Building component demolition sequence safety control method integrating genetic algorithm, Bayesian reasoning and machine learning

Abstract

The invention provides a building component dismantling sequence safety control method fusing a genetic algorithm, Bayesian reasoning and machine learning, a directed acyclic graph of a component dependency relationship is established through topological sorting, and a feasible and low-risk dismantling sequence path is generated in combination with optimization operation of mechanical analog simulation and the genetic algorithm. And a Bayesian reasoning mechanism is used for dynamically sensing the structure risk, and a machine learning model is used for assisting to improve the risk judgment stability, so that the collaborative linkage of disassembly sequence path generation and risk control is realized. Through data interaction and decision linkage, a closed-loop control chain for driving disassembly sequence path regulation and control by risk judgment is formed, and dynamic optimization and whole-process safety guarantee of building structure disassembly are realized through fusion of a genetic algorithm, Bayesian reasoning and machine learning.

Description

Technical Field

[0001] The present invention relates to the field of construction technology, and particularly to a safety control method for the demolition sequence of building components that integrates genetic algorithms, Bayesian inference, and machine learning. Background Art

[0002] With the increasing demands for urban construction and renewal, the demolition of buildings has become an inevitable and important part of modern engineering. Most traditional building demolition methods rely on experience and manual judgment. Especially when dealing with high-rise buildings with complex structures, the safety issues are more prominent. During the demolition process, due to the complex dependency relationships between structural components, incorrect demolition sequences may lead to structural instability or even collapse, posing significant safety hazards. Therefore, how to effectively control risks and optimize the demolition sequence during the demolition process to ensure construction safety is an urgent problem to be solved in current building demolition technology.

[0003] However, existing building demolition methods usually presuppose fixed demolition sequences, lacking monitoring and adjustment of the real-time state of the building structure and being difficult to cope with the stress changes that occur during the demolition process; moreover, although real-time monitoring technologies have been relatively mature, how to effectively utilize real-time data (such as stress, displacement, vibration, etc.) for the assessment of structural instability risks remains a blind spot in the existing technology; in addition, the aging of building structures, changes in material properties, and noise in real-time monitoring data bring a large amount of uncertainty, and existing methods are difficult to handle these uncertainties. Summary of the Invention

[0004] (I) Technical Problems to be Solved Based on this, the present invention provides a safety control method for the demolition sequence of building components that integrates genetic algorithms, Bayesian inference, and machine learning to solve the problems mentioned in the background art, namely, the lack of monitoring and adjustment of the real-time state of the structure due to the fixed demolition sequence, insufficient real-time data analysis capabilities, and the inability to handle uncertain factors.

[0005] (II) Technical Solutions To achieve the above object, the present invention provides a safety control method for the demolition sequence of building components that integrates genetic algorithms, Bayesian inference, and machine learning, including: S1: Analyze the dependency relationships between components in the building structure using topological sorting to construct a directed acyclic graph of the demolition sequence; S2: Randomly select demolition sequence schemes that satisfy topological constraints as the initial population of the genetic algorithm, where ; evaluate the safety of the state of the demolition sequence scheme based on mechanical simulation, and combine the optimization operations of the genetic algorithm to obtain the actually adopted demolition scheme; S3: During the process of the actually adopted disassembly plan, based on the monitoring data of the building structure, the dynamic risk prediction and continuous correction of the building structure state are carried out by combining Bayesian inference and machine learning.

[0006] Further, the S2 includes: S201: In the directed acyclic graph, randomly select disassembly sequence plans that meet the topological constraints as candidate plans, and use the candidate plans as the initial population of the genetic algorithm; The disassembly sequence plan is represented by the component sequence vector in the directed acyclic graph:

[0007] Among them, represents the disassembly sequence plan; represents the component number to be removed in the th step; represents the number of disassembly steps of the disassembly sequence plan, where ; S202: Based on mechanical simulation, evaluate the safety of the state of the candidate plan and calculate the fitness of the candidate plan; In the disassembly sequence plan, for the state after each step of disassembly, use the finite element analysis method to perform mechanical simulation on the stress distribution of the building structure, output the actual stress value of the component to be disassembled in the next step, and calculate the safety factor of :

[0008] Among them, represents the material yield strength of the component to be disassembled in the next step; represents the actual stress of the component to be disassembled in the next step; Then the fitness function of the entire disassembly sequence plan is defined as follows:

