Building aging prediction method and building aging prediction system based on big data modeling

Abstract

The application belongs to the technical field of building engineering, and discloses a building aging prediction method and system based on big data modeling, which comprises the following steps: S1, acquiring multi-source heterogeneous building related data; S2, preprocessing the multi-source heterogeneous building related data to extract multi-dimensional features related to building aging; S3, constructing a building aging knowledge graph; S4, constructing a multi-modal deep learning model; S5, training the multi-modal deep learning model; S6, inputting the processed building related data to be predicted into the trained multi-modal deep learning model for prediction according to steps S1-S3. The application uses a deep learning model to perform deep learning and multi-scale time series analysis on the structure health monitoring data, environmental parameters, material properties and historical maintenance records of the whole life cycle of the building, accurately capturing the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process.

Description

Technical Field

[0001] This invention belongs to the field of building engineering technology, and specifically relates to a building aging prediction method and system based on big data modeling. Background Technology

[0002] With the continuous development of the construction industry, the aging problem of existing buildings is becoming increasingly prominent, and their accurate prediction and effective maintenance have become a research hotspot in the current engineering field. Traditional methods for assessing and predicting building aging mainly rely on empirical judgment or statistical analysis based on limited data. However, faced with increasingly complex and variable building structures and environmental factors, existing technologies still face significant challenges in terms of data integration, model accuracy, and prediction efficiency.

[0003] Specifically, existing solutions struggle to effectively integrate heterogeneous data from multiple sources, including sensors, historical maintenance records, material properties, and environmental monitoring, leading to information fragmentation. Their aging prediction models often lack accuracy and have limited ability to characterize long-term aging trends, making them ill-suited to the stringent requirements of efficiency and accuracy in practical engineering. Furthermore, the lack of big data modeling capabilities makes it difficult for prediction results to fully reflect the true aging state of buildings, thus affecting the scientific rigor and timeliness of maintenance and renovation decisions. Therefore, we propose a building aging prediction method and system based on big data modeling to address these issues. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a building aging prediction method and system based on big data modeling, which solves the problem of low prediction accuracy for building aging.

[0005] This invention is achieved through the following scheme: a building aging prediction method based on big data modeling, comprising the following steps: S1. Obtain multi-source heterogeneous building-related data, including building structural health monitoring data, building environmental climate data, building material property data, building historical maintenance data, and building design data; S2. Preprocess the multi-source heterogeneous building-related data to generate a standardized dataset, and perform feature engineering on the dataset to extract multi-dimensional features related to building aging. S3. Construct a knowledge graph of building aging to represent the relationship between building components, building materials, building environment and building aging phenomena; S4. Construct a multimodal deep learning model and integrate multi-dimensional features and a building aging knowledge graph to learn the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process. S5. Train the multimodal deep learning model; S6. After processing the building-related data to be predicted according to steps S1-S3, input it into the trained multimodal deep learning model for prediction, and output the prediction result.

[0006] A further improvement of the building aging prediction method based on big data modeling in this invention is that the building structural health monitoring data includes the building structure's strain, acceleration, displacement, tilt angle, crack width, temperature, and humidity. The building environment climate data includes ambient temperature, ambient humidity, wind speed, wind direction, rainfall, and solar radiation intensity; The building material property data includes the compressive strength, tensile strength, modulus of elasticity, durability, corrosion rate, and fatigue performance of concrete, steel, brick, stone, and wood. The building's historical maintenance data includes all inspection reports, maintenance dates, maintenance content, maintenance materials, maintenance costs, and fault types. The architectural design data includes design drawings, construction specifications, completion acceptance reports, load conditions, and geographical location.

[0007] A further improvement of the building aging prediction method based on big data modeling in this invention lies in the fact that the preprocessing of multi-source heterogeneous building-related data specifically includes the following steps: S2011. Perform timestamp synchronization, data alignment, missing value filling, and outlier filtering on building structural health monitoring data to form a continuous and complete time-series data stream; S2012. Perform timestamp synchronization, data alignment, data cleaning, and format conversion on building environment and climate data to ensure synchronization with building structural health monitoring data. S2013. Perform structured processing on building material attribute data and convert unstructured text information into quantifiable feature parameters; S2014. Perform event serialization and feature extraction on building history maintenance data, and identify key maintenance event types, occurrence times, and impacts on building performance from text descriptions; S2015. Encode architectural design data and convert design parameters and construction specifications into numerical features; S2016. Normalize all the processed data from steps S2011-S2015 to eliminate the influence between data of different dimensions and orders of magnitude.

[0008] A further improvement of the building aging prediction method based on big data modeling in this invention lies in the fact that the feature engineering processing of the dataset specifically includes the following steps: S2021. Extract time-domain features, frequency-domain features, and time-frequency-domain features from time-series data of building structural health monitoring; S2022. Extract long-term trend characteristics, seasonal characteristics, and extreme event characteristics from built environment and climate data; S2023. Extract material degradation rate parameters and initial performance parameters from building material property data; S2024. Extract maintenance frequency, average repair time, fault recurrence rate, and maintenance cost percentage from building historical maintenance data; S2025. Extract building structure type, number of floors, total area, design life and dimensions of key building components from architectural design data; S2026. Dimensionality reduction of the features extracted in steps S2021-S2025 is performed through principal component analysis or independent component analysis.

[0009] A further improvement of the building aging prediction method based on big data modeling in this invention is that the time-domain features include mean, variance, standard deviation, kurtosis, skewness, waveform factor, impulse factor, margin factor, and kurtosis factor; the frequency-domain features include the main frequency components after fast Fourier transform, peak power spectral density, and energy center frequency; and the time-frequency-domain features include wavelet energy and wavelet entropy after wavelet transform, as well as instantaneous frequency and instantaneous amplitude after Hilbert-Huang transform.

