A multi-objective intelligent optimization decision method for modification and reconstruction of existing buildings

Abstract

The application relates to the technical field of existing building reconstruction, and particularly discloses a multi-target intelligent optimization decision method for existing building decoration and reconstruction, which comprises the following steps: collecting geometric point clouds and continuous time sequence infrared thermal images of existing building decoration and reconstruction targets, fusing and registering the geometric point clouds and the continuous time sequence infrared thermal images to generate an initial digital twin, and updating the geometry and physical state of the initial digital twin by mapping hidden defects detected by intelligent sensors; constructing a decoration engineering knowledge graph, inferring and completing missing hidden engineering semantic correlations through a graph neural network, and generating an existing component dynamic data set; through deep fusion and dynamic mapping of multi-source heterogeneous sensing data, the digital twin capable of continuously updating and truly reflecting the geometric shape and internal damage state of the existing building is constructed, high-fidelity dynamic reconstruction and accurate perception of the existing building information are realized, and the defects of traditional BIM model static distortion are overcome.

Description

Technical Field

[0001] This invention relates to the field of existing building renovation technology, and in particular to a multi-objective intelligent optimization decision-making method for the renovation of existing buildings. Background Technology

[0002] Driven by the dual strategies of urban renewal, the construction industry is undergoing a transformation from incremental expansion to quality improvement of existing buildings. With the large scale of existing buildings, energy-saving and green renovation has become a key path to reduce carbon emissions and achieve sustainable development. As the most frequent and directly related link in the renovation activities, the low-carbon transformation of building decoration and renovation is particularly important. Its renovation activities involve multiple links such as material production, demolition, transportation, construction and waste disposal. The entire process is intensive in resource and energy consumption and carbon emissions, making it a key area of ​​focus for the construction industry to achieve green and low-carbon development.

[0003] In existing technologies, the BIM models relied upon for the renovation and reconstruction of existing buildings are static and incomplete, lacking the ability to dynamically perceive the hidden defects and true remaining value of existing components. This leads to distorted component data on which demolition decisions are based, making it difficult to support accurate assessment of preservation value. Moreover, traditional AHP assessment relies on fixed weights, which cannot adapt to the differentiated value orientations among renovation and reconstruction projects of existing buildings. This results in low efficiency in solving massive component combination schemes under multi-objective conflicts, making it difficult for demolition schemes to reach the global optimum. In addition, the Pareto solution set generated by renovation and reconstruction of existing buildings is large in scale and the merits of each scheme are ambiguous, causing the selected renovation scheme to frequently encounter resource conflicts and carbon emission exceedances in subsequent construction, making it difficult to ensure the attainability of final benefits and the controllability of the implementation process. Therefore, this paper proposes a multi-objective intelligent optimization decision-making method for renovation and reconstruction of existing buildings to solve the above-mentioned problems. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention provides a multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings, which can effectively solve the problems involved in the prior art.

[0005] The objective of this invention can be achieved through the following technical solution: This invention provides a multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings, comprising the following steps: S1. Collect geometric point cloud and continuous temporal infrared thermal image of the target of renovation and renovation of existing buildings, register and fuse to generate an initial digital twin, realize synchronous digital mapping of the geometry and temperature of the existing buildings, and use intelligent sensors (ground penetrating radar) to detect hidden defects and map and update the geometry and physical state of the initial digital twin, so that the hidden defects are transformed from invisible to locatable and quantifiable digital objects. S2. Construct a knowledge graph for decoration engineering, and use graph neural network reasoning to complete the missing semantic associations of hidden engineering, so that the scattered hidden engineering information can form a reasonable association network and generate a dynamic dataset of existing components containing complete semantic features. S3. Based on the reinforcement learning agent, analyze the differentiated value orientation of the project, dynamically adjust the AHP (Analytic Hierarchy Process) evaluation weight according to the project preference, so that the evaluation standard automatically adapts to the project value orientation rather than being rigid and fixed. Using the dynamic data of existing components as input, generate an initial elite population, so that the initial solution set naturally carries the project preference characteristics, and improves the quality of the optimization starting point. S4. Input the initial elite population into the improved NSGA-II algorithm, and perform multi-objective iterative optimization with carbon emissions, cost and construction period as optimization objectives. Automatically seek the optimal balance point among the three conflicting objectives, and quickly output a Pareto optimal demolition scheme set for the renovation of existing buildings, providing a set of high-quality alternative schemes that do not dominate each other for subsequent screening. S5. An improved TOPSIS algorithm (superior-inferior solution distance method) is introduced. Taking the Pareto optimal demolition scheme set as the object, the frequency of construction conflicts and the standard deviation of carbon emission fluctuations in each scheme in the digital twin are used as negative penalty indicators. This allows the conflict risk and emission instability risk during the construction process to be quantified and included in the evaluation. The schedule is calculated, and the superiority and risk level of the scheme are comprehensively reflected by a single quantitative indicator. S6. Based on the schedule, sort the Pareto solutions for the renovation of existing buildings, eliminate inferior solutions with high risk and high volatility, so that the final solution meets the dual criteria of high efficiency and low risk, and uniquely lock in the optimal renovation solution for existing buildings, ensuring that the selected solution has the lowest implementation risk and the most stable overall benefits under multiple conflicting objectives.

[0006] Preferably, in step S1, the process of generating the initial digital twin is as follows: A 3D laser scanner is used to scan the renovation target of the existing building from multiple stations and angles to obtain geometric point cloud data covering the entire space. At the same time, an infrared thermal imager is used to continuously collect time-series thermal image sequences of the same spatial area at preset time intervals. The geometric shape and surface temperature field of the existing building are captured synchronously, and the data integrity is significantly improved. The geometric point cloud data is registered and fused with the temporal infrared thermal image sequence. Using the spatial coordinates of the point cloud as a reference, the thermal image pixels are spatially mapped through an iterative nearest point algorithm. Then, spatiotemporal interpolation is used to generate an initial digital twin that combines three-dimensional geometric contours and dynamic distribution of surface temperature field. The geometric and temperature data are accurately aligned under a unified spatial reference, and the twin information dimensions are richer. In the initial digital twin, temperature anomaly areas on the surface of each existing component are used as candidate identifiers of hidden defects. Smart sensors deployed in the target area of ​​the existing building renovation are used to detect the underground or internal layered echo signals to verify and locate the candidate identifiers, generate a hidden defect distribution map and map it to the corresponding spatial coordinates of the initial digital twin. The smart sensor is a ground penetrating radar. The surface temperature anomaly and the internal echo signal corroborate each other, and the reliability of hidden defect location is significantly enhanced. Based on the distribution map of hidden defects, the geometric accuracy parameters and physical state attributes of each existing component in the initial digital twin are updated to form an updated digital twin that can dynamically reflect the true geometric shape and internal damage of the existing building. The true geometric shape and internal damage of the existing building are completely reproduced in the digital space, and the fidelity of the twin is greatly improved.

[0007] Preferably, the process of updating the geometric and physical states of the initial digital twin is as follows: For each defect in the hidden defect distribution map, its spatial boundary coordinates and defect depth value are extracted to correct the geometric contour accuracy of the corresponding existing component in the initial digital twin. At the same time, the geometric dimensions of existing components with material spalling or cross-sectional reduction are negatively compensated and adjusted to correct the geometric deviation caused by defects one by one, so that the geometric accuracy of the twin is highly consistent with that of the real component. Based on the rate of change and temperature difference amplitude of the surface temperature of each existing component in the time-series thermal imaging sequence, the degree of decay of its thermal inertia index and thermal conductivity is deduced in reverse. Combined with the change characteristics of reflection intensity in the ground penetrating radar echo signal, the remaining mechanical performance level and corrosion damage depth of each existing component are quantitatively evaluated. The thermal performance degradation and mechanical performance decay are quantitatively evaluated simultaneously, so that the physical state of the component can be fully diagnosed. The corrected geometric profile, the assessed residual mechanical performance level, and the corrosion damage depth are used as updated attributes for each existing component. These are written into the component attribute fields corresponding to the initial digital twin, forming an updated digital twin whose geometric and physical states are consistent with the on-site measured data. The updated twin fully reflects the true geometric form and physical state of the existing building, providing a reliable data foundation for subsequent decision-making.

[0008] Preferably, in step S2, the process of generating the dynamic dataset of existing components is as follows: By extracting entity types, entity attributes, and inter-entity relationships in the decoration and renovation field from national building standards and specifications, enterprise process standard databases, and historical existing building renovation projects, a knowledge graph of decoration engineering covering component categories, material properties, connection methods, failure modes, and repair processes is constructed. This unified and structured storage of multi-source heterogeneous knowledge enables the systematic expression of decoration engineering knowledge. The known geometric, physical, and damage attributes of each existing component in the dynamic dataset of existing components are used as the initial features of the nodes and input into the message passing layer of the graph neural network. After multiple rounds of neighborhood aggregation and node state updates, the missing hidden engineering attribute values ​​and their hidden association types with other existing components are inferred and predicted. The hidden engineering attributes and associations are accurately inferred and completed, effectively filling the information blind spots. The missing attribute values ​​and hidden association types obtained by graph neural network inference prediction are backfilled into the null fields corresponding to the existing component dynamic dataset, generating an existing component dynamic dataset containing complete geometric, physical, damage and hidden engineering semantic association features. This allows the existing component data to go from incomplete to complete, providing a semantically complete data foundation for subsequent decision-making.