[0009] Among them, represents the disassembly sequence plan; represents the number of steps of the disassembly sequence plan; represents the safety factor corresponding to the current state ; S203: Determine whether the fitness is within the safe range. If so, use the candidate solution with the maximum fitness as the actual disassembly sequence plan; if not, proceed to the next step; S204: Screen the candidate solution with the highest fitness for the optimization operation of the genetic algorithm, where and ; S205: Use the result of the optimization operation as the new candidate solution, and check whether the new candidate solution meets the topological constraints. If it meets, go to S202; if it does not meet, go to S204;

[0010] Further, the S3 includes: S301: During the process of the actual disassembly plan, obtain the monitoring data of the current state of the building structure ; S302: According to the monitoring data ; calculate the corresponding Bayesian posterior probability as the Bayesian risk score value , where represents the occurrence of a structural instability event; S303: Input the monitoring data into the trained machine learning model to obtain the risk score value output by the machine learning model ; S304: Synthesize the Bayesian risk score value and the risk score value output by the machine learning model , and perform dynamic risk prediction and continuous correction on the current state .

[0011] Further, the definition of the monitoring data of the state in S301 is as follows:

[0012] Among them, represents the maximum displacement of all remaining components in the disassembly sequence, reflecting the most severe deformation degree of the current building structure; represents the maximum value of the ratio of the actual strain of all remaining components in the disassembly sequence to their material yield strain, which is used to reflect the potential failure risk of the most dangerous components; represents the peak acceleration at the key parts of the building structure, reflecting the degree of inertial impact on the building structure, where the key parts are selected by experts through analysis and judgment; It is the maximum horizontal inclination angle of the current building structure as a whole or a typical floor, which is used to reflect the overturning tendency of the structure; It indicates the maximum value of the ratio of residual stress to the ultimate strength of the material at the key parts of the current building structure, and is used to measure the residual bearing capacity of the building structure.

[0013] Furthermore, in S302, based on the monitoring data , use the Bayesian inference formula to evaluate the current state The posterior probability , the formula is as follows

[0014] in, Indicates that the structure has experienced an instability event; Indicates status monitoring data; Represents the static prior probability, which is determined by the hierarchical weight and safety factor of the component to be disassembled in the next step in the directed acyclic graph The smaller the topological depth of the component to be disassembled in the next step, The smaller the value, the higher the prior probability is assigned; represents the likelihood probability, which is obtained by statistics of a large number of simulation samples accumulated during the genetic algorithm optimization process in step S2. In order to avoid the monitoring value in the multidimensional feature space It is difficult to completely overlap with the simulation sample, so it can be expanded appropriately The tolerance range, that is, will be in the multidimensional feature space with The Euclidean distance is less than The simulation sample is used as ,in is a preset positive number; represents the probability of monitoring, which can be generated through statistical analysis; Represents the posterior probability, which serves as the basis for risk assessment of the current state; Receive new monitoring data After that, the static prior probability is no longer used to calculate the posterior probability of the new state. Instead, the posterior probability of the previous state is used as the prior probability of the new state, as shown below: .

[0015] Furthermore, the training of the machine learning model in S303 includes: In each state , collect monitoring data characteristics of building structures , and record the posterior probability obtained by Bayesian inference , build the sample The training set composed of First, for each sample calculate its risk score ; The scoring function comprehensively considers the topological depth and Bayesian posterior probability of the component to be disassembled in the next step. The formula is as follows:

[0016] Among them, represents the topological depth of the component to be disassembled in the next step; is a minimum non-zero constant with a value of 0.001; represents the sliding average of the posterior probabilities in the last times, which is used to reflect the average risk of the current state and the state after the previous disassembly steps; If the number of current disassembly steps is less than , then takes the value of the actual existing number of steps; is a weighting coefficient used to allocate weights between the two types of information representing "structural dependence level" and "instability probability assessment" , and satisfies , is set by empirical data or training and verification results; Take the historically accumulated monitoring data as input, and use the K-means clustering method to perform clustering in the high-dimensional state space to obtain clusters , and the center of each cluster; For each sample , based on and the similarity between the corresponding cluster center and the risk score , calculate the representative score of the sample . The representative score is used to judge the priority of the sample in the training set; The formula for the representative score is as follows:

[0017] Among them, represents the monitoring data features in the sample and the Euclidean distance between the cluster center of the cluster to which it belongs; is a weight coefficient and ; In each cluster, retain several samples with the highest score as representative samples to form a training subset; Input the training subset into the machine learning model for training.