[0010] A further improvement of the building aging prediction method based on big data modeling in this invention lies in the fact that the construction of the building aging knowledge graph specifically includes the following steps: Define the entity types of the building aging knowledge graph, which include buildings, components, materials, environmental factors, aging phenomena, maintenance activities, and damage types; Define the association types of the knowledge graph of building aging, including inclusion, constituent factors, causes of impact, phenomena that lead to the occurrence, areas of occurrence, repair parts, and manifestations; Entities and relationships are extracted from architectural design data, building material attribute data, and building history maintenance data to construct the initial framework of a knowledge graph of building aging. By using natural language processing technology, new entities and relationships can be identified and extracted from unstructured building history maintenance data to expand the knowledge graph of building aging.

[0011] A further improvement of the building aging prediction method based on big data modeling in this invention lies in the following steps in constructing the multimodal deep learning model: S401. Construct a multi-scale temporal convolutional network module to process the temporal characteristics of building structural health monitoring data and building environment climate data, and capture long-term dependencies and local temporal patterns. S402. Construct a graph neural network module to process the knowledge graph of building aging, learn the structured correlation features between components, materials and environmental factors, and capture the propagation path of damage in the structure. S403. Construct a multilayer perceptron network module to process static features such as building material properties, building history maintenance features, and building design features. S404. Construct a feature fusion module to deeply fuse the outputs of the multi-scale temporal convolutional network module, the graph neural network module, and the multilayer perceptron network module. S405. Construct an aging prediction output module to predict building aging indicators based on the output of the feature fusion module.

[0012] A further improvement of the building aging prediction method based on big data modeling in this invention is that the aging indicators include the probability of damage to a certain structural component in the next year, the percentage of peeling area of ​​interior and exterior wall coatings in the next three years, and the predicted maintenance cost of a specific device in the next five years.

[0013] A further improvement of the building aging prediction method based on big data modeling in this invention is that step S5 specifically includes the following steps: S501. Use the multi-dimensional features and the building aging knowledge graph as input to the model; S502. Use the actual observed building aging indicators as the training target for the model; S503. The backpropagation algorithm is used to optimize the model parameters; S504. Use historical datasets for model training and validation, and adjust the model's hyperparameters using a cross-validation strategy.

[0014] A building aging prediction system is provided for implementing the building aging prediction method described above, the system comprising: The data acquisition module is used to acquire multi-source heterogeneous building-related data, including building structural health monitoring data, building environment climate data, building material property data, building historical maintenance data, and building design data. The data processing module is used to preprocess multi-source heterogeneous building-related data to generate a standardized dataset, and to perform feature engineering on the dataset to extract multi-dimensional features related to building aging. The building aging knowledge graph construction module is used to construct a building aging knowledge graph to represent the relationship between building components, building materials, building environment and building aging phenomena. The deep learning model building module is used to build multimodal deep learning models and integrate multi-dimensional features and building aging knowledge graphs to learn the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process. The model training module is used to train the multimodal deep learning model.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a multi-source heterogeneous data fusion architecture and utilizes a deep learning model to perform deep learning and multi-scale time-series analysis on structural health monitoring data, environmental parameters, material properties, and historical maintenance records throughout the building's entire life cycle. This accurately captures the nonlinear dynamic evolution laws and multi-factor coupling effects during the building aging process. Furthermore, it introduces knowledge graph technology to structurally represent the complex relationships between building components, materials, and environmental factors, improving the model's understanding and prediction accuracy of aging mechanisms. This provides a scientific basis and technical support for the maintenance and renovation of existing buildings and has significant practical application value. Attached Figure Description

[0016] Figure 1 A flowchart of the building aging prediction method of the present invention is shown.

[0017] Figure 2 The flowchart illustrating the preprocessing of multi-source heterogeneous building-related data according to the present invention is shown.

[0018] Figure 3 A flowchart illustrating the feature engineering process of the dataset according to the present invention is shown.

[0019] Figure 4 The flowchart illustrating the construction of a multimodal deep learning model according to the present invention is shown.

[0020] Figure 5 A flowchart illustrating the training process of the multimodal deep learning model according to the present invention is shown. Detailed Implementation

[0021] To address the problem of low accuracy in building aging prediction, this invention provides a building aging prediction method and system based on big data modeling. The following detailed description, in conjunction with accompanying drawings, provides further illustration of this big data modeling-based building aging prediction method and system.

[0022] See Figures 1-5 As shown, a building aging prediction method based on big data modeling includes the following steps: S1. Obtain relevant data on multi-source heterogeneous buildings, including building structural health monitoring data, building environmental climate data, building material property data, building historical maintenance data, and building design data; S2. Preprocess the multi-source heterogeneous building-related data to generate a standardized dataset, and perform feature engineering on the dataset to extract multi-dimensional features related to building aging. S3. Construct a knowledge graph of building aging to represent the relationship between building components, building materials, building environment and building aging phenomena; S4. Construct a multimodal deep learning model and integrate multi-dimensional features and a building aging knowledge graph to learn the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process. S5. Train the multimodal deep learning model; S6. After processing the building-related data to be predicted according to steps S1-S3, input it into the trained multimodal deep learning model for prediction, and output the prediction result.

[0023] By constructing a multi-source heterogeneous data fusion architecture, deep learning models are used to perform deep learning and multi-scale time-series analysis on structural health monitoring data, environmental parameters, material properties, and historical maintenance records throughout the building's entire life cycle. This accurately captures the nonlinear dynamic evolution laws and multi-factor coupling effects during the building aging process. Furthermore, knowledge graph technology is introduced to structurally represent the complex relationships between building components, materials, and environmental factors, improving the model's understanding and prediction accuracy of aging mechanisms. This provides a scientific basis and technical support for the maintenance and renovation of existing buildings, and has significant practical application value.

