An intelligent maintenance decision method, system and device for a building curtain wall and a storage medium

Abstract

The application discloses a kind of intelligent maintenance decision methods, systems, equipment and storage medium of building curtain wall, the method includes: obtaining project evaluation data;Multi-modal feature fusion processing is carried out to the project evaluation data after being structured, and multi-modal fusion feature data is obtained;Building curtain wall evaluation dataset is constructed based on multi-modal fusion feature data;Residual quantitative variation auto-encoder model trained is used to process building curtain wall evaluation dataset, and building curtain wall feature identifier is generated;Building curtain wall feature identifier is combined with building curtain wall evaluation dataset based on, and query-identifier training pair set is constructed;Query-identifier training pair set is used to train curtain wall maintenance strategy generation model, and the trained curtain wall maintenance strategy generation model is obtained;The curtain wall maintenance demand of user is obtained and input to the trained curtain wall maintenance strategy generation model, and curtain wall maintenance strategy information is obtained.The application can reduce the safety accidents caused by improper curtain wall maintenance decision.

Description

Technical Field

[0001] This invention belongs to the field of building engineering information technology, specifically relating to an intelligent maintenance decision-making method, system, equipment, and storage medium for building curtain walls. Background Technology

[0002] With the accelerating pace of urbanization, the demand for curtain walls in urban buildings is gradually increasing. my country has now become the world's largest producer and user of curtain walls. However, with the rapid growth of curtain walls, safety accidents caused by improper curtain wall maintenance decisions are also occurring more and more frequently. Therefore, how to effectively maintain and make decisions regarding building curtain walls is particularly important.

[0003] In existing technologies, the maintenance decisions for building curtain walls mainly face the following problems:

[0004] (1) Information silos and fragmentation: Because the knowledge required for building curtain wall maintenance (such as structural parameters, inspection data, and operating procedures) is scattered in documents of different formats, including CAD drawings, PDF reports, images and photos, and spreadsheets, it is difficult for the maintenance party to integrate the required maintenance information from a large amount of data, resulting in low integration efficiency.

[0005] (2) Non-semantic identification and retrieval are difficult because traditional documents for building curtain walls often fail to reflect their semantic content and maintenance relevance to other documents. Therefore, when maintenance problems occur, maintenance personnel find it difficult to locate relevant solutions and historical experience, which can easily exacerbate the safety issues of the curtain wall.

[0006] (3) High knowledge barriers in professional fields. Due to the highly professional and unique nature of the building curtain wall field, it is difficult for the Large Language Model (LLM) to directly extract professional information in this field, which may lead to problems such as inaccurate knowledge and content illusion, thereby affecting the technical accuracy and reliability of curtain wall maintenance strategy information. In addition, most of the main entities responsible for curtain wall maintenance are property management companies. These entities generally lack curtain wall-related professional backgrounds. Faced with complex technical standards and professional test reports, they often find it difficult to accurately understand and flexibly apply them, thus making it impossible to formulate and implement curtain wall maintenance decision-making schemes. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and provide an intelligent maintenance decision-making method for building curtain walls. This method can effectively process multi-source heterogeneous data, retrieve and refine maintenance data, and generate executable curtain wall maintenance plans, thereby effectively reducing safety accidents caused by improper curtain wall maintenance decisions.

[0008] The second objective of this invention is to provide an intelligent maintenance decision-making system for building curtain walls.

[0009] A third objective of this invention is to provide a computer device.

[0010] A fourth objective of this invention is to provide a storage medium.

[0011] The first objective of this invention can be achieved by adopting the following technical solution:

[0012] A smart maintenance decision-making method for building curtain walls includes the following steps:

[0013] S1. Obtain project evaluation data for the building curtain wall; the project evaluation data includes curtain wall design documents, curtain wall construction documents, curtain wall inspection and testing documents, curtain wall acceptance documents, curtain wall use and maintenance documents, and curtain wall management documents;

[0014] S2. The project evaluation data is structured, and then the structured project evaluation data is subjected to multimodal feature fusion processing to obtain multimodal fusion feature data.

[0015] S3. Construct a building curtain wall evaluation dataset based on the multimodal fusion feature data, and divide it into a multimodal fusion training set; the building curtain wall evaluation dataset includes project ID, document type, relevance level, document path and multimodal fusion features;

[0016] S4. Construct a residual quantization variational autoencoder model, and train the residual quantization variational autoencoder model using the multimodal fusion training set to obtain a trained residual quantization variational autoencoder model; the residual quantization variational autoencoder model includes an encoder module, a residual quantization module, and a decoder module.

[0017] S5. The trained residual quantization variational autoencoder model is used to process each group of data in the building curtain wall evaluation dataset to generate the corresponding building curtain wall feature identifier.

[0018] S6. Construct a simulated query corresponding to each group of data in the building curtain wall evaluation dataset, and pair the simulated query with the building curtain wall feature identifier obtained in step S5 to obtain a query-identifier training pair set.

[0019] S7. Construct a curtain wall maintenance strategy generation model. Train the curtain wall maintenance strategy generation model using the query-identifier training set to obtain a trained curtain wall maintenance strategy generation model. The curtain wall maintenance strategy generation model adopts the GR generative recommendation model.

[0020] S8. Obtain the user's curtain wall maintenance requirements and input them into the trained curtain wall maintenance strategy generation model to obtain curtain wall maintenance strategy information.

[0021] Preferably, the specific steps of step S2 are as follows:

[0022] S21. Based on the importance of curtain wall maintenance, the documents in the project assessment data are labeled with relevance levels, and then the labeled documents are classified to obtain the text data, image data and structured data corresponding to the labeled documents.

[0023] S22. Use the T5-base model to extract text from the text data to obtain the corresponding text vectors;

[0024] S23. Using the pre-trained ViT image processing model, image embedding extraction and image content annotation are performed on the image data to obtain an image embedding vector containing the annotation text embedding.

[0025] S24. Use the Pandas tool in the Python tool library to transform the structured data to obtain key-value pair text vectors;

[0026] S25. Perform adaptive fusion processing on the text vector, image embedding vector, and key-value pair text vector to obtain multimodal fusion features;

[0027] S26. Repeat steps S21-S25 to process each file in the project evaluation data to obtain multimodal fusion feature data.

[0028] Preferably, the specific steps of step S25 are as follows:

[0029] S251. Perform modality missing detection on the multimodal feature vectors obtained in steps S21-S24. If a missing feature vector is detected, generate a modality missing marker vector corresponding to the missing modality feature vector, and proceed to step S252. The number of modality missing marker vectors is at most two. If no missing feature vector is detected, perform multimodal feature fusion processing on the text vector, image embedding vector, and key-value pair text vector to obtain multimodal fused features. The multimodal feature vectors include text vectors, image embedding vectors, and key-value pair text vectors.

[0030] S252. If step S251 detects that there is only a single modality feature vector in the multimodal feature vector, then the standardized feature vector of the modality feature vector is used as the basic feature vector; if step S251 detects that there are two modality feature vectors in the multimodal feature vector, then the two modality feature vectors are dynamically weighted and fused according to the correlation level and data type to generate the basic feature vector.

[0031] S253. Map the modality missing marker vector obtained in step S251 to an auxiliary feature vector, and add it element-wise to the basic feature vector obtained in step S252 to obtain the multimodal fusion feature.

[0032] Preferably, the multi-level relevance labeling in step S21 is divided into five levels, including the practical execution level L4, the strategy decision level L3, the background support level L2, the standard reference level L1, and the irrelevant information level L0.

