CN117611015A

CN117611015A - Real-time monitoring system for quality of building engineering

Info

Publication number: CN117611015A
Application number: CN202410087335.5A
Authority: CN
Inventors: 张可鑫; 刘川; 宋明奇; 赵学翰; 郝春松; 李斯; 康俊菲; 荆扬; 梁世超; 殷向开; 刘立朋
Original assignee: Hengshui Yetong Construction Engineering Co ltd
Current assignee: Hengshui Yetong Construction Engineering Co ltd
Priority date: 2024-01-22
Filing date: 2024-01-22
Publication date: 2024-02-27
Anticipated expiration: 2044-01-22
Also published as: CN117611015B

Abstract

The invention relates to the technical field of nondestructive detection, in particular to a real-time monitoring system for building engineering quality. According to the invention, through the combination of a K-means clustering algorithm and a genetic algorithm, load mode identification and iterative optimization are realized in load analysis, an automatic encoder and a convolutional neural network in data reconstruction analyze building structures, a deep reinforcement learning model is utilized in enhanced prediction, dynamic response prediction is optimized, digital twin and nonlinear analysis are combined with a dynamic Bayesian network and a mixed effect model, health state analysis is provided, main component analysis and independent component analysis in information compression simplify data processing, and a strategy-optimized particle swarm optimization algorithm and a state estimation dynamic Bayesian network improve decision efficiency and risk management capability.

Description

Real-time monitoring system for quality of building engineering

Technical Field

The invention relates to the technical field of nondestructive detection, in particular to a real-time monitoring system for building engineering quality.

Background

Nondestructive testing is an important technical field for evaluating properties of materials, components or whole structures without compromising their future usability. This technique is particularly critical in the field of construction engineering, as it allows engineers and inspection personnel to monitor and evaluate the quality and performance of building materials without affecting the structural integrity of the building. Common nondestructive testing methods include ultrasonic testing, X-ray or gamma ray testing, magnetic particle testing, permeation testing, and infrared and thermal imaging techniques. By these methods, potential problems such as cracks, corrosion, structural defects, etc., can be detected, thereby ensuring safety and durability of the building.

The building engineering quality real-time monitoring system is an advanced technical system applied to the field of building engineering, and aims to monitor the health condition and performance of a building structure in real time. Such systems typically include sensors, a data acquisition unit, and analysis software. The sensors are deployed at key locations of the building and collect data in real time, such as parameters of stress, temperature, vibration, etc., which reflect the current state of the building structure. By analyzing this data in real time, the system can discover potential structural problems in time, such as crack propagation, corrosion, or other factors that cause structural weakening. Such monitoring enables timely maintenance and reinforcement, greatly increasing the safety and reliability of the building, while reducing long-term maintenance costs.

Conventional building engineering quality monitoring systems often face a number of challenges in processing large-scale and complex data. Conventional systems often lack efficient data analysis and processing capabilities, resulting in less accurate load analysis and feature extraction, affecting the effectiveness of resource allocation and prediction strategies. In addition, traditional systems perform poorly in terms of dynamic environmental adaptability, and it is difficult to update and optimize response strategies in real time, which results in an insufficient assessment of building health status, increasing the difficulty of risk management. In terms of data storage and transmission, due to the lack of efficient data compression technology, conventional systems often face the problems of large data storage space and low transmission efficiency, and further limit the application range and efficiency of the system.

Disclosure of Invention

The invention aims to solve the defects in the prior art, and provides a real-time monitoring system for the quality of a building engineering.

In order to achieve the above purpose, the present invention adopts the following technical scheme: the real-time monitoring system for the construction engineering quality comprises a load analysis module, a data reconstruction module, an enhanced prediction module, a digital twin module, a nonlinear analysis module, an information compression module, a strategy optimization module and a state estimation module;

The load analysis module adopts a K-means clustering algorithm to perform pattern recognition on the building load based on the data monitored by the sensor network in real time, and uses a genetic algorithm to perform iterative optimization on the recognized pattern, and meanwhile, load allocation is integrated to generate a load optimization scheme;

the data reconstruction module adopts an automatic encoder to perform data dimension reduction processing based on a load optimization scheme, and uses a convolutional neural network to perform depth analysis and feature extraction on building structure images and data so as to generate a structural feature map;

the enhanced prediction module adopts a deep reinforcement learning model to learn environment and structural features based on the structural feature map, and circularly optimizes the dynamic response prediction of the building structure through strategy optimization iteration and self-evaluation adjustment to generate a prediction strategy model;

the digital twin module carries out depth modeling and state prediction on time series data of a building by adopting a dynamic Bayesian network based on a prediction strategy model, updates the model to match new data input, and generates a digital twin entity;

the nonlinear analysis module is based on a digital twin entity, adopts a mixed effect model to carry out nonlinear analysis on the response of building individual and group structures, extracts individual characteristics and group rules, and generates a response analysis result;

The information compression module adopts principal component analysis and independent component analysis to perform data dimension reduction based on response analysis results, and performs pattern recognition and compression storage to generate a compressed information data set;

the strategy optimization module intelligently adjusts and optimizes the load distribution and management strategy by adopting a particle swarm optimization algorithm based on the compressed information data set to generate an optimized regulation strategy;

the state estimation module adopts a dynamic Bayesian network to carry out continuous estimation and risk prediction of the building health state based on the optimized regulation strategy, and generates health state estimation.

As a further aspect of the present invention, the load optimization scheme includes load distribution of floors, suggested load adjustment amplitude, and optimized energy consumption prediction, the structural feature map includes feature vectors of key structural elements, potential risk point distribution, and structural health indexes, the prediction policy model includes future state prediction, suggested maintenance measures, and emergency response policies, the digital twin entity includes a virtual model, a state evolution path, and a predicted maintenance time point, the response analysis result includes individual response characteristics, population response trends, and key impact factors, the compressed information data set includes reduced-dimension key data, information compression rate, and reconstruction error assessment, and the optimization regulation policy includes an optimized load distribution scheme, energy efficiency improvement measures, and expected running cost savings, and the health state assessment includes current health indexes, future risk prediction, and suggested inspection intervals.

As a further scheme of the invention, the load analysis module comprises a load monitoring sub-module, a first pattern recognition sub-module and a genetic optimization sub-module;

the load monitoring submodule is based on real-time monitoring requirements, adopts a distributed sensing network, comprises a temperature sensor, a vibration sensor and a stress sensor, collects load data of a plurality of parts of a building, comprises temperature change, vibration frequency and stress change, performs data summarization and preliminary screening, and generates a load data set;

the first pattern recognition submodule adopts a K-means clustering algorithm based on a load data set, and divides data into K categories by calculating and comparing distances among data points, recognizes differentiated load patterns including normal load and overload, classifies and marks, and generates load pattern classification;

the genetic optimization submodule carries out multi-generation iterative optimization on the load distribution scheme by adopting a genetic algorithm through simulation selection, crossover and mutation operation based on load mode classification, and captures an optimal load distribution strategy based on energy consumption and efficiency to generate a load optimization scheme.