[0018] Furthermore, in step S304, during the demolition process, when the machine learning model has been fully trained, start the "joint discrimination mechanism". The rules are as follows: If , the current risk judgment is recognized, that is, the posterior probability obtained by Bayesian inference is considered correct; If , the current risk judgment is not recognized, that is, there is a problem with the posterior probability obtained by Bayesian inference. The demolition operation is suspended, and after manual review, a secondary judgment is made; Among them, is a preset threshold.

[0019] Furthermore, the S1 includes: S101: Obtain and analyze relevant data of the building structure, determine the interdependent relationship of each component during the demolition process, and establish a dependency matrix; S102: Based on the dependency matrix, perform topological sorting on the components and draw a directed acyclic graph.

[0020] Furthermore, the optimization operations of the genetic algorithm in the S204 include crossover and mutation.

[0021] (III) Beneficial Effects As can be seen from the above technical solutions, a safety control method for the demolition sequence of building components proposed by the present invention has the following beneficial effects: 1. By screening the disassembly sequence path through topological sorting, optimizing the short-step disassembly plan using the genetic algorithm, and evaluating the stability of each disassembly state in combination with mechanical simulation, it is ensured that each disassembly operation is within the controllable range, significantly reducing the risk of structural instability caused by improper disassembly sequence and improving the structural safety during the disassembly process.

[0022] 2. Collect real-time monitoring status data, and use the Bayesian inference model to recursively update the instability probability of the current state. At the same time, the machine learning model quickly gives an auxiliary judgment to achieve double verification. Once the risk exceeds the threshold, the operation is automatically suspended and an optimized plan is regenerated to ensure risk closed-loop control and achieve dynamic identification and instant response to risks.

[0023] 3. The present invention can continuously receive on-site data and dynamically learn historical judgment results, enabling the system to have the ability of self-learning and model evolution. It can effectively cope with structural response changes caused by unpredictable factors such as material aging, environmental disturbances, and construction errors, improve the robustness and reliability of the overall system, and enhance the adaptability to uncertain factors during the disassembly process. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings. The drawings are schematic and should not be construed as limiting the present invention in any way. In the drawings: Figure 1Flow chart of the building demolition sequence safety control method integrating genetic algorithm, Bayesian inference and machine learning according to the present invention; Figure 2 Flow chart of the demolition plan with the minimum risk formed by combining mechanical simulation and genetic algorithm according to the present invention; Figure 3 Schematic diagram of risk discrimination by Bayesian inference-machine learning joint according to the present invention. Specific embodiments

[0025] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0026] As Figure 1 shown, the present invention provides a building demolition sequence safety control method integrating topological sorting, optimization algorithm, Bayesian inference and machine learning, including: S1: Analyze the dependency relationship between components in the building structure by topological sorting to construct a directed acyclic graph of the demolition sequence; the specific steps include: S101: Obtain and analyze the relevant data of the building structure, determine the mutual dependency relationship between components during the demolition process and establish a dependency matrix; To ensure the safety of the demolition process, it is first necessary to obtain the relevant data of the building. These data include but are not limited to design drawings (i.e., the design drawings of the building, including the load-bearing structure and the layout of non-load-bearing walls on each floor) and structural details (including material type, connection method between components, component size and strength, etc.). Through these data, a comprehensive analysis of the structural characteristics of the building is carried out to determine the mutual dependency relationship between components during the demolition process.

[0027] The dependency relationship between each component is determined based on the following principles: 1. Load-bearing structure dependency principle; there is a dependency relationship between the load-bearing structure components of the building. Beams depend on columns, floors depend on beams, secondary beams depend on main beams, shear walls depend on the surrounding frames, and structural coupling beams depend on walls. The dependent components must be demolished later.

[0028] 2. Upper structure priority demolition principle; in a multi-story building, the load-bearing components of the upper floors must be demolished first to avoid affecting the stability of the lower floors.

[0029] 3. Principle of demolishing lateral force resisting members later; Lateral force resisting members play a key role in providing the overall lateral stability of a building. During demolition, lateral force resisting members should be demolished as late as possible compared to other members under the same circumstances. Otherwise, it may lead to local instability or even overall collapse.