[0033] Preferably, the specific steps of step S5 are as follows:

[0034] S51. Obtain the multimodal fusion features contained in any set of data in the building curtain wall evaluation dataset, and denote them as semantic embedding vectors;

[0035] S52. Input the semantic embedding vector into the encoder module of the trained residual quantization variational autoencoder model. The encoder module maps the semantic embedding vector to a low-dimensional vector to obtain a low-dimensional latent vector.

[0036] S53. Then, the low-dimensional latent vector is input into the residual quantization module. The residual quantization module uses a multi-layer concatenated codebook to recursively quantize the low-dimensional latent vector, and finally generates a preliminary feature identifier.

[0037] S54. The decoder module is used to decode and reconstruct the preliminary feature identifier obtained in step S53 for verification. If the verification is successful, the preliminary feature identifier is output as the building curtain wall feature identifier. If the verification fails, the data obtained in step S51 is removed from the building curtain wall evaluation dataset.

[0038] S55. Repeat steps S51-S54 to process each group of data in the building curtain wall evaluation dataset.

[0039] Preferably, the total loss function of the residual quantization variational autoencoder model is expressed as follows:

[0040] L ID =L RQ-VAE +λ1L al +λ2L rel ;

[0041] Among them, L ID Let L be the total loss function. RQ-VAE Let L be the residual quantization loss function, λ1 and λ2 be the weighting coefficients, and L be the weighting coefficients. al For the alignment loss function, L rel This is the loss function for multi-level correlation learning.

[0042] Preferably, the specific steps of step S8 are as follows:

[0043] S81. Obtain the user's curtain wall maintenance requirements, input the curtain wall maintenance requirements into the trained curtain wall maintenance strategy generation model described in step S7, and generate multiple candidate building curtain wall feature identifiers.

[0044] S82. Based on the multiple candidate building curtain wall feature identifiers obtained in step S81, the building curtain wall evaluation dataset is searched to obtain a preliminary matching document set;

[0045] S83. Based on the relevance level, the preliminary matched document set is filtered and sorted to obtain the core reference document set;

[0046] S84. Based on the relevance level, perform hierarchical information extraction on the core reference document set, and mark each extracted information with its corresponding document identifier source information to obtain a structured core information set;

[0047] S85. Perform identifier tracing verification and curtain wall domain mandatory rule verification on all structured core information in the structured core information set obtained in step S84, and delete the structured core information that fails the verification.

[0048] S86. Integrate the verified structured core information according to the curtain wall maintenance industry standard template to generate standardized and executable curtain wall maintenance strategy information.

[0049] A smart maintenance decision-making system for building curtain walls, used to implement the smart maintenance decision-making method for the building curtain walls, the system comprising:

[0050] The project evaluation data module is used to obtain project evaluation data for building curtain walls. The project evaluation data includes curtain wall design documents, curtain wall construction documents, curtain wall inspection and testing documents, curtain wall acceptance documents, curtain wall use and maintenance documents, and curtain wall management documents.

[0051] The preprocessing module is used to perform structured processing on the project evaluation data to obtain structured project evaluation data.

[0052] The multimodal feature fusion module is used to perform multimodal feature fusion processing on the structured project evaluation data to obtain multimodal fused feature data;

[0053] The building curtain wall evaluation dataset generation module is used to construct a building curtain wall evaluation dataset based on the multimodal fusion feature data, and to divide the dataset into a multimodal fusion training set. The building curtain wall evaluation dataset includes project ID, document type, relevance level, document path, and multimodal fusion features.

[0054] The residual quantization variational autoencoder model building module is used to build the initial residual quantization variational autoencoder model.

[0055] The residual quantization variational autoencoder model training module is used to train the residual quantization variational autoencoder model using the multimodal fusion training set to obtain a trained residual quantization variational autoencoder model; the residual quantization variational autoencoder model includes an encoder module, a residual quantization module, and a decoder module.

[0056] The building curtain wall feature identifier module is used to process each group of data in the building curtain wall evaluation dataset using a trained residual quantization variational autoencoder model to generate the corresponding building curtain wall feature identifier.

[0057] The query-identifier training pair set construction module is used to construct a simulated query corresponding to each group of data in the building curtain wall evaluation dataset. The simulated query is then paired with the building curtain wall feature identifier obtained by the building curtain wall feature identifier module to obtain the query-identifier training pair set.

[0058] The curtain wall maintenance strategy generation model building module is used to build the initial curtain wall maintenance strategy generation model;

[0059] The curtain wall maintenance strategy generation model training module is used to train the curtain wall maintenance strategy generation model using a query-identifier training set to obtain a trained curtain wall maintenance strategy generation model.

[0060] The curtain wall maintenance strategy information generation module is used to obtain the user's curtain wall maintenance requirements and input them into the trained curtain wall maintenance strategy generation model to obtain curtain wall maintenance strategy information.

[0061] A computer device includes a processor and a memory for storing processor-executable programs, wherein when the processor executes the programs stored in the memory, it implements the intelligent maintenance decision-making method for building curtain walls.

[0062] A storage medium storing a program that, when executed by a processor, implements the intelligent maintenance decision-making method for the building curtain wall.

[0063] The present invention has the following advantages and effects compared with the prior art:

[0064] (1) The intelligent maintenance decision-making method for building curtain walls of the present invention can effectively perform structured processing on the multi-source heterogeneous data (such as structural parameters, detection data, and operation specifications) contained in the project evaluation data, and convert the processed multi-source heterogeneous data into multimodal fusion feature data for training the residual quantization variational autoencoder model. The residual quantization variational autoencoder model generates semantic building curtain wall feature identifiers based on the multimodal fusion feature data, and constructs a query-identifier training set for training the curtain wall maintenance strategy generation model based on the building curtain wall feature identifiers. This enables the trained curtain wall maintenance strategy generation model to effectively retrieve and integrate the corresponding curtain wall maintenance schemes based on the user's input information, so that the user can make curtain wall maintenance decisions. This overcomes the problem in the prior art that the maintenance party has difficulty integrating the required maintenance information from a large amount of data, resulting in low integration efficiency, and improves the overall safety of building curtain wall maintenance.

[0065] (2) The intelligent maintenance decision-making method for building curtain walls of the present invention overcomes the problems of inaccurate knowledge and illusion that exist when general LLM is directly applied to the curtain wall field in the prior art by effectively combining the residual quantization variational autoencoder model and the curtain wall maintenance strategy generation model, thus ensuring the technical accuracy and reliability of generating curtain wall maintenance strategy information.

[0066] (3) The intelligent maintenance decision system for building curtain walls of the present invention, through the effective cooperation of the curtain wall maintenance strategy information generation module and other modules, can generate accurate and easy-to-understand curtain wall maintenance strategy information based on the actual curtain wall maintenance needs input by the user, thereby overcoming the problem in the prior art that users have difficulty in formulating and implementing curtain wall maintenance decision schemes due to professional knowledge barriers. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating an intelligent maintenance decision-making method for building curtain walls provided in Embodiment 1 of the present invention.

[0068] Figure 2 This is a schematic diagram of the residual quantization variational autoencoder model provided in Embodiment 1 of the present invention;

[0069] Figure 3 This is a schematic diagram of the structure of an intelligent maintenance decision-making system for building curtain walls provided in Embodiment 2 of the present invention;

[0070] Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of the present invention.