As a further scheme of the invention, the data reconstruction module comprises a self-coding sub-module, a convolution analysis sub-module and a feature extraction sub-module;

The self-coding submodule adopts an automatic encoder algorithm based on a load optimization scheme, input load data is compressed into low-dimensional characteristic representation through an encoder by setting a neural network structure of the encoder and the decoder, then the data is reconstructed through a decoder, and network parameters are optimized through back propagation during the process, so that compression and characteristic extraction of the data are carried out, and compression load characteristics are generated;

the convolution analysis submodule adopts a convolution neural network to set a multi-layer convolution and pooling layer structure based on compression load characteristics, carries out layer-by-layer convolution and subsampling on input compression characteristics, extracts local characteristics and gradually combines the local characteristics into global characteristics, gradually refines key structural characteristics of a building through nonlinear activation and normalization of the recognition capability of an enhancement model, and generates building characteristic classification;

the feature extraction submodule is used for comprehensively applying edge detection and texture analysis technologies based on building feature classification, identifying and marking structural details and potential risk areas of a building, and extracting key feature points based on the health state of the building by comparing and analyzing similarity and difference of multi-area features to generate a structural feature map.

As a further scheme of the invention, the enhanced prediction module comprises a feature learning sub-module, a strategy iteration sub-module and a self-evaluation sub-module;

The characteristic learning submodule carries out comprehensive characteristic learning by adopting a convolutional neural network and a cyclic neural network based on a structural characteristic map, wherein the convolutional neural network extracts spatial characteristics from building data, the spatial characteristics comprise structural shapes and spatial distribution, and the cyclic neural network processes time sequence data to capture the change trend of structural states along with time so as to generate comprehensive environmental characteristics;

the strategy iteration submodule carries out strategy iteration by adopting a strategy gradient or Q learning method in reinforcement learning based on comprehensive environmental characteristics, simulates the influence of differentiated actions on a prediction target by establishing a reward mechanism, and gradually adjusts and optimizes action strategies by multiple rounds of iterative optimization to generate an improved action strategy;

the self-evaluation sub-module adopts a model self-evaluation method based on an improved action strategy, comprises the steps of continuously monitoring the prediction performance and generalization capability of a model, evaluating and adjusting the model in real time, and adjusting the weight and parameters of the neural network by analyzing the deviation between a prediction result and an actual situation to generate a prediction strategy model.

As a further scheme of the invention, the digital twin module comprises a time sequence modeling sub-module, a state prediction sub-module and a model updating sub-module;

The time sequence modeling submodule is based on a prediction strategy model of the digital twin module, adopts a dynamic Bayesian network algorithm to carry out structural modeling on building time sequence data, codes time sequence dependency relations among data points by constructing a probability graph model of a time sequence, carries out probability inference and prediction of the model, and generates a building data time dependency model;

the state prediction submodule is used for extracting and analyzing features of future states of the building by adopting a random forest algorithm based on a building data time dependent model, constructing a plurality of decision trees, carrying out combined analysis on the differentiated features, predicting the future states and generating predicted building state analysis;

the model updating submodule is used for updating the model in real time by using an online learning method based on the prediction building state analysis, gradually adjusting model parameters when receiving new data, maintaining the adaptability and response speed of the model to new conditions, and generating a digital twin entity.

As a further scheme of the invention, the nonlinear analysis module comprises a mixed effect modeling sub-module, a population analysis sub-module and a feature identification sub-module;

The mixed effect modeling submodule is based on a digital twin entity, adopts a generalized linear mixed model to carry out statistical modeling on response data of an individual structure of a building, establishes statistical association between an individual and a group by combining a fixed effect and a random effect, allows the model to capture unique changes at an individual level, and simultaneously refers to a trend at a group level to generate individual response characteristic analysis;

the group analysis submodule analyzes building group data from different layers by adopting a multi-level model based on individual response characteristic analysis, processes a complex data set by constructing a data structure comprising a plurality of layers, respectively analyzes the data on multiple layers to reveal interrelationships and differences among individuals and inside groups, and generates group response rule extraction;

the feature recognition submodule extracts based on a group response rule, analyzes the extracted features by adopting principal component analysis and cluster analysis, simplifies a data structure by dimension reduction processing, highlights important features, classifies similar data points into groups, distinguishes and recognizes key features and modes of individual and group responses, and generates a response analysis result.

As a further scheme of the invention, the information compression module comprises a data dimension reduction sub-module, a second pattern recognition sub-module and a compression storage sub-module;

the data dimension reduction submodule adopts principal component analysis based on response analysis results, identifies and extracts key principal components as new data representation through statistical analysis of correlation among variables in original data, and further adopts independent component analysis, and separates original signals from mixed signals through maximization of statistical independence of non-Gaussian source signals to generate a dimension reduction feature set;

the second pattern recognition submodule adopts a support vector machine based on the dimension reduction feature set, classifies the dimension reduced data by constructing a decision boundary, and simultaneously uses a neural network to recognize key patterns and trends in the data set by training internal patterns of data learning data so as to generate a pattern recognition data set;

the compression storage submodule adopts a lossless compression technology based on a pattern recognition data set, encodes the data through Huffman coding or LZW algorithm, reduces the storage of repeated data, adopts a lossy compression technology, reduces the data quantity through reducing the data precision, retains the key information and the characteristics of the data, and generates a compression information data set.

As a further scheme of the invention, the policy optimization module comprises a load adjustment sub-module, a policy adjustment sub-module and a real-time update sub-module;

the load adjustment sub-module adopts a particle swarm optimization algorithm based on the compressed information data set, searches in a solution space by setting potential solutions represented by particles, adjusts a flight path according to personal experience and group experience by each particle, captures an optimal solution of load distribution, and generates an intelligent load distribution scheme;

the strategy adjustment submodule gradually reduces the temperature from the initial temperature based on an intelligent load distribution scheme by adopting a simulated annealing algorithm, and each temperature stage carries out small-range random disturbance on the current solution, determines whether to accept a new solution or not through probability, avoids local optimization and generates an optimization management strategy;

the real-time updating sub-module is based on an optimization management strategy, uses a genetic algorithm to simulate selection, crossover and mutation, circularly generates a new strategy combination, evaluates the fitness, selects an optimal strategy for iteration, carries out continuous optimization of the strategy, and generates an optimal regulation strategy.

As a further scheme of the invention, the state estimation module comprises a state monitoring sub-module, a risk prediction sub-module and a health assessment sub-module;

The state monitoring submodule executes real-time data collection based on an optimized regulation strategy, continuously monitors vibration, temperature and pressure data through a sensor network deployed at key structural points of a building, tracks physical state changes of the building, determines data change trend of key monitoring points and generates a real-time state data set;

the risk prediction submodule performs dynamic Bayesian network analysis based on the real-time state data set, updates probability distribution according to the real-time data, identifies potential structural risk modes by calculating time correlation among data points, predicts short-term and long-term risk trends, and generates risk prediction analysis;

the health evaluation sub-module is used for carrying out building health evaluation based on risk prediction analysis, utilizing time sequence analysis, mining long-term data trend and change mode, analyzing by combining an expert system with industry standard and historical data, evaluating the overall health condition of the building, and generating health state evaluation.