[0030] 4. Principle of preferentially demolishing vulnerable and high-risk areas: When multiple members are closely connected, demolition should start from the vulnerable or most dangerous areas first to avoid accidental collapse or instability during the demolition process.

[0031] Define a dependency matrix in combination with the interdependence of each member . Among them, =1 indicates that after member is demolished, member can be demolished. Traverse all building members to establish a dependency matrix.

[0032] S102: Based on the dependency matrix, perform topological sorting on the members and draw a directed acyclic graph.

[0033] Perform topological sorting on the members through the dependency matrix to generate a demolition sequence that conforms to the dependency relationship. The resulting directed acyclic graph (DAG) shows the precedence and dependency relationships between the members, providing a clear reference path for subsequent demolition operations.

[0034] After completing the construction of the directed acyclic graph, further introduce the concept of topological depth to measure the hierarchical influence degree of each member in the overall structural domination relationship. Topological depth reflects the degree to which a member is controlled by how many upstream members and its relative "rear" or "dominated" position in the entire disassembly path. For each member node in the directed acyclic graph, define its topological depth as the length of the longest path from any source node to , that is, starting from any source node in the directed acyclic graph, the number of edges included in the longest path to . A source node represents a node without predecessors and can be directly demolished without considering the demolition sequence of other parts.

[0035] The present invention uses the topological depth of the members as one of the importance scores to measure them in the directed acyclic graph, and then participates in the estimation and modeling of the prior instability probability.

[0036] S2: In the directed acyclic graph, randomly select disassembly sequence schemes that meet the topological constraints as the initial population of the genetic algorithm; Based on mechanical simulation, evaluate the safety of the states of the disassembly sequence schemes, and combine with the optimization operations of the genetic algorithm to obtain the actually adopted disassembly scheme; In a directed acyclic graph, randomly select one or more disassembly sequence schemes (i.e., candidate schemes) that satisfy topological constraints (conforming to the precedence and dependency relationships between components) as the initial population of the genetic algorithm. Each candidate scheme consists of several disassembly steps. Then, use the mechanical simulation method to evaluate the structural state and instability risk of each candidate scheme, and finally select the optimal disassembly order scheme through the fitness function. Since the goal of the building disassembly order problem is to find a set of component disassembly orders that satisfy the topological dependency relationship and are overall optimally safe, this problem belongs to the constrained permutation optimization problem.

[0037] As Figure 2 shown, S2 specifically includes: S201: In a directed acyclic graph, randomly select disassembly sequence schemes that satisfy topological constraints (conforming to the precedence and dependency relationships between components) as candidate schemes, and use the candidate schemes as the initial population of the genetic algorithm; The disassembly order of building components is modeled as a permutation optimization problem, and the gene of the genetic algorithm is the component order vector in the directed acyclic graph :

[0038] Among them, represents the component number to be removed in the th step. The length of is , representing a sequence of removed components (i.e., disassembly sequence scheme). Different sequences will affect the stress distribution and instability risk of the building structure.

[0039] The encoding method of the genetic algorithm uses the permutation vector of component numbers to represent the disassembly order, ensuring that each disassembly sequence scheme corresponds to a complete and non-repeating disassembly sequence.

[0040] S202: Based on mechanical simulation, conduct a safety assessment of the state of the candidate scheme and calculate the fitness of the candidate scheme; In each disassembly scheme, for the state after each step of disassembly, use the finite element analysis (FEM) method to conduct a mechanical simulation of the stress distribution of the building structure. The mechanical simulation considers the influence of external loads and boundary conditions on the structural stress, outputs the actual stress value of the component to be disassembled in the next step, and calculates its safety factor. The formula for the safety factor is as follows: (1) Among them, Indicates the component to be disassembled in the next step The yield strength of the material; Indicates the component to be disassembled in the next step The actual stress.

[0041] During the state evaluation process after each disassembly step, the instability risk of the component to be disassembled in the next step is judged by the safety factor. If the safety factor of a certain state is close to 1, it means that the building structure is close to instability; if the safety factor of a certain state is much greater than 1, it means that the building structure is in a stable state.