[0071] Figure 5 This is a schematic diagram of the structure of a storage medium provided in Embodiment 4 of the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0074] In the description of this invention, it should be understood that the terms "upper", "lower", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0075] Example 1

[0076] like Figure 1-2 As shown, this embodiment provides an intelligent maintenance decision-making method for building curtain walls, including the following steps:

[0077] S1. Obtain project evaluation data for the building curtain wall; the project evaluation data includes curtain wall design documents, curtain wall construction documents, curtain wall inspection and testing documents, curtain wall acceptance documents, curtain wall use and maintenance documents, and curtain wall management documents;

[0078] Specifically, the curtain wall design documents mainly describe the basic information of the curtain wall, such as design grade, service life, scale, type, dimensions, materials, and structure. These documents primarily include: curtain wall design drawings, structural calculation sheets, and material lists.

[0079] The curtain wall construction documents mainly describe the precautions, construction process, and specifications for curtain wall construction, and are usually completed by the construction unit. These documents mainly include: curtain wall construction plans and schedule charts, construction guidelines, and on-site process data documents for curtain wall construction (mainly including construction photos of various parts, charts of regular supervision and inspection records by supervisors / team leaders, and construction record charts), etc.

[0080] The aforementioned curtain wall inspection and testing documents mainly describe the inspection and testing data of various indicators of the curtain wall, such as material performance and structural performance, and are usually completed by a third-party inspection and testing organization. These documents mainly include: curtain wall inspection and testing reports (containing information such as: inspection location, inspection materials, inspection methods, inspection results, judgment basis, and judgment results).

[0081] The curtain wall acceptance documents mainly describe the results of the acceptance of each sub-project of the curtain wall and the overall project. They are usually completed by the construction unit or the general contractor. These documents mainly include: acceptance reports (including acceptance checklists, acceptance results, and other information).

[0082] The curtain wall usage and maintenance documents mainly describe the management methods for the curtain wall during the completion and acceptance phase, and are primarily completed by the curtain wall project acceptance unit or the curtain wall user. These documents mainly include: curtain wall maintenance manuals (including precautions for curtain wall use, maintenance time limits and methods, etc.), curtain wall maintenance record reports (including periodic inspection record charts, maintenance record charts, etc.), and safety assessment reports (mainly including: expert assessment reports and curtain wall safety appraisal reports), etc.

[0083] The aforementioned curtain wall management documents mainly describe the basic construction, use, maintenance methods, and requirements of curtain walls, and are usually issued by local administrative authorities. These documents primarily include: local curtain wall management regulations (issued by administrative departments), curtain wall design standards, curtain wall construction standards, curtain wall inspection and testing standards, curtain wall appraisal standards, and other management methods and technical specifications.

[0084] Specifically, in this embodiment, project evaluation data for 500 building curtain walls were obtained, and the files for each project contained the necessary data for the curtain wall evaluation.

[0085] S2. The project evaluation data is structured, and then the structured project evaluation data is subjected to multimodal feature fusion processing to obtain multimodal fusion feature data.

[0086] Specifically, the process of step S2 is as follows:

[0087] S21. Based on the importance of curtain wall maintenance, the documents in the project assessment data are labeled with relevance levels, and then the labeled documents are classified to obtain the text data, image data and structured data corresponding to the labeled documents.

[0088] Specifically, the multi-level relevance labeling in step S21 is divided into five levels, which include the practical execution level L4, the strategy decision level L3, the background support level L2, the standard reference level L1, and the irrelevant information level L0.

[0089] Specifically, the classification description of multi-level relevance labeling is as follows:

[0090] The practical execution level L4 consists of core documents that directly guide maintenance operations, such as maintenance manuals, maintenance record reports, and safety assessment reports; example materials include: "Glass Curtain Wall Sealant Replacement Operation Manual".

[0091] Strategic decision level L3 refers to key documents that influence maintenance decisions; such as inspection and testing reports, acceptance reports, etc.; example data is "2024 Glass Curtain Wall Sealant Aging Test Report".

[0092] Background support level L2 provides auxiliary documents for maintenance background; such as construction documents (e.g., construction process records), design documents, etc.; example data such as "Glass Curtain Wall Sealant Construction Process Record".

[0093] Standard reference level L1 provides reference documents for maintenance standards, such as curtain wall management documents (maintenance standards), etc.; example data such as GB 16776-2025 "Silicone structural sealants for building".

[0094] Irrelevant information level L0 refers to documents that are not related to maintenance; such as structural design drawings of other buildings, construction records of non-curtain wall types, etc. Example data: "Project Indoor Floor Tile Construction Record".

[0095] S22. Use the T5-base model to extract text from the text data to obtain the corresponding text vectors;

[0096] S23. Using the pre-trained ViT image processing model, image embedding extraction and image content annotation are performed on the image data to obtain an image embedding vector containing the annotation text embedding.

[0097] S24. Use the Pandas tool in the Python tool library to transform the structured data to obtain key-value pair text vectors;

[0098] S25. Perform adaptive fusion processing on the text vector, image embedding vector, and key-value pair text vector to obtain multimodal fusion features; the specific steps of step S25 are as follows:

[0099] S251. Perform modality missing detection on the multimodal feature vectors obtained in steps S21-S24. If a missing feature vector is detected, generate a modality missing marker vector corresponding to the missing modality feature vector, and proceed to step S252. The number of modality missing marker vectors is at most two. If no missing feature vector is detected, perform multimodal feature fusion processing on the text vector, image embedding vector, and key-value pair text vector to obtain multimodal fused features. The multimodal feature vectors include text vectors, image embedding vectors, and key-value pair text vectors.

[0100] S252. If step S251 detects that there is only a single modality feature vector in the multimodal feature vector, then the standardized feature vector of the modality feature vector is used as the basic feature vector; if step S251 detects that there are two modality feature vectors in the multimodal feature vector, then the two modality feature vectors are dynamically weighted and fused according to the correlation level and data type to generate the basic feature vector.

[0101] S253. Map the modality missing marker vector obtained in step S251 to an auxiliary feature vector, and add it element by element to the basic feature vector obtained in step S252 to obtain the multimodal fusion feature.

[0102] S26. Repeat steps S21-S25 to process each file in the project evaluation data to obtain multimodal fusion feature data.

[0103] S3. Construct a building curtain wall evaluation dataset based on the multimodal fusion feature data, and divide it into a multimodal fusion training set; the building curtain wall evaluation dataset includes project ID, document type, relevance level, document path and multimodal fusion features;

[0104] S4. Construct a residual quantization variational autoencoder model, and train the residual quantization variational autoencoder model using the multimodal fusion training set to obtain a trained residual quantization variational autoencoder model.

[0105] Specifically, the residual quantization variational autoencoder model includes an encoder module, a residual quantization module, and a decoder module;

[0106] The encoder module adopts a 3-layer fully connected MLP network structure. The input layer dimension of the encoder module is set to 768 dimensions (consistent with the dimension of the multimodal fusion features), the hidden layer dimension is set to 512 dimensions, the activation function of the hidden layer is ReLU, and the output layer dimension is set to 256 dimensions (consistent with the dimension of the codebook vector), the activation function of the output layer is Tanh.

[0107] Specifically, the weights of the encoder module are initialized through multimodal fusion features during the model pre-training stage, and further optimized by combining correlation level loss during the fine-tuning stage, so as to align the latent vector space with the semantic distribution of the curtain wall maintenance data.

[0108] The residual quantization module is equipped with a 4-layer cascaded codebook, with each layer corresponding to four semantic levels: curtain wall type, data type, core content, and details.