Compared with the prior art, the invention has the advantages and positive effects that:

according to the invention, through the combination of the K-means clustering algorithm and the genetic algorithm, more accurate building load pattern recognition and iterative optimization are realized in the load analysis module, and the resource allocation efficiency is improved. The automatic encoder and the convolutional neural network in the data reconstruction module are used for deeply analyzing the building structure, and higher accuracy is provided for feature extraction. The enhanced prediction module optimizes the dynamic response prediction of the building structure by utilizing the deep reinforcement learning model, and enhances the accuracy and adaptability of the prediction. The digital twin module and the nonlinear analysis module combine the dynamic Bayesian network and the mixed effect model to provide a more comprehensive and deep analysis of the health status of the building. The principal component analysis and the independent component analysis in the information compression module effectively simplify data processing and improve data storage and transmission efficiency. The particle swarm optimization algorithm of the strategy optimization module and the dynamic Bayesian network of the state estimation module jointly improve the decision efficiency and the risk management capability of the whole system.

Drawings

FIG. 1 is a system flow diagram of the present invention;

FIG. 2 is a schematic diagram of a system framework of the present invention;

FIG. 3 is a flow chart of a load analysis module according to the present invention;

FIG. 4 is a flow chart of a data reconstruction module according to the present invention;

FIG. 5 is a flow chart of an enhanced prediction module of the present invention;

FIG. 6 is a flow chart of a digital twinning module of the present invention;

FIG. 7 is a flow chart of a nonlinear analysis module according to the present invention;

FIG. 8 is a flow chart of an information compression module of the present invention;

FIG. 9 is a flow chart of a policy optimization module of the present invention;

fig. 10 is a flow chart of a state estimation module according to the present invention.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

In the description of the present invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like indicate orientations or positional relationships based on the orientation or positional relationships shown in the drawings, merely to facilitate describing the present invention and simplify the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and therefore should not be construed as limiting the present invention. Furthermore, in the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

Example 1

Referring to fig. 1 to 2, a real-time monitoring system for building engineering quality includes a load analysis module, a data reconstruction module, an enhanced prediction module, a digital twin module, a nonlinear analysis module, an information compression module, a strategy optimization module, and a state estimation module;

the load analysis module adopts a K-means clustering algorithm to perform pattern recognition on the building load based on the data monitored by the sensor network in real time, and adopts a genetic algorithm to perform iterative optimization on the recognized pattern, and meanwhile, load allocation is synthesized to generate a load optimization scheme;

the state estimation module is used for carrying out continuous estimation and risk prediction on the building health state by adopting a dynamic Bayesian network based on the optimized regulation strategy, and generating health state assessment.

The load optimization scheme comprises load distribution of floors, suggested load adjustment amplitude and optimized energy consumption prediction, the structural feature map comprises feature vectors of key structural elements, potential risk point distribution and structural health indexes, the prediction strategy model comprises future state prediction, suggested maintenance measures and emergency response strategies, the digital twin entity comprises a virtual model, a state evolution path and a predicted maintenance time point, the response analysis result comprises individual response characteristics, group response trend and key influence factors, the compressed information data set comprises key data after dimension reduction, information compression rate and reconstruction error assessment, the optimized regulation strategy comprises an optimized load distribution scheme, energy efficiency improvement measures and expected running cost saving, and the health state assessment comprises current health indexes, future risk prediction and suggested inspection intervals.

In the load analysis module, real-time data collected through a sensor network is used for carrying out pattern recognition on building loads by using a K-means clustering algorithm. In this process, the algorithm first randomly selects K initial centroids and then assigns data points to the nearest centroids according to euclidean distance to form K clusters. Next, the center point of each cluster is recalculated, and the process iterates until the allocation of data points within the cluster is no longer changing. At this time, the identified patterns are iteratively optimized by using a genetic algorithm, including selection, crossover and mutation operations, to improve accuracy and stability of the patterns. Through the steps, the load modes are accurately divided, and then an optimization scheme is generated by combining load allocation, so that the energy consumption is effectively reduced, and the running efficiency of the building is improved.

In the data reconstruction module, based on a load optimization scheme, an automatic encoder is adopted to perform data dimension reduction processing. The automatic encoder compresses data by an encoder and then reconstructs the data by a decoder with the aim of making the reconstructed data as close as possible to the original data. Meanwhile, a Convolutional Neural Network (CNN) is used for carrying out depth analysis and feature extraction on the building structure image and data. The CNN automatically extracts key features of the image through the multi-layer convolution layer and the pooling layer, so that the complexity and subjectivity of manually selecting the features are avoided. In this way, the generated structural feature map shows the feature vector, potential risk point distribution and structural health index of the key structural elements in detail, and provides important basis for subsequent analysis.

In the enhanced prediction module, the deep reinforcement learning model is utilized to learn the environment and the structural features based on the structural feature map. In this process, the model learns the optimal strategy through interactions with the environment to maximize long term benefits. Policy optimization iterations and self-evaluation adjustments help models improve continuously, thereby circularly optimizing the dynamic response predictions of the building structure. The generated prediction strategy model comprises future state prediction, maintenance measures and emergency response strategies, and scientific decision support is provided for building management.

In the digital twin module, a dynamic Bayesian network is adopted to carry out depth modeling and state prediction on time series data of a building. The dynamic bayesian network is capable of handling uncertainty and dynamic changes in time series data by continually updating the model to match new data inputs to generate digital twin entities. The virtual model not only reflects the current state of the building, but also predicts the future maintenance time point through the state evolution path, and provides a powerful tool for long-term health management of the building.

In the nonlinear analysis module, based on the digital twin entity, the response of the building individual and group structure is subjected to nonlinear analysis by adopting a mixed effect model. The mixed effect model can simultaneously consider the fixed effect and the random effect, and more accurately describe the complex response behavior of the building structure. Through the analysis, individual response characteristics, group response trend and key influence factors can be extracted, and scientific basis is provided for safety evaluation and maintenance of the building structure.

In the information compression module, based on the response analysis result, data dimension reduction is performed using Principal Component Analysis (PCA) and Independent Component Analysis (ICA). PCA finds the principal components of the data by linear transformation, while ICA finds the independent components to reveal the intrinsic structure of the data. The two methods work together to effectively extract and compress key information and generate a compressed information data set. This data set not only reduces storage requirements, but also increases the efficiency of data processing.

In the policy optimization module, a particle swarm optimization algorithm is adopted to intelligently adjust and optimize the load distribution and management policies based on the compressed information data set. Particle swarm optimization algorithms find the optimal solution by modeling the social behavior of the bird swarm, each particle in the algorithm represents a potential solution, and updates its own location by following the current optimal particle and the individual's historical optimal location. The method can efficiently find the optimal load distribution scheme, the energy efficiency improvement measure and the expected running cost saving scheme, and provides a powerful tool for energy management and optimization of the building.

In the state estimation module, based on an optimized regulation strategy, a dynamic Bayesian network is adopted to carry out continuous estimation and risk prediction of the building health state. By the method, the health condition of the building can be monitored and evaluated in real time, potential risks can be predicted in time, and a health state evaluation report is generated. These reports not only include current health metrics and future risk predictions, but also provide suggested inspection intervals, providing a scientific basis for long-term maintenance and safety management of the building.