[0009] Preferably, in step S3, the process of generating an initial elite population using existing component dynamic data as input is as follows: We collect case data from multiple historical projects in the field of existing building renovation and reconstruction, including the value orientation labels, renovation scale, structural type, corresponding AHP weight allocation scheme and final implementation effect score of each project. We construct a training dataset containing samples of multiple types of projects, and use the feature parameters of each project in the training dataset as the state input, the weight allocation scheme of the AHP criterion layer of the project as the action output, and the final implementation effect score of the project as the reward value. We use a deep Q network for offline training to generate a reinforcement learning agent that can map project features to the optimal weights. This systematically learns and solidifies historical project experience to form a reusable weight inference capability. The differentiated value orientation of the current existing building renovation and renovation project, as well as the parameters of project scale and structural type, are input into the trained reinforcement learning agent. The agent outputs the weight combination of each criterion layer of AHP adapted to the current project through forward inference. The fixed weights in the traditional AHP are replaced by the weight combination of each criterion layer of AHP. The differentiated retention value of each existing component in the dynamic dataset of existing components is comprehensively scored, so that the scoring criteria are automatically adapted with the project value orientation, avoiding the evaluation bias caused by fixed weights. Existing components with comprehensive scores higher than a preset threshold are selected to form a high-value retention candidate set. The demolition measures corresponding to each existing component in the high-value retention candidate set are encoded into binary gene sequences. The feasibility of the gene sequences is verified in combination with on-site construction space constraints and safety regulations. The gene sequences that pass the verification are combined to generate an initial elite population. High-value components are accurately selected and encoded as feasible solutions, and the optimization starting point is directly aimed at high-quality solutions that can be implemented.

[0010] Preferably, in step S4, the multi-objective iterative optimization process is as follows: A multi-objective fitness function is constructed with the total carbon emissions throughout the entire life cycle of renovation and renovation of existing buildings, the comprehensive renovation cost, and the total construction period as the three optimization objectives. The spatial constraints in the updated digital twin are encoded as penalty function terms and added to the fitness function of each objective. The three objectives of carbon emissions, cost, and construction period are uniformly quantified, so that spatial constraints are directly integrated into the starting point of optimization. The initial elite population is input into the improved NSGA-II algorithm. Fitness evaluation is performed on each chromosome in the population. Based on the three target values ​​of each chromosome, a fast non-dominated sort is performed. At the same time, the crowding distance of each chromosome in the same non-dominated layer is calculated. Parent individuals are selected with an elite retention strategy that prioritizes non-dominated level and then crowding distance. Excellent individuals are selected layer by layer according to comprehensive performance, so that the diversity and convergence of the population are guaranteed simultaneously. Perform simulated binary crossover operation based on adaptive crossover probability and polynomial mutation operation based on adaptive mutation probability on the selected parent individuals to generate offspring population. After merging the parent and offspring, repeat non-dominated sorting and crowding distance screening. Iterate until the termination condition is met, and output Pareto optimal dismantling scheme set. The crossover and mutation are dynamically adjusted according to the evolutionary state. When the iteration terminates, output a set of non-dominated high-quality schemes for subsequent screening.

[0011] Preferably, the adjustment process for the adaptive crossover probability and the adaptive mutation probability is as follows: Set upper and lower limits for crossover probability. Calculate the crossover probability of the current generation in a non-linear decreasing manner based on the ratio of the current iteration generation to the preset maximum iteration generation. This allows the NSGA-II algorithm to maintain a high crossover rate in the early stages of iteration to explore the solution space in a wide area, and to reduce the crossover rate in the later stages of iteration to protect the superior gene structure. The crossover rate is automatically adjusted as the iteration progresses. The average crowding distance of the first non-dominant frontier individual in the current generation is calculated and the ratio of it to the historical maximum average crowding distance is used as a population diversity index. The mutation probability of the current generation is dynamically adjusted with the population diversity index as feedback. When the population diversity decreases, the mutation rate is adaptively increased to introduce new genes, and when the diversity is sufficient, the mutation rate is kept low to avoid the degradation of excellent solutions. The mutation intensity is also adaptively adjusted in real time with the population diversity. The calculated crossover and mutation probabilities of the current generation are applied to the crossover and mutation operations of the current generation, respectively. At the end of each iteration, the historical maximum average crowding distance is updated to achieve adaptive adjustment of the crossover and mutation parameters with the evolution state of the algorithm. The crossover and mutation parameters are adaptively adjusted in tandem with the evolution state, thereby improving the optimization efficiency and solution quality of the algorithm.

[0012] Preferably, in step S5, the calculation process for the pasting progress is as follows: Each Pareto-optimal demolition scheme is mapped back to the updated digital twin, and a 4D construction process simulation is performed. The number of spatial conflicts and the frequency of process interference in the construction simulation of each scheme are counted hourly. At the same time, the instantaneous carbon emission values ​​in each simulation period are recorded and their standard deviation is calculated. Each scheme undergoes a complete construction process simulation in a virtual environment, exposing potential conflicts and emission fluctuations in advance. The total carbon emissions, comprehensive renovation costs, and total construction period of each scheme are used as positive evaluation indicators. The total frequency of spatial conflicts and process interference is used as the construction conflict penalty indicator. The standard deviation of carbon emissions is used as the fluctuation penalty indicator. Together, an improved TOPSIS evaluation matrix is ​​constructed. Positive benefits and negative risks are uniformly included in the evaluation framework, so that the merits and demerits of the schemes can be comprehensively measured. The improved TOPSIS evaluation matrix is ​​vector normalized. Based on the positive and negative ideal solutions of each objective, the weighted Euclidean distance between each scheme and the positive ideal solution and the negative ideal solution are calculated. Then, the tracking progress of each scheme relative to the ideal solution is calculated, and the distance between each scheme and the ideal state is accurately quantified. Based on the tracking progress, a single quantitative basis is provided for the sorting of schemes.

[0013] Preferably, the process of constructing the improved TOPSIS evaluation matrix and calculating the patching progress is as follows: By setting the construction conflict penalty index and the fluctuation penalty index as negative attributes, the characteristic that the larger the value of the scheme is, the higher the risk of implementation is embedded in the normalization process of the evaluation matrix. This makes the magnitude of the negative attribute value after normalization inversely correlated with the evaluation quality, directly transforming conflict and fluctuation risks into evaluation disadvantages, so that high-risk schemes are at a disadvantage from the beginning of the evaluation. The adaptive AHP weight combination output by the reinforcement learning agent is linearly mapped to the weight coefficients of each evaluation objective. The attribute values ​​in the normalized evaluation matrix are weighted, so that the differentiated value orientation of the project runs through the entire evaluation process of the improved TOPSIS. The project value orientation is extended from weight setting to the entire evaluation process, ensuring that the evaluation results are internally consistent with the project objectives. Based on the weighted evaluation matrix, the positive and negative ideal solutions for each attribute are determined. The weighted Euclidean distance between each scheme and the positive ideal solution and the weighted Euclidean distance between each scheme and the negative ideal solution are calculated. The ratio of the two is defined as the tracking progress. The tracking progress comprehensively reflects the distance relationship between the scheme and the positive and negative ideal solutions, and supports the sorting and selection of schemes with a single indicator.

[0014] Preferably, in step S6, the process of uniquely identifying the optimal renovation plan for an existing building is as follows: All Pareto optimal demolition schemes in the set are sorted in descending order of their demolition progress values ​​to form an initial priority sequence. Each scheme is then clearly ranked according to its overall performance to ensure that the optimization results are easy to screen in the future. Starting from the head of the initial priority sequence, each item is traversed and inferior solutions are eliminated if the construction conflict penalty index exceeds the project's preset upper limit, the carbon emission fluctuation standard deviation exceeds the allowable threshold, or the progress is lower than the set lower limit. A candidate solution sequence after double risk filtering is generated, and solutions with excessive conflict, excessive emission fluctuation, and low progress are eliminated one by one, which significantly improves the quality of the candidate sequence. The scheme with the highest progress in the candidate scheme sequence after risk filtering is selected as the final and only optimal renovation scheme for the existing building renovation. The construction sequence, resource scheduling plan and carbon emission control strategy of the scheme are written into the execution layer of the updated digital twin to form a traceable closed-loop decision result. The construction sequence, resource plan and carbon emission control strategy of the optimal scheme are fully written into the execution layer, and the decision result is directly executed and traced in the twin.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This multi-objective intelligent optimization decision-making method for the renovation and renovation of existing buildings constructs a digital twin that can continuously update and accurately reflect the geometric shape and internal damage state of existing buildings through deep fusion and dynamic mapping of multi-source heterogeneous sensing data. This achieves high-fidelity dynamic reconstruction and accurate perception of existing building information, overcoming the defects of static distortion in traditional BIM models.

[0016] 2. This multi-objective intelligent optimization decision-making method for the renovation and renovation of existing buildings constructs a knowledge graph of decoration engineering and uses graph neural networks to reason and complete the attributes of existing components and the relationships between components. This achieves intelligent reasoning and complete completion of the semantic features of hidden works, solves the problem of incomplete information caused by missing data of existing buildings, and enhances the integrity of the dataset and the robustness of supporting decision-making.