[0109] The decoder module adopts a 3-layer fully connected MLP network structure symmetrical to the encoder module. The input layer dimension of the decoder module is set to 256 dimensions (consistent with the dimension of the quantized latent vector), the hidden layer dimension is set to 512 dimensions, the activation function of the hidden layer is ReLU, the output layer dimension is set to 768 dimensions (consistent with the dimension of the original multimodal fusion features), and the activation function of the output layer is Linear.

[0110] It should be noted that, in this embodiment, the function of the decoder module is to reconstruct the quantized latent vectors back to the original multimodal feature space. The reconstruction loss ensures that key semantic information is not lost during the quantization process, thereby driving the optimization of the model.

[0111] Specifically, the training parameters of the residual quantization variational autoencoder model are set as follows:

[0112] Key parameter settings:

[0113] The residual quantization variational autoencoder model has 4 codebook layers (m), and the semantics of each codebook layer are as follows: the first layer is the curtain wall type, the second layer is the data type, the third layer is the core content, and the fourth layer is the details; the size K of each codebook layer is set to 256.

[0114] Specifically, in this embodiment, the codebook initialization method adopts K-means clustering and is based on the document embedding of the training set; this setting is to facilitate the model to converge quickly.

[0115] Standard training and hyperparameter settings:

[0116] The optimizer for the residual quantization variational autoencoder model is AdamW; the batch size is set to 128, the number of training epochs is 300, λ1=0.01, λ2=0.001; and a training termination mechanism is set.

[0117] Specifically, in the pre-training phase: using only L RQ-VAE The residual quantization variational autoencoder model is trained for 50 epochs to ensure initial convergence of the codebook; in the fine-tuning stage: an alignment loss L is added. al Correlation loss L rel The residual quantization variational autoencoder model was then fine-tuned and trained for 250 rounds, and the "identifier semantic consistency" was evaluated using a validation set after every 10 rounds to ensure that the identifier prefix overlap rate of documents with the same relevance level met the requirements.

[0118] It should be noted that after the residual quantization variational autoencoder model is trained, the Sinkhorn algorithm needs to be used to process duplicate identifiers to ensure that each curtain wall document corresponds to a unique identifier.

[0119] Model verification indicator settings:

[0120] Core indicator: semantic consistency of identifiers (qualified when the prefix coincidence rate of L4 document identifiers is ≥80%); auxiliary indicator: basic error (L RQ-VAE <0.1 is qualified).

[0121] The total loss function of the residual quantized variational autoencoder model is expressed as follows:

[0122] L ID =L RQ-VAE +λ1L al +λ2L rel ;

[0123] Where, L ID is the total loss function, L RQ-VAE is the residual quantization loss function, λ1 and λ2 are weight coefficients, L al is the alignment loss function, L rel is the multi-level correlation learning loss function;

[0124] Specifically, the expression of L RQ-VAE as the residual quantization loss function is as follows:

[0125] ;

[0126] Where, L re is the basic reconstruction loss function, L rq is the codebook loss function.

[0127] Specifically, the expression of L re as the basic reconstruction loss function is as follows:

[0128] ;

[0129] Specifically, the expression of L rq as the codebook loss function is as follows:

[0130] ;

[0131] Where, sg[•] denotes the stop-gradient operation, which prevents gradient updates to the quantized embeddings during the backpropagation process; r l is the residual vector of the l-th layer; c l is the codebook of the l-th layer (a set of codebook vectors containing K 256-dimensional vectors).

[0132] Specifically, the expression of the multi-level correlation learning loss function is as follows:

[0133] Specifically, L rqThe first item in the guaranteed codebook vector With the corresponding residuals Approximately; the second term is determined by hyperparameters. Weighted, constrained residuals to remain close to the selected codebook entries.

[0134] Specifically, L al The alignment loss function is represented as follows:

[0135] ;

[0136] Where, ω j This indicates the alignment weight.

[0137] Specifically, the multi-level correlation learning loss function L rel The specific representation is as follows:

[0138] ;

[0139] Where m is the total number of codebook layers (m=4), and l is the current codebook layer (l=1,2,...,m). The weights are hierarchical and decrease with increasing layer number (β1=1.0, β2=0.75, β3=0.5, β4=0.25); Tl is the training triple combination of the l-th layer: where a is the anchor document, p is the positive sample document with a higher relevance than a, and n is the negative sample document with a lower relevance than a. These are the residual vectors of the anchor point, positive sample, and negative sample in the l-th layer, respectively. is the margin parameter, which is fixed at 0.2 in this scheme; sim(·,·) is the cosine similarity.

[0140] S5. The trained residual quantization variational autoencoder model is used to process each group of data in the building curtain wall evaluation dataset to generate the corresponding building curtain wall feature identifier.

[0141] The specific steps of step S5 are as follows:

[0142] S51. Obtain the multimodal fusion features contained in any set of data in the building curtain wall evaluation dataset, and denote them as semantic embedding vectors;

[0143] S52. Input the semantic embedding vector into the encoder module of the trained residual quantization variational autoencoder model. The encoder module maps the semantic embedding vector to a low-dimensional vector to obtain a low-dimensional latent vector.

[0144] Specifically, the core purpose of step S52 is to compress the semantic embedding into a low-dimensional latent vector while retaining the core semantic information, thus preparing for subsequent residual quantization.

[0145] S53. Then, the low-dimensional latent vector is input into the residual quantization module. The residual quantization module uses a multi-layer concatenated codebook to recursively quantize the low-dimensional latent vector, and finally generates a preliminary feature identifier.

[0146] Specifically, the residual quantization module process described in step S53 is illustrated in the following example:

[0147] Taking a glass curtain wall sealant replacement manual as an example:

[0148] After the residual quantization module receives the 256-dimensional latent vector Z output by the encoder module, its working steps are as follows:

[0149] 1) Level 1 quantization (corresponding to the first segment of the identifier:<a_na> (Curtain wall type)

[0150] Calculate the Euclidean distance between the hidden vector Z and all vectors in the first-level codebook: d 1j =||Zc 1j ||2, j=1~256; find the codebook vector c with the smallest distance. 1na Its index n a This is the value of the first segment identifier; calculate the first level residual: r1 = Zc 1na ;

[0151] Semantic meaning: n a =12 corresponds to "glass curtain wall", n a =15 corresponds to "stone curtain wall", n a =18 corresponds to "metal curtain wall"; Example output:<a_12> ;

[0152] 2) Second-level quantization (corresponding to the second segment of the identifier:<b_nb> (Data type)

[0153] Calculate the Euclidean distance between the residual r1 of layer 1 and all vectors of the codebook of layer 2: d 2j =||r1-c 2j ||2, j=1~256; find the codebook vector c with the smallest distance. 2nb Its index nb is the value of the second segment identifier; calculate the second level residual: r2 = r1 - c 2nb ;

[0154] Semantic meaning: nb=45 corresponds to "Curtain Wall Usage and Maintenance Documents", nb=32 corresponds to "Curtain Wall Inspection and Testing Documents", and nb=22 corresponds to "Curtain Wall Construction Documents"; Example output:<b_45> ;

[0155] 3) Third-level quantization (corresponding to the third segment of the identifier:<c_nc> (Core content)

[0156] Calculate the Euclidean distance between the residual r² of layer 2 and all vectors of the codebook in layer 3: d 3j =||r2-c 3j ||2, j=1~256; find the codebook vector c with the smallest distance. 3nc Its index nc is the value of the third segment identifier; calculate the third level residual: r3 = r2 - c 3nc ;