[0042] Fitness function Used to evaluate the overall safety of each disassembly sequence plan. In each plan, the current state (that is, the component to be disassembled in the next step is ) The corresponding safety factor is , and the safety of each state is interrelated. Therefore, the fitness function should reflect the coherence and cumulative risk of multiple states. The definition of the fitness function is as follows: (2) Wherein, Represents the disassembly sequence plan; Represents the number of steps of the disassembly sequence plan; Represents the current state The corresponding safety factor; The solution of is shown in formula (1).

[0043] In this way of multiplying step by step, the fitness function can reflect the coherent risk of each state in the disassembly plan. If the safety factor of a certain step is low, the fitness value will drop significantly, thus guiding to avoid high-risk plans.

[0044] S203: Judge whether the fitness is within the safe range. If so, take the candidate plan with the maximum fitness as the actually adopted disassembly sequence plan (that is, the disassembly plan with the minimum risk); if not, go to the next step; The safe range is set artificially according to experience.

[0045] S204: Screen the Candidate plans with the highest fitness for the optimization operation of the genetic algorithm; The optimization operations include crossover operation and mutation operation. Crossover adopts the PMX (Partially Matched Crossover) algorithm, and mutation is to randomly exchange, insert or reverse the elements (that is, the construction numbers) in the demolition sequence according to a certain ratio to generate a new demolition sequence.

[0046] S205: Use the result of the optimization operation as a new candidate solution, and check whether the new candidate solution meets the topological constraints. If it meets, go to S202; if it does not meet, go to S204.

[0047] To ensure that all components comply with the domination and dependence relationships among building components, after generating a new disassembly sequence plan in each generation of genetic optimization operations, topological constraint checking is performed on the new disassembly sequence plan.

[0048] After multiple rounds of iterative optimization by the genetic algorithm, select the disassembly sequence plan with the highest fitness value as the actually adopted disassembly sequence plan. This plan ensures the maximization of safety in each state during the disassembly process and the minimum risk of instability.

[0049] S3: During the process of the actually adopted disassembly plan, through the monitoring data of the building structure, combined with Bayesian inference and machine learning, dynamically predict and continuously correct the state of the building structure.

[0050] During the demolition process, to achieve precise identification and dynamic response control of the structural instability risk during building demolition, a combined enhanced risk discrimination mechanism combining Bayesian inference and machine learning is introduced on the basis of "real-time monitoring". S3 specifically includes: S301: During the process of the disassembly plan with the minimum risk, obtain the monitoring data of the building structure in the current state ; ; Obtain the monitoring data of the building structure (including all remaining components in the disassembly sequence plan) in the current state by deploying multiple types of sensors ; the monitoring data is a multi-dimensional vector, which is defined as follows: ; where, (3) where, represents the maximum displacement of all remaining components in the disassembly sequence (i.e., the disassembly sequence plan), reflecting the most serious deformation degree of the building structure in the current state; represents the maximum value of the ratio of the actual strain of all remaining components in the disassembly sequence to their material yield strain, used to reflect the potential failure risk of the most dangerous components; represents the peak acceleration at the key parts of the building structure, reflecting the degree of inertial impact on the building structure, where the key parts are selected by experts through analysis and judgment; is the maximum inclination angle of the current building structure as a whole or a typical floor in the horizontal direction, used to reflect the overturning trend of the structure; Represents the maximum ratio of the residual stress to the ultimate strength of the material at the key parts of the current building structure, and is used to measure the remaining load-bearing capacity of the building structure.

[0051] S302: According to the monitored data ; Calculate the corresponding Bayesian posterior probability As the Bayesian risk score value ; Based on the multi-dimensional vector , Use the Bayesian inference formula to evaluate the current state 's posterior probability , The formula is as follows: (4) Where, Indicates that the structure has an instability event; Indicates the state 's monitored data; Represents the static instability probability that is independent of the monitored data (i.e., the prior probability), which is determined by the hierarchical weight and safety factor of the component to be disassembled in the next step in the directed acyclic graph The smaller the topological depth of the component to be disassembled in the next step and the smaller the value, the higher the prior instability probability is given; Indicates the probability of obtaining the monitored data when the instability occurs (i.e., the likelihood probability), which is obtained by statistically analyzing a large number of simulation samples accumulated during the genetic algorithm optimization process in step S2. To avoid the monitored values in the multi-dimensional feature space from being exactly coincident with the simulation samples, the tolerance range of can be appropriately expanded. That is to say, the simulation samples with an Euclidean distance less than from in the multi-dimensional feature space are regarded as , where is a preset positive number; Represents the total probability of the current monitored data appearing in all states (i.e., the monitoring probability), which can be generated through statistical analysis; Represents the posterior probability after combining prior knowledge and state monitoring, and is used as the basis for risk assessment of the current state.