[0017] 3. This multi-objective intelligent optimization decision-making method for the renovation and renovation of existing buildings utilizes a reinforcement learning agent to dynamically generate optimal evaluation weights based on the project's differentiated preferences, thereby achieving a precise response to the project's core demands. Based on this, it efficiently selects high-value retention objects from a massive number of components, generating high-quality initial solutions, and realizing adaptive analysis and strategy generation with a differentiated value orientation for the project. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the workflow of a multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to the present invention; Figure 2 This is a flowchart of the optimization calculation process of the improved non-dominated sorting genetic algorithm II (NSGA-II) of this invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings, comprising the following steps: S1. Collect geometric point clouds and continuous temporal infrared thermal images of existing building renovation targets. After registration and fusion, generate an initial digital twin, achieving synchronous digital mapping of the existing building's geometry and temperature. Use intelligent sensors (ground penetrating radar) to detect hidden defects and update the geometric and physical state of the initial digital twin, transforming invisible defects into locatable and quantifiable digital objects. A 3D laser scanner is used to scan the existing building renovation targets from multiple stations and angles, acquiring geometric point cloud data covering the entire space. Simultaneously, an infrared thermal imager continuously collects temporal thermal image sequences of the same spatial area at preset time intervals. The geometric shape and surface temperature field of the existing building are captured synchronously, significantly improving data integrity. The geometric point cloud data and temporal infrared thermal image sequences are registered and fused. Using the point cloud spatial coordinates as a reference, an iterative nearest-point algorithm is used to spatially map the thermal image pixels. Then, spatiotemporal interpolation is used to generate a digital twin that combines three-dimensional geometric contours and dynamic surface temperature field data. The initial digital twin of the building is precisely aligned with geometric and temperature data under a unified spatial benchmark, resulting in richer information dimensions. In the initial digital twin, temperature anomaly areas on the surface of each existing component are used as candidate identifiers for hidden defects. Smart sensors deployed in the target area of ​​the existing building renovation project detect the underground or internal layer echo signals to verify and locate the candidate identifiers, generating a hidden defect distribution map and mapping it to the corresponding spatial coordinates of the initial digital twin. The smart sensor is a ground-penetrating radar, and the surface temperature anomalies and internal echo signals corroborate each other, significantly enhancing the reliability of hidden defect location. Based on the hidden defect distribution map, the geometric accuracy parameters and physical state attributes of each existing component in the initial digital twin are updated to form an updated digital twin that can dynamically reflect the true geometric shape and internal damage status of the existing building. The true geometric shape and internal damage status of the existing building are completely reproduced in digital space, greatly improving the fidelity of the twin. In this embodiment, a control network is deployed within the target area of ​​the existing building renovation project, with the spacing between control points controlled between 10m and 15m. A Leica RTC360 3D laser scanner is used to perform multi-angle scanning at each station at three horizontal azimuth angles (0°, 120°, and 240°) and three vertical elevation angles (-45°, 0°, and +45°). The overlap rate between adjacent stations is no less than 30%, acquiring geometric point cloud data covering the entire building space. The scanning time for each station is controlled between 2 and 3 minutes, and the effective scanning radius of a single station does not exceed 50m. FLIR scanning is also used simultaneously. The T1040 infrared thermal imager continuously acquires temporal thermal image sequences of the same spatial area at 15-minute intervals from 6:00 AM to 6:00 PM daily. Emissivity is set according to the surface material of each component: 0.92 for concrete, 0.18 to 0.25 for metal, 0.85 to 0.90 for wood, and 0.90 to 0.95 for ceramic tile. Thermal image data from each measurement is stored in R-JPEG format, and the ambient temperature and relative humidity at the time of acquisition are recorded simultaneously. The geometric point cloud data acquired from each station is imported into Cyclone. The 360Plus software platform was used for stitching and registration. Using the coordinates of the first station as the global reference, an iterative nearest-point algorithm was employed for automatic registration of point cloud data from multiple stations. The iteration termination condition was set to a root mean square error of less than 2mm, and the maximum number of iterations was set to 200. After registration, the point cloud data was uniformly thinned and filtered at a sampling interval of 0.02m to remove noise and outliers, generating a three-dimensional geometric contour model of the building. The time-series infrared thermal image sequence was then imported into FLIR. The Tools analysis platform uses point cloud spatial coordinates as the spatial reference and selects at least six uniformly distributed feature corner points in each thermal image frame as control points. It uses an iterative nearest-point algorithm to establish a homography transformation matrix between the thermal image pixel coordinate system and the point cloud spatial coordinate system, mapping each pixel of the thermal image to its corresponding spatial coordinate in the point cloud. Then, it uses cubic spline interpolation to perform temporal interpolation for missing moments between adjacent thermal image frames, and inverse distance weighted interpolation to perform spatial domain fusion for overlapping areas between adjacent stations. This generates a digital twin of the dynamic temperature change of each component surface over time, with a spatial resolution of 0.05m × 0.05m and a temporal resolution of one complete temperature field data set every 15 minutes. From the initial digital twin, it extracts the surface temperature distribution data of each existing component. Areas where the temperature difference between different regions on the same component surface exceeds 3.0℃ and lasts for more than 60 minutes are identified as temperature anomaly areas and used as candidate identifiers for hidden defects. The spatial coordinates and boundary range of each candidate identifier are recorded, and GSSI is deployed at the corresponding locations of each candidate identifier. The SIR-4000 ground-penetrating radar is equipped with shielded antennas with center frequencies of 400MHz and 900MHz. The 400MHz antenna has a detection depth set from 1.0m to 4m.The detection depth of the 900MHz antenna was set to 0.3m to 1.0m, the measurement line spacing was set to 0.10m, the number of sampling points was set to 512 points / s, and the scanning rate was set to 100 scans / s. Automatic gain control and background removal filtering were used to preprocess the echo signal. The two-way travel time and reflection intensity values ​​of the reflection layer at each measurement point were extracted. Based on the propagation speed of electromagnetic waves in concrete (0.10m / ns to 0.12m / ns), the two-way travel time was converted into defect depth. Anomaly zones with reflection intensity values ​​lower than 1 / 3 of the average background reflection intensity and a duration of not less than 0.05m were identified as verified hidden defects. A hidden defect distribution map containing defect type, spatial boundary coordinates, and depth values ​​was generated and mapped one by one to the corresponding spatial coordinate positions of the initial digital twin. Based on this, the process of updating the geometry and physical state of the initial digital twin is as follows: For each defect in the hidden defect distribution map, its spatial boundary coordinates and defect depth values ​​are extracted to correct the geometric contour accuracy of the corresponding existing components in the initial digital twin. At the same time, for existing components with material spalling or cross-sectional reduction, negative compensation adjustments are made to the geometric dimensions. The geometric deviations caused by defects are corrected one by one, so that the geometric accuracy of the twin is highly consistent with that of the real components. Based on the rate of change of surface temperature and temperature difference amplitude of each existing component in the time-series thermal imaging sequence, the attenuation degree of its thermal inertia index and thermal conductivity is inversely deduced, combined with the ground-penetrating radar echo signal. The variation characteristics of reflection intensity in the image are used to quantitatively assess the remaining mechanical performance level and corrosion damage depth of each existing component. The degradation of thermal performance and the decay of mechanical performance are simultaneously quantitatively assessed, so that the physical state of the components can be fully diagnosed. The corrected geometric profile, the assessed remaining mechanical performance level and corrosion damage depth are used as the updated attributes of each existing component and written into the component attribute fields corresponding to the initial digital twin. This forms an updated digital twin whose geometric and physical states are consistent with the on-site measured data. The updated twin fully reflects the true geometric form and physical state of the existing building, providing a reliable data foundation for subsequent decision-making. In this embodiment, each defect record in the hidden defect distribution map is read, and the spatial boundary coordinate set and defect depth value corresponding to the defect are extracted. The spatial boundary coordinate set is a closed area defined by the three-dimensional coordinates of the defect center point and the expansion radius of the defect in the horizontal and vertical directions. The defect depth value is the maximum reflection layer conversion depth of the ground-penetrating radar line within this area. Based on the spatial boundary coordinates, the set of mesh nodes affected by the defect is retrieved and located in the component geometric mesh of the initial digital twin. For material spalling defects, the retrieved mesh nodes are offset inward along the normal direction of the component surface where the node is located. The offset amount is 0.85 times the ground-penetrating radar measured defect depth value, and the offset is made at the defect edge. A transition zone with a thickness gradually increasing from zero to the maximum offset is set in the edge region. The transition zone width is taken as the range extending outward from the defect boundary by 0.08m to avoid abrupt changes in the corrected geometric surface. For cross-section reduction defects, the original cross-sectional dimensions of the components within the defect coverage area are extracted. The width or height parameter value of the cross section is reduced in the component modeling parameters according to the measured reduction amount. The reduction amount is taken as the arithmetic mean of the defect depth measured by ground penetrating radar at different measuring points of the cross section. After each geometric correction operation is completed, the corrected mesh node coordinates and cross-sectional parameters are temporarily stored in the memory cache. After all defects are processed, the original geometric data in the initial digital twin is replaced uniformly. The geometric correction accuracy is controlled within ±2mm. For each existing component with hidden defects, the measured temperature values ​​of the component surface during the daily period from 6:00 to 18:00 are extracted from the time-series thermal imaging. The temperature change rate within that period is calculated by dividing the temperature difference between two adjacent measurements by the 15-minute time interval. The temperature difference between the abnormal temperature area and the adjacent normal area on the same component surface at the same time is used as the temperature difference amplitude. The temperature change rate and temperature difference amplitude are input into a thermal parameter inversion model pre-stored in a database. This thermal parameter inversion model contains a library of component surface temperature response curves under different thermal inertia indices and thermal conductivity. By comparing the measured temperature response curves with the preset curves in the curve library hourly, the appropriate curve is selected. The thermal inertia index and thermal conductivity corresponding to the curve with the smallest residual are taken as the current estimated values ​​of the component. Then, the relative deviation between the estimated value and the initial factory standard value of the component material is taken as the degree of thermal conductivity attenuation. At the same time, the ground-penetrating radar echo signal at the corresponding measuring point of the component is retrieved, and the reflection intensity value of each reflection layer is extracted. Taking the reflection intensity of the first layer as the benchmark, the layer-by-layer attenuation rate of each deep layer reflection intensity relative to the reflection intensity of the first layer is calculated. Layers with an attenuation rate greater than 40% are identified as mechanical property degradation layers. According to the proportion of the degradation layer to the total detection depth, the remaining mechanical property level is divided into level one, level two and level three. The corrosion damage depth is taken as the difference between the converted depth value corresponding to the bottom of the degradation layer and the converted depth value of the top of the defect.Each existing component that has undergone geometric correction and physical evaluation is assigned a unique component identification code. This component identification code maintains a one-to-one correspondence with the component ID field in the initial digital twin. The corrected set of mesh node coordinates is written into the geometric attribute field corresponding to the component ID in the form of a point cloud coordinate array, overwriting the original coordinate data in that field. At the same time, the corrected cross-sectional width and cross-sectional height parameters are written into the cross-sectional parameter subfield of the component. The remaining mechanical performance level is written into the mechanical performance subfield under the physical state attribute field in the form of enumerated values. Level 1 corresponds to a remaining mechanical performance of not less than 85% of the original design value, Level 2 corresponds to a remaining mechanical performance of not less than 85% of the original design value, and Level 3 corresponds to a remaining mechanical performance of not less than 85% of the original design value. The remaining mechanical properties are between 60% and 85% of the original design values. Level 3 corresponds to remaining mechanical properties lower than 60% of the original design values. Corrosion damage depth, measured in millimeters, is written to the damage depth subfield under the physical state attribute field. After all attribute fields are written, a write record containing the component identification code, correction timestamp, and all updated attribute values ​​is generated and stored in the historical version table of the project database. Once all existing component update attributes have been written, a refresh command is sent to the digital twin engine to reload the geometric and physical attribute data of each component and perform model reconstruction, generating an updated digital twin whose geometric and physical states are consistent with the field measurement data.