[0157] Semantic meaning: nc=89 corresponds to "Sealant Maintenance", nc=67 corresponds to "Insulating Glass Inspection", and nc=55 corresponds to "Bolt Tightening"; Example output:<c_89> ;

[0158] 4) Fourth-level quantization (corresponding to the fourth segment of the identifier:<d_nd> ,detail);

[0159] Calculate the Euclidean distance between the residual r3 of layer 3 and all vectors of the codebook of layer 4: d 4j =||r3-c 4j ||2, j=1~256; find the codebook vector c with the smallest distance. 4nd Its index nd is the value of the 4th segment identifier; calculate the final residual: r4 = r3 - c 4nd ;

[0160] Semantic meaning: nd=23 corresponds to "Sealant Replacement Operation", nd=24 corresponds to "Sealant Aging Detection", and nd=25 corresponds to "Sealant Selection Criteria"; Example output:<d_23> ;

[0161] 5) Initial identifier generation;

[0162] The indexes obtained from the four levels of quantization are concatenated in sequence to obtain the initial four-segment hierarchical identifier for the document:<a_12><b_45><c_89><d_23> (Semantic interpretation: Glass curtain wall → Usage and maintenance documents → Sealant maintenance → Replacement operation).

[0163] Specifically, the residual quantization module performs 4-level cascaded recursive residual quantization on the 256-dimensional latent vector Z. Each quantization level corresponds to a semantic level, generating a discrete index for that level, which is then concatenated into a 4-segment primary feature identifier. The residual quantization uses a "layer-by-layer quantization of residuals" approach, simulating the quantization effect of a large codebook with multiple small codebooks: the first codebook captures the coarsest-grained global semantics (curtain wall type), the second codebook captures the second-coarsest-grained category semantics (data type), the third codebook captures the medium-grained core semantics (core content), and the fourth codebook captures the finest-grained detailed semantics (specific parameters / steps).

[0164] S54. The decoder module is used to decode and reconstruct the preliminary feature identifier obtained in step S53 for verification. If the verification is successful, the preliminary feature identifier is output as the building curtain wall feature identifier. If the verification fails, the data obtained in step S51 is removed from the building curtain wall evaluation dataset.

[0165] The specific process of step S54 is as follows:

[0166] First, the preliminary feature identifier obtained in step S53 is input into the decoder module, and the reconstruction error between the reconstruction output and the original multimodal fusion feature is calculated (in this example, the preset threshold is set to 0.05). If the reconstruction error is less than the preset threshold, the verification is deemed successful, and the preliminary feature identifier is output as the building curtain wall feature identifier. If the reconstruction error is greater than or equal to the preset threshold, the quantization of the sample is deemed to have failed, and it is removed from the building curtain wall evaluation dataset.

[0167] Specifically, the building curtain wall feature identifier is a hierarchical identifier consisting of four segments, as illustrated in the following example:

[0168] (<a_12><b_45><c_89><d_23> );

[0169] Each segment corresponds to a semantic level; each semantic level is defined by...<x_n> Let x represent the codebook level, including four items: a, b, c, and d; n represent the category number, n∈[1:256]; then the identifier format for each curtain wall data is: <a_n a > <b_n b > <c_n c > <d_n d >,n a n b n c n d ∈[1:256];

[0170] S55. Repeat steps S51-S54 to process each group of data in the building curtain wall evaluation dataset.

[0171] S6. Construct a simulated query corresponding to each group of data in the building curtain wall evaluation dataset, and pair the simulated query with the building curtain wall feature identifier obtained in step S5 to obtain a query-identifier training pair set.

[0172] The specific steps of step S6 are as follows:

[0173] S61. Obtain the document title, key attributes, and document summary corresponding to each group of data in the building curtain wall evaluation dataset, and construct a corresponding simulation query based on the document title, key attributes, and document summary;

[0174] Specifically, the method for generating the simulated query is as follows:

[0175] For the i-th group of data in the building curtain wall evaluation dataset, extract its document title t. i Key attribute set A i (Including curtain wall type, data type, core testing indicators) and document summary i Based on the document title t i Key attribute set A i and document summaries i Simulated queries can be generated in the following three ways:

[0176] (1) Title-based query: directly query by document title t i As a query q i (1) , i.e. q i (1) =t i ;

[0177] (2) Attribute combination query: Randomly select 1 to 3 attributes from the key attribute set Ai and combine them in natural language to generate an attribute combination query q. i (2) For example, for a set of glass curtain wall inspection data, the key attributes are {curtain wall type: concealed frame glass curtain wall, inspection item: sealant bonding strength, building name: XX Building}, then the query "XX Building concealed frame glass curtain wall sealant bonding strength inspection" can be generated.

[0178] (3) Summary rewriting query: for document summary s i Extract and rewrite key information to generate a summary rewriting query. i (3) Specifically, extract document summaries. i The core entities (building name, curtain wall type, type of defect, and test results) are combined into a natural language query.

[0179] Each set of data generates 3 simulated queries, corresponding to the three methods mentioned above (q). i (k) The k-th simulated query for the i-th data set (k=1,2,3).

[0180] S62. Pair each simulated query generated in step S61 with the corresponding building curtain wall feature identifier obtained in step S5 to obtain a query-identifier training pair set.

[0181] The specific process of step S62 is as follows:

[0182] Specifically, for the i-th data set, its building curtain wall feature identifier is ID. i (like<a_12><b_45><c_89><d_23> Then, the following training pairs are constructed:

[0183] (q i (1) ID i ), (q i (2) ID i ), (q i (3) ID i )

[0184] Perform the above operation on all N sets of data to obtain the training set D. train ={(q j ID j )}, where j=1,2,3,...N;

[0185] S63. The query-identifier training set is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0186] Specifically, in this embodiment, the query-identifier training set contains at least 100 core queries (such as "annual inspection items for glass curtain walls", "sealing adhesive aging and replacement cycle", etc.), and each query corresponds to 5 to 10 related documents at the L2 to L4 level;

[0187] S7. Construct a curtain wall maintenance strategy generation model. Train the curtain wall maintenance strategy generation model using the query-identifier training set to obtain a trained curtain wall maintenance strategy generation model. The curtain wall maintenance strategy generation model adopts the GR generative recommendation model.

[0188] Specifically, in this embodiment, the GR generative recommendation model uses the T5 architecture model as its backbone; its training objective is: given a natural language query, to autoregressively generate the corresponding building curtain wall feature identifier sequence;

[0189] Specifically, the model training process is as follows:

[0190] S71. Constructing training data:

[0191] The query-identifier training pairs obtained in step S6 are used as training data; wherein, the input of each training sample is a query text, and the output is the building curtain wall feature identifier corresponding to the query (represented by the multi-layer codebook index sequence generated in step S5).

[0192] Data augmentation processing is performed on the training data, including: random sampling of multiple queries for the same document, synonym replacement and random deletion of query text, in order to expand the number of training samples;

[0193] S72, Model Architecture Design:

[0194] The curtain wall maintenance strategy generation model uses T5-base as the backbone network and includes two parts: an encoder and a decoder.

[0195] The encoder receives a token sequence of the query text, passes it through a multi-layer Transformer self-attention mechanism, and outputs a contextual semantic representation of the query.