[0024] Among them, the building structure health monitoring data includes the building structure's strain, acceleration, displacement, tilt angle, crack width, temperature, and humidity; Building environment climate data includes ambient temperature, ambient humidity, wind speed, wind direction, rainfall, and solar radiation intensity; Building material property data includes compressive strength, tensile strength, modulus of elasticity, durability, corrosion rate, fatigue performance, production batch information, and quality inspection report data for concrete, steel, brick, stone, and wood. Building history maintenance data includes all inspection reports, maintenance dates, maintenance content, maintenance materials, maintenance costs, and types of malfunctions; Architectural design data includes design drawings, construction specifications, completion acceptance reports, load conditions, and geographical location.

[0025] Furthermore, by deploying a sensor array on the building structure, data on strain, acceleration, displacement, tilt angle, crack width, temperature, and humidity of the building structure are collected in real time; the sensor array includes strain sensors, acceleration sensors, displacement sensors, tilt sensors, crack gauges, temperature sensors, and humidity sensors. The system uses meteorological monitoring equipment to collect real-time data on ambient temperature, humidity, wind speed, wind direction, rainfall, and solar radiation intensity in the area where the building is located. By comprehensively collecting various information related to building aging, it is found that the aging process is influenced by a combination of factors, and a single data source often cannot provide a complete representation. Therefore, it is necessary to obtain data from multiple dimensions and multiple sources to construct a comprehensive and in-depth dataset.

[0026] Specifically, in this embodiment, the building structural health monitoring data includes, but is not limited to, the building's design drawings, construction records, material testing reports, and real-time data collected by the structural health monitoring system. Design drawings provide initial static information such as the building's geometric dimensions, structural layout, and component connection methods. Construction records contain key process data such as concrete pouring time, curing conditions, rebar type and arrangement, and special construction techniques; this data is crucial for assessing the initial performance and potential defects of materials. Material testing reports provide quantitative indicators such as concrete strength grade, steel yield strength, coating thickness, and adhesion, reflecting the intrinsic properties of building materials. More importantly, a network of structural health monitoring sensors deployed at key structural locations acquires dynamic data in real time. For example, strain sensors monitor the deformation of components under load, displacement sensors monitor the settlement and lateral movement of the overall structure or local components, and vibration sensors monitor the building's response frequency and amplitude characteristics under wind loads, earthquakes, or human activities. These sensors continuously collect data at a preset sampling frequency (e.g., 100 to 1000 times per second) and transmit it to a central data aggregation platform via wired or wireless communication protocols (e.g., Industrial Ethernet protocol or 5G mobile communication technology). The data format is typically time-series numerical data, including timestamps, sensor identifiers, measured values, and units. Data verification mechanisms include sensor self-checks, data range checks, and comparisons with historical data trends to ensure data accuracy and integrity. Typical value ranges include, for example, strain values ​​for concrete members within ±1,000 microstrains, and structural vibration frequencies within 0.1 Hz to 50 Hz.

[0027] Building environmental climate data includes, but is not limited to, weather station data, satellite remote sensing data, and environmental sensors deployed on the building's exterior. Weather stations provide long-term historical meteorological data for the building's location, such as daily maximum / minimum temperatures, average humidity, wind speed, wind direction, rainfall, and solar radiation intensity. This data is typically recorded hourly or daily and acquired through standard meteorological data interfaces. Temperature and humidity sensors, anemometers, and rain gauges deployed on the building's exterior provide real-time environmental parameters for the building's specific microclimate zone. For example, temperature probes installed on exposed areas such as building facades and roofs monitor surface temperature fluctuations in real time to capture the effects of sunlight and wind cooling on material thermal stress. Humidity sensors monitor relative humidity, which is crucial for assessing the risk of moisture erosion and mold growth in materials. This environmental data is typically collected every minute to every hour, and the data type is numerical time-series data. Geographic Information System (GIS) data provides static geographic information such as the building's geological conditions, soil type, groundwater level, and topographic slope, as well as historical seismic activity records of the surrounding area. This information is essential for assessing foundation settlement and seismic risk. Data verification is conducted through cross-referencing with data from nearby weather stations and checking for consistency with historical data. Typical value ranges include, for example, ambient temperatures between -40°C and 50°C, and relative humidity between 10% and 98%.

[0028] Building usage data includes, but is not limited to, building energy management systems, occupancy counting systems, and building management systems. Energy management systems record building energy consumption patterns, such as the operating time, power consumption, and duration of equipment like air conditioning, lighting, and elevators. This data indirectly reflects the operating load and usage intensity of internal building equipment, thus affecting the fatigue life and wear of related components. Occupancy counting systems monitor occupancy density and flow patterns in different areas of the building, such as daily peak traffic and area occupancy rates at different times. High-density areas may lead to higher localized wear or load accumulation. Building management systems record the usage of internal building facilities, such as elevator operation counts, fire protection system self-inspection records, and access control system opening and closing records. This data provides a reference for equipment lifespan and maintenance cycles. Data collection frequencies range from per minute to per day, and data types include discrete events or aggregated statistics. Data validation is achieved by comparing data with the building's design capacity and historical usage patterns. For example, elevators may operate between 5,000 and 20,000 times per month.

[0029] Building history maintenance data includes, but is not limited to, maintenance management systems, manual inspection reports, repair work order records, and historical damage photos. The maintenance management system records the date, content, materials used, cost, problems found, and solutions for each maintenance session. For example, it may include the location of concrete crack repairs, the type of repair material, and the post-repair effect evaluation. Manual inspection reports describe in detail the type, extent, and location of damage visually observed by inspectors, accompanied by photos and text descriptions. Repair work orders record information such as the time of equipment failure, the symptoms, the repair personnel, the repair time, and the parts replaced. Historical damage photos and video recordings provide a visual representation of the damage's evolution. This data is typically in unstructured text, semi-structured tables, and image formats, requiring additional structuring. Data verification ensures accuracy by cross-referencing maintenance records from different sources, such as repair work orders, with actual maintenance reports. Typical maintenance records include, but are not limited to, "Local cracking of the roof waterproofing layer, grouting performed," and "Peeling of exterior wall paint, approximately ten square meters, repainted."

