem to ensure design optimization, construction efficiency and operation and maintenance management. Traditional BIM information retrieval mostly relies on keyword search or simple spatial query, with defects such as slow response, low accuracy, and difficulty in spanning multi-source heterogeneous data, making it difficult to meet the complex and changeable query requirements in actual engineering projects. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for quickly retrieving building mechanical and electrical BIM model information to solve the deficiencies in the prior art, and to be able to achieve precise positioning and rapid retrieval of efficient, multi-dimensional, and multi-modal information in large-scale complex BIM models.

[0004] An embodiment of the present application provides a method for quickly retrieving building mechanical and electrical BIM model information, and the method includes: According to the multi-modal retrieval instruction input by the user, parse the semantic intention and spatial orientation through the spatial topological relationship map, and generate a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints; Based on the composite retrieval parameter, use a hierarchical graph convolutional network to construct a dynamic search space, and the hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model; Input the candidate model index into a multi-objective optimization engine, and use a dynamic pruning algorithm driven by reinforcement learning to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a candidate model set with a confidence level higher than the preset confidence threshold after pruning; Based on the candidate model set, expand the associated equipment nodes through knowledge graph embedding completion technology to generate an enhanced retrieval result set containing implicit associated equipment.

[0005] Optionally, the step of, according to the multi-modal retrieval instruction input by the user, parse the semantic intention and spatial orientation through the spatial topological relationship map, and generate a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints, includes: According to the voice, gesture, and text retrieval instructions input by the user, a multimodal fusion network is used to separately extract the voice text vector, gesture trajectory coordinates, and text keyword vector; Based on the voice text vector and the text keyword vector, a bidirectional attention mechanism is used to calculate the semantic association weight, and an intent classifier is used to identify the device type, functional attributes, and operation requirements, generating a structured semantic intent description; According to the gesture trajectory coordinates and the BIM model space coordinate system, a dynamic projection algorithm is used to map the gesture trajectory into a three-dimensional space range constraint, and the device orientation hierarchy is parsed in combination with the floor topology relationship map, generating a spatial orientation code; The structured semantic intent description and the spatial orientation code are input into an adversarial training network, and semantic-spatial ambiguity is eliminated through feature adversarial alignment, outputting composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints.

[0006] Optionally, based on the composite retrieval parameters, a hierarchical graph convolutional network is used to construct a dynamic search space. The hierarchical graph convolutional network extracts the topological connection features and semantic attribute features of the device pipelines in the BIM model, generating a candidate model index structure with multi-dimensional feature encoding, including: According to the connection relationship of the device pipelines in the BIM model, a hierarchical topological graph structure is constructed, where the devices are nodes, the pipeline connections are edges, and the floor partitions are supernodes, generating a multi-granularity topological graph; The device model, function parameters, and operation and maintenance record text are extracted, and semantic embedding vectors are generated through a pre-trained language model and bound to the node attributes of the multi-granularity topological graph, generating an enhanced topological graph with semantic attributes; On the enhanced topological graph, hierarchical graph convolutional operations are performed. In the first layer, the features of directly adjacent devices are aggregated, in the second layer, the features of cross-floor supernodes are aggregated, and in the third layer, the features of system-level functional groups are aggregated, generating a multi-scale feature matrix; Based on the semantic keywords and spatial constraints in the composite retrieval parameters, the node weights in the multi-scale feature matrix are calculated, and the contribution degrees of the features at each level are dynamically adjusted through a gating mechanism, generating weighted feature encoding; The weighted feature encoding is concatenated with the device spatial coordinates, and dimensionality reduction and compression are performed through an autoencoder to construct a candidate model index structure with multi-dimensional feature encoding.

[0007] Optionally, the candidate model index is input into a multi-objective optimization engine, and a dynamic pruning algorithm driven by reinforcement learning is used to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, a set of candidate models with a confidence level higher than the preset confidence threshold after pruning is output, including: According to the candidate model index structure, the initial search path tree is constructed with the device with the highest semantic similarity as the starting point, and a search state space including path node sequences and branch weights is generated; A deep Q network is used to define the state-action space, where the state is the current path node feature, the action is to select the next hop node or terminate the search, and the Q value table is updated through the Bellman equation to generate a search strategy; During the search process, the topological distance weight and semantic similarity weight are calculated in real time, and the joint loss value is generated through linear weighting; When the joint loss value of the search path exceeds the preset loss threshold, the pruning strategy is triggered to retain the Top-K path branches with the lowest loss value, prune redundant branches, and generate a streamlined search path set; Traverse the simplified search path set, extract the path endpoint device nodes, sort them by joint loss value, and output a set of candidate models whose confidence is higher than the preset confidence threshold.

[0008] Optionally, based on the candidate model set, the associated device nodes are expanded by knowledge graph embedding completion technology to generate an enhanced search result set containing implicit associated devices, including: Map the device nodes in the candidate model set to the pre-built electromechanical knowledge graph, learn the embedded representation of the device nodes in the functional, physical and logical dimensions through the graph neural network, and generate the device relationship vector; Based on the equipment relationship vector, the random walk algorithm is used to traverse the knowledge graph, extract the implicit association paths that have engineering coupling relationships with the candidate equipment, and generate a list of potential associated equipment; Calculate the semantic relevance and topological proximity between the potential associated devices and the candidate models, retain the devices with a relevance greater than 0.7 through threshold filtering, and generate a hidden associated subset; The implicitly associated subset is merged with the original candidate model set, sorted by comprehensive association degree, and an enhanced retrieval result set containing implicitly associated devices is output.

[0009] Optionally, the initial search path tree is constructed based on the candidate model index structure, taking the device with the highest semantic similarity as the starting point, and generating a search state space including a path node sequence and branch weights, including: According to the device semantic embedding information in the candidate model index, the cosine similarity and the Jaccard coefficient are fused through a hybrid similarity model to generate a device semantic association matrix representing the semantic similarity between devices. Extract the device node with the highest semantic similarity from the device semantic association matrix as the search starting point, create an initial path tree with the node as the root node based on the topological connection relationship of the BIM model, and generate the root node attributes and hierarchical tags; According to the direct connections, cross-floor connections, and system-level connection relationships of devices in the topology graph, expand the child nodes of the path tree in the order of hierarchical priorities. Record the topological hop count and connection type of the parent and child nodes during each expansion to generate multi-level path branches. Based on the device semantic association matrix and topological hop count, calculate the initial weights of each path branch through a dynamic attenuation function to generate a weighted search state space.

[0010] Optionally, based on the device relationship vector, traverse the knowledge graph using the random walk algorithm to extract implicit association paths that have an engineering coupling relationship with the candidate device, and generate a list of potential associated devices, including: According to the embedding values in the functional, physical, and logical dimensions of the device relationship vector, calculate the transition probability of the edges in the knowledge graph, set the preference weight for the walking direction, and generate the initialization parameters for the random walk with preference. Starting from the candidate device node, dynamically select the next-hop node based on the transition probability. After each transfer, adjust the subsequent transfer probability according to the semantic consistency of the current path. If the consistency is lower than the preset consistency threshold, restart the walk to generate highly associated path segments. For the path segments generated by the walk, calculate the path confidence through a pre-trained graph attention network, filter out the segments with a confidence lower than 0.85, and retain the set of valid paths. Extract the nodes that do not appear in the original candidate set and share functions or physical constraints with the candidate device from the set of valid paths to generate a list of potential associated devices.

[0011] Another embodiment of the present application provides a rapid retrieval system for building mechanical and electrical BIM model information. The system includes: A parsing module for parsing the semantic intention and spatial orientation through the spatial topology relationship graph according to the multi-modal retrieval instruction input by the user, and generating composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints. A construction module for constructing a dynamic search space based on the composite retrieval parameters using a hierarchical graph convolutional network. The hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the device pipelines in the BIM model. An optimization module for inputting the candidate model index into a multi-objective optimization engine, and using a dynamic pruning algorithm driven by reinforcement learning to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a set of candidate models with a confidence higher than the preset confidence threshold after pruning. A generation module for expanding the associated device nodes based on the set of candidate models through knowledge graph embedding completion technology to generate an enhanced retrieval result set that includes implicit associated devices.

[0012] Another embodiment of the present application provides a storage medium, in which a computer program is stored. Wherein, the computer program is configured to execute the method described in any one of the above when running.

[0013] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the method described in any one of the above.

[0014] Compared with the prior art, a method for quickly retrieving building electrical and mechanical BIM model information provided by the present invention generates composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints according to a multi-modal retrieval instruction input by a user; based on the composite retrieval parameters, a hierarchical graph convolutional network is used to construct a dynamic search space to generate a candidate model index structure with multi-dimensional feature encoding; the candidate model index is input into a multi-objective optimization engine, and a dynamic pruning algorithm driven by reinforcement learning is used to optimize the search path in real time, and a set of candidate models with a confidence level higher than a preset confidence threshold after pruning is output; based on the set of candidate models, the knowledge graph embedding completion technology is used to expand associated device nodes to generate an enhanced retrieval result set including implicit associated devices, so as to be able to achieve efficient, multi-dimensional, and multi-modal information accurate positioning and rapid retrieval in a large-scale complex BIM model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a hardware structure block diagram of a computer terminal for a method for quickly retrieving building electrical and mechanical BIM model information provided by an embodiment of the present invention; Figure 2 It is a schematic flow chart of a method for quickly retrieving building electrical and mechanical BIM model information provided by an embodiment of the present invention; Figure 3 It is a schematic structural diagram of a system for quickly retrieving building electrical and mechanical BIM model information provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be construed as a limitation of the present invention.