[0196] The decoder generates the codebook indexes of the building curtain wall feature identifiers layer by layer in an autoregressive manner; specifically, the decoder generates the index c of the l-th layer codebook in the l-th step. l Its generation process uses the indices c1, c2, ..., c already generated in the previous l-1 steps. l−1 And the encoder output is a condition;

[0197] The size of the output layer vocabulary of the decoder is set to be consistent with the codebook size of each layer of RQ-VAE in step S4 (i.e., 256 codebook vectors per layer).

[0198] S73, Model Training:

[0199] The Teacher Forcing strategy is used for training; during training, the input to each step of the decoder is a prefix of the real identifier sequence, and the loss function is the sum of the cross-entropy losses of the codebook index predictions of each layer:

[0200] ;

[0201] Where L is the total number of identifier layers (consistent with the number of RQ-VAE layers in step S4, i.e., 4 layers), q is the encoded representation of the query text, and c l This is the index of the actual codebook at level l;

[0202] The training hyperparameters are set as follows: learning rate 3×10 −4 Batch size is 32; maximum training epochs are 50; optimizer is AdamW, weight decay coefficient is 0.01;

[0203] S74. Model Evaluation and Selection:

[0204] The model is evaluated using the test set defined in step S63; the evaluation metrics include:

[0205] Recall@K: The percentage of the Top-K generated candidate identifiers that match the correct document identifier;

[0206] MRR (Mean Reciprocal Rank): The mean of the reciprocals of the first occurrence position of the correct document identifier;

[0207] The model checkpoint with the best combined performance of Recall@10 and MRR was selected as the trained curtain wall maintenance strategy generation model.

[0208] S8. Obtain the user's curtain wall maintenance requirements and input them into the trained curtain wall maintenance strategy generation model to obtain curtain wall maintenance strategy information.

[0209] S81. Obtain the user's curtain wall maintenance requirements, and input the curtain wall maintenance requirements into the trained curtain wall maintenance strategy generation model described in step S7 to generate multiple candidate building curtain wall feature identifiers; the specific process is as follows:

[0210] (1) Obtain the user's curtain wall maintenance requirements, input the curtain wall maintenance requirements into the trained curtain wall maintenance strategy generation model in step S7, and the encoder generates a contextual semantic representation of the maintenance query based on the curtain wall maintenance requirements;

[0211] (2) The decoder generates identifiers layer by layer in an autoregressive manner based on the context semantic representation of the maintenance query. When generating the codebook index of the l-th layer, the decoder limits the candidate output range of the model to 256 valid indices in the codebook of the l-th layer of RQ-VAE trained in step S4, excludes the remaining invalid outputs, and ensures that each generated index corresponds to a real codebook vector.

[0212] (3) To obtain multiple candidate results, a multi-path parallel search strategy is adopted: at each step of decoding, the B generating paths with the highest current probability are retained simultaneously (preferably B=10), rather than retaining only a single optimal path; each path is independently extended to subsequent layers until all paths generate complete L-layer identifiers; finally, B candidate building curtain wall feature identifiers {id1,id2,…,id B} and their corresponding generation probabilities {p1,p2,…,p} B}

[0213] S82. Based on the B candidate building curtain wall feature identifiers obtained in step S81, the building curtain wall evaluation dataset is searched to obtain a preliminary matching document set; the specific steps of step S82 are as follows:

[0214] S821. For each candidate building curtain wall feature identifier id i The identifier index is used to find documents that exactly match it and add them to the candidate document set.

[0215] S822. If the number of completely matched documents is less than the preset threshold N min (Preferred N) min If the number of documents retrieved is 5, the matching criteria are gradually relaxed: initially, only the first L-1 levels of the identifier's index are required to be consistent. If this is still insufficient, the requirement is further relaxed to the first L-2 levels of the index, and so on, until the number of retrieved documents reaches N. min ;

[0216] S823. Perform deduplication on the recalled candidate documents to obtain a preliminary set of matching documents D={d1,d2,…,d…} M}, where M≥N min ;

[0217] Specifically, compared with traditional full-text search, the use of multiple candidate building curtain wall feature identifiers to search the building curtain wall evaluation dataset in step S82 can effectively improve the search speed. The search speed is more than 100 times faster than traditional full-text search, and the search time for a single request is less than 0.1 seconds, with a search accuracy of more than 90%.

[0218] S83. Based on the relevance level, the preliminary matched document set is filtered and sorted to obtain the core reference document set;

[0219] Specifically, the top-K documents (ideally K=5) are selected as reference documents; the ranking is based on a comprehensive score.

[0220] ;

[0221] Where, p gen (d j Sim(q,d) represents the probability of generating the identifier corresponding to the document in step S81 (if the same document is matched by multiple candidate identifiers, the highest generation probability is taken); j To maintain the required text and documents d j The semantic similarity is calculated by comparing the T5-base encoded vector of the maintenance requirement text with that of the document d. j The cosine similarity of the multimodal fusion feature vector (obtained in step S2) is obtained; α is the balance coefficient (preferably α=0.6).

[0222] Specifically, the core reference document set only selects the top 5 documents in the sorted list; this setting is to ensure that the generated strategy focuses on actionable steps while filtering out low-priority background documents.

[0223] S84. Based on the relevance level, perform hierarchical information extraction on the core reference document set, and mark each extracted information with its corresponding document identifier source information to obtain a structured core information set;

[0224] Specifically, in step S84, different extraction strategies are used for documents with different relevance levels in the core reference document set, as follows:

[0225] Practical execution level L4 document: Full text extraction operation steps, tool list, material requirements, and safety precautions;

[0226] Strategy Decision Level 3: Extract problem severity, maintenance cycle, and acceptance criteria;

[0227] Background support level L2: Extract key points of construction technology and material parameters;

[0228] Standard Reference Level L1: Only the corresponding standard number and mandatory clauses are extracted.

[0229] Specifically, in step S84, for each piece of extracted information, the four-segment identifier and page number of its source document are marked, and the extracted information is organized into a structured JSON format; this setting is to solve the problem of untraceable information source.

[0230] S85. Perform identifier tracing verification and curtain wall domain mandatory rule verification on all structured core information in the structured core information set obtained in step S84, and delete the structured core information that fails the verification.

[0231] Specifically, the process of performing identifier tracing verification and curtain wall domain mandatory rule verification on the structured core information set in step S85 is as follows:

[0232] Identifier source verification: Before generating each piece of key information, the curtain wall maintenance strategy generation model must match the source marker in the structured information. If the source of the corresponding document identifier cannot be found, the information is deleted directly.

[0233] Domain rule verification: The information generated by the curtain wall maintenance strategy generation model is compared with a predefined mandatory rule base for the curtain wall domain (the rule base includes mandatory clauses from standards such as the "Technical Specification for Glass Curtain Wall Engineering" JGJ102 and the "Technical Specification for Maintenance and Repair of Existing Curtain Walls" T / CECS 863-2021). If a mandatory rule is violated, automatic correction is performed and the basis is indicated. The automatic correction adopts a combination of rule matching and pre-trained language models: for clauses with clearly defined numerical values ​​(such as minimum safety distance, maximum allowable deformation, etc.), direct replacement is performed through rule matching; for clauses involving scheme selection, correction suggestions are generated based on the specification context through a pre-trained language model.

[0234] Specifically, after repeated verification, the dual verification mechanism in step S85 can reduce the error rate (illusion rate) of key information to below 4.1%, thereby effectively solving the problems of inaccurate knowledge and content illusion caused by the difficulty of large language models in the existing technology to directly extract professional information in this field.

[0235] S86. Integrate the verified structured core information according to the curtain wall maintenance industry standard template to generate standardized and executable curtain wall maintenance strategy information.