Referring to fig. 3, the load analysis module includes a load monitoring sub-module, a first pattern recognition sub-module, and a genetic optimization sub-module;

the load monitoring submodule is based on real-time monitoring requirements, adopts a distributed sensing network and comprises a temperature sensor, a vibration sensor and a stress sensor, collects load data of a plurality of parts of a building, including temperature change, vibration frequency and stress change, and performs data summarization and preliminary screening to generate a load data set;

the first pattern recognition submodule is used for classifying and labeling the data into K categories by calculating and comparing the distances among data points based on a load data set and adopting a K-means clustering algorithm, and recognizing differentiated load patterns comprising normal load and overload to generate load pattern classification;

the genetic optimization submodule classifies based on load modes, adopts a genetic algorithm, performs multi-generation iterative optimization on a load distribution scheme through simulation selection, crossover and mutation operation, captures an optimal load distribution strategy based on energy consumption and efficiency, and generates a load optimization scheme.

In the load monitoring submodule, real-time monitoring is realized through a distributed sensing network, and mainly relates to a temperature sensor, a vibration sensor and a stress sensor. These sensors are distributed at key locations in the building and collect data about temperature changes, vibration frequencies and stress changes. For example, a temperature sensor records temperature data for each inspection point, a vibration sensor measures the vibration frequency of the building structure, and a stress sensor captures the magnitude of stress experienced by the structure. These data are stored in time series, typically including time stamps of the data points and corresponding measurements. And (3) primarily screening the collected data, removing abnormal values or noise, and ensuring the quality and accuracy of the data. Then, a load data set is generated through data integration and analysis processing. This dataset reflects the load conditions of the building structure at different points in time in detail, providing the basis for subsequent pattern recognition and load optimization.

In the first pattern recognition sub-module, a load data set is processed by adopting a K-means clustering algorithm, so that the recognition of a load pattern is realized. The K-means clustering algorithm first randomly selects K centroids as initial cluster centers, then calculates the distance from each data point to these centroids, and assigns the data point to the nearest centroid to form K clusters. Next, the algorithm updates the centroid position for each cluster, i.e., moves the centroid to the mean position of all points of the cluster. This process is repeated until the centroid position no longer changes significantly or a preset number of iterations is reached. In this process, the K-means algorithm can effectively classify data into different categories, such as regular load and overload, by considering various indicators of building load, such as temperature, vibration, and stress. Finally, the generated load pattern classification clearly marks the characteristics of various loads, and key information is provided for further load optimization.

In the genetic optimization sub-module, a genetic algorithm is adopted to optimize a load distribution scheme based on load pattern classification. Genetic algorithms simulate natural selection and genetic principles, searching the solution space through selection, crossover and mutation operations. Initially, the algorithm randomly generates a set of load distribution schemes, i.e., an initial population. In each iteration, the performance of each scheme is evaluated according to fitness functions (energy consumption and efficiency), and the scheme with better performance is selected to be reserved to the next generation. The interleaving operation allows the two schemes to exchange part of the features, creating a new scheme. The variation operation is to randomly change the pattern to increase the diversity. After multiple generations of iteration, the genetic algorithm can find the optimal or near-optimal load distribution strategy, and the strategy optimizes the energy consumption and the efficiency while ensuring the safety of the building. The generated load optimization scheme not only reduces the running cost of the building, but also improves the energy use efficiency.

It is assumed that in the load monitoring sub-module, the distributed sensing network collects the following data during the day: temperature data measured by the temperature sensor at different floors (such as 20 ℃ and 22 ℃), vibration frequencies recorded by the vibration sensor (such as 50Hz and 55 Hz) and stress values captured by the stress sensor (such as 100N/m and 120N/m). The data are initially screened and sorted to form a load data set which comprises time stamps and index values.

In a first pattern recognition sub-module, the data are processed using a K-means clustering algorithm. Setting the cluster number K as 2, the algorithm randomly selects two centroids first, and then distributes data points to the nearest centroids according to the temperature, the vibration frequency and the stress value to form two clusters. After multiple iterations, two stable clusters are finally formed, representing normal load and overload, respectively.

In the genetic optimization sub-module, a series of load distribution schemes are initially randomly generated based on the two load patterns. Each solution is evaluated by a fitness function and iteratively optimized by selection, crossover and mutation operations. Assuming that after 50 iterations, the algorithm determines an optimal load distribution strategy, which guarantees building safety while achieving optimal optimization of energy consumption.

Finally, the generated load optimization scheme comprises a specific load distribution scheme, predicted energy consumption saving amount and the like, and provides important decision support for energy management, operation and maintenance of the building.

Referring to fig. 4, the data reconstruction module includes a self-encoding sub-module, a convolution analysis sub-module, and a feature extraction sub-module;

the convolution analysis submodule adopts a convolution neural network to set a multi-layer convolution and pooling layer structure based on compression load characteristics, carries out layer-by-layer convolution and sub-sampling on the input compression characteristics, extracts local characteristics and gradually combines the local characteristics into total local characteristics, gradually refines key structural characteristics of a building through nonlinear activation and normalization to enhance the recognition capability of a model, and generates building characteristic classification;

In the self-encoding sub-module, data of the load optimization scheme is processed through an automatic encoder algorithm. An automatic encoder is a neural network, and consists of an encoder and a decoder. The function of the encoder is to compress the high-dimensional input data into a low-dimensional representation of the features, from which the decoder reconstructs the original data. In a specific operation, the network structure of the automatic encoder is first determined, including the number of layers of the encoder and decoder and the number of neurons per layer. The input data is a multidimensional data set obtained based on a load optimization scheme, for example, time-series data including a plurality of indicators of temperature, vibration frequency, stress, and the like.

The encoder section compresses the data layer by layer, each layer of neural network transforms the data by weights and biases, and introduces nonlinearities by activating the function to extract more abstract features. These layer-by-layer extracted features are compressed into a low-dimensional feature representation. The decoder section then starts from this low-dimensional representation, increases the dimension of the data layer by layer, and eventually reconstructs the output as close as possible to the original data. During reconstruction, the data transformation is also performed by weight, bias and activation functions.

During the whole training process of the automatic encoder, the network parameters are optimized by using a back propagation algorithm. The difference between the reconstructed data and the original data is measured by calculating the difference, using an loss function (e.g., mean square error), and then adjusting the weights and offsets of the network by gradient descent to minimize the reconstruction error. Through repeated iterative training, the network learns how to effectively compress and reconstruct data, and compression load characteristics are generated. These features preserve the most critical information while reducing the data dimension, providing an important basis for subsequent analysis.

In the convolutional analysis sub-module, further analysis is performed using a Convolutional Neural Network (CNN) based on the compression loading characteristics. The CNN processes the input data through the multi-layer convolution and pooling layers to effectively extract and identify local features and gradually combine them into global features. In a specific implementation, the structure of the CNN is first set, including the number and order of the convolution layers, the pooling layers, and the size and number of filters (convolution kernels) used for each layer.

Each convolution layer convolves the input data with a filter, extracts local features, and adds nonlinearity by activating a function (e.g., reLU). And the pooling layer downsamples the convolved feature map, so that the dimension and complexity of the data are reduced. As the network layer deepens, convolutional neural networks can progressively abstract more complex global features from simple local features. In addition, normalization layers are often included in the network to stabilize the training process and improve the generalization ability of the model. Through superposition and synergy of these layers, CNNs are able to extract key features about building structures from compressed load features, generating building feature classifications. These classifications reflect the structural characteristics of the building and are of great value in guiding maintenance and optimization of the building.