[0017] An embodiment of the present invention first provides a method for quickly retrieving building electrical and mechanical BIM model information. This method can be applied to an electronic device, such as a computer terminal, specifically, such as an ordinary computer, etc.

[0018] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 It is a hardware structure block diagram of a computer terminal for a method for quickly retrieving building electrical and mechanical BIM model information provided by an embodiment of the present invention. As Figure 1As shown in the figure, the computer device includes a processor, a memory, and a network interface connected via a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0019] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which when executed, can cause the processor to execute any one of the rapid retrieval methods for building mechanical and electrical BIM model information.

[0020] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0021] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, it can cause the processor to execute any one of the rapid retrieval methods for building mechanical and electrical BIM model information.

[0022] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in the figure is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0023] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0024] See Figure 2 , an embodiment of the present invention provides a rapid retrieval method for building mechanical and electrical BIM model information, which may include the following steps: S201, according to the multi-modal retrieval instruction input by the user, parse the semantic intention and spatial orientation through the spatial topology relationship map, and generate a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints; This step analyzes multimodal instructions such as the speech, gestures, and text input by the user, and combines with the spatial topology relationship map of the BIM model (such as floor distribution, equipment connection relationship) to identify the user's semantic needs (such as "find the air conditioning unit on the second floor") and spatial orientation (such as the three-dimensional area circled by the gesture). The semantic intention analysis uses a bidirectional attention mechanism to extract keywords, and the spatial orientation mapping converts the gesture trajectory into three-dimensional coordinate constraints through a dynamic projection algorithm. Finally, through an adversarial training network, the semantic and spatial features are aligned to generate composite retrieval parameters, solving the problem of semantic and spatial separation in traditional retrieval, converting the user's vague intention into precise structured retrieval conditions, and improving the multi-dimensional adaptability of the retrieval input. For example, when the user circles a certain area with a gesture and describes "distribution box" verbally, the system can automatically associate all distribution equipment in that area, avoiding the inefficiency of global search.

[0025] S202. Based on the composite retrieval parameters, use a hierarchical graph convolutional network to construct a dynamic search space. The hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model. The hierarchical graph convolutional network abstracts the equipment pipelines in the BIM model into a multi-granularity topological graph (equipment as nodes, connections as edges, floors as super-nodes), and aggregates adjacent features (such as directly connected equipment), cross-layer features (such as cross-floor pipelines), and system-level features (such as the overall topology of the HVAC system) layer by layer. At the same time, semantic attributes such as equipment models and operation and maintenance records are fused, and the weights of different-level features are dynamically adjusted through a gating mechanism. Finally, a multi-dimensional index integrating topology and semantics is generated, breaking through the limitations of traditional BIM retrieval that only relies on labels or coordinates. Through the joint encoding of topology and semantics, an index structure with high representational ability is constructed. For example, when retrieving "valves connected to a certain water pump", the network can automatically capture indirectly connected equipment and associate its functional parameters, improving the accuracy of complex queries.

[0026] S203. Input the candidate model index into a multi-objective optimization engine, and use a dynamic pruning algorithm driven by reinforcement learning to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a set of candidate models whose confidence after pruning is higher than the preset confidence threshold. The multi-objective optimization engine uses the device with the highest semantic similarity as the search starting point to construct an initial path tree, and dynamically selects search actions (expanding nodes or terminating) through a Deep Q-Network (DQN). During the search process, the combined loss value of the topological distance (such as the pipeline hop count) and the semantic similarity (such as the device function matching degree) is calculated in real time. When the loss exceeds the threshold, the low-weight branches are pruned, and the Top-K high-confidence paths are retained. Finally, an optimized candidate set is output to solve the problem of low search efficiency in large-scale BIM models, and the search depth and breadth are dynamically balanced through reinforcement learning. For example, when searching for "power distribution equipment with high failure rate", the algorithm can quickly prune non-critical paths, focus on the devices associated with historical operation and maintenance records, and shorten the response time.

[0027] S204. Based on the candidate model set, expand the associated device nodes through knowledge graph embedding completion technology to generate an enhanced retrieval result set containing implicit associated devices.

[0028] Map the candidate devices to a pre-constructed electromechanical knowledge graph (including functional, physical, and logical associations), learn the relationship embedding vectors of the devices through a graph neural network, and use the random walk algorithm to extract implicit association paths (such as devices sharing the same control logic but having no direct connection). Through threshold filtering (such as association degree > 0.7) and comprehensive sorting, the implicit devices are merged into the retrieval results, breaking through the explicit association limit, mining the engineering coupling relationships between devices, and providing more comprehensive decision support. For example, when retrieving a certain air conditioning unit, the results can be extended to the humidity sensor that it depends on but is not directly connected, assisting maintenance personnel in predicting potential failures.

[0029] Specifically, according to the multi-modal retrieval instructions input by the user, parse the semantic intention and spatial orientation through the spatial topological relationship graph to generate composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints, including: According to the voice, gesture, and text retrieval instructions input by the user, separately extract the voice text vector, gesture trajectory coordinates, and text keyword vector through a multi-modal fusion network; The multi-modal fusion network adopts a channel-separated processing architecture, and designs dedicated feature extraction modules for voice, gesture, and text respectively: Voice processing channel: Use a pre-trained Wav2Vec 2.0 model (including 12 layers of Transformer, with a hidden layer dimension of 768) to convert the voice input into a frame-level acoustic feature sequence (frame length 25ms, frame shift 10ms).

[0030] Aggregate the features through a time average pooling layer to output a fixed-length voice text vector (for example, dimension 256). For example, for the user voice instruction "Find the air conditioning unit on the third floor", the vector [0.34, -0.82,..., 1.02] is obtained after processing.

[0031] Gesture processing channel: Capture the three-dimensional coordinate sequence of the gesture trajectory through the Leap Motion sensor (sampling rate 120Hz), and use the cubic spline interpolation algorithm to smooth the trajectory noise to generate a spatio-temporally continuous gesture trajectory coordinate matrix (dimension N×3, where N is the number of sampling points).

[0032] Use the Dynamic Time Warping (DTW) algorithm to align gesture paths at different speeds. For example, the gesture of drawing a rectangle in the air by the user is normalized to a coordinate sequence of 100 sampling points.

[0033] Text processing channel: The input text instruction (such as "emergency ventilation equipment") is encoded by the BERT-base model (12 layers, 768 dimensions), and the sentence vector at the CLS token position (dimension 768) is extracted.

[0034] Select the Top-5 keywords (such as "ventilation", "equipment", "emergency") through TF-IDF weighting. Each keyword is independently encoded as a Word2Vec vector (dimension 300), and the mean value is taken to generate a text keyword vector (dimension 300).

[0035] The multi-modal features are dimensionally aligned in the middle layer: the speech vector passes through a fully connected layer (input 256→output 300), the text keyword vector remains unchanged, and the gesture trajectory coordinates are encoded as a 300-dimensional vector through an LSTM network (hidden unit 128). Finally, the feature vectors of the three modalities are output: the speech text vector (300D), the gesture trajectory coordinate matrix (N×3), and the text keyword vector (300D).

[0036] Based on the speech text vector and the text keyword vector, use a bidirectional attention mechanism to calculate the semantic association weights, and identify the device type, functional attributes, and operation requirements through an intent classifier to generate a structured semantic intent description; The bidirectional attention mechanism enhances semantic understanding through cross-modal interaction: Attention matrix construction: Calculate the similarity matrix (300×300) between the speech text vector (Q) and the text keyword vector (K), using scaled dot-product attention (scaling factor √300).

[0037] Softmax normalize the similarity matrix row by row to obtain the attention weight matrix A. For example, the attention weight of the keyword "ventilation" to the speech segment "third floor air conditioner" is 0.83.

[0038] Feature fusion: Multiply the weight matrix A by the text keyword vector (V) to generate an attention-weighted fusion vector (300D).

[0039] Concatenate the fusion vector with the original speech vector (total dimension 600), reduce the dimension through a fully connected layer (600→300), and output the semantic enhancement vector.

[0040] The intent classifier adopts a multi-task learning framework: Device type recognition: A fully connected layer (300→128→64) is connected to the Softmax output. The classification labels include 20 types of devices such as "air conditioning unit", "power distribution cabinet", etc., and the cross-entropy loss function is used.

[0041] Functional attribute recognition: A bidirectional LSTM (hidden unit 64) processes the semantic enhancement vector and outputs sequence labels (such as "cooling capacity ≥ 5000W", "fire protection level A").

[0042] Operation requirement parsing: A conditional random field (CRF) decodes operation verbs (such as "find", "maintain", "replace") and combines context constraints.