[0052] In addition, set the instability risk threshold , when , trigger the risk response mechanism, pause the current demolition operation, prompt for temporary reinforcement and data review, and then return to step S2 to regenerate a new demolition sequence plan. This process will be repeated to ensure that each demolition state remains within the controllable risk range.

[0053] When In order to improve the temporal continuity and prediction robustness of state risk assessment, the present invention introduces a recursive update mechanism in Bayesian reasoning. After that, the static prior instability probability is no longer used to calculate the posterior probability of the new state. Instead, the posterior probability of the previous state is used as the prior probability of the new state, as shown below: (5) By dynamically adjusting the prior probability, the continuity and traceability of the risk assessment results in the time dimension are guaranteed, and emergencies that occur during the dismantling process can be responded to more flexibly.

[0054] S303: The monitoring data Input the trained machine learning model to obtain the machine learning risk score value ; In order to improve the efficiency of risk identification in long-term operation, the present invention uses machine learning methods to continuously learn the Bayesian reasoning process. ), collect monitoring data characteristics of building structures , and record the risk score value (i.e., posterior probability) obtained by Bayesian inference ; Build sample The training set composed of.

[0055] In order to construct a high-quality training set that is representative and discriminative, the present invention proposes a risk-guided cluster sampling mechanism, which is used to screen out representative samples with high risk focus and state coverage from the historical samples accumulated during the disassembly process for model training of machine learning.

[0056] First, for each sample Calculate its risk score The scoring function comprehensively considers the topological depth and Bayesian posterior probability of the component to be disassembled in the next step, and the formula is as follows: (6) in, Indicates the topological depth of the component to be disassembled in the next step; is a very small non-zero constant, with a value of 0.001; Indicates recent The sliding average of the posterior probability is used to reflect the difference between the current state and the previous state. The average risk of the state after the first step of disassembly is calculated, so as to improve the sensitivity of the training sample sampling strategy to the risk evolution trend in subsequent sampling. The default value is 4; if the current number of disassembly steps is insufficient ,but The value is the actual number of steps. is a weighting coefficient used to represent the "structural dependence level" and "buckling probability assessment" to allocate weights between two types of information, and satisfy , is set by empirical data or training verification results and can be set to , by default. This default setting indicates that the posterior risk is the main reference factor, but the topological position still has an influence.

[0057] Take the eigenvector of the monitoring data accumulated historically as the input, and use the K-means clustering method to perform clustering in the high-dimensional state space to obtain clusters , and the center of each cluster. For each sample , based on its similarity to the corresponding cluster center and the risk score , calculate the representative score of the sample. The representative score is used to judge the priority (i.e., weight) of the sample in the training set. The formula for the representative score is as follows: (7) where represents the eigenvector of the monitoring data in sample and the Euclidean distance between the center of the cluster to which it belongs; is the weight coefficient and , the default value is set to 0.3, and

[0058] In each cluster, retain several (in this embodiment, 1000) samples with the highest scores as representative samples to form a training subset. This training subset ensures both the wide coverage of samples and the focused expression of high-risk areas, providing an efficient training data basis for the machine learning model (in this embodiment, the machine learning model is the GBDT model). The training subset is used to train the Gradient Boosting Decision Tree (GBDT) model to gradually learn the risk discrimination pattern in Bayesian inference.

[0059] The GBDT model approximates the Bayesian output result through multiple rounds of ensemble iteration of weak classifiers, has good non-linear modeling ability and small sample learning effect, can effectively simulate the complex mapping relationship implicit in the Bayesian inference process, and forms a fast and efficient risk auxiliary identifier after accumulating enough samples. After training, the GBDT model can independently output the machine learning risk score value .

[0060] S304: Synthesize the above-mentioned Bayesian risk score value and the risk score value output by the machine learning model to perform dynamic risk prediction and continuous correction on the current state

[0061] As Figure 3 shown, during the demolition process, when the machine learning model has been fully trained, start the "joint discrimination mechanism", and the rules are as follows: If , then recognize the current risk judgment (that is, consider the posterior probability obtained by Bayesian inference to be correct); If , then do not recognize the current risk judgment (that is, consider that there is a problem with the posterior probability obtained by Bayesian inference), suspend the demolition operation, and after manual review, perform a second judgment; wherein is a preset threshold value.