[0030] The preprocessing of multi-source heterogeneous building-related data specifically includes the following steps: S2011. Perform timestamp synchronization, data alignment, missing value filling, and outlier filtering on building structural health monitoring data to form a continuous and complete time-series data stream; S2012. Perform timestamp synchronization, data alignment, data cleaning, and format conversion on building environment and climate data to ensure synchronization with building structural health monitoring data. S2013. Perform structured processing on building material attribute data and convert unstructured text information into quantifiable feature parameters; S2014. Perform event serialization and feature extraction on building history maintenance data, and identify key maintenance event types, occurrence times, and impacts on building performance from text descriptions; S2015. Encode architectural design data and convert design parameters and construction specifications into numerical features; S2016. Normalize all the processed data from steps S2011-S2015 to eliminate the influence between data of different dimensions and orders of magnitude.

[0031] The specific steps involved in feature engineering of the dataset are as follows: S2021. Extract time-domain features, frequency-domain features, and time-frequency-domain features from time-series data of building structural health monitoring; S2022. Extract long-term trend characteristics, seasonal characteristics, and extreme event characteristics from built environment and climate data; S2023. Extract material degradation rate parameters and initial performance parameters from building material property data; S2024. Extract maintenance frequency, average repair time, fault recurrence rate, and maintenance cost percentage from building historical maintenance data; S2025. Extract building structure type, number of floors, total area, design life and dimensions of key building components from architectural design data; S2026. Dimensionality reduction of the features extracted in steps S2021-S2025 is performed through principal component analysis or independent component analysis.

[0032] The time-domain features include mean, variance, standard deviation, kurtosis, skewness, waveform factor, impulse factor, margin factor, and kurtosis factor; the frequency-domain features include the main frequency components after the fast Fourier transform, the peak power spectral density, and the energy center frequency; the time-frequency domain features include wavelet energy and wavelet entropy after the wavelet transform, and the instantaneous frequency and instantaneous amplitude after the Hilbert-Huang transform; and the extreme event features include cumulative sunshine hours, cumulative rainfall, and the duration of extreme temperatures.

[0033] Further, the acquired data undergoes cleaning, starting with missing value handling: For missing values ​​in building structural health monitoring data due to transient sensor malfunctions or communication interruptions, time-series interpolation methods, such as linear interpolation based on sliding windows, polynomial interpolation, or spline interpolation, are used to fill in the missing values ​​based on adjacent valid data points. For a small number of missing values ​​in non-time-series data (e.g., material properties), the mean, median, or mode can be used for filling. When the missing value ratio is too high (e.g., exceeding 20%), the availability of the data source needs to be assessed, or more complex model-based methods (e.g., principal component analysis reconstruction) should be considered for filling. The second step is outlier detection and handling: Extreme outliers in the data are identified using statistical methods, such as the three-standard-deviation rule, box plot analysis, or machine learning methods, such as isolation forests and local anomaly factor algorithms. Identified outliers are processed according to their properties, such as replacing them with the mean or median of neighboring points, or directly marking them as missing values ​​before interpolation. Finally, data denoising is performed: for signals with random noise acquired by sensors, digital filtering techniques, such as Butterworth filters and Kalman filters, are used to smooth the signal and remove high-frequency noise while retaining valid information. For example, structural vibration signals may require low-pass filtering to eliminate high-frequency environmental noise. Data verification is conducted throughout the process; the processed data needs to be visually inspected to confirm the cleaning effect.

[0034] Data integration and alignment: This addresses the fusion of data from different sources and time scales. First, timestamp alignment: all time-series data is unified to a preset sampling frequency and time base. For example, if environmental data is collected hourly while structural health monitoring data is collected every second, high-frequency data needs to be downsampled or low-frequency data upsampled and interpolated to ensure all data are aligned in the time dimension. Common alignment methods include nearest neighbor matching, linear interpolation, or cubic spline interpolation. Second, spatial coordinate system transformation: for data involving spatial location (e.g., damage sites, sensor locations), it is unified into a unified three-dimensional coordinate system of the building for spatial correlation analysis. For data of different granularities, such as daily average temperature and instantaneous strain, aggregation or decomposition is required to ensure logical consistency of features. The data structure is unified into a multi-dimensional time-series tensor; for example, the structural modal data of a building can be represented as a three-dimensional tensor equal to the time step multiplied by the number of sensors multiplied by the feature dimension.

[0035] Feature extraction and construction involves transforming raw data into advanced features that effectively characterize the aging state of buildings. For time-series data, such as structural strain and temperature fluctuations, statistical features (mean, variance, kurtosis, skewness, maximum, minimum, rate of change), frequency domain features (dominant frequency after Fourier transform and wavelet transform, energy distribution), and temporal domain features (autocorrelation function, cross-correlation function) are extracted. For example, the natural frequency drift in structural vibration signals is an important indicator of structural deterioration. For image data (e.g., damage photographs), convolutional neural networks are used to extract visual features, such as crack length, width, area, and color changes. For text data (e.g., maintenance records), natural language processing techniques are used to extract semantic features such as keywords, damage types, and repair measures. Simultaneously, interactive features are constructed, such as the cumulative effect of temperature and humidity (accumulating or weighting daily temperature and humidity data to form cumulative exposure) and strain response features under load (e.g., changes in the ratio of strain to load). Furthermore, domain knowledge is used to construct features based on physical models or empirical formulas, such as concrete carbonation depth and steel corrosion rate. The dimensions and number of these features will increase significantly, requiring management.