[0043] For example, the input instruction "repair the leaking pipe on the fifth floor" is processed to generate a structured semantic intent description: { "device type": "water supply and drainage pipe", "functional attributes": ["material = PVC", "pipe diameter = DN50"], "operation requirement": "repair" }

[0044] According to the gesture trajectory coordinates and the BIM model space coordinate system, map the gesture trajectory to a three-dimensional space range constraint through a dynamic projection algorithm, and combine the floor topology relationship map to analyze the device orientation hierarchy to generate a spatial orientation code; The dynamic projection algorithm realizes the accurate mapping of the gesture space to the BIM coordinate system: Coordinate system alignment: Establish a transformation matrix between the Leap Motion coordinate system (origin at the sensor center) and the BIM model world coordinate system (origin at the southwest corner of the building). Collect 4 groups of corresponding point coordinates through a calibration board, and use the least squares method to solve the rotation matrix R (3×3) and the translation vector T (3×1).

[0045] For example, the sensor coordinate point (0,0,0) corresponds to the BIM coordinate (102.3, 56.7, 0.0). Calculate the rotation angle θ = 32° and the translation amount T = [102.3, 56.7, 0]^T.

[0046] Trajectory mapping: Multiply the gesture trajectory coordinate sequence by R and add T to convert it to the BIM coordinate system. Determine the floor for the Z-axis coordinate (height): The height of each floor is 4.2 meters, and Z = 12.6 meters corresponds to the 3rd floor.

[0047] Extract the minimum bounding cuboid for the gesture-defined area (such as a rectangular box) to generate spatial range constraint parameters: { "min_x": 45.2, "max_x": 52.8, "min_y": 30.1, "max_y": 38.4, "floor": 3 }。

[0048] The floor topology relationship graph is stored in a graph structure: Nodes: floors (L1 - L10), partitions (such as L3 - Area A), devices; Edges: include relationships such as "is contained in", "adjacent to", "vertically connected", etc.

[0049] Parse the device orientation hierarchy through SPARQL query. For example, the L3 - Area A corresponding to the gesture area contains the device "AHU - 3A - 01" to generate a spatial orientation code: "Spatial orientation": { "floor": 3, "partition": "Area A", "coordinate range": [45.2 ≤ X ≤ 52.8, 30.1 ≤ Y ≤ 38.4], "adjacent devices": ["AHU - 3A - 01", "Pump - 3A - 02"] }。

[0050] Input the structured semantic intention description and the spatial orientation code into the adversarial training network, and eliminate semantic - spatial ambiguity through feature adversarial alignment to output composite retrieval parameters that fuse semantic keywords and three - dimensional coordinate constraints.

[0051] The adversarial training network consists of a generator (G) and a discriminator (D): Generator design: Input: structured semantic vector (300D) + spatial orientation vector (200D); Structure: 4 - layer fully connected (500 → 256 → 128 → 64), output composite retrieval parameters (64D); Objective: Generate a unified feature representation that confuses the discriminator.

[0052] Discriminator design: Input: composite retrieval parameters (64D); Structure: 3 - layer convolution (kernel size 3, channels 16 → 32 → 64) + fully connected layer (64 → 2); Objective: Distinguish the feature source (semantic / spatial), cross - entropy loss function.

[0053] Adversarial training process: Feature adversarial alignment: When training D, fix G, input positive samples (matched semantic - space pairs) and negative samples (random combinations), and learn to distinguish true and false features; When training G, fix D, and through the Gradient Reversal Layer (GRL), maximize the discrimination error of D, forcing G to generate indistinguishable unified features.

[0054] Loss function: Generator loss: L_G = L_adv (adversarial loss)+0.5*L_recon (reconstruction loss, MAE); Discriminator loss: L_D = CE (cross - entropy).

[0055] Compound retrieval parameter generation: After 50 rounds of training (batch size 32, learning rate 0.001), an example of the output parameters is: { "Semantic weight": [0.85, 0.76, 0.91], / / corresponding to "air - conditioning unit", "cooling capacity ≥ 5000W", "maintenance" "Spatial weight": [0.93, 0.88], / / corresponding to "L3 - Area A", "coordinate range" "Fusion threshold": 0.78 }

[0056] These parameters will be used to construct the dynamic search space of the subsequent hierarchical graph convolutional network to ensure that the retrieval results meet both semantic and spatial constraints.

[0057] Specifically, based on the compound retrieval parameters, use the hierarchical graph convolutional network to construct a dynamic search space. The hierarchical graph convolutional network generates a candidate model index structure with multi - dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model, including: According to the connection relationship of the equipment pipelines in the BIM model, construct a hierarchical topological graph structure, where the equipment is the node, the pipeline connection is the edge, and the floor partition is the super - node, generating a multi - granularity topological graph; The construction of the hierarchical topological graph structure extracts the entity information of equipment (such as fans, pumps) and pipelines (such as air ducts, water pipes) from the geometric and logical data of the BIM model. By parsing the IFC (Industry Foundation Classes) standard file, extract the globally unique ID, spatial coordinates and connection relationship of the equipment instances. For example, an air - conditioning unit (ID: AHU - 01) on a certain floor is connected to multiple terminal air vents (ID: Diffuser - 45~60) through an air duct (ID: Duct - 123), forming a basic topological relationship of "equipment - pipeline - equipment".

[0058] Hierarchical Division Rules: Device Level: A single device serves as an atomic node, such as a fan coil unit (FCU), valve (Valve); Pipeline Level: The pipeline serves as an edge, storing connection direction, pipe diameter, and medium type (such as chilled water, steam); Floor Level: Floor partitions (such as F1 - core tube, F2 - office area) are used as supernodes to aggregate all devices in that area; System Level: Divided by function (such as air conditioning system, water supply and drainage system), clustering cross - floor devices into functional groups.

[0059] Multi - granularity Topology Map Generation: Supernode Attributes: Floor nodes contain area, height, and device density (e.g., the F1 node contains 20 devices with a density of 0.5 devices / ㎡); Edge Attributes: Pipeline edges are labeled with flow rate (such as 500 m³ / h) and pressure loss (such as 2 kPa); Cross - level Connection: Connections between supernodes represent cross - floor pipelines (such as the riser from F1 to F2).

[0060] For example, a building BIM model is parsed into a topology map containing 3 floor supernodes, 58 device nodes, and 120 pipeline edges. Among them, the air conditioning system supernode contains 15 device nodes and 30 duct edges.

[0061] Extract device models, functional parameters, and operation and maintenance record texts, generate semantic embedding vectors through a pre - trained language model, and bind them to the node attributes of the multi - granularity topology map to generate an enhanced topology map with semantic attributes; The generation of semantic embedding vectors uses a variant of the pre - trained language model BERT - RoBERTa (Robustly Optimized BERT) due to its strong representation ability for professional terms. The input texts include: Device Model: Such as "Carrier 19XR 500kW Centrifugal Chiller"; Functional Parameters: Such as "Cooling Capacity 500RT, COP = 6.2"; Operation and Maintenance Records: Such as "The bearing was replaced in 2023, and the vibration value dropped to 2.1 mm / s".

[0062] Text Pre - processing and Vectorization: Word Segmentation: Use a professional dictionary (such as the HVAC glossary) to enhance the word - segmentation effect and avoid incorrect splitting of "COP"; Sequence Truncation: For long texts (such as operation and maintenance records), use a sliding window (length 512 tokens, step size 256); Vector Generation: Extract the vector of the [CLS] token through the RoBERTa-base model (with a 768-dimensional hidden layer) as the semantic embedding.

[0063] Attribute Binding: Device Node: Associate the 768-dimensional vector with the topology node. For example, the AHU-01 node is attached with vector V1; Super Node: Aggregate the mean value of the vectors of the subordinate devices. For example, the vector of Floor F1 is the average of V1~V20; Edge Attribute: Attach the medium type encoding to the pipeline edge (e.g., cold water = 0.3, hot water = 0.7).

[0064] Example of Enhanced Topology Graph: The semantic vector of a certain pump node (ID: Pump-05) contains information such as "power 22kW", "head 35m", "motor maintenance in 2022", etc., which together with the topological connection relationship form enhanced features.

[0065] On the enhanced topology graph, hierarchical graph convolution operations are adopted. Among them, the first layer aggregates the features of directly adjacent devices, the second layer aggregates the features of cross-floor super nodes, and the third layer aggregates the features of system-level functional groups to generate a multi-scale feature matrix; The Hierarchical Graph Convolution Network (Hierarchical GCN) adopts three-layer progressive feature aggregation: The First Layer (Device Level): Adjacency Definition: The direct connection of device nodes (e.g., pump → valve); Aggregation Function: GraphSAGE mean aggregation, and the formula is simplified as: node feature = its own feature + mean of neighbor features.

[0066] Example: The chiller (feature V1) aggregates the features of the adjacent pump (V2) and cooling tower (V3), and the output is V1' = (V1 + V2 + V3) / 3.

[0067] The Second Layer (Floor Level): Adjacency Definition: The cross-floor connection between super nodes (e.g., the riser from F1 → F2); Aggregation Function: GAT (Graph Attention Network) weighted aggregation, with the number of attention heads = 4; Example: The F1 super node (feature V_F1) aggregates the features of the F2 node (V_F2), and the weight is calculated by the matching degree of the device types on the two floors (e.g., the floor with a higher proportion of air conditioning systems has a higher weight).

[0068] The Third Layer (System Level): Adjacency Definition: The logical relationship between devices within the functional group (e.g., the chiller and fan coil unit in the air conditioning system); Aggregation function: Cluster-GCN block aggregation, block size = 32; Example: The supernode of the air conditioning system aggregates the features of 15 devices to generate the system-level feature vector V_sys.