[0021] S2. Construct a knowledge graph for decoration engineering. Use graph neural networks to infer and complete missing semantic relationships in hidden works, forming a reasonable network of relationships from scattered hidden works information. This generates a dynamic dataset of existing components containing complete semantic features, providing a semantically complete and clearly correlated data foundation for subsequent decision-making. Extract entity types, entity attributes, and inter-entity relationships from national building standards and specifications, enterprise process standard libraries, and historical existing building renovation projects. Construct a knowledge graph for decoration engineering covering component categories, material properties, connection methods, failure modes, and repair processes. This unifies and structures the storage of multi-source heterogeneous knowledge, enabling a systematic expression of decoration engineering knowledge and transforming the existing component dynamic dataset into a coherent whole. The known geometric, physical, and damage attributes of each existing component are used as initial features of the nodes and input into the message passing layer of the graph neural network. After multiple rounds of neighborhood aggregation and node state updates, the missing hidden engineering attribute values ​​and their hidden association types with other existing components are inferred and predicted. The hidden engineering attributes and associations are accurately inferred and completed, effectively filling information blind spots. The missing attribute values ​​and hidden association types obtained by the graph neural network inference and prediction are backfilled into the corresponding null fields of the existing component dynamic dataset, generating an existing component dynamic dataset containing complete geometric, physical, damage, and hidden engineering semantic association features. This makes the existing component data go from incomplete to complete, providing a semantically complete data foundation for subsequent decision-making. In this embodiment, based on the current building decoration and renovation engineering quality acceptance specifications and the industry-standard classification system for decoration and renovation sub-projects, major components such as light steel keel partition walls, stone veneer, metal ceilings, tile laying, wood veneer, and glass partitions are extracted as entity type nodes. Each type of component is bound with entity attribute fields such as material density, thermal conductivity, compressive strength, fire rating, design service life, and surface treatment method. Based on the construction practices in the current construction quality acceptance specifications and typical node detail drawings, standard connection relationships such as support, anchoring, bonding, hanging, caulking, and snap-fit ​​connections between components are extracted. By collecting process records and completion acceptance data of existing building renovation projects completed in the past five years, typical failure modes such as hollowing of the decorative layer, keel corrosion, peeling of the adhesive layer, aging of sealant, and cracking of the decorative panel were extracted, along with corresponding repair processes such as replacement repair, grouting reinforcement, partial replacement, full layer redo, and surface sealing. Entities, attributes, and relationships were structured and stored according to the resource description framework triplet format. The Neo4j graph database was used as the storage engine to complete the physical database construction of the decoration engineering knowledge graph, forming a knowledge representation system covering all elements of decoration and renovation engineering. The graph neural network model was set as a three-layer graph attention network. The network architecture uses an input layer with the same dimension as the number of known attribute fields in the existing dynamic dataset of components, resulting in an initial feature vector of 128 dimensions per node. The hidden layer dimension is set to 256 dimensions, and the output layer dimension is set to 192 dimensions, equal to the total number of missing attributes to be predicted. During model training, the complete component attribute triples stored in the decoration engineering knowledge graph are used as supervision labels. 80% of the node data in the graph is divided into a training set, 10% into a validation set, and 10% into a test set. The Adam optimizer is used for parameter learning, with an initial learning rate set to 0.001. After each iteration, the mean absolute error on the validation set is used as the training set. The early stopping mechanism monitors the following: when the validation set error fails to decrease for 15 consecutive training epochs, training is terminated and the current model parameters are saved. The message passing layer performs three rounds of neighborhood aggregation. In each round of aggregation, each node sends a message packet containing its own feature vector to its neighboring nodes. The receiving node sums all the received neighbor message packets with attention weights, performs a residual connection with its current features, and updates the output features for this round using the ReLU activation function. After three rounds of aggregation, the missing attribute prediction vector of each node is output, along with the probability value of a hidden association between the node and other nodes in the graph, where the probability value exceeds 0.A 75-fold association pair is determined to be a valid hidden association type. The 192-dimensional prediction vectors corresponding to each existing component node in the graph neural network output layer are read. According to the pre-established feature dimension and attribute field mapping table, the value of each dimension is written into the null value field of the corresponding component record in the existing component dynamic dataset. For continuous attribute fields, the predicted floating-point value is directly written; for categorical attribute fields, the category label corresponding to the highest probability value in the prediction vector is used. For hidden association types, based on the node association probability matrix output by the graph neural network, for association pairs with a probability value exceeding 0.75, a hidden association list field is added to the existing component dynamic dataset. The target node ID and association type name of the associated pair are appended to the list field as key-value pairs. After each component record is populated, it is checked whether any fields in the record are still empty or have values ​​outside the reasonable range. For records that fail the check, the record is marked as awaiting manual review and exported separately to the review checklist file. Records that pass the check directly overwrite the original records in the existing component dynamic dataset. After all component records have been populated and checked, the updated complete dataset is exported in JSON format and stored in the component data set of the project database. The total number of records in the dataset remains consistent with the initial existing component dynamic dataset.