[0062] Considering that in the initial stage of applying the building component demolition sequence safety control method of the present invention that combines genetic algorithm, Bayesian inference and machine learning, there is not enough historical samples accumulated for training the machine learning model, a phased discrimination mechanism is set, and the rules are as follows: 1. When the number of historical samples used to train the machine learning model does not reach the preset threshold value , only use the posterior probability of Bayesian inference to discriminate risks; 2. When the number of historical samples used to train the machine learning model reaches the preset threshold value , and the mean absolute error MAE is less than 0.05, adopt the above-mentioned joint discrimination mechanism of Bayesian inference and machine learning.

[0063] The present invention provides a sequential safety control method throughout the whole process of building disassembly. By establishing a directed acyclic graph of component dependencies through topological sorting, combining the optimization operations of mechanical simulation and genetic algorithm to generate a feasible and low-risk disassembly sequence path, and dynamically perceiving structural risks with a Bayesian inference mechanism, supplemented by a machine learning model to improve the stability of risk discrimination, the coordination and linkage between the generation of disassembly sequence paths and risk control are realized. The data interaction and decision linkage of the present invention form a closed-loop control chain that drives the regulation of the disassembly sequence path by risk judgment, and realizes the dynamic optimization and full-process safety guarantee of building structure demolition by integrating genetic algorithm, Bayesian inference and machine learning.

[0064] ​As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A safety control method for the demolition sequence of building components that integrates genetic algorithms, Bayesian inference, and machine learning, characterized in that, Including: S1: Analyze the dependency relationships between components in the building structure using topological sorting, and construct a directed acyclic graph of the disassembly sequence; S2: Randomly select disassembly sequence schemes that meet the topological constraints in the directed acyclic graph as the initial population of the genetic algorithm, where ; perform a safety assessment on the state of the disassembly sequence scheme based on mechanical simulation, and combine the optimization operations of the genetic algorithm to obtain the actually adopted disassembly scheme; S3: During the process of the actually adopted disassembly plan, through the monitoring data of the building structure, combine Bayesian inference and machine learning to conduct dynamic risk prediction and continuous correction of the building structure state.

2. The method according to claim 1, wherein The S2 includes: S201: Randomly select disassembly sequence schemes that meet the topological constraints in the directed acyclic graph as candidate schemes, and use the candidate schemes as the initial population of the genetic algorithm; The component sequence vector in the directed acyclic graph for the disassembly sequence plan is represented as: Among them, represents the disassembly sequence plan; represents the component number to be removed in the th step, where represents the number of disassembly steps of the disassembly sequence plan; S202: Based on mechanical simulation, evaluate the safety of the state of the candidate plan and calculate the fitness of the candidate plan; In the disassembly sequence plan, for the state after each step of disassembly , the finite element analysis method is used to conduct a mechanical simulation of the stress distribution of the building structure, and the component to be disassembled in the next step 's actual stress value is obtained, and the safety factor of is calculated : Among them, represents the material yield strength of the component to be disassembled in the next step ; represents the actual stress of the component to be disassembled in the next step ; Then the fitness function of the entire disassembly sequence plan is defined as follows: Among them, represents the disassembly sequence plan; represents the number of steps of the disassembly sequence plan; represents the current state corresponding safety factor; S203: Determine whether the fitness is within the safe range. If so, use the candidate plan with the maximum fitness as the actually adopted disassembly sequence plan; if not, go to the next step; S204: Screen the candidate solutions with the highest fitness for genetic algorithm optimization operations, where and ; S205: Use the result of the optimization operation as a new candidate plan, and check whether the new candidate plan meets the topological constraints. If it meets, go to S202; if it does not meet, go to S204.

3. The method according to claim 1, characterized in that, The S3 includes: S301: During the process of the actually adopted disassembly plan, obtain the current state of the monitoring data of the building structure ; S302: According to the monitoring data ; calculate the corresponding Bayesian posterior probability as the Bayesian risk score value , where represents that a structural instability event occurs; S303: Input the monitoring data into the trained machine learning model to obtain the risk score value output by the machine learning model ; S304: Synthesize the above-mentioned Bayesian risk score value and the risk score value output by the machine learning model , and perform dynamic risk prediction and continuous correction on the current state .