[0069] Multi-scale feature matrix generation: Device-level features: Preserve the original resolution (e.g., 58×768); Floor-level features: Dimension reduction to 58×256; System-level features: Further compressed to 58×64; Concatenation: Concatenate the results of the three layers along the feature dimension to generate a 58×(768 + 256 + 64) = 58×1088-dimensional matrix.

[0070] Based on the semantic keywords and spatial constraints in the composite retrieval parameters, calculate the node weights in the multi-scale feature matrix, and dynamically adjust the contribution degrees of features at each level through a gating mechanism to generate weighted feature encodings; Weight calculation is divided into two parts: semantic matching and spatial constraint: Semantic matching weight: Keyword vectorization: Input the user query (e.g., "efficient chiller") into the same RoBERTa model to generate a 768-dimensional query vector Q; Similarity calculation: Use cosine similarity to measure the matching degree between Q and the device semantic vector. For example, the similarity of a certain chiller is 0.85; Spatial constraint weight: Coordinate filtering: Bounded by the three-dimensional space range specified by the user (e.g., X∈[100,200], Y∈[50,150], Z∈[0,30]); Distance attenuation: Use an exponential decay function, with the weight at the center point being 1.0 and the weight at the edge being 0.2; Gating mechanism: Input: Features at each level (device-level F1, floor-level F2, system-level F3); Gating signal: Learn the weight parameters α1, α2, α3 (initial value 0.33) through a fully connected layer; Dynamic adjustment: Correct the α values according to the semantic / spatial weight ratio in the retrieval parameters (e.g., semantic 70%, spatial 30%); Output: Weighted feature F = α1*F1 + α2*F2 + α3*F3.

[0071] Example: When the user searches for "air conditioning equipment on the F2 floor", the spatial weight is increased to 0.6, and the system-level feature weight α3 is increased from 0.3 to 0.5.

[0072] Concatenate the weighted feature encoding with the device space coordinates, and reduce the dimension and compress through an autoencoder to construct a candidate model index structure for multi-dimensional feature encoding.

[0073] The structure of the Autoencoder is designed as follows: Encoder: A 4-layer fully connected network (input 1088 + 3 = 1091 dimensions → 512 → 256 → 128 → 64); Decoder: A symmetric structure (64 → 128 → 256 → 512 → 1091); Loss function: Mean Squared Error (MSE) + Feature Similarity Loss (Cosine Similarity).

[0074] Spatial coordinate concatenation: Coordinate normalization: Normalize the global coordinates of the BIM model (e.g., X = 12345.67) to the interval [0, 1]; Concatenation: Combine the device feature vector (1088 dimensions) with the coordinates (3 dimensions) into a 1091-dimensional input vector.

[0075] Example of dimensionality reduction and compression: The features of a certain chiller are compressed into a 64-dimensional vector V_enc = [0.23, -0.56,..., 0.78] by the encoder, and the error after decoding and restoration is < 5%.

[0076] Index construction: Hierarchical index: Establish an inverted index by system, floor, and device type; Approximate Nearest Neighbor (ANN): Use the HNSW (Hierarchical Navigable Small World) algorithm to accelerate retrieval; Dynamic update: When the BIM model is modified, trigger incremental encoding to update the index.

[0077] Final candidate index structure: Contains 64-dimensional feature vectors and spatial coordinates of 58 devices, supports semantic + spatial joint retrieval, and the response time is < 100ms.

[0078] Specifically, input the candidate model index into a multi-objective optimization engine, and use a reinforcement learning-driven dynamic pruning algorithm to optimize the search path in real-time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a set of candidate models whose confidence after pruning is higher than the preset confidence threshold, including: Based on the candidate model index structure, start from the device with the highest semantic similarity to construct an initial search path tree, and generate a search state space containing the path node sequence and branch weights; When constructing the initial search path tree, it is first necessary to determine the starting device for the search. Based on the semantic embedding information of the devices in the candidate model index (such as device function descriptions, model parameters, etc.), a hybrid similarity model is used to calculate the semantic correlation degree between devices. The hybrid similarity model combines cosine similarity and Jaccard coefficient: Cosine similarity: Measures the direction consistency of the device semantic embedding vectors (such as 512-dimensional vectors). For example, the cosine similarity between an air conditioner unit and a fresh air unit can reach 0.85 due to their similar functions; Jaccard coefficient: Used to compare the proportion of the intersection of device attributes. For example, the common attributes (flow rate, head) of two pumps account for 80%, and the coefficient is 0.8.

[0079] The final semantic similarity is calculated through weighted fusion (cosine weight 0.7, Jaccard weight 0.3). Assuming that a certain device has the highest similarity (such as 0.92) with the retrieval keyword, it is selected as the search starting point.

[0080] Construction process of the search path tree: Root node initialization: Set the starting device as the root node, and record its attributes (such as coordinates, function type) and level markers (such as "L1 - air conditioning system").

[0081] Level expansion rules: Direct connection layer (L1): Traverse the devices directly physically connected to the root node (such as the air ducts and valves connected to the air conditioner unit) to generate the first-level child nodes; Cross-floor connection layer (L2): Through the floor topology relationship map of the BIM model, identify the associated devices across floors (such as the distribution box corresponding to the air conditioner unit located on the lower floor) to generate the second-level child nodes; System-level connection layer (L3): Based on the functional group association (such as the control logic associated devices of the entire HVAC system), generate the third-level child nodes.

[0082] Branch weight calculation: Topological hop count attenuation: Define the attenuation coefficient λ = 0.8. For each additional hop (such as from L1 to L2), the weight is multiplied by λ. For example, the initial weight of the L1 branch is 1.0, the weight of the L2 branch is 0.8, and the weight of the L3 branch is 0.64; Connection type addition: The weight of physical connections (such as pipes) is +0.2, and the weight of logical connections (such as control links) is +0.1.

[0083] The finally generated search state space contains all the path branches of the tree structure and their weights. For example, the root node is "Air conditioner unit A", and its child nodes include "Air duct B (weight 1.2)", "Valve C (weight 1.1)", "Distribution box D (weight 0.88)", etc., forming a multi-level branch network.

[0084] Define the state-action space using a Deep Q-Network (DQN). Here, the state is the feature of the current path node, and the action is to select the next-hop node or terminate the search. Update the Q-value table through the Bellman equation to generate a search strategy. Construction and training of the Deep Q-Network (DQN): Definition of the state space: Node feature vector: It includes device semantic embedding (512 dimensions), spatial coordinates (3 dimensions), topological level (1 dimension), and the weight of the current path (1 dimension), totaling 517 dimensions. Path context: Record the features of the last 3 nodes to form a 1551-dimensional temporal state vector.

[0085] Definition of the action space: Action 1: Select the next-hop node (selected from the list of child nodes of the current node). Action 2: Terminate the search and return the current path.

[0086] Design of the reward function: Positive reward: Finding a device that matches the retrieval target (+1.0). Negative rewards: The path weight is lower than the threshold (-0.5), and the search step limit is exceeded (-0.2). Neutral reward: Exploring a new node (+0.1).

[0087] Bellman equation update mechanism: Q-value update formula (non-mathematical description): New Q-value = Old Q-value + Learning rate α × (Immediate reward + Discount factor γ × Maximum Q-value of the next state - Old Q-value).

[0088] Parameter settings: Learning rate α = 0.001, Discount factor γ = 0.9, Experience replay buffer capacity of 10,000 entries, Batch size of 32.

[0089] Example of the training process: Initialization: Randomly select the search starting point, execute an action (such as selecting the child node "Air duct B"), and record the state transition (S→S') and the reward. Experience replay: After storing 1000 transition records, randomly sample 32 records to update the Q-network. Policy generation: After 1000 iterations, the Q-network can accurately predict high-return paths (such as selecting "Distribution box D" can quickly find the associated control module).

[0090] Dynamic pruning strategy: Real-time Q-value monitoring: If the Q-value of the current path continuously decreases for 5 steps, it is determined as an inefficient path and pruning is triggered. Path backtracking: After pruning, return to the nearest high-Q-value node to re-explore.

[0091] During the search process, calculate the topological distance weight and semantic similarity weight in real time, and generate a combined loss value through linear weighting; When the combined loss value of the search path exceeds the preset loss threshold, trigger the pruning strategy, retain the Top-K path branches with the lowest loss value, prune redundant branches, and generate a refined search path set; Calculation of the combined loss function: Topological distance weight: Hop count penalty: For each additional hop in the path, the distance weight increases by 0.1. For example, the weight of a 3-hop path is 0.3; Cross-level penalty: The weight of the hop count across floors (L2) is multiplied by 1.2, and the weight of the system-level (L3) hop count is multiplied by 1.5.

[0092] Semantic similarity weight: Dynamic decay: For each node passed in the path, the semantic similarity decreases by a decay rate of β = 0.9. For example, the starting similarity is 0.92, and the second hop is 0.92 × 0.9 = 0.828; Node matching degree: If the current node attribute matches the retrieval keyword (such as containing "frequency conversion"), the weight is increased by 0.15.

[0093] Combined loss value: Calculation formula (non-mathematical): Loss value = topological distance weight × 0.6 + (1 - semantic similarity weight) × 0.4; Threshold setting: The preset loss threshold is 0.7, and paths with a loss value exceeding this are determined to be inefficient paths.