In the feature extraction sub-module, edge detection and texture analysis technologies are comprehensively applied based on building feature classification. Edge detection techniques are used to identify and label the contours and boundaries of building structures, while texture analysis techniques are used to analyze and identify the texture features of building surfaces. These techniques deal with building feature classification data from convolution analysis sub-modules that already contain preliminary feature information for the building structure.

In a specific operation, the edge detection technique determines edges by identifying regions in the image where the brightness changes significantly, and common algorithms include Sobel, canny, and the like. Texture analysis describes the characteristics of the texture by calculating the statistical properties of various regions in the image, such as contrast, uniformity, etc. By these techniques, details of the building structure and potentially risk areas can be accurately marked. Then, key feature points of the health state of the building, such as cracks, corrosion and the like, are identified through comparison and analysis of the features of the plurality of areas. The finally generated structural feature map shows the key structural elements and potential risk points of the building in detail, and provides important basis for health monitoring and maintenance of the building.

It is envisaged that in the self-encoding submodule, the input data is time series data comprising temperature (20 ℃, 22 ℃), vibration (50 Hz, 55 Hz) and stress (100N/m, 120N/m < mu >). After compression and reconstruction by an automatic encoder, the resulting low-dimensional features are compact representations of the data, e.g., feature vectors down to 3 dimensions. In the convolution analysis sub-module, these features are further processed to identify key structural features of the building, such as specific vibration modes or stress distributions. Finally, in the feature extraction sub-module, the structural details and potential risk areas of the building are marked through edge detection and texture analysis technology, and the generated structural feature map contains detailed description and position information of the areas, so that visual guidance is provided for maintenance of the building.

Referring to fig. 5, the enhanced prediction module includes a feature learning sub-module, a strategy iteration sub-module, and a self-evaluation sub-module;

the feature learning submodule carries out comprehensive feature learning by adopting a convolutional neural network and a cyclic neural network based on the structural feature map, wherein the convolutional neural network extracts spatial features from building data, the spatial features comprise structural shapes and spatial distribution, the cyclic neural network processes time sequence data, and the change trend of structural states along with time is captured to generate comprehensive environmental features;

The strategy iteration sub-module adopts a strategy gradient or Q learning method in reinforcement learning to carry out strategy iteration based on comprehensive environmental characteristics, simulates the influence of differentiated actions on a prediction target by establishing a reward mechanism, and gradually adjusts and optimizes action strategies by multiple rounds of iterative optimization to generate an improved action strategy;

the self-evaluation sub-module adopts a model self-evaluation method based on an improved action strategy, comprises the steps of continuously monitoring the prediction performance and generalization capability of the model, evaluating and adjusting the model in real time, and generating a prediction strategy model by analyzing the deviation between a prediction result and the actual situation and adjusting the weight and parameters of the neural network.

In the feature learning sub-module, comprehensive feature learning is achieved by combining a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN). The data format of the sub-module is mainly multidimensional data based on a structural feature map, and the multidimensional data comprises spatial features (such as shape and distribution) and time sequence data (such as change of structural state along with time) of a building structure. CNNs are used to process spatial features, extracting spatial features in building data, such as shape and spatial distribution of structures, etc., through multiple convolution and pooling layers. Each convolution layer extracts local features of the image by applying a plurality of filters, while the pooling layer is used to reduce the spatial dimensions of the features and extract more abstract features. These features are abstracted and integrated layer by layer to form a comprehensive description of the spatial characteristics of the building structure.

Meanwhile, the RNN is particularly suitable for processing time series data, and can effectively capture dynamic changes of the data along with time. Here, the RNN remembers the information of the previous time step through the cyclic connection of its internal state (hidden layer), thereby capturing the trend of the change in the state of the structure at each time step. This structure of RNNs enables them to understand and predict the time-dependent relationship in sequence data. In this way, the combination of CNNs and RNNs enables the sub-module to extract and integrate features of spatial and temporal dimensions from the building data, generating integrated environmental features. These features provide a solid basis for a comprehensive analysis and understanding of the building environment.

In the strategy iteration submodule, a strategy gradient or Q learning method in reinforcement learning is adopted based on comprehensive environmental characteristics. This process involves establishing a rewarding mechanism to simulate the effect of different actions on the predicted targets. In the strategy gradient approach, the goal is to learn directly a strategy that can map observed states to actions taken. By testing different actions and receiving rewards or penalties, the model gradually learns how to select the optimal action to maximize the total rewards. Q learning is a value-based method that estimates the value (Q value) of each state-action pair, and model learning how to select the action that will bring the highest expected reward. Whether strategy gradients or Q learning are employed, the process involves multiple rounds of iterative optimization, by which the model can gradually adjust and optimize its action strategy. The resulting improved action strategy provides more intelligent and efficient decision support for the management and operation of the building.

In the self-evaluation sub-module, a model self-evaluation method is adopted to evaluate and adjust the model in real time based on an improved action strategy. The process comprises continuously monitoring the prediction performance and generalization capability of the model, and analyzing the deviation of the prediction result from the actual situation. Through real-time evaluation, the defects of the model in prediction, such as low prediction accuracy or poor adaptability to new conditions, can be timely found. For these problems, improvements are made by adjusting the weights and parameters of the neural network. The self-evaluation mechanism ensures that the model is not only effective in the initial training stage, but also continuously keeps high efficiency and accuracy in the actual application process. The generated prediction strategy model is an important tool for accurately predicting the building state and environment, and has important value for guiding maintenance, optimization operation and risk management of the building.

It is envisaged that in the feature learning sub-module, the structural feature map contains spatial feature data of the building, such as the shape (circles, squares) and spatial distribution (centers, edges) of a certain structure, and time series data, such as the change of structural stress over time (gradual increase from 100N/m to 150N/m). The CNN extracts patterns of shape and distribution from the spatial features, while the RNN analyzes the trend of stress over time. In the strategy iteration sub-module, these integrated features are used to train a reinforcement learning model that learns through multiple rounds of iteration how to make optimal decisions in different building states. In the self-evaluation sub-module, the performance of the model is evaluated and adjusted by comparing the predicted stress variation with the actual observed data to improve the accuracy of the prediction. The finally generated prediction strategy model can provide accurate predictions about building structure health conditions and environmental changes, and provide guidance for building management and maintenance.

Referring to fig. 6, the digital twin module includes a time series modeling sub-module, a state prediction sub-module, and a model update sub-module;

the state prediction submodule is used for extracting and analyzing features of future states of the building by adopting a random forest algorithm based on a building data time dependent model, carrying out combined analysis on the differentiated features by constructing a plurality of decision trees, predicting the future states and generating predicted building state analysis;

In the time series modeling sub-module, a predictive strategy model of the digital twin module is processed by a dynamic Bayesian network algorithm. The data format of the sub-module is mainly time series data of a building, including time-varying data of temperature, vibration frequency, stress and the like. A dynamic bayesian network is a probabilistic graphical model for processing time series data that captures time series dependencies among data points. In an implementation, a network structure is first defined, including state variables, observation variables, and probabilistic relationships therebetween. The state variables represent the internal states of the building, while the observed variables are actual measurements of these states.