[0036] Feature Standardization and Normalization: To avoid the adverse effects of differences in the dimensions and numerical ranges of different features on deep learning model training, this sub-step standardizes or normalizes all features. Common methods include zero-mean normalization (Z-score normalization), which subtracts the mean from the feature value and divides by the standard deviation to make it follow a standard normal distribution with a mean of zero and a variance of one; or min-max normalization, which linearly maps the feature values ​​to a preset interval (e.g., zero to one). The choice of method depends on the distribution characteristics of the features and the specific requirements of the model. This helps to accelerate model convergence and improve the model's generalization ability. The processed feature data is stored as a feature matrix in a uniform format, with each sample corresponding to one row and each feature corresponding to one column.

[0037] By adopting the above design, the acquired multi-source heterogeneous building-related data is transformed into a unified, high-quality, and informative feature representation suitable for deep learning model training. The original data typically suffers from inconsistent formats, missing data, anomalies, and noise, and there are significant differences in spatiotemporal scales between different modalities. Therefore, this step is a crucial preliminary step for achieving accurate aging prediction.

[0038] The specific steps involved in constructing a knowledge graph of building aging are as follows: Define the entity types of the building aging knowledge graph. Entity types include buildings, components, materials, environmental factors, aging phenomena, maintenance activities, and damage types. Define the association types of the knowledge graph of building aging, including: including, constituted by, affected by, caused by, occurring in, repaired, and manifested as; Entities and relationships are extracted from architectural design data, building material attribute data, and building history maintenance data to construct the initial framework of a knowledge graph of building aging. By using natural language processing technology, new entities and relationships can be identified and extracted from unstructured building history maintenance data to expand the knowledge graph of building aging.

[0039] The specific steps involved in constructing a multimodal deep learning model are as follows: S401. Construct a multi-scale temporal convolutional network module to process the temporal characteristics of building structural health monitoring data and building environment climate data, and capture long-term dependencies and local temporal patterns. S402. Construct a graph neural network module to process the knowledge graph of building aging, learn the structured correlation features between components, materials and environmental factors, and capture the propagation path of damage in the structure. S403. Construct a multilayer perceptron network module to process static features such as building material properties, building history maintenance features, and building design features. S404. Construct a feature fusion module to deeply fuse the outputs of the multi-scale temporal convolutional network module, the graph neural network module, and the multilayer perceptron network module. S405. Construct an aging prediction output module to predict building aging indicators based on the output of the feature fusion module.

[0040] Furthermore, in this embodiment, specialized encoders are designed and implemented for different types and modalities of feature data to map the original features to a unified low-dimensional, semantically rich feature space. For example, for temporal feature modalities (such as structural strain time series, environmental temperature time series, and equipment runtime time series), a Temporal Convolutional Network (TCN) or a Long Short-Term Memory (LSTM) network is used as the encoder. The TCN effectively captures local dependencies and multi-scale temporal patterns in long sequences through causal convolutions and dilated convolutional layers, while avoiding the gradient vanishing / exploding problem of recurrent neural networks and supporting parallel computation. The output of a TCN encoder can be represented as:

[0041] in, The input time-series feature data is represented by W and b, which represent the weights and biases of the convolution kernel, respectively. Conv represents a series of causal dilation convolution operations.

[0042] For static structural feature modes (such as material strength, design load, and component size), a multi-layer perceptron (MLP) is used as an encoder to map these static attributes into feature vectors of fixed length.

[0043] For image feature modalities (such as damaged photos and BIM model renderings), a pre-trained convolutional neural network (CNN), such as a residual network (ResNet) or a vision transformer (VisionTransformer), is used as a feature extractor to extract high-level semantic features from the image.

[0044] Each modal encoder is carefully designed to maximize the representational power of the modal information. The encoder output is a series of feature vectors with the same dimension, which represent abstract representations of the modal data.

[0045] Modal Feature Fusion Mechanism Design: This mechanism aims to effectively fuse the feature vectors output by different modal encoders to capture the synergistic effects and complementary information between modalities. A single modality often cannot fully reflect the aging state of a building; therefore, feature fusion is crucial for improving prediction accuracy. This invention employs a fusion strategy based on an attention mechanism. Specifically, a cross-modal attention module is introduced, which dynamically assigns weights to features of different modalities based on the importance of the current prediction task. This attention mechanism allows the model to prioritize modal information that contributes most to the current aging prediction during the fusion process. For example, when predicting structural damage, the model may assign higher weights to structural strain features; while when predicting exterior wall material peeling, environmental temperature and humidity accumulation features may be given higher weights. The calculation process of the attention mechanism can be expressed as follows:

[0046] in, The encoded feature representing the m-th modality; It represents the set of all modality-coded features; Query, Key, and Value are all linear transformation functions; It is the attention weight of the m-th modality; It is the final fused feature vector; through this mechanism, the fused feature vector can comprehensively consider all modal information and highlight the contribution of key modalities.

[0047] Aging Prediction Decoder Construction: The fused feature vector is input into a decoder network to generate the final building aging prediction result. The decoder is typically a regression network consisting of one or more fully connected layers, with a structure including multiple linear layers and non-linear activation functions. Its function is to transform abstract fused features into specific, quantifiable aging metrics. These aging metrics can be multi-dimensional, for example: The probability of structural damage at a specific point in the future (e.g., the next year, three years, or five years); The performance degradation rate of the main materials (e.g., concrete, steel, coatings); The remaining fatigue life of specific components (such as beams, columns, and floor slabs); Forecasted future maintenance costs; Overall building health index.

[0048] The decoder's output layer uses different activation functions depending on the type of prediction task. For example, for probabilistic prediction tasks, the output layer uses the sigmoid activation function to restrict the output value to between zero and one; for continuous numerical prediction tasks, the output layer can directly output the predicted value. The decoder design should ensure that it can capture the complex nonlinear mapping between fused features and aging metrics.