[0094] Execution of the pruning strategy: Loss monitoring: Calculate the loss value of each path branch in real time. For example: Path 1: 3 hops, loss value 0.65; Path 2: 5 hops, loss value 0.82 (exceeds the threshold).

[0095] Pruning operation: Top-K retention: K = 5, retain the 5 paths with the lowest loss value; Redundant path elimination: Delete paths with a loss value > 0.7 or a repetition rate > 80%.

[0096] Path backtracking optimization: For the pruned paths, use the greedy algorithm for local optimization (such as replacing high-loss nodes); If all paths are pruned, expand the search scope (such as adding L3-level nodes).

[0097] Example scenario: Retrieval target: Search for air-conditioning equipment related to "frequency conversion control"; Search process: The initial path passes through "Air conditioner unit A → Distribution box D → Control module E", with a loss value of 0.58 (retained); Another path "Air conditioner unit A → Air duct B → Air outlet F" has a loss value of 0.75 (pruned). Result: The retained path points to "Control module E", and the associated frequency converter device is added to the candidate set.

[0098] Traverse and streamline the search path set, extract the device nodes at the end points of the paths, sort them according to the combined loss value, and output the candidate model set with a confidence level higher than the preset confidence threshold.

[0099] Result generation and sorting: End point node extraction: Obtain the device instance from the last node of each path. For example, the end point of path 1 is "Frequency converter G", and the end point of path 2 is "Sensor H"; Confidence calculation: Basic confidence: 1 - combined loss value. For example, a loss value of 0.6 corresponds to a confidence of 0.4; Attribute bonus: If the device label exactly matches the retrieval keyword (such as the name contains "frequency conversion"), the confidence is increased by 0.3; Topology bonus: The end point node is located on the critical path (such as the main power supply path), and the confidence is increased by 0.2.

[0100] Threshold filtering: The preset confidence threshold is 0.5, and devices below this value are excluded. For example: Frequency converter G: Confidence 0.72 (retained); Sensor H: Confidence 0.45 (excluded).

[0101] Output of the candidate model set: Sorting rule: Sort in descending order of confidence, and in case of the same score, sort in ascending order of topological distance; Result example: Frequency converter G (confidence 0.72, located at L2 level); Control module E (confidence 0.68, L3 level); Distribution box D (confidence 0.55, L1 level).

[0102] Dynamic optimization mechanism: Feedback learning: The click-through rate and dwell time of the user on the retrieval results are used as reinforcement signals to inversely update the reward function of the DQN; Threshold adaption: Dynamically adjust the confidence threshold according to the historical retrieval success rate (such as when the success rate < 70%, the threshold is reduced to 0.45).

[0103] Specifically, based on the candidate model set, expand the associated device nodes through knowledge graph embedding completion technology to generate an enhanced retrieval result set containing implicit associated devices, including: Map the device nodes in the candidate model set to a pre - constructed electromechanical knowledge graph, and learn the embedding representations of the device nodes in the functional, physical, and logical dimensions through a graph neural network to generate device relationship vectors; The construction of the electromechanical knowledge graph is based on industry standards (such as the IFC standard) and historical engineering data, and includes three core relationships: functional, physical, and logical: Functional relationship: The role of the device in the system (e.g., an air - handling unit belongs to the HVAC system and has a "supply air" relationship with the air duct); Physical relationship: The spatial connection between devices (e.g., a water pump is directly connected to a valve through a pipe, and the distance ≤ 2 meters); Logical relationship: Control dependency (e.g., the start signal of a fire pump depends on the alarm of a smoke sensor).

[0104] The graph neural network (GNN) architecture uses GraphSAGE (Sampling Aggregation Graph Network), and the specific parameters are as follows: Input features: Device attributes (model, power, installation location coordinates) and relationship types (functional / physical / logical); Hidden layer: 3 layers, with 256 dimensions for each layer, and the activation function is ReLU; Output: A 128 - dimensional device relationship vector, representing the semantic relevance of the device in the multi - dimensional relationship.

[0105] Training process: Data preparation: Sample 100,000 groups of device relationship triples (head entity, relationship, tail entity) from the knowledge graph, e.g., (air - handling unit A, supply air, air duct B); Negative sampling: Generate 5 negative samples for each positive sample (e.g., randomly replace the tail entity with an irrelevant device); Loss function: Use cross - entropy loss, the optimization goal is the relationship classification accuracy of distinguishing positive and negative samples, the learning rate is 0.001, and the number of training epochs is 100.

[0106] Example: After the air - handling unit node is processed by the GNN, the cosine similarity of its relationship vector with the air duct node in the functional dimension is 0.92, the similarity with the adjacent water pump node in the physical dimension is 0.85, and the similarity with the temperature sensor in the logical dimension is 0.78.

[0107] Based on the device relationship vectors, use the random - walk algorithm to traverse the knowledge graph, extract the implicit association paths that have engineering coupling relationships with the candidate devices, and generate a list of potential associated devices; The random - walk algorithm mines the implicit associations between devices (such as indirect control dependencies or cross - system collaborations) by setting the transition probability preferences: Random Walk Parameter Settings: Random Walk Length: 10 steps (can cover indirect relationships of more than 3 levels); Restart Probability: 0.15 (to prevent deviating too far from the initial node); Transition Probability Weights: Functional Relationship 60%, Physical Relationship 30%, Logical Relationship 10%.

[0108] Path Generation: Starting from a candidate device (such as air conditioner unit A), select the next hop according to the similarity of the current node relationship vector. For example, if the similarity in the functional dimension is higher than that in the physical dimension, then preferentially jump to duct B along the "air supply" relationship; Update the path semantic consistency score after each random walk step (e.g., the score of the path "air conditioner unit A → duct B → fire damper C" is 0.8 because it conforms to the interlock logic of the HVAC and fire protection systems).

[0109] Path Filtering: Use a pre-trained graph attention network (GAT) to evaluate the path confidence. The input of the GAT model is the path node relationship vector, and the output is a confidence score from 0 to 1, with the threshold set to 0.85; Filter low-confidence paths (e.g., the confidence of "air conditioner unit A → lighting switch D" is 0.2 because there is no functional association).

[0110] Example: Starting from air conditioner unit A, random walks may discover the implicit path "air conditioner unit A → duct B → fire damper C → smoke sensor E", revealing the indirect interlock relationship between the air conditioning system and the fire protection system, and generating a list of potential associated devices: [duct B, fire damper C, smoke sensor E].

[0111] Calculate the semantic association degree and topological proximity degree between potential associated devices and the candidate model, and retain devices with an association degree greater than 0.7 through threshold filtering to generate an implicit association subset; Calculation of Semantic Association Degree: Functional Association Degree: Based on the cosine similarity of device relationship vectors (e.g., the functional similarity between air conditioner unit A and fire damper C is 0.75); Text Association Degree: Use the BERT model to calculate the similarity of device description texts (e.g., the text association degree between "fire damper" and "air volume control" is 0.68); Comprehensive Semantic Association Degree: Weighted average (Functional 70% + Text 30%), for example, 0.75×0.7 + 0.68×0.3 = 0.729.

[0112] Calculation of Topological Proximity Degree: Physical Distance: Calculate the shortest path hop count between devices in the BIM model (e.g., there are 2 pipe nodes between air conditioner unit A and fire damper C, hop count = 2); System Hierarchy Distance: Calculate according to the system attribution hierarchy (e.g., belonging to the same HVAC system is 0, across systems is 1); Comprehensive topological proximity: Normalized formula (1 - number of hops / maximum number of hops). Assuming the maximum number of hops = 5, a 2-hop corresponds to a proximity of 0.6.

[0113] Threshold filtering: Set the comprehensive correlation threshold to 0.7 (e.g., for the fire damper C, the correlation degree 0.729 > 0.7, so it is retained); forcefully eliminate devices with a topological proximity < 0.4 (such as devices across floors without direct connections).

[0114] Example: The fire damper C in the list of potentially associated devices is retained due to its comprehensive correlation degree of 0.729, while the lighting switch D (correlation degree 0.35) is filtered. Finally, a latent association subset is generated: [duct B, fire damper C].

[0115] Merge the latent association subset with the original candidate model set, sort by the comprehensive correlation degree, and output an enhanced retrieval result set containing latent associated devices.

[0116] Data merging: De-duplicate and merge the original candidate set (such as [air conditioning unit A, water pump X]) with the latent association subset ([duct B, fire damper C]); add relationship labels to the latent associated devices (such as "Fire damper C: There is a fire protection linkage with the air conditioning unit A").

[0117] Sorting rules: Comprehensive correlation degree: Weight 60% (such as 0.729 for the fire damper C); Topological proximity: Weight 30% (such as 0.6 for the fire damper C); Device criticality: Weight 10% (such as the fire protection equipment priority coefficient 1.2).

[0118] Sorting score = 0.6×0.729 + 0.3×0.6 + 0.1×1.2 = 0.725.

[0119] Result display: The output format is a JSON array, including device ID, name, correlation degree, and relationship description; supports visual interaction (such as highlighting associated devices in the BIM model and showing the relationship path).

[0120] Example: The enhanced retrieval result set is sorted by score as follows: air conditioning unit A (score 0.85); fire damper C (score 0.725); duct B (score 0.68); water pump X (score 0.65).