The state of each point in time during the construction of the network is not only dependent on the state of the previous point in time, but is also affected by the current observations. By setting transition probabilities (transitions between states) and observation probabilities (state-to-observation mappings), the network is able to describe the evolution of states over time. In the model training stage, the network parameters, namely the specific values of the learning state transition and the observation probability, are learned according to the existing time series data by using methods such as maximum likelihood estimation or Bayesian inference. After training is completed, the model can make probabilistic inference and prediction of future states. The process enables the model to understand and predict the change rule of the building state along with time through learning historical data, and generates a building data time dependent model. The model provides important decision support for the state monitoring and early warning of the building, and helps a manager to know the future development trend of the building.

In the state prediction sub-module, based on a building data time dependent model, a random forest algorithm is adopted to extract and analyze features of future states. Random forests are an integrated learning method that analyzes and predicts data by constructing multiple decision trees. Each decision tree is trained on a random subset of the data, enabling capture of different features and patterns in the data. In a specific operation, a plurality of sample subsets are randomly extracted from the original data set first, and then a decision tree is built on each subset. When constructing decision trees, the splitting is performed by randomly selecting features, so that each tree focuses on different aspects of the data. When state prediction is performed, input data is passed to each tree and the prediction results of all trees are aggregated, typically by voting or averaging, to obtain the final prediction. The random forest improves the accuracy and the robustness of prediction in this way, and the generated prediction building state analysis can be used for guiding maintenance and management work of a building and planning countermeasures in advance.

In the model updating sub-module, the model is updated in real time by using an online learning method based on the prediction building state analysis. Online learning is a dynamic learning process in which a model gradually adjusts its parameters as new data is received. In a specific implementation, an update rule of a model parameter is set first, for example, information of new data is gradually integrated by an incremental learning mode. When new observation data arrives, the model adjusts its state predictions and time dependencies based on these data. The continuous updating process ensures the adaptability and response speed of the model to new conditions, so that the model can timely reflect the latest change of the building state. The generated digital twin entity is a dynamic representation of building status and performance, continually updated from new data, providing an accurate depiction of the current and future status of the building, and providing a powerful tool for long-term health monitoring and management of the building.

In the time series modeling module, it is assumed that the data processed is temperature and vibration data for each hour of the building in a week. The dynamic bayesian network learns the law of temperature and vibration over time by analyzing these data. In the state prediction sub-module, a random forest algorithm predicts the temperature and vibration trend of the next day based on these time dependent models. Finally, in the model updating sub-module, when new day data arrives, the model is updated according to the latest data so as to ensure the accuracy and timeliness of prediction. The digital twin entity generated in this way can continuously reflect the real-time state of the building, and provides real-time data support for maintenance and management of the building.

Referring to fig. 7, the nonlinear analysis module includes a mixed effect modeling sub-module, a population analysis sub-module, and a feature recognition sub-module;

the group analysis submodule analyzes building group data from different layers by adopting a multi-level model based on individual response characteristic analysis, processes a complex data set by constructing a data structure comprising a plurality of layers, respectively analyzes the data on multiple layers to reveal the interrelationship and the difference between individuals and inside the group, and generates group response rule extraction;

In the mixed effect modeling sub-module, statistical modeling is carried out on the building individual structure response data of the digital twin entity through a generalized linear mixed model. The data format handled by this sub-module typically includes response indicators of stress, vibration, and temperature of the building structure, which data changes over time and reflects different conditions of the building. The generalized linear hybrid model combines a fixed effect, which represents a common characteristic common to all buildings, and a random effect, which captures the unique changes of individual buildings on these common characteristics.

In practice, the structure of the model is first determined, including selecting variables that reflect the response of the building, such as stress, vibration, etc., and defining the fixed and random effect portions of these variables. The fixed effect is partially related to establishing a statistical model of the characteristics common to the population of buildings, while the random effect is set for each individual building to capture its unique response characteristics. Model parameters are then estimated by maximum likelihood estimation or similar methods, which include coefficients that account for variables and variance components of random effects. After model estimation is completed, the model estimation method can be used for predicting and analyzing the response of the individual building. Such statistical modeling not only can analyze the unique response characteristics of each building, but also can reveal the overall trend of the building response at the population level, and the generated individual response feature analysis is critical to understanding and optimizing the structural response of the building.

And in the group analysis submodule, building group data of different layers are analyzed by adopting a multi-level model based on individual response characteristic analysis. The data formats here include the geographical location of the building, year of construction, materials used, etc., which are present at different levels, such as individual buildings, groups of buildings, entire cities, etc. The multi-level model is capable of analyzing data at different levels, respectively, revealing interrelationships and differences between individuals and within a population. In practice, a data structure is built that includes multiple levels, and then a statistical model is built at each level. The model considers the interaction and influence among different layers, and through the layering analysis, the influence of each layer of factors on the building response can be more accurately understood. The generated group response rule extraction helps to formulate management and maintenance policies for the entire building group or specific sub-groups.

In the feature recognition sub-module, based on the extraction of the group response rule, the extracted features are further analyzed by adopting Principal Component Analysis (PCA) and cluster analysis. At this stage, the data format includes various statistical indicators and characteristic data derived from the population analysis. Principal component analysis converts raw data into a set of linearly independent variables, called principal components, by linear transformation, thereby simplifying the data structure and highlighting important features. Cluster analysis is the grouping of individuals based on data characteristics to identify buildings or groups with similar characteristics. In the specific operation, firstly, PCA is used for carrying out dimension reduction processing on the data, so that the complexity of the data is reduced, and then, a clustering algorithm (such as K-means or hierarchical clustering) is applied to divide the data into different groups according to the principal component result. The processing not only can identify and distinguish key characteristics and modes of individual and group responses, but also is helpful for understanding the overall behavior characteristics of building groups, and the generated response analysis results have important significance for guiding maintenance and improvement strategies of the buildings.

In the mixed effect modeling module, the processed data includes temperature changes (e.g., 30 ℃ average in summer and 5 ℃ average in winter) and vibration response data (e.g., frequency between 50 Hz and 60 Hz) for a building during different seasons. Through generalized linear mixed model analysis, the response characteristics of the building in different seasons can be known and compared with other buildings, so that the individual and group response characteristics of the building are revealed. In the group analysis sub-module, building groups of different geographic locations and year of construction are further analyzed, revealing differences between the different groups. Finally, in the feature recognition sub-module, the data are processed using PCA and cluster analysis to identify major factors that have significant impact on building response, such as the impact of geographic location on temperature changes, and the impact of different construction years on vibration response. These analysis results provide important data support for maintenance and optimization of the building.

Referring to fig. 8, the information compression module includes a data dimension reduction sub-module, a second pattern recognition sub-module, and a compression storage sub-module;

based on a response analysis result, the data dimension reduction submodule adopts principal component analysis, identifies and extracts key principal components as new data representation by statistically analyzing the relativity among variables in the original data, and further adopts independent component analysis, separates the original signals from the mixed signals by maximizing the statistical independence of non-Gaussian source signals, and generates a dimension reduction feature set;

The second pattern recognition submodule adopts a support vector machine based on the dimension reduction feature set, classifies the dimension reduced data by constructing a decision boundary, and simultaneously uses a neural network to recognize key patterns and trends in the data set by training internal patterns of data learning so as to generate a pattern recognition data set;

the compression storage sub-module adopts a lossless compression technology based on a pattern recognition data set, encodes the data through Huffman coding or LZW algorithm, reduces the storage of repeated data, adopts a lossy compression technology, reduces the data quantity through reducing the data precision, reserves the key information and the characteristics of the data, and generates a compression information data set.