[0022] S3. Based on reinforcement learning agents, the system analyzes the differentiated value orientation of projects and dynamically adjusts the AHP evaluation weights according to project preferences. This ensures that the evaluation criteria automatically adapt to the project's value orientation rather than being rigidly fixed. Using existing component dynamic data as input, an initial elite population is generated, naturally carrying project preference characteristics in the initial solution set, thus improving the quality of the optimization starting point. Case data from multiple historical projects in the field of existing building renovation and reconstruction are collected, including each project's value orientation labels, renovation scale, structural type, corresponding AHP weight allocation scheme, and final implementation effect score. A training dataset containing multiple types of project samples is constructed. The feature parameters of each project in the training dataset are used as state input, the AHP criterion layer weight allocation scheme of the project is used as action output, and the final implementation effect score of the project is used as reward value. A deep Q-network is used for offline training to generate a reinforcement learning agent that can map project features to optimal weights. This systematically learns and solidifies historical project experience, forming... The system enables reusable weighted reasoning. It inputs the differentiated value orientation, project scale, and structural type parameters of existing building renovation projects into the trained reinforcement learning agent. The agent then outputs a weight combination of each criterion layer of the AHP (Advanced Hierarchical Process) adapted to the current project through forward reasoning. This combination replaces the fixed weights in the traditional AHP, and performs a comprehensive evaluation of the differentiated retention value of each existing component in the dynamic dataset of existing components. This allows the evaluation criteria to automatically adapt to the project's value orientation, avoiding evaluation bias caused by fixed weights. Existing components with comprehensive scores higher than a preset threshold are selected to form a high-value retention candidate set. The demolition measures corresponding to each existing component in the high-value retention candidate set are encoded into binary gene sequences. The feasibility of the gene sequences is verified by combining on-site construction space constraints and safety regulations. The gene sequences that pass the verification are combined to generate an initial elite population, accurately selecting and encoding high-value components as feasible solutions. The optimization starting point directly targets high-quality solutions that can be implemented. It should be noted that during the construction of the training dataset, the case data of each historical project were standardized and coded. Value-oriented labels were represented as one-hot vectors according to four types: economic priority, construction period priority, carbon emission priority, and comprehensive balance. The scale of renovation was input as continuous features using building area and total number of floors. The structural type was coded as discrete features according to four categories: frame structure, shear wall structure, frame-shear wall structure, and steel structure. In the AHP criterion layer weight allocation scheme, the weight values ​​of the three criterion layers—physical state and integrity, material quality and performance, and reuse potential—were all recorded as floating-point numbers between 0 and 1, and the sum of the three was normalized to equal 1. The final implementation effect score was determined by the owner, supervisor, and third party after project completion. The comprehensive score, jointly assessed by multiple evaluation agencies, is out of 100. All feature parameters, weight allocation schemes, and performance scores are associated by project ID and stored as a structured dataset. The training sample consists of no fewer than 200 historical projects, covering four building types: residential, office, hotel, and commercial complexes. The deep Q-network employs a three-layer fully connected neural network structure. The number of neurons in the input layer matches the total dimension of the feature parameters, the number of neurons in the hidden layer is set to 256, and the number of neurons in the output layer matches the size of the discrete action space of the AHP criterion layer weight combination. The action space is constructed by discretizing the weights of each criterion layer with a step size of 0.05, generating predefined weight combinations as the optional action set. During training, a set of predefined weight combinations is used. The system employs a replay mechanism with a replay buffer capacity of 10,000 transfer samples. During each training session, 128 samples are randomly selected for mini-batch gradient descent. The target network update frequency is set to synchronize every 200 steps. The discount factor is set to 0.95. The exploration rate, starting from an initial value of 1.0, decays exponentially with the number of training steps to a final value of 0.01. The decay rate is set to decrease to 0.95 times the current value every 1000 steps. When the exploration rate drops below 0.01, the decay stops and is fixed at 0.01. Training termination conditions are set as follows: the average reward value fluctuates within ±2 points for 500 consecutive steps, or the total training steps reach 50,000 steps, at which point the system forcibly terminates and saves the current network parameters. During the agent's forward inference phase, the differentiated value of the current item is... Value-guided labels are converted into one-hot vectors according to the same encoding rules as the training dataset. The building area and total number of floors in the project scale parameters are normalized. The structure type encoding vector is concatenated with the above features to form a complete 152-dimensional input feature vector. This feature vector is forward-propagated through a trained deep Q-network. Each neuron in the network output layer corresponds to a predefined discrete weight combination action. The weight combination corresponding to the neuron with the largest Q value in the output layer is selected as the optimal AHP criterion layer weight combination for the current project. In this combination, the weights for physical state and integrity are between 0.30 and 0.55, the weights for material quality and performance are between 0.25 and 0.45, and the weight for utilization potential is 0.Between 15 and 0.35, the sum of the three values ​​equals 1. After obtaining the adaptive weight combination, for each component record in the existing component dynamic dataset, the following four index values ​​are read: crack level score, deformation rate score, corrosion degree score, and node connection reliability score under the physical state and integrity dimension; material durability score, metal material priority score, and special material value score under the material quality and performance dimension; and new scenario adaptability score and transportation cost ratio score under the reuse potential dimension. The score values ​​of each index within each dimension are multiplied by the corresponding index's pre-calculation value. After assigning weights, the scores for each dimension are summed. Then, each dimension score is multiplied by its corresponding adaptive weight from the criterion layer, and summed again to obtain the comprehensive retention value score for the component. The comprehensive score has a maximum of 100 points. The threshold is set to 70 points for economic priority, 65 points for carbon emission priority, 75 points for schedule priority, and 70 points for comprehensive balance. The filtering operation iterates through all records in the existing component dynamic dataset, extracting component records whose comprehensive scores reach or exceed the corresponding thresholds and adding them to the high-value retention candidate set. During the coding phase, ... For each component in the candidate set, a corresponding binary gene fragment is generated based on its component type and available dismantling methods. The length of the gene fragment is determined by the number of available dismantling measures for that component. Each bit in the fragment corresponds to a dismantling measure; a value of 1 indicates that the measure is used, and a value of 0 indicates that it is not used. Only one bit in the gene fragment of the same component is set to 1 to ensure that each type of component selects only one dismantling measure at any given time. The gene fragments of all components are sequentially assembled according to the order of the components in the candidate set to form a complete binary gene sequence. The total length of this sequence is equal to the sum of the number of available dismantling measures for all components in the candidate set. During the feasibility verification phase, each gene sequence is decoded into a specific combination of dismantling measures. Each sequence is then checked against the minimum working space requirements, maximum lifting weight limit, construction passage width conditions, and safety regulations regarding edge protection, adjacent component protection, and dust control. Gene sequences that do not meet any of these constraints are directly marked as infeasible and removed. Gene sequences that pass the feasibility verification are retained as valid individuals. The collection of all valid individuals constitutes the initial elite population, with a population size controlled between 50 and 80 individuals.

[0023] S4. Input the initial elite population into the improved NSGA-II algorithm, and perform multi-objective iterative optimization with carbon emissions, cost and construction period as optimization objectives. Automatically seek the optimal balance point among the three conflicting objectives, and quickly output a set of Pareto optimal demolition schemes for renovation and renovation of existing buildings, providing a set of high-quality alternative schemes that do not dominate each other for subsequent screening.

[0024] S5. An improved TOPSIS algorithm (superior-inferior solution distance method) is introduced. Taking the Pareto optimal demolition scheme set as the object, the frequency of construction conflicts and the standard deviation of carbon emission fluctuations in the digital twin of each scheme are used as negative penalty indicators. This allows the conflict risk and emission instability risk during the construction process to be quantified and included in the evaluation. The schedule is calculated, and a single quantitative indicator is used to comprehensively reflect the superiority and inferiority of the scheme and the level of risk.

[0025] S6. Based on the schedule, sort the Pareto solutions for the renovation of existing buildings, eliminate inferior solutions with high risk and high volatility, so that the final solution meets the dual criteria of high efficiency and low risk, and uniquely lock in the optimal renovation solution for existing buildings. This ensures that the selected solution has the lowest implementation risk and the most stable comprehensive benefits under multiple conflicting objectives, so that the decision result has clear uniqueness and feasibility, and avoids implementation deviations caused by fuzzy selection.

[0026] Example 2, as Figure 1 , Figure 2 As shown, based on Example 1, this invention provides a technical solution: In S4, the multi-objective iterative optimization process is as follows: A multi-objective fitness function is constructed with the total carbon emissions throughout the entire lifecycle of existing building renovation, comprehensive renovation costs, and total construction period as the three optimization objectives. The spatial constraints in the updated digital twin are encoded as penalty function terms and added to the fitness function of each objective. The three objectives of carbon emissions, cost, and construction period are uniformly quantified, allowing spatial constraints to be directly integrated into the optimization starting point. The initial elite population is input into the improved NSGA-II algorithm, and fitness evaluation is performed on each chromosome in the population. A fast non-dominated sorting is then performed based on the three objective values ​​of each chromosome. The crowding distance of each chromosome within the same non-dominated layer is calculated. Parent individuals are selected using an elite retention strategy that prioritizes non-dominated level and then crowding distance. Excellent individuals are selected layer by layer based on comprehensive performance, ensuring both population diversity and convergence. Simulated binary crossover based on adaptive crossover probability and polynomial mutation based on adaptive mutation probability are performed on the selected parent individuals to generate offspring population. After merging the parent and offspring, the non-dominated sorting and crowding distance selection are repeated until the termination condition is met. The Pareto optimal dismantling scheme set is output. Crossover and mutation are dynamically adjusted according to the evolutionary state. When the iteration terminates, a set of non-dominated high-quality schemes are output for subsequent selection. In this embodiment, the total carbon emissions throughout the entire life cycle are calculated using the sum of the carbon dioxide equivalents generated during the building material production, construction, transportation, and demolition stages as the target value of the fitness function. Specifically, carbon emissions during the building material production stage are determined by multiplying the production carbon emission factor of each material in the existing component dynamic dataset by its usage. Carbon emissions during the construction stage are calculated based on the energy consumption of construction machinery shifts and the corresponding carbon emission factor. Carbon emissions during the transportation stage are determined based on the transportation distance of building materials and waste and the energy consumption intensity of transportation vehicles. Carbon emissions during the demolition stage are calculated based on the energy consumption of the demolition process and the carbon emission factor corresponding to the waste disposal method. The comprehensive renovation cost consists of four items: material procurement cost, labor cost, machinery rental fee, and waste disposal cost. The sum of these factors serves as the target value of the fitness function. All costs are calculated based on current market prices and the quantity data in the existing component dynamic dataset. The total construction period is calculated by multiplying the standard time quota for each demolition and renovation process by the corresponding quantity, and combining this with the logical relationships and overlap times between processes to obtain the total number of calendar days, which is then used as the target value of the fitness function. Spatial constraints are encoded as penalty function terms, specifically including minimum working space constraints, maximum lifting weight constraints, construction passage width constraints, and adjacent component protection constraints. If any constraint is not satisfied, a pre-set large positive penalty term is added to each target fitness value of the scheme, placing schemes that do not meet the constraints at a disadvantage in subsequent non-dominated ranking, thus guiding the algorithm to prioritize searching for schemes that meet the constraints. A feasible solution with spatial constraints; the initial elite population is set to 60 individuals, each individual representing a binary gene sequence, which, after decoding, corresponds to a complete set of demolition measures. During fitness assessment, the aforementioned multi-objective fitness function is called for each individual in the population to calculate three objective values: total life-cycle carbon emissions, comprehensive renovation cost, and total construction period. A penalty function term is appended to each objective value. Rapid non-dominated sorting divides the population into different non-dominated layers based on the three objective values. The first non-dominated layer contains all individuals not dominated by any other individual, the second non-dominated layer contains individuals not dominated by the remaining individuals after removing the first layer, and so on. Crowding distance calculation is performed within the same non-dominated layer. For each individual, the crowding distance is calculated as the sum of the distances between adjacent individuals in each target dimension. The crowding distance for boundary individuals is set to infinity. When selecting a parent individual, two individuals are randomly selected from the current population, and their non-dominance levels are compared. The individual with the lower level wins. If the levels are the same, the crowding distance is compared, and the individual with the larger distance wins. This process is repeated until the maximum number of parent individuals is selected, and the number of parent individuals is set to 50% of the population size. A simulated binary crossover operation is applied between two parent gene sequences. The crossover probability is adaptively adjusted according to the current iteration generation. When the iteration generation is in the first 30%, the crossover probability is set to 0.90 to promote wide-area exploration of the population. When the iteration generation is in the last 30%, the crossover probability is reduced to 0.45. To protect the superior gene structure, the crossover probability in the intermediate stage is determined by linear interpolation. After the crossover operation is completed, a polynomial mutation operation is performed on each position of the offspring gene sequence. The mutation probability is adaptively adjusted according to the diversity index of the current population. The diversity index is calculated as the ratio of the average crowding distance of the first non-dominated layer to the historical maximum average crowding distance. When this ratio is lower than 0.30, the mutation probability is increased to 0.15 to introduce new genes and prevent premature convergence. When this ratio is higher than 0.70, the mutation probability is reduced to 0.05 to avoid the degeneration of superior genes. The offspring population is then merged with the parent population. Subsequently, a fast non-dominated sorting and crowding distance calculation are performed on the merged population. Individuals with high non-dominated levels and large crowding distances are selected to form a new generation of parent populations, maintaining the same population size as the initial population. The iteration termination condition is set as follows: the maximum number of iterations reaches 200 generations, or the average target value change rate of the first non-dominated frontier is less than 1% within 30 consecutive generations. After terminating the iteration, all individuals belonging to the first non-dominated layer in the current population are output as the Pareto optimal demolition scheme set, and the schemes in the Pareto optimal demolition scheme set are non-dominated in terms of carbon emissions, cost, and construction period. Based on this, the adjustment process of adaptive crossover probability and adaptive mutation probability is as follows: set upper and lower limits for crossover probability, calculate the crossover probability of the current generation in a non-linear decreasing manner according to the ratio of the current iteration generation to the preset maximum iteration generation, so that the NSGA-II algorithm maintains a high crossover rate in the early stage of iteration to explore the solution space in a wide area, reduces the crossover rate in the later stage of iteration to protect the excellent gene structure, and automatically adjusts the crossover rate with the progress of iteration. Calculate the average crowding distance of the first non-dominant frontier individual in the current generation and use the ratio of it to the historical maximum average crowding distance as a population diversity index. Use the population diversity index as feedback to dynamically adjust the mutation probability of the current generation, so that when the population diversity decreases, the mutation rate is adaptively increased to introduce new genes, and when the diversity is sufficient, the mutation rate is kept low to avoid the degradation of excellent solutions. Moreover, the mutation intensity is adaptively adjusted in real time with the population diversity. The calculated crossover probability and mutation probability of the current generation are applied to the crossover operation and mutation operation of the current generation, respectively. At the end of each iteration, the historical maximum average crowding distance is updated to realize the adaptive adjustment of crossover and mutation parameters with the evolution state of the algorithm. The crossover and mutation parameters are adaptively adjusted in synergy with the evolution state, and the algorithm optimization efficiency and solution set quality are improved simultaneously. In this embodiment, the upper limit of the crossover probability is set to 0.90, and the lower limit is set to 0.45. Based on the ratio r of the current iteration generation to the preset maximum iteration generation, the crossover probability of the current generation is calculated in a non-linear decreasing manner. In the early stage of iteration, when r ≤ 0.30, the crossover probability remains at 0.90; in the later stage of iteration, when r ≥ 0.70, the crossover probability decreases to 0.45. The intermediate stage follows a quadratic function curve of r for a smooth transition, ensuring that the crossover probability remains at a high level in the early stage of iteration to promote wide-area exploration of the solution space by the population, and gradually decreases as iteration progresses to protect the discovered superior gene structures. The average crowding distance of the first non-dominated frontier individual in the current generation is calculated, and the historical maximum average crowding distance recorded from the algorithm's inception to the previous generation is retrieved. The ratio ρ between the two is used as a population diversity index. The closer the population diversity index value is to 1, the closer the current population diversity is to the historical optimum; the smaller the value, the more the population tends to converge or even fall into a local optimum. The mutation probability of the current generation is dynamically adjusted using ρ as a feedback quantity. When ρ is below 0.30, the population diversity is considered severely insufficient, and the mutation probability is set to 0.15 to significantly increase the mutation rate and introduce new genes. When ρ is above 0.70, the population diversity is considered sufficient, and the mutation probability is reduced to 0.05 to avoid the destruction of good solutions by the high mutation rate. The specific value of the mutation probability of the current generation is determined by linear interpolation in the intermediate region. The crossover probability and mutation probability calculated under the current iteration are applied to the simulated binary crossover operation and polynomial mutation operation of the current generation, respectively. In the crossover stage of each iteration, the crossover probability is used to determine whether the parent gene sequence has undergone crossover recombination, and in the mutation stage, the mutation probability is used to determine whether the gene positions in the offspring gene sequence have been flipped. After the population update and non-dominated sorting are completed in each iteration, the average crowding distance of the first non-dominated frontier of the current generation is calculated. If this value exceeds the historical maximum average crowding distance, the historical maximum average crowding distance is updated to the current value to ensure that the historical maximum average crowding distance always remains the peak value of all iterations.