[0121] Specifically, according to the candidate model index structure, starting from the device with the highest semantic similarity, construct an initial search path tree, and generate a search state space including a sequence of path nodes and branch weights, including: Based on the device semantic embedding information in the candidate model index, the cosine similarity and Jaccard coefficient are fused through a hybrid similarity model to generate a device semantic association matrix representing the semantic similarity between devices; The device semantic embedding information in the candidate model index is derived from a pre-trained language model (such as BERT), which encodes text information such as device models, function descriptions, and operation and maintenance records into 768-dimensional vectors. For example, the description of an air conditioner unit "Variable frequency multi-connected unit, cooling capacity 12kW, energy efficiency ratio 3.5" is encoded into a 768-dimensional vector, where the first 128 dimensions mainly capture the "variable frequency" characteristic, the middle 256 dimensions correspond to the "cooling capacity" parameter, and the remaining dimensions reflect details such as the "energy efficiency ratio".

[0122] The design goal of the hybrid similarity model is to comprehensively consider the multi-dimensional semantic associations of devices: Cosine similarity: It is used to measure the global semantic similarity between device model vectors. For example, the cosine similarity between the model vectors of two air conditioner units in a 768-dimensional space is 0.92, indicating a high degree of functional similarity.

[0123] Jaccard coefficient: It is used to calculate the intersection-to-union ratio of the set of device function parameters. For example, if the set of function parameters of device A is {cooling, heating, variable frequency} and that of device B is {cooling, silent}, then the Jaccard coefficient is the size of the intersection {cooling} (1) divided by the size of the union {cooling, heating, variable frequency, silent} (4), resulting in 0.25.

[0124] The hybrid weight assignment adopts a dynamic strategy: For structured texts such as device models and technical specifications, the weight of the cosine similarity is set to 0.7; For discrete attributes such as function parameters and operation and maintenance records, the weight of the Jaccard coefficient is set to 0.3.

[0125] The final similarity calculation formula is: Hybrid similarity = 0.7 * Cosine similarity + 0.3 * Jaccard coefficient.

[0126] For example, if the cosine similarity of the models between device A and device B is 0.85 and the Jaccard coefficient of the function parameters is 0.4, then the hybrid similarity is 0.7×0.85 + 0.3×0.4 = 0.715. The similarity calculation results for all device pairs form an N×N semantic association matrix (N is the total number of devices), and the matrix element value range is [0,1].

[0127] Extract the device node with the highest semantic similarity from the device semantic association matrix as the search starting point, and based on the topological connection relationship of the BIM model, create an initial path tree with this node as the root node, and generate the root node attributes and hierarchical markers; Assume that the semantic similarity between device D (refrigeration unit) and the target retrieval instruction in the semantic association matrix is the highest (0.95), then D is selected as the root node. The construction of the initial path tree is divided into the following steps: Root node attribute extraction: Spatial attribute: The three-dimensional coordinates of device D (x = 120.5m, y = 45.3m, z = 5F); Functional attribute: Refrigerating capacity 15kW, supply voltage 380V; Topological attribute: Directly connected devices (duct E, water pump F).

[0128] Hierarchical marking rule: Level 0: Root node (device D); Level 1: Directly connected devices (duct E, water pump F); Level 2: Cross-floor connected devices (such as distribution box G located on the 6th floor); Level 3: System-level associated devices (such as building automation system H).

[0129] Topological connection relationship parsing is based on IFC (Industry Foundation Classes) data of the BIM model: Extract the IfcRelConnectsPathElements entity from the IFC file to obtain the physical connection relationship between devices; Parse the IfcSpace entity to determine the floor where the device is located and construct the cross-floor connection relationship.

[0130] For example, device D is connected to the air conditioning terminal I through duct E and is also connected to the distribution box G through a cable. The root node of the initial path tree is D, the level 1 child nodes are E and F, the level 2 child node is G, and the level 3 child node is H.

[0131] According to the direct connection, cross-floor connection, and system-level connection relationships of devices in the topological graph, expand the path tree child nodes in the order of hierarchical priority. Record the topological hop count and connection type of the parent and child nodes each time an expansion is made to generate multi-level path branches; Hierarchical priority rule: Direct connection (level 1) is expanded first: Ensure that the search preferentially covers physically adjacent devices; Cross-floor connection (level 2) is expanded secondarily: Handle cross-regional associations; System-level connection (level 3) is expanded last: Handle logical associations.

[0132] Path expansion process: Level 1 expansion: Starting from the root node D, traverse its directly connected devices E and F: The hop count of device E (duct) is 1, and the connection type is "duct connection"; The hop count of device F (water pump) is 1, and the connection type is "water pipe connection".

[0133] Level 2 expansion: Starting from device F, it is found that it is cross-floor connected to the distribution box G (6th floor): The hop count accumulates to 2, and the connection type is marked as "cable cross - layer connection". Record the path branch D → F → G.

[0134] Hierarchy 3 extension: Starting from distribution box G, associated with building automation system H: The hop count accumulates to 3, and the connection type is marked as "system - level logical association". Record the path branch D → F → G → H.

[0135] Path branch storage structure: Each path branch is stored as a JSON object, for example: { "path": ["D", "F", "G", "H"], "hops": 3, "types": ["water pipe connection", "cable cross - layer connection", "system - level logical association"], "weights": [0.9, 0.7, 0.5] }

[0136] Based on the device semantic association matrix and topological hop count, calculate the initial weight of each path branch through a dynamic attenuation function to generate a weighted search state space.

[0137] The dynamic attenuation function is designed in an exponential attenuation form, and the attenuation factor is related to the topological hop count and semantic similarity: Among them, α is the attenuation coefficient (default 0.8), and the hop count starts from 1.

[0138] Weight calculation example: Path D → E: Hop count = 1, semantic similarity = 0.85; Weight = 0.85 × 0.8^(1 - 1)=0.85.

[0139] Path D → F → G: Hop count = 2, semantic similarity = 0.7 (similarity between F and G); Weight = 0.7 × 0.8^(2 - 1)= 0.7 × 0.8 = 0.56.

[0140] Path D → F → G → H: Hop count = 3, semantic similarity = 0.6 (similarity between G and H); Weight = 0.6 × 0.8^(3 - 1) = 0.6 × 0.64 = 0.384.

[0141] Search state space construction: Node table: Records the attributes and levels of each device; Edge table: Stores connection relationships, hop counts, and weights; Path weight index: Sorts path branches in descending order of weight for quick retrieval.

[0142] For example, the search state space contains the following high-weight paths: D→E (weight 0.85); D→F (weight 0.78); D→E→I (weight 0.65).

[0143] Specifically, based on the device relationship vector, use the random walk algorithm to traverse the knowledge graph, extract implicit association paths with engineering coupling relationships with candidate devices, and generate a list of potential associated devices, including: According to the embedding values in the functional, physical, and logical dimensions of the device relationship vector, calculate the transition probability of the edges in the knowledge graph, set the preference weight for the walking direction, and generate the initial parameters for the random walk with preference. The device relationship vector is generated by a graph neural network (such as GraphSAGE) and contains embedding values in three dimensions: function, physical, and logic. For example, the embedding vector of a certain air conditioner device may be: Functional dimension (512 dimensions): Features such as refrigeration power and energy efficiency ratio; Physical dimension (512 dimensions): Installation position coordinates and pipeline connection methods; Logical dimension (512 dimensions): Control logic dependencies and linked device lists.

[0144] Calculation of transition probability: Weight assignment for edge types: Functional association edge (such as the linkage between an air conditioner and a fresh air system): Weight 0.4; Physical connection edge (such as the direct connection between an air conditioner and an air duct): Weight 0.3; Logical dependency edge (such as the start and stop of an air conditioner being controlled by a thermostat): Weight 0.3.

[0145] Similarity calculation: For adjacent node embedding vectors, calculate the cosine similarity by weighting according to dimensions. For example, the physical dimension similarity between an air conditioner and an air duct = 0.85, the functional dimension = 0.6, and the logical dimension = 0.3. Then the comprehensive similarity = 0.85×0.3 + 0.6×0.4 + 0.3×0.3 = 0.56.

[0146] Normalization of transition probability: Convert the similarity values into a probability distribution through Softmax. If a node has 3 edges with similarities of 0.56, 0.42, and 0.30 respectively, then the transition probabilities are [0.56 / (0.56 + 0.42 + 0.30) = 0.44, 0.33, 0.23].

[0147] Setting of walking direction preference: Physical connection first: In the initial walking stage (the first 3 steps), the weight of physical edges is increased by 50%; Functional association second: In the 4th to 6th steps, the weight of functional edges is increased by 30%; Logical Dependency Supplement: The subsequent steps restore the original weights.

[0148] Example of Initialization Parameters: { "max_steps": 10, / / Maximum number of walking steps "restart_threshold": 0.6, / / Path consistency threshold, below which the walk is restarted "preference_weights": [0.5, 0.3, 0.2] / / Initial preference weights for physical, functional, and logical edges }.

[0149] Starting from the candidate device node, the next-hop node is dynamically selected based on the transition probability. After each transfer, the subsequent transfer probability is adjusted according to the semantic consistency of the current path. If the consistency is lower than the preset consistency threshold, the walk is restarted to generate highly correlated path segments; Dynamic Walk Process: Path Initialization: Starting from the candidate device node (such as air conditioner unit A), record the initial path segment as [Air conditioner A].