[0049] By using multimodal deep learning technology, preprocessed heterogeneous features are deeply integrated to uncover the complex nonlinear laws in the building aging process and ultimately output accurate aging prediction results.

[0050] Among them, aging indicators include the probability of damage to a certain structural component in the next year, the percentage of peeling area of ​​interior and exterior wall paint in the next three years, and the predicted maintenance cost of a specific piece of equipment in the next five years.

[0051] Step S5 specifically includes the following steps: S501. Use multi-dimensional features and building aging knowledge graph as inputs to the model; S502. Use the actual observed building aging indicators as the training target for the model; S503. The backpropagation algorithm is used to optimize the model parameters; S504. Use historical datasets for model training and validation, and adjust the model's hyperparameters using a cross-validation strategy.

[0052] Specifically, in this embodiment, through iterative optimization, the constructed multimodal deep learning model can accurately learn the predictive patterns of building aging from the input data. The first step is loss function design: selecting an appropriate loss function based on the nature of the prediction task. For example, for regression prediction tasks (predicting continuous aging indicators), mean squared error (MSE) or mean absolute error (MAE) is often used; for classification prediction tasks (predicting aging risk levels), cross-entropy loss is used. The loss function defines the difference between the model's predicted values ​​and the true values. The loss function is typically expressed as:

[0053] in, It is the sample size; This is the true aging value; These are model predictions; These are all the trainable parameters of the model.

[0054] Secondly, optimizer selection is crucial. Variations of gradient descent, such as Adaptive Moments Estimation (Adam), Stochastic Gradient Descent (SGD), or Adaptive Gradient (Adagrad), are used to minimize the loss function. The optimizer calculates the gradient of the loss function with respect to the model parameters and updates the parameters along the negative direction of the gradient, gradually improving model performance. Hyperparameter tuning is a critical step in the training process, including learning rate, batch size, number of network layers, number of hidden units, and regularization coefficients. The selection of these hyperparameters significantly impacts model performance and is typically explored systematically using methods such as grid search, random search, or Bayesian optimization. To evaluate the model's generalization ability and prevent overfitting, a K-Fold Cross-Validation strategy is employed. The dataset is divided into training, validation, and test sets. The model is trained on the training set, hyperparameter tuning and model selection are performed on the validation set, and finally, the model's final performance is evaluated on a separate test set. During training, regularization techniques (such as L1 / L2 regularization and Dropout) are often introduced to further improve the model's generalization ability.

[0055] Further, in step S6, the building data to be predicted is input, and the building aging prediction result is generated through the aging prediction model. This step is the application stage of the method of the present invention, using the trained model to predict the aging of new buildings, specifically including the following steps: Data collection and preprocessing for buildings to be predicted: For specific buildings requiring aging prediction, collect multi-source heterogeneous related data following the same process described in steps S1-S3, and perform rigorous data cleaning, integration, feature extraction, and standardization. Ensure that the data to be predicted has the same format, dimensions, and feature representation as the data used during model training to guarantee the effectiveness of model inference.

[0056] Prediction Model Loading and Inference: The multimodal deep learning aging prediction model trained and optimized in step S5 is loaded into the inference engine. The inference engine can be a dedicated hardware accelerator (e.g., a graphics processing unit) or a high-performance computing platform. Then, the preprocessed building data to be predicted is input into the loaded model. The model follows its predetermined forward propagation path, sequentially passing through the modal encoder, modal feature fusion mechanism, and aging prediction decoder, to calculate and generate the corresponding prediction output. The entire inference process is typically completed in milliseconds to seconds, meeting the prediction efficiency requirements of practical applications.

[0057] Generate building aging prediction indicators: The output of the model inference is the building aging prediction result, which corresponds in form to the aging indicators defined by the decoder output in step S405. For example, the output is a vector containing the probability of damage to a structural component in the next year, the percentage of peeling area of ​​interior and exterior wall paint in the next three years, and the predicted maintenance cost of a specific piece of equipment in the next five years. These prediction results are quantifiable and interpretable values ​​that can directly support subsequent assessments and decisions.

[0058] After step S6 is completed, the following steps are also included: Accuracy Evaluation of Prediction Results: The model's predicted building aging results are evaluated from multiple dimensions to verify their consistency with actual aging conditions. Evaluation metrics include, but are not limited to: For continuous numerical prediction tasks, calculating the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). For example, calculating the mean squared error between the predicted material degradation rate and the actual detected degradation rate. For classification prediction tasks (e.g., aging risk level prediction), calculating accuracy, precision, recall, and F1 score. The model's predictive performance on different time scales and aging indicators is verified through comparative analysis with historical actual aging data.

[0059] Aging Risk Level Classification: Based on the generated quantitative building aging prediction indicators, a series of scientifically reasonable thresholds are set to classify the aging state of buildings into different risk levels. For example, multiple levels can be set such as "low risk," "medium risk," "high risk," or "emergency repair." For instance, when the predicted probability of structural damage exceeds 20%, it can be classified as "high risk"; when the predicted maintenance cost growth rate exceeds 30% of the historical average, it can be classified as "medium risk." The risk level classification should combine industry standards, safety regulations, and economic considerations to provide decision-makers with clear action priorities.

[0060] Visualization: The visualization presents the predicted building aging results and their corresponding risk levels in an intuitive and interactive way. The visualization interface may include: time-series graphs showing the changing trends of future aging indicators, such as a curve showing the predicted concrete strength decay over time; heat maps or color-coded 3D building models visually displaying the aging risk distribution of different areas or components, for example, high-risk areas are highlighted in red; and data dashboards displaying summary statistics, historical comparisons, and real-time status of key aging indicators. Users can filter and view different prediction time ranges and different component types through the interactive interface to meet personalized decision-making needs. The visualization module aims to lower the information comprehension threshold and improve decision-making efficiency.