[0027] In S5, the calculation process for the schedule is as follows: Each scheme in the Pareto optimal demolition scheme set is mapped back to the updated digital twin, a 4D construction process simulation is performed, and the number of spatial conflicts and the frequency of process interferences in the construction simulation of each scheme are counted hourly. At the same time, the instantaneous carbon emission values ​​in each simulation period are recorded and their standard deviations are calculated. Each scheme undergoes a complete construction process simulation in the virtual environment to expose potential conflicts and emission fluctuations in advance. The total carbon emissions over the entire life cycle of each scheme, the comprehensive renovation cost, and the total construction period are used as positive evaluation indicators, and the total frequency of spatial conflicts and process interferences is used as the evaluation metric. Using carbon emission standard deviation as the fluctuation penalty index and construction conflict penalty index as the construction conflict penalty index, an improved TOPSIS evaluation matrix is ​​constructed. Positive benefits and negative risks are uniformly incorporated into the evaluation framework, so that the merits of the schemes can be comprehensively measured. The improved TOPSIS evaluation matrix is ​​vector normalized. Based on the positive and negative ideal solutions of each objective, the weighted Euclidean distance between each scheme and the positive ideal solution and the weighted Euclidean distance between each scheme and the negative ideal solution are calculated. Then, the tracking progress of each scheme relative to the ideal solution is calculated, and the distance between each scheme and the ideal state is accurately quantified. Based on the tracking progress, a single quantitative basis is provided for the sorting of schemes. In this embodiment, each solution in the Pareto optimal demolition solution set is written back to the construction execution layer of the updated digital twin. The construction process simulation module built into the digital twin engine is invoked to advance the virtual construction of each process hourly. During the simulation, geometric and spatial constraint data in the digital twin are used to detect whether the movement trajectories, operating radii, and material stacking areas of each construction machine overlap or intrude. Whenever two or more construction activities occur simultaneously in the same spatial location and within the same time period, a spatial conflict is accumulated. Simultaneously, based on the technological connections and resource occupancy status between each process, it is detected whether there are any conflicts caused by the incomplete completion of the previous process. In cases where subsequent processes cannot start as planned, each occurrence of such a test result constitutes a process interference. Instantaneous carbon emissions are calculated by summing the carbon dioxide equivalent generated by mechanical energy consumption, material consumption, and transportation activities in each time period of the simulation on a second-by-second basis, recording the instantaneous emission intensity on an hourly basis. After the simulation, the standard deviation of carbon emissions for the entire construction cycle is calculated based on all instantaneous records. In constructing the improved TOPSIS evaluation matrix, each scheme is considered a row vector within the matrix. The total carbon emissions over the entire life cycle, comprehensive renovation cost, and total construction period are the three positive indicators in the first three columns of the matrix, and the construction conflict penalty indicator is the second column. The fourth column includes the carbon emission fluctuation penalty index as the fifth column. The values ​​of each positive index are directly taken from the adaptive evaluation calculation results. The total life-cycle carbon emissions are recorded in tons of CO2 equivalent, the comprehensive renovation cost in tens of thousands of yuan, and the total construction period in calendar days. The construction conflict penalty index is obtained by statistically analyzing the total frequency of all spatial conflicts and process interferences during the 4D simulation. This frequency directly reflects the potential management difficulties and rework risks caused by competition for spatial resources and disordered process sequences during actual construction. The carbon emission fluctuation penalty index uses the standard deviation of instantaneous carbon emission values ​​within the simulation period; a larger value indicates higher carbon emissions during construction. The more drastic the fluctuations, the more vector normalization is applied to the constructed improved TOPSIS evaluation matrix. This involves normalizing the attribute values ​​in each column using vector norm normalization, which is achieved by taking the square root of the sum of the squares of the attribute values ​​in each column and dividing each attribute value by this denominator to obtain the normalized attribute value. In the normalization process, positive evaluation indicators and negative penalty indicators use a unified normalization rule to ensure that all attributes are on the same numerical scale. The positive ideal solution is composed of the optimal value in each column of normalized attribute values. For positive evaluation indicators, the minimum value among all options is taken; for negative penalty indicators, the minimum value among all options is also taken; and for negative ideal solutions, the maximum value among all columns is taken.When calculating the weighted Euclidean distance, the weight coefficients of each objective dimension are determined by the adaptive AHP weight combination output by the reinforcement learning agent after linear mapping transformation. The calculation of the tracking progress is achieved by dividing the weighted Euclidean distance between each scheme and the negative ideal solution by the sum of the weighted Euclidean distance between the scheme and the positive ideal solution and the weighted Euclidean distance between the scheme and the negative ideal solution. The larger the tracking progress value, the closer the scheme is to the ideal state in terms of overall benefits. It should also be noted that the process of constructing the improved TOPSIS evaluation matrix and calculating the schedule is as follows: the construction conflict penalty index and the fluctuation penalty index are set as negative attributes. The feature that the larger the value of the negative attribute is, the higher the risk of the scheme implementation is embedded in the normalization process of the evaluation matrix. This makes the magnitude of the negative attribute value after normalization inversely related to the evaluation quality. The conflict and fluctuation risks are directly transformed into evaluation disadvantages, so that high-risk schemes are in a disadvantageous position at the beginning of the evaluation. The adaptive AHP weight combination output by the reinforcement learning agent is converted into the weight coefficient of each evaluation objective through linear mapping. The attribute values ​​in the normalized evaluation matrix are weighted, so that the project's differentiated value orientation runs through the entire evaluation process of the improved TOPSIS. The project value orientation is extended from the weight setting to the entire evaluation process, ensuring that the evaluation results are internally consistent with the project objectives. Based on the weighted evaluation matrix, the positive ideal solution and the negative ideal solution of each attribute are determined respectively. After calculating the weighted Euclidean distance between each scheme and the positive ideal solution and the weighted Euclidean distance between each scheme and the negative ideal solution, the ratio of the two is defined as the schedule. The schedule comprehensively reflects the distance relationship between the scheme and the positive and negative ideal solutions, and supports the sorting and selection of schemes with a single index. In this embodiment, during normalization, the total frequency of spatial conflicts and process interference accumulated in the 4D simulation, as well as the standard deviation of instantaneous carbon emissions, are included in the fourth and fifth columns of the evaluation matrix, respectively. Specifically, for the construction conflict frequency in the fourth column, the frequency of each scheme is divided by the maximum frequency among all schemes in this column, so that the normalized value falls within the range of 0 to 1. The higher the frequency, the closer the normalized value is to 1, and the lower it is automatically in the subsequent ideal solution determination. For the carbon emission standard deviation in the fifth column, the same normalization rule is used, and the maximum value among the standard deviations of each scheme is used as the denominator for standardization, ensuring that schemes with more drastic carbon emission fluctuations show higher normalization in the evaluation matrix. Numerical values ​​are then explicitly embedded into the improved TOPSIS framework, where higher negative attribute values ​​indicate higher risk. This makes the evaluation results favor solutions with smoother construction processes and fewer spatial conflicts. When converting the adaptive AHP weight combination output by the reinforcement learning agent into weight coefficients for each evaluation objective via linear mapping, the weights of physical state and integrity, material quality and performance, and reuse potential output by the agent based on the current project value orientation are used as a basis. A pre-defined mapping matrix converts these three into weight allocations for the corresponding five evaluation indicators. In the mapping rules, the weights of physical state and integrity are mapped to the weight coefficients of total life-cycle carbon emissions, and the weights of material quality and performance are mapped to the weight coefficients of comprehensive life-cycle carbon emissions. The weighting coefficient for renovation costs is then mapped to the weighting coefficient for the total construction period using the potential weighting. Simultaneously, the arithmetic mean of the weights of the first three criterion layers is used as the weighting coefficient for the construction conflict penalty indicator, and the normalized standard deviation of the three criterion layer weights is used as the weighting coefficient for the carbon emission fluctuation penalty indicator, ensuring that the sum of the weights of the five evaluation indicators is strictly equal to 1. After mapping, each weighting coefficient is multiplied by the corresponding column attribute value in the aforementioned standardized evaluation matrix, ensuring that the project's differentiated value orientation is integrated from the setting of evaluation indicator weights throughout the entire evaluation process of the improved TOPSIS. When determining the positive and negative ideal solutions for each attribute based on the weighted evaluation matrix, for the five evaluation indicators, the sum of the weights of each column is uniformly used. The minimum value among the weighted normalized attribute values ​​forms the positive ideal solution vector, and the maximum value among the weighted normalized attribute values ​​in each column forms the negative ideal solution vector. Since all indicators (including positive evaluation indicators and negative penalty indicators) after normalization represent the better evaluation direction with smaller values, the positive ideal solution corresponds to the global minimum value of each dimension, and the negative ideal solution corresponds to the global maximum value of each dimension. Based on this, when calculating the weighted Euclidean distance between each solution and the positive ideal solution, the square of the difference between the solution and the corresponding dimension value of the positive ideal solution is calculated for each of the five dimensions, multiplied by the weight coefficient of that dimension, and then summed and the square root is taken. The weighted Euclidean distance of the negative ideal solution is calculated in the same way, only the positive ideal solution is replaced with the negative ideal solution.After calculating the two distances, the weighted Euclidean distance between each scheme and the negative ideal solution is divided by the sum of the weighted Euclidean distances between the scheme and the positive ideal solution and the negative ideal solution. The resulting ratio is defined as the matching progress of the scheme. The closer the matching progress value is to 1, the closer the scheme is to the ideal state in terms of comprehensive benefits and the lower the implementation risk. The expression for pasting progress is as follows: ; ; ; In the formula: For the first The progress of each evaluation plan; For the first The weighted Euclidean distance between each solution and the ideal solution; The evaluation indicators are numbered and correspond to five evaluation dimensions: total carbon emissions over the entire life cycle, comprehensive renovation cost, total construction period, total frequency of construction conflicts, and standard deviation of carbon emissions. It refers to the number of evaluation indicators; For the first The positive ideal solution of an evaluation index is the minimum value among the weighted normalized attribute values ​​of all schemes on that index. For the first The weighted Euclidean distance between each solution and the negative ideal solution; For the first The first scheme is in the Normalized attribute values ​​on each evaluation indicator; and They are the first The positive and negative ideal values ​​of a normalized objective; The weights are pre-set by decision-makers based on the importance of each evaluation indicator.