4. The method according to claim 3, wherein The status in S301 Monitoring data is defined as follows: Among them, represents the maximum displacement of all remaining components in the demolition sequence, reflecting the most severe deformation degree of the building structure in the current state; represents the maximum value of the ratio of the actual strain of all remaining components in the demolition sequence to their material yield strain, used to reflect the potential failure risk of the most dangerous components; represents the peak acceleration at the key parts of the building structure, reflecting the degree of inertial impact on the building structure, where the key parts are selected by experts through analysis and judgment; is the maximum inclination angle of the current building structure as a whole or a typical floor in the horizontal direction, used to reflect the overturning trend of the structure; represents the maximum value of the ratio of the residual stress to the ultimate strength of the material at the key parts of the current building structure, used to measure the remaining load-bearing capacity of the building structure.

5. The method according to claim 4, wherein In S302, based on the monitoring data , use the Bayesian inference formula to evaluate the posterior probability of the current state , and the formula is as follows ​ Among them, indicates that a structural instability event occurs; represents the state monitoring data; represents the static prior probability, which is determined by the hierarchical weight and safety factor of the component to be disassembled in the directed acyclic graph The smaller the topological depth of the component to be disassembled in the next step, the smaller the value, the higher the prior probability is assigned; represents the likelihood probability, which is obtained by statistically analyzing a large number of simulation samples accumulated during the genetic algorithm optimization process in step S2. To avoid the monitoring values in the multi-dimensional feature space from being difficult to completely coincide with the simulation samples, the tolerance range of can be appropriately expanded. That is to say, the simulation samples whose Euclidean distance from in the multi-dimensional feature space is less than are regarded as , where is a preset positive number; represents the monitoring probability, which can be generated through statistical analysis; represents the posterior probability, which is used as the basis for risk assessment of the current state; After receiving new monitoring data instead of using a static prior probability to calculate the posterior probability of the new state, the posterior probability of the previous state is used as the prior probability of the new state, as follows: 。 6. The method according to claim 5, characterized in that, The training of the machine learning model in S303 includes: At each state collect the monitoring data characteristics of the building structure and record the posterior probability obtained by Bayesian inference to construct a training set composed of samples; First, for each sample calculate its risk score ; The scoring function comprehensively considers the topological depth and Bayesian posterior probability of the component to be disassembled next, and the formula is as follows: in, Indicates the topological depth of the component to be disassembled in the next step; is a very small non-zero constant, with a value of 0.001; Indicates recent The sliding average of the posterior probability is used to reflect the difference between the current state and the previous state. The average risk of the state after the disassembly step; if the current number of disassembly steps is insufficient ,but The value is the actual number of steps. is the weighting coefficient used to represent the "structural dependency level" and "instability probability assessment" Assign weights between the two types of information and satisfy , Set by empirical data or training verification results; Using the monitoring data accumulated historically as input, clustering is performed in the high-dimensional state space by means of the K-means clustering method to obtain clusters , and the center of each cluster; for each sample , based on the similarity to the corresponding cluster center and the risk score , the representative score of the sample is calculated, and the representative score is used to judge the priority of the sample in the training set; the formula for the representative score is as follows: Among them, represents the monitoring data features in the sample and the cluster center of the belonging cluster of the Euclidean distance; is the weight coefficient and ; In each cluster, retain the samples with the highest scores as representative samples to form a training subset; input the training subset into a machine learning model for training.

7. The method according to claim 6, characterized in that, In S304, during the demolition process, when the machine learning model has been fully trained, start the "joint discrimination mechanism", and the rules are as follows: If , the current risk judgment is recognized, that is, the posterior probability obtained by Bayesian inference is considered correct; If , the current risk judgment is not recognized, that is, it is considered that there is a problem with the posterior probability obtained by Bayesian inference. The demolition operation is suspended, and after manual review, a secondary judgment is made; wherein, is a preset threshold value.

8. The method according to claim 1, wherein The S1 includes: S101: Obtain and analyze the relevant data of the building structure, determine the mutual dependency relationships of each component during the demolition process, and establish a dependency matrix; S102: Based on the dependency matrix, perform topological sorting on the components and draw a directed acyclic graph.

9. The method according to claim 2, characterized in that, The optimization operation of the genetic algorithm in S204 includes crossover and mutation.

Images