[0236] The industry standard template for curtain wall maintenance includes the following sections: Project Overview and Maintenance Requirements; Current Status Assessment and Defect Diagnosis of Curtain Wall; Maintenance Scheme Design and Technical Measures; Maintenance Cycle Plan and Inspection Arrangement; Safety Precautions and Emergency Plan; Maintenance Basis (Referenced Standards and Regulations).

[0237] The integration process is as follows: the information extracted from each level is filled in according to the order of the template chapters. For chapters in the template that require the integration of information from multiple reference documents, a pre-trained language model is used to summarize and fuse multiple pieces of information to generate a coherent and complete strategy text.

[0238] Specifically, the standardized and executable curtain wall maintenance strategy information includes project overview (project name, location, problem description), maintenance basis (referenced test report number, standard number, document identifier), detailed operation steps (step-by-step numbering, with a list of tools and materials), safety protection measures, quality acceptance standards, and subsequent maintenance recommendations.

[0239] Specifically, the output example of the curtain wall maintenance strategy information is as follows:

[0240] For example: User maintenance requirement (input): How to maintain the aging of glass curtain wall sealant?

[0241] Model output (final maintenance strategy information):

[0242] (1) Maintenance object:

[0243] Sealant around glass curtain wall panels and sealing strips for operable sashes.

[0244] (2) Maintenance basis:

[0245] The procedures shall be carried out in accordance with the curtain wall inspection report, maintenance manual, and relevant national standards for sealants.

[0246] (3) Problem Judgment:

[0247] If the sealant cracks, hardens, powders, or detaches, it is considered to have failed due to aging and needs to be replaced entirely.

[0248] (4) Maintenance steps:

[0249] Cleaning old adhesive: Use special tools to remove aged sealant and clean the gaps until there is no dust or oil.

[0250] Surface preparation: Dry and degrease the bonding surfaces to ensure bonding strength.

[0251] Application of sealant: Use qualified neutral silicone weather-resistant sealant, apply it continuously and fully, and smooth the surface.

[0252] Curing and maintenance: Allow to cure naturally for more than 24 hours, during which time avoid rain and external contact.

[0253] (5) Safety requirements:

[0254] Safety belts must be worn when working at heights, safety warning zones must be set up, and workers must be certified to work at heights.

[0255] (6) Acceptance criteria:

[0256] The sealant is continuous without bubbles or breaks, adheres firmly to the substrate, and shows no signs of detachment or cracking.

[0257] (7) Recommendations for the next maintenance:

[0258] It is recommended to conduct a comprehensive inspection of the sealant every 5 years and replace it every 8–10 years depending on its aging condition.

[0259] Example 2

[0260] like Figure 3 As shown, this embodiment provides an intelligent maintenance decision-making system for building curtain walls, used to implement the intelligent maintenance decision-making method for building curtain walls as described in Embodiment 1. The system includes:

[0261] The project evaluation data module is used to obtain project evaluation data for building curtain walls. The project evaluation data includes curtain wall design documents, curtain wall construction documents, curtain wall inspection and testing documents, curtain wall acceptance documents, curtain wall use and maintenance documents, and curtain wall management documents.

[0262] The preprocessing module is used to perform structured processing on the project evaluation data to obtain structured project evaluation data.

[0263] The multimodal feature fusion module is used to perform multimodal feature fusion processing on the structured project evaluation data to obtain multimodal fused feature data;

[0264] The building curtain wall evaluation dataset generation module is used to construct a building curtain wall evaluation dataset based on the multimodal fusion feature data, and to divide the dataset into a multimodal fusion training set. The building curtain wall evaluation dataset includes project ID, document type, relevance level, document path, and multimodal fusion features.

[0265] The residual quantization variational autoencoder model building module is used to build the initial residual quantization variational autoencoder model.

[0266] The residual quantization variational autoencoder model training module is used to train the residual quantization variational autoencoder model using the multimodal fusion training set to obtain a trained residual quantization variational autoencoder model; the residual quantization variational autoencoder model includes an encoder module, a residual quantization module, and a decoder module.

[0267] The building curtain wall feature identifier module is used to process each group of data in the building curtain wall evaluation dataset using a trained residual quantization variational autoencoder model to generate the corresponding building curtain wall feature identifier.

[0268] The query-identifier training pair set construction module is used to construct a simulated query corresponding to each group of data in the building curtain wall evaluation dataset. The simulated query is then paired with the building curtain wall feature identifier obtained by the building curtain wall feature identifier module to obtain the query-identifier training pair set.

[0269] The curtain wall maintenance strategy generation model building module is used to build the initial curtain wall maintenance strategy generation model;

[0270] The curtain wall maintenance strategy generation model training module is used to train the curtain wall maintenance strategy generation model using a query-identifier training set to obtain a trained curtain wall maintenance strategy generation model.

[0271] The curtain wall maintenance strategy information generation module is used to obtain the user's curtain wall maintenance requirements and input them into the trained curtain wall maintenance strategy generation model to obtain curtain wall maintenance strategy information.

[0272] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. The system can be applied to the intelligent maintenance decision-making method for building curtain walls described in Embodiment 1 above.

[0273] Example 3

[0274] like Figure 4As shown, this embodiment provides a computer device, which includes a processor 102, a memory, an input device 103, a display 104, and a network interface 105 connected via a system bus 101. The processor 102 provides computing and control capabilities. The memory includes a non-volatile storage medium 106 and internal memory 107. The non-volatile storage medium 106 stores an operating system, computer programs, and a database. The internal memory 107 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium 106. When the computer programs are executed by the processor 102, they implement the intelligent maintenance decision-making method for building curtain walls described in Embodiment 1.

[0275] Example 4

[0276] like Figure 5 As shown, this embodiment provides a storage medium storing a program. When the program is executed by a processor, it implements the intelligent maintenance decision-making method for building curtain walls described in Embodiment 1.

[0277] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0278] In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this embodiment, the computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0279] The computer-readable storage medium described above can be used to write computer programs for executing this embodiment in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Java, Python, and C++—and conventional procedural programming languages—such as C or similar programming languages. The program can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0280] The above embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any changes or other equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.

Claims

1. A smart maintenance decision-making method for building curtain walls, characterized in that, Includes the following steps: S1. Obtain project evaluation data for the building curtain wall; the project evaluation data includes curtain wall design documents, curtain wall construction documents, curtain wall inspection and testing documents, curtain wall acceptance documents, curtain wall use and maintenance documents, and curtain wall management documents; S2. The project evaluation data is structured, and then the structured project evaluation data is subjected to multimodal feature fusion processing to obtain multimodal fusion feature data. S3. Construct a building curtain wall evaluation dataset based on the multimodal fusion feature data, and divide it into a multimodal fusion training set; the building curtain wall evaluation dataset includes project ID, document type, relevance level, document path and multimodal fusion features; S4. Construct a residual quantization variational autoencoder model, and train the residual quantization variational autoencoder model using the multimodal fusion training set to obtain a trained residual quantization variational autoencoder model; the residual quantization variational autoencoder model includes an encoder module, a residual quantization module, and a decoder module. S5. The trained residual quantization variational autoencoder model is used to process each group of data in the building curtain wall evaluation dataset to generate the corresponding building curtain wall feature identifier. S6. Construct a simulated query corresponding to each group of data in the building curtain wall evaluation dataset, and pair the simulated query with the building curtain wall feature identifier obtained in step S5 to obtain a query-identifier training pair set. S7. Construct a curtain wall maintenance strategy generation model. Train the curtain wall maintenance strategy generation model using the query-identifier training set to obtain a trained curtain wall maintenance strategy generation model. The curtain wall maintenance strategy generation model adopts the GR generative recommendation model. S8. Obtain the user's curtain wall maintenance requirements and input them into the trained curtain wall maintenance strategy generation model to obtain curtain wall maintenance strategy information.