[0028] In S6, the process of uniquely identifying the optimal renovation scheme for existing building renovation is as follows: All schemes in the Pareto optimal demolition scheme set are sorted in descending order of their installation progress values ​​to form an initial priority sequence. Each scheme is clearly sorted according to its comprehensive performance to ensure that the optimization results are easy to screen later. Starting from the head of the initial priority sequence, each scheme is traversed item by item to eliminate inferior solutions that exceed the project's preset upper limit for construction conflict penalty indicators, exceed the allowable threshold for carbon emission fluctuation standard deviation, or have installation progress lower than the set lower limit. This generates a candidate scheme sequence after double risk filtering. Schemes that exceed conflict limits, have excessive emission fluctuations, or have low installation progress are eliminated item by item to significantly improve the quality of the candidate sequence. The scheme with the highest installation progress in the candidate scheme sequence after risk filtering is selected as the final and unique optimal renovation scheme for existing building renovation. The construction sequence, resource scheduling plan, and carbon emission control strategy of this scheme are written into the execution layer of the updated digital twin to form a traceable closed-loop decision result. The construction sequence, resource plan, and carbon emission control strategy of the optimal scheme are fully written into the execution layer, and the decision result is directly executed and traced in the twin. In this embodiment, the output installation progress data of each scheme is read, and all schemes are arranged in descending order of installation progress value. Schemes with the same installation progress value are then ordered in ascending order of construction conflict penalty index. If the conflict index is still the same, they are ordered again in ascending order of carbon emission standard deviation. Finally, an initial priority sequence is generated and stored in the priority sorting table of the project database. At the same time, a unique sorting batch number and timestamp are assigned to this sequence. The project's preset risk control parameters are read: the upper limit of the construction conflict penalty index is set according to the project site space conditions. For space-constrained renovation projects... The limit is set at 30 times, and the upper limit is set at 50 times for projects with ample space. The allowable threshold for the standard deviation of carbon emission fluctuations is set according to the environmental protection department's control requirements for the stability of carbon emissions during the construction period. Generally, it is taken as 1.5 times the standard deviation of the historical statistical data of instantaneous carbon emissions of similar projects in the project area. The lower limit for the application schedule is uniformly set at 0.45. Schemes below this value are considered to have failed to meet the comprehensive benefit standards. During the traversal, starting from the first scheme at the head of the sequence, the application schedule, conflict frequency, and carbon emission standard deviation of each scheme are read one by one and compared with the three thresholds. Schemes with all parameters within the limit are retained; those exceeding any limit are eliminated. The eliminated scheme's number and the reason for exceeding the limit are recorded and stored separately in the elimination scheme log for later analysis. After traversal, all retained schemes are reorganized into a candidate scheme sequence, the number of which equals the number of valid retained schemes. The top-ranked scheme from the candidate scheme sequence is extracted as the optimal scheme, satisfying the triple conditions of highest dismantling progress, conflict frequency not exceeding the upper limit, and carbon emission fluctuation not exceeding the threshold. After extraction, the binary gene sequence of this scheme is parsed to reconstruct the specific combination of dismantling measures and processes. The system prioritizes the work sequence, plans the configuration of construction machinery, and sets the planned start and completion times for each work process. The parsed results are then encapsulated according to the standard data interface format of the digital twin execution layer and written into the construction sequence table, resource scheduling plan, and carbon emission control strategy fields of the execution layer. After the writing operation is completed, a reload command is sent to the execution layer to reload the above decision data into the updated digital twin and perform a full verification. Once the verification is successful, a decision record containing the scheme number, decision timestamp, and key indicator values ​​is generated and stored in the decision result table of the project database.

[0029] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-objective intelligent optimization decision-making method for the renovation and remodeling of existing buildings, characterized in that, Includes the following steps: S1. Collect geometric point cloud and continuous temporal infrared thermal image of the target of renovation and decoration of existing buildings, register and fuse to generate an initial digital twin, and use intelligent sensors to detect hidden defects and map and update the geometric and physical state of the initial digital twin. S2. Construct a knowledge graph of decoration engineering, and use graph neural network reasoning to complete the missing semantic associations of hidden engineering and generate a dynamic dataset of existing components. S3. Based on the reinforcement learning agent's analysis of the project's differentiated value orientation, dynamically adjust the AHP evaluation weight according to the project's preferences, and use the existing component dynamic data as input to generate an initial elite population. S4. Input the initial elite population into the improved NSGA-II algorithm, and perform multi-objective iterative optimization with carbon emissions, cost and construction period as optimization objectives to quickly output the Pareto optimal demolition scheme set for the renovation of existing buildings. S5. An improved TOPSIS algorithm is introduced, taking the Pareto optimal demolition scheme set as the object. The frequency of construction conflicts and the standard deviation of carbon emission fluctuations in the digital twin of each scheme are used as negative penalty indicators to calculate the schedule. The single quantitative indicator comprehensively reflects the quality and risk level of the scheme. S6. Sort the Pareto scheme set of existing building renovation and transformation according to the progress of the renovation, and uniquely identify the optimal renovation scheme of the existing building renovation and transformation.

2. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 1, characterized in that: In step S1, the process of generating the initial digital twin is as follows: A 3D laser scanner is used to scan the existing building renovation target from multiple stations and angles to obtain geometric point cloud data covering the entire space. At the same time, an infrared thermal imager is used to continuously collect time-series thermal image sequences of the same spatial area at preset time intervals. The geometric point cloud data is registered and fused with the temporal infrared thermal image sequence. Based on the spatial coordinates of the point cloud, the thermal image pixels are spatially mapped by the iterative nearest point algorithm, and then spatiotemporal interpolation is used to generate an initial digital twin that combines three-dimensional geometric contours and dynamic distribution of surface temperature field. In the initial digital twin, the temperature anomaly areas on the surface of each existing component are used as candidate identifiers of hidden defects. The intelligent sensors deployed in the target area of ​​the existing building renovation are used to detect the underground or wall internal layer echo signals to verify and locate the candidate identifiers, generate a hidden defect distribution map and map it to the corresponding spatial coordinates of the initial digital twin. The intelligent sensor is a ground penetrating radar. Based on the distribution map of hidden defects, the geometric accuracy parameters and physical state attributes of each existing component in the initial digital twin are updated to form an updated digital twin that reflects the true geometric shape and internal damage status of the existing building.

3. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 2, characterized in that: The process of updating the geometric and physical state of the initial digital twin is as follows: For each defect in the hidden defect distribution map, its spatial boundary coordinates and defect depth value are extracted to correct the geometric contour accuracy of the corresponding existing component in the initial digital twin. At the same time, the geometric dimensions of existing components with material spalling or cross-sectional reduction are negatively compensated and adjusted. Based on the rate of change and temperature difference amplitude of the surface temperature of each existing component in the time-series thermal imaging sequence, the attenuation degree of its thermal inertia index and thermal conductivity is inversely deduced. Combined with the change characteristics of reflection intensity in the ground penetrating radar echo signal, the residual mechanical performance level and corrosion damage depth of each existing component are quantitatively evaluated. The corrected geometric profile, the assessed residual mechanical property level, and the corrosion damage depth are used as updated attributes for each existing component and written into the component attribute fields corresponding to the initial digital twin, forming an updated digital twin whose geometric and physical states are consistent with the field measurement data.

4. The multi-objective intelligent optimization decision-making method for renovation and remodeling of existing buildings according to claim 1, characterized in that: In S2, the process of generating the dynamic dataset of existing components is as follows: By extracting entity types, entity attributes, and inter-entity relationships in the decoration and renovation field from national building standards and specifications, enterprise process standard databases, and historical existing building renovation projects, a knowledge graph of decoration engineering covering component categories, material properties, connection methods, failure modes, and repair processes is constructed. The known geometric, physical, and damage attributes of each existing component in the existing component dynamic dataset are used as the initial features of the nodes and input into the message passing layer of the graph neural network. After multiple rounds of neighborhood aggregation and node state updates, the missing hidden engineering attribute values ​​and the hidden association types between them and other existing components are inferred and predicted. The missing attribute values ​​and hidden association types obtained from graph neural network inference prediction are backfilled into the null fields corresponding to the existing component dynamic dataset, generating an existing component dynamic dataset containing complete geometric, physical, damage, and hidden engineering semantic association features.

5. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 1, characterized in that: In step S3, the process of generating an initial elite population using existing component dynamic data as input is as follows: We collected case data from multiple historical projects in the field of existing building renovation and reconstruction, constructed a training dataset containing samples of multiple types of projects, and used the feature parameters of each project in the training dataset as the state input, the AHP criterion layer weight allocation scheme of the project as the action output, and the final implementation effect score of the project as the reward value. We used a deep Q network for offline training to generate a reinforcement learning agent that can map project features to the optimal weights. The differentiated value orientation of the current existing building renovation and renovation project, as well as the parameters of project scale and structural type, are input into the trained reinforcement learning agent. The agent outputs the weight combination of each criterion layer of AHP adapted to the current project through forward inference. The weight combination of each criterion layer of AHP replaces the fixed weight in the traditional AHP and performs a comprehensive score on the differentiated retention value of each existing component in the dynamic dataset of existing components. Existing components with comprehensive scores higher than a preset threshold are selected to form a high-value retention candidate set. The demolition measures corresponding to each existing component in the high-value retention candidate set are encoded into binary gene sequences. The feasibility of the gene sequences is verified in combination with on-site construction space constraints and safety regulations. The gene sequences that pass the verification are combined to generate an initial elite population.

6. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 1, characterized in that: In S4, the multi-objective iterative optimization process is as follows: A multi-objective fitness function is constructed with the total carbon emissions throughout the entire life cycle of renovation and renovation of existing buildings, the comprehensive renovation cost, and the total construction period as the three optimization objectives. The spatial constraints in the updated digital twin are encoded as penalty function terms and added to the fitness function of each objective. The initial elite population is input into the improved NSGA-II algorithm. Fitness evaluation is performed on each chromosome in the population. The chromosome is sorted quickly according to the three major target values. At the same time, the crowding distance of each chromosome in the same non-dominated layer is calculated. Parent individuals are selected with an elite retention strategy that prioritizes non-dominated level and then crowding distance. Perform simulated binary crossover based on adaptive crossover probability and polynomial mutation based on adaptive mutation probability on the selected parent individuals to generate the offspring population. After merging the parent and offspring, repeat the non-dominated sorting and crowding distance screening. Iterate until the termination condition is met, and output the Pareto optimal demolition scheme set.

7. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 6, characterized in that: The adjustment process for the adaptive crossover probability and adaptive mutation probability is as follows: Set upper and lower limits for crossover probability, and calculate the crossover probability of the current generation in a non-linear decreasing manner based on the ratio of the current iteration generation to the preset maximum iteration generation, so that the NSGA-II algorithm maintains a high crossover rate in the early stage of iteration and reduces the crossover rate in the later stage of iteration. The average crowding distance of the first non-dominant frontier individual in the current generation is calculated and the ratio of it to the historical maximum average crowding distance is used as a population diversity index. The mutation probability of the current generation is dynamically adjusted using the population diversity index as feedback, so that the mutation rate is adaptively increased when the population diversity decreases and the mutation rate is kept low when the diversity is sufficient. The calculated crossover and mutation probabilities of the current generation are applied to the crossover and mutation operations of the current generation, respectively. The historical maximum average crowding distance is updated at the end of each iteration, so as to achieve adaptive adjustment of the crossover and mutation parameters as the algorithm evolves.

8. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 1, characterized in that: In step S5, the calculation process for the application progress is as follows: Map each Pareto optimal demolition scheme back to the updated digital twin, perform 4D construction process simulation, and count the number of spatial conflicts and process interferences that occur in the construction simulation of each scheme hourly. At the same time, record the instantaneous carbon emission values ​​in each simulation period and calculate their standard deviation. The total carbon emissions throughout the entire life cycle of each scheme, the comprehensive renovation cost, and the total construction period are used as positive evaluation indicators. The total frequency of spatial conflicts and process interference is used as the construction conflict penalty indicator, and the standard deviation of carbon emissions is used as the fluctuation penalty indicator. Together, an improved TOPSIS evaluation matrix is ​​constructed. The improved TOPSIS evaluation matrix is ​​vector normalized. Based on the positive and negative ideal solutions of each objective, the weighted Euclidean distance between each scheme and the positive ideal solution and the weighted Euclidean distance between each scheme and the negative ideal solution are calculated. Then, the tracking progress of each scheme relative to the ideal solution is calculated.

9. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 8, characterized in that: The process of constructing the improved TOPSIS evaluation matrix and calculating the patching progress is as follows: The construction conflict penalty index and the fluctuation penalty index are set as negative attributes. The feature that the larger the value of the negative attribute index is, the higher the risk of the implementation of the scheme is embedded in the normalization process of the evaluation matrix. This makes the magnitude of the negative attribute value after normalization inversely correlated with the evaluation quality. The adaptive AHP weight combination output by the reinforcement learning agent is converted into the weight coefficients of each evaluation target through linear mapping. The attribute values ​​in the normalized evaluation matrix are weighted, so that the differentiated value orientation of the project is carried out throughout the entire evaluation process of the improved TOPSIS. Based on the weighted evaluation matrix, the positive and negative ideal solutions for each attribute are determined. The weighted Euclidean distance between each scheme and the positive ideal solution and the weighted Euclidean distance between each scheme and the negative ideal solution are calculated, and the ratio of the two is defined as the matching progress.

10. The multi-objective intelligent optimization decision-making method for the renovation and reconstruction of existing buildings according to claim 1, characterized in that: In S6, the process of uniquely identifying the optimal renovation solution for existing buildings is as follows: All schemes in the Pareto optimal demolition scheme set are sorted in descending order of their installation progress values ​​to form an initial priority sequence; Starting from the head of the initial priority sequence, each item is traversed sequentially to eliminate inferior solutions that exceed the project's preset upper limit for construction conflict penalty indicators, exceed the allowable threshold for carbon emission fluctuation standard deviation, or fall below the set lower limit for schedule, thereby generating a candidate solution sequence after double risk filtering. The scheme with the highest progress in the candidate scheme sequence after risk filtering is selected as the final and only optimal renovation scheme for the existing building renovation. The construction sequence, resource scheduling plan and carbon emission control strategy of this scheme are written into the execution layer of the updated digital twin.

Images