[0061] Maintenance Strategy Recommendations: Based on accurate building aging predictions and categorized aging risk levels, the system automatically generates a series of targeted maintenance strategy recommendations. These recommendations include: preventative maintenance plans, such as scheduling exterior wall renovations before predicted paint peeling reaches a certain level; priority repair area recommendations, such as clearly identifying components or parts requiring priority inspection and repair based on high-risk areas in the 3D model; and resource allocation optimization recommendations, such as optimizing the allocation of maintenance resources based on the urgency of different maintenance tasks, required material and labor costs. These recommendations aim to guide property managers and engineering technicians in developing more scientific and efficient maintenance and renovation plans, thereby extending the building's lifespan, ensuring resident safety, and optimizing operating costs.

[0062] A building aging prediction system is provided to implement the building aging prediction method described above. The system includes: The data acquisition module is used to acquire multi-source heterogeneous building-related data, including building structural health monitoring data, building environmental climate data, building material property data, building historical maintenance data, and building design data. (This module is responsible for acquiring multi-source heterogeneous building-related data from various sensors, databases, and interfaces. It internally includes multiple data interface units, such as sensor interface units for connecting structural health monitoring sensors, network interface units for acquiring environmental meteorological data, and database interface units for importing historical data from building management and maintenance management systems. This module also has data caching and preliminary verification functions to ensure the real-time performance and preliminary quality of the raw data. For example, the sensor interface units can support multiple industrial communication protocols, such as Modbus and CAN bus protocols, to ensure compatibility with different types of sensors.) The data processing module preprocesses multi-source, heterogeneous building-related data to generate a standardized dataset. It then performs feature engineering on the dataset to extract multi-dimensional features related to building aging. (This module deeply processes the raw data acquired by the data acquisition module to generate feature representations suitable for model training and inference. It is further divided into a data cleaning unit, a feature engineering unit, and a data storage unit. The data cleaning unit performs operations such as missing value imputation, outlier detection and removal, and noise filtering to improve data purity. The feature engineering unit is responsible for data integration and alignment, high-dimensional feature extraction and construction, and feature standardization and normalization, transforming the raw data into a structured feature matrix usable for deep learning models. The data storage unit is a distributed database or big data warehouse used to efficiently store massive amounts of raw data and processed feature data, supporting fast querying and retrieval. For example, a Hadoop distributed file system combined with a NoSQL database can be used for large-scale data storage.) The building aging knowledge graph construction module is used to construct a building aging knowledge graph to represent the relationship between building components, building materials, building environment and building aging phenomena. The deep learning model building module is used to build multimodal deep learning models and integrate multi-dimensional features and building aging knowledge graphs to learn the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process. The model training module is used to train multimodal deep learning models (this module is the core intelligent part of the system, responsible for building, training, optimizing, and managing multimodal deep learning aging prediction models. This module includes a model building unit, a training optimization unit, and a model library. The model building unit provides a flexible framework to support the construction of various deep learning network structures, including temporal convolutional networks, long short-term memory networks, fully connected networks, convolutional neural networks, and attention mechanisms. The training optimization unit utilizes large-scale parallel computing resources (such as graphics processing unit clusters) to execute the model training process, achieving loss function minimization and model parameter optimization. The model library is used to store trained models of different versions and prediction tasks, facilitating model version control, loading, and updates. This module also provides model performance monitoring and evaluation functions to ensure continuous model optimization). The aging prediction module is responsible for loading pre-trained models and performing real-time or batch aging prediction inference on new building data to be predicted. Its core is the prediction inference engine, which efficiently performs forward propagation calculations of the model and generates building aging prediction indicators. This module is designed with a high-concurrency, low-latency architecture to meet the requirements of prediction response speed in practical applications. For example, using inference optimization engines such as TensorRT for model deployment can significantly improve inference speed. The Results Analysis and Display module is responsible for in-depth analysis of the aging prediction results generated by the model and presenting them to users in an intuitive and user-friendly manner. This module includes an evaluation unit, a visualization interface, and a decision support unit. The evaluation unit quantitatively assesses the accuracy of the prediction results based on preset evaluation indicators. The visualization interface provides rich charts and interactive interfaces, such as 3D building models, time series graphs, and heat maps, intuitively displaying aging trends, risk distribution, and key indicators. The decision support unit automatically generates maintenance strategy suggestions and resource allocation optimization plans based on the prediction results and risk levels, providing users with a scientific basis for decision-making. This module provides user access via web services or desktop applications.

[0063] These modules communicate through standardized application programming interfaces (APIs) to ensure smooth data flow and collaborative operation among functional modules. The entire system runs on a server cluster with powerful computing and storage capabilities to support the demands of large data volumes and highly complex models. The system possesses excellent scalability and fault tolerance, capable of adapting to ever-increasing data volumes and evolving model algorithms.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] The present invention has been described in detail above with reference to the accompanying drawings and embodiments. Those skilled in the art can make various modifications to the present invention based on the above description. Therefore, certain details in the embodiments should not be construed as limiting the present invention, and the scope of protection of the present invention shall be defined by the appended claims.

Claims

1. A method for predicting building aging based on big data modeling, characterized in that, Includes the following steps: S1. Obtain multi-source heterogeneous building-related data, including building structural health monitoring data, building environmental climate data, building material property data, building historical maintenance data, and building design data; S2. Preprocess the multi-source heterogeneous building-related data to generate a standardized dataset, and perform feature engineering on the dataset to extract multi-dimensional features related to building aging. S3. Construct a knowledge graph of building aging to represent the relationship between building components, building materials, building environment and building aging phenomena; S4. Construct a multimodal deep learning model and integrate multi-dimensional features and a building aging knowledge graph to learn the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process. S5. Train the multimodal deep learning model; S6. After processing the building-related data to be predicted according to steps S1-S3, input it into the trained multimodal deep learning model for prediction, and output the prediction result.