[0150] Next-Hop Selection: According to the transition probability distribution of the current node, select the next node according to the roulette wheel algorithm. For example, the neighbor nodes of air conditioner A include duct B (probability 0.44), thermostat C (0.33), and fresh air unit D (0.23). The random number 0.5 falls within the range of duct B, so duct B is selected.

[0151] Semantic Consistency Evaluation: Calculate the semantic consistency of the path segment [Air conditioner A → Duct B]: Functional Consistency: Both the air conditioner and the duct belong to the ventilation system, score +0.8; Physical Consistency: Direct connection, score +1.0; Logical Consistency: No direct control relationship, score +0.2; Comprehensive Score = 0.8×0.4 + 1.0×0.3 + 0.2×0.3 = 0.68, higher than the threshold 0.6, continue walking.

[0152] Transition Probability Adjustment: If the current path score is high (e.g., >0.7), then increase the weight of the same type of edge by 10%. For example, the weight of the physical edge increases from 0.3 to 0.33.

[0153] Restart Mechanism: If the score is <0.6 after a certain transfer (e.g., Air conditioner A → Thermostat C → Lighting device E, score 0.45), then terminate the current path and restart the walk.

[0154] Example of Path Segment: Valid path: Air conditioner A → Air duct B → Fresh air unit D → Fire damper F, score sequence [0.68, 0.72, 0.65]; Invalid path: Air conditioner A → Thermostat C → Lighting switch G, terminated due to score 0.45.

[0155] For the path segments generated by random walk, calculate the path confidence through a pre-trained graph attention network, filter out the segments with confidence lower than 0.85, and retain the set of valid paths; Graph attention network (GAT) modeling: Network structure: 3-layer GAT, with 4 heads per layer, hidden layer dimension 256, output dimension 128.

[0156] Input features: Concatenation of the functional, physical, and logical embedding vectors of each node in the path segment (1536 dimensions); Attention calculation: Calculate the attention coefficients for the edges between path nodes to focus on key associations. For example, in the path Air conditioner A → Air duct B → Fresh air D, the attention coefficient between the air conditioner and the air duct is 0.9, and the coefficient between the air duct and the fresh air is 0.7.

[0157] Confidence prediction: Map the path features to confidence values (0 - 1) through a fully connected layer. During training, cross-entropy loss is used, with positive samples being the highly associated paths manually labeled (such as air conditioner - air duct - fresh air), and negative samples being the irrelevant paths generated by random walk (such as air conditioner - lighting switch).

[0158] Filtering and retention rules: Confidence threshold: 0.85, only retain paths with a value higher than this; Path length constraint: Minimum 3 hops, maximum 6 hops; Diversity guarantee: The same device can appear in at most 5 paths.

[0159] Example: High-confidence path: Air conditioner A → Air duct B → Fresh air D, confidence 0.92; Low-confidence path: Air conditioner A → Thermostat C → Distribution box H, confidence 0.78, filtered out.

[0160] Extract the nodes that do not appear in the original candidate set and share functional or physical constraints with the candidate devices from the set of valid paths to generate a list of potential associated devices.

[0161] Rules for extracting associated devices: Not in the original candidate set: Exclude the retrieved devices (such as air conditioner A, air duct B); Functional constraint matching: At least one of the device function labels is the same as the candidate device (such as "ventilation", "refrigeration"); Physical constraint matching: Installation location in the same area (such as Floor 2, East Area) or pipeline direct / indirect connection.

[0162] Extraction process: Traverse the valid paths: For the path of air conditioner A → air duct B → fresh air D → fire damper F, extract fresh air D and fire damper F; Relevance verification: Fresh air D: The function label is "ventilation", which belongs to the same system as the "refrigeration" of air conditioner A, and the physical positions are adjacent (distance < 5 meters), and it is added to the list; Fire damper F: The function label is "fire protection", which has no direct functional association with the air conditioner, but is physically connected to the air duct B, and is added to the list (subject to subsequent filtering).

[0163] Threshold filtering: Calculate the semantic association degree between the device and the candidate model (such as the cosine similarity between fresh air D and air conditioner A is 0.75) and the topological proximity (hop count = 2). The comprehensive association degree = 0.75 × 0.6 + (1 / 2) × 0.4 = 0.65, which is lower than the threshold of 0.7, so it is excluded.

[0164] Example of the final list: Potential associated devices: Fresh air unit D (association degree 0.75), air volume regulating valve E (association degree 0.72); Excluded devices: Fire damper F (association degree 0.65), distribution box H (association degree 0.58).

[0165] It can be seen that according to the multi-modal retrieval instruction input by the user, a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints is generated; based on the composite retrieval parameter, a hierarchical graph convolutional network is used to construct a dynamic search space, and a candidate model index structure with multi-dimensional feature encoding is generated; the candidate model index is input into the multi-objective optimization engine, and a dynamic pruning algorithm driven by reinforcement learning is used to optimize the search path in real time, and a set of candidate models with a confidence level higher than the preset confidence threshold after pruning is output; based on the set of candidate models, the knowledge graph embedding completion technology is used to expand the associated device nodes, and an enhanced retrieval result set including implicit associated devices is generated, so as to be able to achieve accurate positioning and fast retrieval of efficient, multi-dimensional, and multi-modal information in a large-scale complex BIM model.

[0166] Another embodiment of the present invention provides a rapid retrieval system for building mechanical and electrical BIM model information. Refer to Figure 3 , the system may include: Parsing module 301, configured to parse the semantic intention and spatial orientation through the spatial topology relationship graph according to the multi-modal retrieval instruction input by the user, and generate a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints; Construction module 302, configured to construct a dynamic search space based on the composite retrieval parameter by using a hierarchical graph convolutional network. The hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model; The optimization module 303 is configured to input the candidate model index into the multi-objective optimization engine, and use a reinforcement learning-driven dynamic pruning algorithm to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, it outputs a set of candidate models whose confidence after pruning is higher than the preset confidence threshold; The generation module 304 is configured to expand associated device nodes based on the set of candidate models through knowledge graph embedding completion technology, and generate an enhanced retrieval result set including implicit associated devices.

[0167] It can be seen that according to the multi-modal retrieval instruction input by the user, a composite retrieval parameter that fuses semantic keywords and three-dimensional coordinate constraints is generated; based on the composite retrieval parameter, a hierarchical graph convolutional network is used to construct a dynamic search space, and a candidate model index structure with multi-dimensional feature encoding is generated; the candidate model index is input into the multi-objective optimization engine, and a reinforcement learning-driven dynamic pruning algorithm is used to optimize the search path in real time, and a set of candidate models whose confidence after pruning is higher than the preset confidence threshold is output; based on the set of candidate models, associated device nodes are expanded through knowledge graph embedding completion technology, and an enhanced retrieval result set including implicit associated devices is generated, so as to be able to achieve accurate positioning and fast retrieval of efficient, multi-dimensional, and multi-modal information in a large-scale complex BIM model.

[0168] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and wherein the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0169] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for executing the following steps: S201, according to the multi-modal retrieval instruction input by the user, parse the semantic intention and spatial orientation through the spatial topological relationship graph, and generate a composite retrieval parameter that fuses semantic keywords and three-dimensional coordinate constraints; S202, based on the composite retrieval parameter, use a hierarchical graph convolutional network to construct a dynamic search space, and the hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the device pipelines in the BIM model; S203, input the candidate model index into the multi-objective optimization engine, and use a reinforcement learning-driven dynamic pruning algorithm to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a set of candidate models whose confidence after pruning is higher than the preset confidence threshold; S204, based on the set of candidate models, expand associated device nodes through knowledge graph embedding completion technology, and generate an enhanced retrieval result set including implicit associated devices.

[0170] It can be seen that according to the multimodal retrieval instruction input by the user, a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints is generated; based on the composite retrieval parameter, a hierarchical graph convolutional network is used to construct a dynamic search space, and a candidate model index structure with multi-dimensional feature encoding is generated; the candidate model index is input into a multi-objective optimization engine, and a dynamic pruning algorithm driven by reinforcement learning is used to optimize the search path in real time, and a set of candidate models with a confidence level higher than the preset confidence threshold after pruning is output; based on the set of candidate models, the knowledge graph embedding completion technology is used to expand the associated device nodes, and an enhanced retrieval result set including implicit associated devices is generated, so as to achieve efficient, multi-dimensional, and multi-modal information accurate positioning and rapid retrieval in a large-scale complex BIM model.

[0171] An embodiment of the present invention further provides an electronic device, including a memory and a processor, where a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0172] Specifically, the above electronic device may further include a transmission device and an input / output device, where the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0173] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201, according to the multimodal retrieval instruction input by the user, parse the semantic intention and spatial orientation through the spatial topology relationship graph, and generate a composite retrieval parameter that combines semantic keywords and three-dimensional coordinate constraints; S202, based on the composite retrieval parameter, use a hierarchical graph convolutional network to construct a dynamic search space, and the hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model; S203, input the candidate model index into a multi-objective optimization engine, use a dynamic pruning algorithm driven by reinforcement learning to optimize the search path in real time, and output a set of candidate models with a confidence level higher than the preset confidence threshold after pruning by calculating the joint loss function of the topological distance weight and the semantic similarity weight; S204, based on the set of candidate models, expand the associated device nodes through the knowledge graph embedding completion technology, and generate an enhanced retrieval result set including implicit associated devices.