In the data dimension reduction sub-module, the response analysis result is processed by Principal Component Analysis (PCA) and Independent Component Analysis (ICA). The data format of the sub-modules is typically multi-dimensional and includes metrics of the response of various building structures, such as temperature, vibration frequency, stress, etc. First, principal component analysis is used to statistically analyze the correlation between variables in the raw data. In the implementation of PCA, the covariance matrix of the dataset is first calculated to determine the correlation between the variables. Then, the covariance matrix is decomposed by eigenvalues to find the principal eigenvector, i.e., principal component, representing the direction of maximum variance in the data. The data is projected onto these principal components, thereby achieving dimension reduction. In this way, the PCA recognizes and extracts key principal components as new data representations, effectively reducing the dimensionality of the data while retaining the most important information.

The independent component analysis is then used to separate the original signal from the mixed signal. The aim of ICA is to find a linear transformation that makes the transformed signals as statistically independent as possible. In a specific operation, it is first assumed that the data is a linear mixture of some unknown independent source signals, and then the transformation matrix is adjusted by an optimization algorithm until the statistical independence of the output signals is maximized. By ICA, more meaningful and independent features can be separated from the multidimensional data, generating a set of dimension-reducing features. This dimension reduction feature set simplifies the complexity of the original data set, making subsequent processing and analysis more efficient.

In the second pattern recognition sub-module, further pattern recognition is performed using a Support Vector Machine (SVM) and a neural network based on the reduced-dimension feature set. The SVM realizes the classification of the dimension-reduced data by constructing an optimal decision boundary between data points. In practice, data is first mapped to a high-dimensional space by selecting an appropriate kernel function, such as a linear kernel, radial Basis Function (RBF) kernel, etc., in which a hyperplane is constructed to maximize the separation between the different classes. Meanwhile, the neural network is used to learn the intrinsic pattern of data through training data. In particular, neural networks can learn and model complex data relationships through multi-layer structures and nonlinear activation functions, thereby identifying key patterns and trends in the data set. Through the combination of the SVM and the neural network, the submodule can effectively identify and classify key modes from the data subjected to dimension reduction, and a mode identification data set is generated. These data sets are critical to understanding the response characteristics of the building in depth, providing important decision support for maintenance and management of the building.

In the compression storage sub-module, data is compressed using lossless compression and lossy compression techniques based on the pattern recognition data set. Lossless compression techniques, such as huffman coding or LZW algorithms, reduce the storage of duplicate data by intelligently encoding the data. Specifically, huffman coding achieves data compression by assigning shorter codes to elements in the data set that occur more frequently, and assigning longer codes to elements that occur less frequently. The LZW algorithm compresses the data by constructing a dictionary to represent the different string patterns that appear in the data. Meanwhile, lossy compression techniques reduce the amount of data, such as reducing the number of bits of data or simplifying the representation of data, by reducing the accuracy of the data, while striving to preserve key information and features of the data. In this way, compressing the information dataset both reduces the storage space requirements and retains critical information for analysis and decision making. This is particularly important where large amounts of building data need to be processed and stored, contributing to improved data processing efficiency and reduced storage costs.

In the data dimension reduction sub-module, it is assumed that the raw dataset includes a series of building temperature, humidity and energy consumption data. By PCA, these data are reduced in dimension to several principal components that contain the most critical information. The ICA then further separates out more independent features from these mixed principal components, such as factors that primarily affect energy consumption. In a second pattern recognition sub-module, SVMs and neural networks are used to recognize and classify patterns in these dimension-reduction feature sets, such as distinguishing high-energy and low-energy buildings. Finally, in the compressed storage sub-module, these pattern recognition data sets are compressed by huffman coding and lossy compression techniques to facilitate efficient storage and processing. The compressed information data set generated in this way not only reduces the storage space requirement, but also provides important data support for building energy efficiency analysis.

Referring to fig. 9, the policy optimization module includes a load adjustment sub-module, a policy adjustment sub-module, and a real-time update sub-module;

the load allocation submodule adopts a particle swarm optimization algorithm based on the compressed information data set, searches in a solution space by setting potential solutions represented by particles, adjusts a flight path according to personal experience and group experience by each particle, captures an optimal solution of load allocation, and generates an intelligent load allocation scheme;

the strategy adjustment submodule gradually reduces the temperature from the initial temperature based on an intelligent load distribution scheme by adopting a simulated annealing algorithm, carries out random disturbance in a small range on the current solution at each temperature stage, determines whether to accept a new solution or not through probability, avoids local optimization, and generates an optimization management strategy;

based on the optimization management strategy, the real-time updating sub-module uses a genetic algorithm to simulate selection, crossover and mutation, circularly generates new strategy combinations, evaluates the fitness, selects an optimal strategy for iteration, carries out continuous optimization of the strategy, and generates an optimal regulation strategy.

In the load adjuster module, a Particle Swarm Optimization (PSO) algorithm is used to process the compressed information data set to generate an intelligent load distribution scheme. The data format of the sub-module mainly comprises energy usage data of the building, environmental parameters, historical load data and the like, and the data form a multidimensional point set in the knowledge space. The particle swarm optimization algorithm searches the solution space by simulating the social behavior of the shoal or the shoal to find the optimal solution. In practice, a population of particles is first initialized, each particle representing one potential solution in the solution space, i.e., a particular load distribution strategy.

Each particle adjusts its own flight speed and direction based on its own experience (personal best position) and the experience of the population (global best position). In particular, the velocity update of the particles takes into account the current velocity, the distance of the particles from the personal best location, and the distance of the particles from the global best location. In this way, particles can be efficiently searched in the solution space and gradually get closer to the optimal solution. The iterative process of particle swarm optimization continues until a predetermined number of iterations is reached or the quality of the solution meets certain criteria. The intelligent load distribution scheme can optimize the energy use of the building, improve the energy efficiency and reduce the energy consumption and the cost.

In the strategy adjustment sub-module, based on an intelligent load distribution scheme, a simulated annealing algorithm is adopted for optimization. The simulated annealing algorithm is a probability search algorithm, simulates the annealing process in the object, starting from a higher initial "temperature" and gradually "cooling" and performing a small range of random perturbations on the current solution at each temperature stage. In a specific operation, a higher initial temperature is set first, and a random disturbance is performed on the current solution at the temperature, and then whether to accept the new solution is determined according to the quality of the new solution and a specific probability decision criterion (such as a Metropolis criterion). With the gradual decrease of the temperature, the algorithm gradually reduces the search range of the solution, thereby improving the search efficiency. The simulated annealing algorithm can effectively avoid the local optimal solution in the mode, and finally an optimized management strategy is generated. This strategy is of great significance for coping with complex and varying load demand scenarios, such as seasonal variations, daytime load fluctuations, etc.