2. The intelligent maintenance decision-making method for building curtain walls according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Based on the importance of curtain wall maintenance, the documents in the project assessment data are labeled with relevance levels, and then the labeled documents are classified to obtain the text data, image data and structured data corresponding to the labeled documents. S22. Use the T5-base model to extract text from the text data to obtain the corresponding text vectors; S23. Using the pre-trained ViT image processing model, image embedding extraction and image content annotation are performed on the image data to obtain an image embedding vector containing the annotation text embedding. S24. Use the Pandas tool in the Python tool library to transform the structured data to obtain key-value pair text vectors; S25. Perform adaptive fusion processing on the text vector, image embedding vector, and key-value pair text vector to obtain multimodal fusion features; S26. Repeat steps S21-S25 to process each file in the project evaluation data to obtain multimodal fusion feature data.

3. The intelligent maintenance decision-making method for building curtain walls according to claim 2, characterized in that, The specific steps of step S25 are as follows: S251. Perform modality missing detection on the multimodal feature vectors obtained in steps S21-S24. If a missing feature vector is detected, generate a modality missing marker vector corresponding to the missing modality feature vector, and proceed to step S252. The number of modality missing marker vectors is at most two. If no missing feature vector is detected, perform multimodal feature fusion processing on the text vector, image embedding vector, and key-value pair text vector to obtain multimodal fused features. The multimodal feature vectors include text vectors, image embedding vectors, and key-value pair text vectors. S252. If step S251 detects that there is only a single modality feature vector in the multimodal feature vector, then the standardized feature vector of the modality feature vector is used as the basic feature vector; if step S251 detects that there are two modality feature vectors in the multimodal feature vector, then the two modality feature vectors are dynamically weighted and fused according to the correlation level and data type to generate the basic feature vector. S253. Map the modality missing marker vector obtained in step S251 to an auxiliary feature vector, and add it element-wise to the basic feature vector obtained in step S252 to obtain the multimodal fusion feature.

4. The intelligent maintenance decision-making method for building curtain walls according to claim 2, characterized in that, The relevance level labeling in step S21 includes practical execution level L4, strategy decision level L3, background support level L2, standard reference level L1, and irrelevant information level L0.

5. The intelligent maintenance decision-making method for building curtain walls according to claim 1, characterized in that, The specific steps of step S5 are as follows: S51. Obtain the multimodal fusion features contained in any set of data in the building curtain wall evaluation dataset, and denote them as semantic embedding vectors; S52. Input the semantic embedding vector into the encoder module of the trained residual quantization variational autoencoder model. The encoder module maps the semantic embedding vector to a low-dimensional vector to obtain a low-dimensional latent vector. S53. Then, the low-dimensional latent vector is input into the residual quantization module. The residual quantization module uses a multi-layer concatenated codebook to recursively quantize the low-dimensional latent vector, and finally generates a preliminary feature identifier. S54. The decoder module is used to decode and reconstruct the preliminary feature identifier obtained in step S53 for verification. If the verification is successful, the preliminary feature identifier is output as the building curtain wall feature identifier. If the verification fails, the data obtained in step S51 is removed from the building curtain wall evaluation dataset. S55. Repeat steps S51-S54 to process each group of data in the building curtain wall evaluation dataset.

6. The intelligent maintenance decision-making method for building curtain walls according to claim 1, characterized in that, The total loss function of the residual quantization variational autoencoder model is expressed as follows: L ID =L RQ-VAE +λ1L al +λ2L rel ; Among them, L ID Let L be the total loss function. RQ-VAE Let L be the residual quantization loss function, λ1 and λ2 be the weighting coefficients, and L be the weighting coefficients. al For the alignment loss function, L rel This is the loss function for multi-level correlation learning.

7. The intelligent maintenance decision-making method for building curtain walls according to claim 1, characterized in that, The specific steps of step S8 are as follows: S81. Obtain the user's curtain wall maintenance requirements, input the curtain wall maintenance requirements into the trained curtain wall maintenance strategy generation model described in step S7, and generate multiple candidate building curtain wall feature identifiers. S82. Based on the multiple candidate building curtain wall feature identifiers obtained in step S81, the building curtain wall evaluation dataset is searched to obtain a preliminary matching document set; S83. Based on the relevance level, the preliminary matched document set is filtered and sorted to obtain the core reference document set; S84. Based on the relevance level, perform hierarchical information extraction on the core reference document set, and mark each extracted information with its corresponding document identifier source information to obtain a structured core information set; S85. Perform identifier tracing verification and curtain wall domain mandatory rule verification on all structured core information in the structured core information set obtained in step S84, and delete the structured core information that fails the verification. S86. Integrate the verified structured core information according to the curtain wall maintenance industry standard template to generate standardized and executable curtain wall maintenance strategy information.

8. An intelligent maintenance decision-making system for building curtain walls, used to implement the intelligent maintenance decision-making method for building curtain walls as described in any one of claims 1-7, characterized in that, The system includes: The project evaluation data module is used to obtain project evaluation data for building curtain walls. The project evaluation data includes curtain wall design documents, curtain wall construction documents, curtain wall inspection and testing documents, curtain wall acceptance documents, curtain wall use and maintenance documents, and curtain wall management documents. The preprocessing module is used to perform structured processing on the project evaluation data to obtain structured project evaluation data. The multimodal feature fusion module is used to perform multimodal feature fusion processing on the structured project evaluation data to obtain multimodal fused feature data; The building curtain wall evaluation dataset generation module is used to construct a building curtain wall evaluation dataset based on the multimodal fusion feature data, and to divide the dataset into a multimodal fusion training set. The building curtain wall evaluation dataset includes project ID, document type, relevance level, document path, and multimodal fusion features. The residual quantization variational autoencoder model building module is used to build the initial residual quantization variational autoencoder model. The residual quantization variational autoencoder model training module is used to train the residual quantization variational autoencoder model using the multimodal fusion training set to obtain a trained residual quantization variational autoencoder model; the residual quantization variational autoencoder model includes an encoder module, a residual quantization module, and a decoder module. The building curtain wall feature identifier module is used to process each group of data in the building curtain wall evaluation dataset using a trained residual quantization variational autoencoder model to generate the corresponding building curtain wall feature identifier. The query-identifier training pair set construction module is used to construct a simulated query corresponding to each group of data in the building curtain wall evaluation dataset. The simulated query is then paired with the building curtain wall feature identifier obtained by the building curtain wall feature identifier module to obtain the query-identifier training pair set. The curtain wall maintenance strategy generation model building module is used to build the initial curtain wall maintenance strategy generation model; The curtain wall maintenance strategy generation model training module is used to train the curtain wall maintenance strategy generation model using a query-identifier training set to obtain a trained curtain wall maintenance strategy generation model. The curtain wall maintenance strategy information generation module is used to obtain the user's curtain wall maintenance requirements and input them into the trained curtain wall maintenance strategy generation model to obtain curtain wall maintenance strategy information.

9. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the intelligent maintenance decision-making method for building curtain walls as described in any one of claims 1-7.

10. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the intelligent maintenance decision-making method for building curtain walls as described in any one of claims 1-7.

Images