[0174] It can be seen that according to the multimodal retrieval instruction input by the user, a composite retrieval parameter that fuses semantic keywords and three-dimensional coordinate constraints is generated; based on the composite retrieval parameter, a hierarchical graph convolutional network is used to construct a dynamic search space, and a candidate model index structure with multi-dimensional feature encoding is generated; the candidate model index is input into a multi-objective optimization engine, and a dynamic pruning algorithm driven by reinforcement learning is used to optimize the search path in real time, and a set of candidate models with a confidence level higher than the preset confidence threshold after pruning is output; based on the set of candidate models, the associated device nodes are extended through knowledge graph embedding completion technology to generate an enhanced retrieval result set containing implicit associated devices, so as to achieve efficient, multi-dimensional, and multi-modal information accurate positioning and fast retrieval in a large-scale complex BIM model.

[0175] The structure, features, and effects of the present invention have been described in detail based on the embodiments shown in the drawings. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope defined by the drawings. Any changes made according to the concept of the present invention, or equivalent embodiments modified into equivalent changes, still fall within the spirit covered by the specification and the drawings, and should be within the protection scope of the present invention.

Claims

1. A method for quickly retrieving information of a building mechanical and electrical BIM model, characterized in that, The method includes: According to the multimodal retrieval instruction input by the user, parse the semantic intention and spatial orientation through the spatial topological relationship graph, and generate composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints; Based on the composite retrieval parameters, use a hierarchical graph convolutional network to construct a dynamic search space. The hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model; Input the candidate model index into a multi-objective optimization engine, and use a dynamic pruning algorithm driven by reinforcement learning to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a set of candidate models whose confidence after pruning is higher than the preset confidence threshold; Based on the set of candidate models, use knowledge graph embedding completion technology to expand associated equipment nodes, and generate an enhanced retrieval result set that includes implicitly associated equipment.

2. The method according to claim 1, wherein The step of, according to the multimodal retrieval instruction input by the user, parsing the semantic intention and spatial orientation through the spatial topological relationship graph, and generating composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints, includes: According to the voice, gesture, and text retrieval instructions input by the user, separate and extract the voice text vector, gesture trajectory coordinates, and text keyword vector through a multimodal fusion network; Based on the voice text vector and the text keyword vector, calculate the semantic association weight using a bidirectional attention mechanism, and identify the equipment type, functional attributes, and operation requirements through an intention classifier to generate a structured semantic intention description; According to the gesture trajectory coordinates and the BIM model space coordinate system, map the gesture trajectory to a three-dimensional space range constraint through a dynamic projection algorithm, and combine the floor topological relationship graph to analyze the equipment orientation hierarchy to generate a spatial orientation encoding; Input the structured semantic intention description and the spatial orientation encoding into an adversarial training network, eliminate semantic-spatial ambiguity through feature adversarial alignment, and output composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints.

3. The method according to claim 2, characterized in that, The step of, based on the composite retrieval parameters, using a hierarchical graph convolutional network to construct a dynamic search space. The hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the equipment pipelines in the BIM model, includes: According to the connection relationship of the equipment pipelines in the BIM model, construct a hierarchical topological graph structure, where the equipment is the node, the pipeline connection is the edge, and the floor partition is the super node, to generate a multi-granularity topological graph; Extract the equipment model, functional parameters, and operation and maintenance record text, generate semantic embedding vectors through a pre-trained language model, and bind them to the node attributes of the multi-granularity topological graph to generate an enhanced topological graph with semantic attributes; On the enhanced topological graph, perform hierarchical graph convolutional operations. In the first layer, aggregate the features of directly adjacent equipment, in the second layer, aggregate the features of cross-floor super nodes, and in the third layer, aggregate the features of system-level functional groups to generate a multi-scale feature matrix; Based on the semantic keywords and spatial constraints in the composite retrieval parameters, calculate the node weights in the multi-scale feature matrix, and dynamically adjust the feature contribution degrees of each layer through a gating mechanism to generate weighted feature encoding. The weighted feature codes are concatenated with the device space coordinates, and the dimensionality reduction and compression are performed through the autoencoder to construct a candidate model index structure for multi-dimensional feature coding.

4. The method according to claim 3, wherein The candidate model index is input into the multi-objective optimization engine, and the search path is optimized in real time using a dynamic pruning algorithm driven by reinforcement learning. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, a set of candidate models with a confidence level higher than a preset confidence threshold after pruning is output, including: According to the candidate model index structure, the initial search path tree is constructed with the device with the highest semantic similarity as the starting point, and a search state space including path node sequences and branch weights is generated; A deep Q network is used to define the state-action space, where the state is the current path node feature, the action is to select the next hop node or terminate the search, and the Q value table is updated through the Bellman equation to generate a search strategy; During the search process, the topological distance weight and semantic similarity weight are calculated in real time, and the joint loss value is generated through linear weighting; When the joint loss value of the search path exceeds the preset loss threshold, the pruning strategy is triggered to retain the Top-K path branches with the lowest loss value, prune redundant branches, and generate a streamlined search path set; Traverse the simplified search path set, extract the path endpoint device nodes, sort them by joint loss value, and output a set of candidate models whose confidence is higher than the preset confidence threshold.

5. The method according to claim 4, wherein Based on the candidate model set, the associated device nodes are expanded by the knowledge graph embedding completion technology to generate an enhanced search result set containing implicit associated devices, including: Map the device nodes in the candidate model set to the pre-built electromechanical knowledge graph, learn the embedded representation of the device nodes in the functional, physical and logical dimensions through the graph neural network, and generate the device relationship vector; Based on the equipment relationship vector, the random walk algorithm is used to traverse the knowledge graph, extract the implicit association paths that have engineering coupling relationships with the candidate equipment, and generate a list of potential associated equipment; Calculate the semantic relevance and topological proximity between the potential associated devices and the candidate models, retain the devices with a relevance greater than 0.7 through threshold filtering, and generate a hidden associated subset; The implicitly associated subset is merged with the original candidate model set, sorted by comprehensive association degree, and an enhanced retrieval result set containing implicitly associated devices is output.

6. The method according to claim 4, characterized in that, The method constructs an initial search path tree based on the candidate model index structure and takes the device with the highest semantic similarity as the starting point to generate a search state space including a path node sequence and branch weights, including: According to the device semantic embedding information in the candidate model index, the cosine similarity and the Jaccard coefficient are fused through a hybrid similarity model to generate a device semantic association matrix representing the semantic similarity between devices. Extract the device node with the highest semantic similarity from the device semantic association matrix as the search starting point, create an initial path tree with the node as the root node based on the topological connection relationship of the BIM model, and generate the root node attributes and hierarchical tags; According to the direct connection, cross-floor connection and system-level connection relationship of the equipment in the topology diagram, the path tree child nodes are expanded in the order of hierarchical priority. The topological hop count and connection type of the parent and child nodes are recorded each time the expansion is performed to generate multi-level path branches. Based on the device semantic association matrix and topological hop count, calculate the initial weights of each path branch through a dynamic decay function to generate a weighted search state space.

7. The method according to claim 5, characterized in that, Based on the device relationship vectors, traverse the knowledge graph using the random walk algorithm, extract the implicit association paths that have engineering coupling relationships with the candidate devices, and generate a list of potential associated devices, including: According to the embedding values in the functional, physical, and logical dimensions of the device relationship vectors, calculate the transition probabilities of the edges in the knowledge graph, set the preference weights for the walking directions, and generate the initialized parameters for the random walk with preferences. Starting from the candidate device node, dynamically select the next-hop node based on the transition probabilities. After each transfer, adjust the subsequent transfer probabilities according to the semantic consistency of the current path. If the consistency is lower than the preset consistency threshold, restart the walk to generate high-association path segments. For the path segments generated by the walk, calculate the path confidence through a pre-trained graph attention network, filter out the segments with a confidence lower than 0.85, and retain the set of valid paths. Extract the nodes that do not appear in the original candidate set and share functions or physical constraints with the candidate devices from the set of valid paths to generate a list of potential associated devices.

8. A rapid information retrieval system for building mechanical and electrical BIM models, characterized in that, The system includes: A parsing module for parsing the semantic intent and spatial orientation through the spatial topology relationship graph according to the multimodal retrieval instruction input by the user, and generating composite retrieval parameters that fuse semantic keywords and three-dimensional coordinate constraints. A construction module for constructing a dynamic search space based on the composite retrieval parameters using a hierarchical graph convolutional network. The hierarchical graph convolutional network generates a candidate model index structure with multi-dimensional feature encoding by extracting the topological connection features and semantic attribute features of the device pipelines in the BIM model. An optimization module for inputting the candidate model index into a multi-objective optimization engine, and using a dynamic pruning algorithm driven by reinforcement learning to optimize the search path in real time. By calculating the joint loss function of the topological distance weight and the semantic similarity weight, output a set of candidate models with a confidence higher than the preset confidence threshold after pruning. A generation module for expanding the associated device nodes through knowledge graph embedding completion technology based on the set of candidate models, and generating an enhanced retrieval result set that includes implicit associated devices.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, wherein the computer program is set to execute the method according to any one of claims 1-7 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is set to run the computer program to execute the method according to any one of claims 1-7.

Images