In the real-time updating sub-module, the strategy is continuously optimized by using a genetic algorithm based on the optimized management strategy. Genetic algorithm is a search algorithm simulating natural selection and genetic mechanism, and new strategy combination is circularly generated through selection, crossover and mutation operation. In a specific implementation, a set of strategy populations is first initialized, each strategy representing a potential solution. By evaluating the fitness of the strategies, namely the capacity of meeting load requirements and optimizing energy efficiency, strategies with higher fitness are selected to enter the next generation. The interleaving operation allows the two policies to exchange part of their features, creating a new policy combination. The mutation operation is to randomly alter the strategy to introduce new genetic diversity. This cycle is repeated until an optimal load deployment strategy is found. By means of the method, the real-time updating sub-module can ensure that the load allocation strategy is adapted to the continuously changing environment and requirements, and the generated optimal regulation strategy can effectively improve the energy efficiency and the operation performance of the building.

It is assumed that in the load adjuster module, the data includes an energy consumption record of an office building per hour in a week. Through particle swarm optimization, a series of potential energy allocation schemes are generated. In the strategy adjustment submodule, the schemes are optimized through a simulated annealing algorithm, and finally a strategy for reducing the energy consumption in the peak period is found. Then, in the real-time updating sub-module, the strategy is continuously adjusted through a genetic algorithm to adapt to the energy demand change of the next week. The finally generated optimal regulation strategy not only reduces energy consumption, but also balances the load distribution of the whole office building.

Referring to fig. 10, the state estimation module includes a state monitoring sub-module, a risk prediction sub-module, and a health assessment sub-module;

the risk prediction sub-module performs dynamic Bayesian network analysis based on the real-time state data set, updates probability distribution according to the real-time data, identifies potential structural risk modes by calculating time correlation among data points, predicts short-term and long-term risk trends, and generates risk prediction analysis;

the health evaluation sub-module is used for carrying out building health evaluation based on risk prediction analysis, utilizing time sequence analysis to mine long-term data trend and change mode, analyzing by combining an expert system with industry standard and historical data, evaluating the overall health condition of the building, and generating health state evaluation.

In the state monitoring sub-module, real-time data collection is performed, and the sensors are responsible for continuously monitoring key indexes such as vibration, temperature, pressure and the like through a sensor network deployed at key structural points of the building, so that the physical state change of the building is tracked. The data formats generated by these sensors typically include time series data, with each data point including a time stamp and corresponding measurement value. In a specific implementation, the sensor network is deployed first, ensuring that all critical monitoring points are covered. The sensors then collect data in real time, including vibration frequency, temperature changes, stress levels, etc., and transmit these data to a central processing system.

In the central processing system, the collected data is collated and stored while analysis of the trend of the data is performed. Data analysis mainly includes identifying abnormal patterns, trend changes, and periodic changes of data. For example, by comparing historical data to real-time data, structural problems or potential damage may be identified. This process continues, generating real-time state data sets that are critical to maintenance and safety management of the building, provide an immediate view of the current state of the building, and serve as a basis for risk assessment and prediction.

In the risk prediction sub-module, dynamic bayesian network analysis is performed, based on a real-time state dataset. A dynamic bayesian network is a probabilistic graphical model for processing time series data that captures the time correlation between data points. In practice, a probability distribution model is first built based on real-time data, including definitions of state variables and observation variables, and probability relationships among each other. These data are then analyzed by a dynamic bayesian network to calculate transition probabilities between different states, thereby identifying potential structural risk patterns.

By this method, the network can continuously update its probability estimates based on real-time data, providing short-term and long-term risk trend predictions. Such risk prediction analysis is important for timely finding and coping with potential problems with building structures, and can help building managers take preventive measures to avoid or mitigate potential damage.

In the health assessment sub-module, health assessment of the building is performed based on the risk prediction analysis. The process involves mining long-term data trends and patterns of changes using time series analysis and evaluating the overall health of the building in combination with expert systems and industry standards. Time series analysis identifies long-term trends and periodic patterns by in-depth analysis of collected historical and real-time data, which are critical to understanding the health status of a building.

And the expert system is used for reading the analysis result by combining the industry standard and the historical data to provide a comprehensive health evaluation report. These reports include not only the current health of the building, but also the predicted future occurrence of problems and risks. In this way, the health status assessment generated by the health assessment submodule provides a scientific basis for maintenance and management of the building, is beneficial to implementing an effective maintenance strategy, prolongs the service life of the building and ensures the safety of the building.

In the condition monitoring sub-module, it is assumed that the sensor network collects vibration data on a bridge, which data indicates that the bridge has increased in vibration frequency over a specified period of time. The risk prediction sub-module analyzes the data via a dynamic bayesian network to identify a structural risk pattern, such as vibration enhancement due to increased traffic load. Finally, in the health assessment sub-module, the current health condition of the bridge is obtained by combining time series analysis and expert system assessment, but additional monitoring is recommended during peak hours to prevent potential risks. The analysis results provide important guidance for maintenance and management of the bridge.

The present invention is not limited to the above embodiments, and any equivalent embodiments which can be changed or modified by the technical disclosure described above can be applied to other fields, but any simple modification, equivalent changes and modification made to the above embodiments according to the technical matter of the present invention will still fall within the scope of the technical disclosure.

Claims

1. The utility model provides a building engineering quality real-time supervision system which characterized in that: the system comprises a load analysis module, a data reconstruction module, an enhancement prediction module, a digital twin module, a nonlinear analysis module, an information compression module, a strategy optimization module and a state estimation module;

2. The real-time monitoring system for construction quality according to claim 1, wherein: the load optimization scheme comprises load distribution of floors, suggested load adjustment amplitude and optimized energy consumption prediction, the structural feature map comprises feature vectors of key structural elements, potential risk point distribution and structural health indexes, the prediction strategy model comprises future state prediction, suggested maintenance measures and emergency response strategies, the digital twin entity comprises a virtual model, a state evolution path and a predicted maintenance time point, the response analysis result comprises individual response characteristics, group response trend and key influence factors, the compressed information data set comprises key data after dimension reduction, information compression rate and reconstruction error assessment, the optimized regulation strategy comprises an optimized load distribution scheme, energy efficiency improvement measures and expected operation cost saving, and the health state assessment comprises current health indexes, future risk prediction and suggested inspection intervals.

3. The real-time monitoring system for construction quality according to claim 1, wherein: the load analysis module comprises a load monitoring sub-module, a first pattern recognition sub-module and a genetic optimization sub-module;

4. The real-time monitoring system for construction quality according to claim 1, wherein: the data reconstruction module comprises a self-coding sub-module, a convolution analysis sub-module and a feature extraction sub-module;

5. The real-time monitoring system for construction quality according to claim 1, wherein: the enhanced prediction module comprises a feature learning sub-module, a strategy iteration sub-module and a self-evaluation sub-module;

6. The real-time monitoring system for construction quality according to claim 1, wherein: the digital twin module comprises a time sequence modeling sub-module, a state prediction sub-module and a model updating sub-module;

7. The real-time monitoring system for construction quality according to claim 1, wherein: the nonlinear analysis module comprises a mixed effect modeling sub-module, a group analysis sub-module and a characteristic identification sub-module;

8. The real-time monitoring system for construction quality according to claim 1, wherein: the information compression module comprises a data dimension reduction sub-module, a second mode identification sub-module and a compression storage sub-module;

9. The real-time monitoring system for construction quality according to claim 1, wherein: the strategy optimization module comprises a load adjustment sub-module, a strategy adjustment sub-module and a real-time update sub-module;

10. The real-time monitoring system for construction quality according to claim 1, wherein: the state estimation module comprises a state monitoring sub-module, a risk prediction sub-module and a health assessment sub-module;