2. The building aging prediction method based on big data modeling as described in claim 1, characterized in that, The building structure health monitoring data includes the building structure's strain, acceleration, displacement, tilt angle, crack width, temperature, and humidity; The building environment climate data includes ambient temperature, ambient humidity, wind speed, wind direction, rainfall, and solar radiation intensity; The building material property data includes the compressive strength, tensile strength, modulus of elasticity, durability, corrosion rate, and fatigue performance of concrete, steel, brick, stone, and wood. The building's historical maintenance data includes all inspection reports, maintenance dates, maintenance content, maintenance materials, maintenance costs, and fault types. The architectural design data includes design drawings, construction specifications, completion acceptance reports, load conditions, and geographical location.

3. The building aging prediction method based on big data modeling as described in claim 2, characterized in that, The preprocessing of multi-source heterogeneous building-related data specifically includes the following steps: S2011. Perform timestamp synchronization, data alignment, missing value filling, and outlier filtering on building structural health monitoring data to form a continuous and complete time-series data stream; S2012. Perform timestamp synchronization, data alignment, data cleaning, and format conversion on building environment and climate data to ensure synchronization with building structural health monitoring data. S2013. Perform structured processing on building material attribute data and convert unstructured text information into quantifiable feature parameters; S2014. Perform event serialization and feature extraction on building history maintenance data, and identify key maintenance event types, occurrence times, and impacts on building performance from text descriptions; S2015. Encode architectural design data and convert design parameters and construction specifications into numerical features; S2016. Normalize all the processed data from steps S2011-S2015 to eliminate the influence between data of different dimensions and orders of magnitude.

4. The building aging prediction method based on big data modeling as described in claim 3, characterized in that, The feature engineering process for the dataset specifically includes the following steps: S2021. Extract time-domain features, frequency-domain features, and time-frequency-domain features from time-series data of building structural health monitoring; S2022. Extract long-term trend characteristics, seasonal characteristics, and extreme event characteristics from built environment and climate data; S2023. Extract material degradation rate parameters and initial performance parameters from building material property data; S2024. Extract maintenance frequency, average repair time, fault recurrence rate, and maintenance cost percentage from building historical maintenance data; S2025. Extract building structure type, number of floors, total area, design life and dimensions of key building components from architectural design data; S2026. Dimensionality reduction of the features extracted in steps S2021-S2025 is performed through principal component analysis or independent component analysis.

5. The building aging prediction method based on big data modeling as described in claim 4, characterized in that, The time-domain features include mean, variance, standard deviation, kurtosis, skewness, waveform factor, impulse factor, margin factor, and kurtosis factor. The frequency-domain features include the main frequency components after fast Fourier transform, peak power spectral density, and energy center frequency. The time-frequency-domain features include wavelet energy and wavelet entropy after wavelet transform, as well as instantaneous frequency and instantaneous amplitude after Hilbert-Huang transform.

6. The building aging prediction method based on big data modeling as described in claim 2, characterized in that, The specific steps involved in constructing the knowledge graph of building aging are as follows: Define the entity types of the building aging knowledge graph, which include buildings, components, materials, environmental factors, aging phenomena, maintenance activities, and damage types; Define the association types of the knowledge graph of building aging, including inclusion, constituent factors, causes of impact, phenomena that lead to the occurrence, areas of occurrence, repair parts, and manifestations; Entities and relationships are extracted from architectural design data, building material attribute data, and building history maintenance data to construct the initial framework of a knowledge graph of building aging. By using natural language processing technology, new entities and relationships can be identified and extracted from unstructured building history maintenance data to expand the knowledge graph of building aging.

7. The building aging prediction method based on big data modeling as described in claim 1, characterized in that, The construction of the multimodal deep learning model specifically includes the following steps: S401. Construct a multi-scale temporal convolutional network module to process the temporal characteristics of building structural health monitoring data and building environment climate data, and capture long-term dependencies and local temporal patterns. S402. Construct a graph neural network module to process the knowledge graph of building aging, learn the structured correlation features between components, materials and environmental factors, and capture the propagation path of damage in the structure. S403. Construct a multilayer perceptron network module to process static features such as building material properties, building history maintenance features, and building design features. S404. Construct a feature fusion module to deeply fuse the outputs of the multi-scale temporal convolutional network module, the graph neural network module, and the multilayer perceptron network module. S405. Construct an aging prediction output module to predict building aging indicators based on the output of the feature fusion module.

8. The building aging prediction method based on big data modeling as described in claim 7, characterized in that, The aging indicators include the probability of damage to a structural component in the next year, the percentage of peeling area of ​​interior and exterior wall paint in the next three years, and the predicted maintenance cost of a specific piece of equipment in the next five years.

9. The building aging prediction method based on big data modeling as described in claim 7, characterized in that, Step S5 specifically includes the following steps: S501. Use the multi-dimensional features and the building aging knowledge graph as input to the model; S502. Use the actual observed building aging indicators as the training target for the model; S503. The backpropagation algorithm is used to optimize the model parameters; S504. Use historical datasets for model training and validation, and adjust the model's hyperparameters using a cross-validation strategy.

10. A building aging prediction system for implementing the building aging prediction method as described in claim 1, characterized in that, The system includes: The data acquisition module is used to acquire multi-source heterogeneous building-related data, including building structural health monitoring data, building environment climate data, building material property data, building historical maintenance data, and building design data. The data processing module is used to preprocess multi-source heterogeneous building-related data to generate a standardized dataset, and to perform feature engineering on the dataset to extract multi-dimensional features related to building aging. The building aging knowledge graph construction module is used to construct a building aging knowledge graph to represent the relationship between building components, building materials, building environment and building aging phenomena. The deep learning model building module is used to build multimodal deep learning models and integrate multi-dimensional features and building aging knowledge graphs to learn the nonlinear dynamic evolution law and multi-factor coupling effect in the building aging process. The model training module is used to train the multimodal deep learning model.

Images