Method for quickly making demolition scheme of earthquake-damaged building based on genetic algorithm and reinforcement learning

Abstract

The present application belongs to the field of earthquake engineering and civil engineering, and particularly relates to a method for quickly formulating a demolition scheme of a damaged building based on a genetic algorithm and reinforcement learning. The method comprises the following steps: S1: establishing a data information base of a target damaged building to be demolished, mainly including various building structure types, structure plane arrangement, building site categories, and beam column member sizes; S2: establishing a finite element model for structure progressive collapse analysis according to the collected information; S3: determining the range of members allowed to be demolished during the collapse analysis of the demolished structure according to the collected information; S4: calculating a single demolition effect according to the finite element model for progressive collapse analysis; and S5: selecting a genetic algorithm or a Q-learning reinforcement learning algorithm according to whether the influence of the demolition sequence needs to be considered in actual conditions to obtain an optimal demolition scheme. The present application can be used for quickly formulating a demolition scheme of a severely damaged building in post-earthquake repair and reconstruction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of earthquake engineering and civil engineering, and specifically relates to a method for quickly formulating a demolition scheme for a damaged building based on a genetic algorithm and reinforcement learning. BACKGROUND

[0002] After an earthquake, severely damaged building structures cannot continue to be used and become "ruins" standing. In order to quickly repair and rebuild the earthquake-stricken area, it is necessary to quickly formulate an effective demolition scheme for the damaged building. The accurate identification of weak points is an indispensable and important prerequisite for the demolition of damaged buildings. This step requires the integration of professional knowledge in the engineering field and in-depth analysis of the complexity of the target building structure. Through systematic structural analysis, engineers and demolition experts can determine potential weak points, including material properties, structural strength, load distribution, and construction connections. This detailed analysis not only helps to accurately identify possible damage points, but also helps to assess which areas are most likely to be damaged, so as to develop the most appropriate demolition strategy to ensure the success and safety of the demolition project of the damaged building. However, a wrong building demolition scheme may result in serious property losses and prevent the target structure from collapsing as expected in one go, which means that additional engineering and resources are needed to clean up the debris and delay the progress of post-disaster repair and reconstruction. Therefore, the present application quickly formulates a demolition scheme for a damaged building through a genetic algorithm, improves the accuracy of weak point identification, and the feasibility of the demolition scheme is the main breakthrough of the present application.

[0003] The identification of weak points plays a crucial role in the demolition of damaged buildings, as it directly affects the safety and efficiency of the demolition process. Traditional methods of weak point identification usually rely on empirical methods or numerical simulation methods of mechanical analysis. Although these two methods are widely used, there are some problems. Traditional weak point identification methods usually require a lot of modeling and analysis work, which can lead to complexity and time-consuming problems. Traditional finite element analysis methods require detailed three-dimensional models and complex calculations in building demolition projects, which can consume a lot of time and resources. Model accuracy is crucial to the effectiveness of traditional methods. Traditional numerical simulation methods require accurate material properties and structural parameters in building demolition projects, otherwise they may lead to inaccurate predictions. Traditional methods are inefficient in handling large-scale data. Numerical simulation-based methods may require a large amount of calculation time when dealing with large-scale building demolition projects, which may not be suitable for quick decision-making and planning in actual engineering.

[0004] In recent years, machine learning and data-driven development has rapidly developed. Using machine learning methods or data-driven methods to quickly identify weaknesses can not only improve the efficiency and accuracy of engineering, but also significantly save tedious modeling and analysis work. The advantage of this method is that it can use large-scale data and intelligent algorithms to automatically identify potential weaknesses, thereby significantly shortening the time and resources required for the identification process. SUMMARY

[0005] In view of the problems of insufficient precision and low efficiency of the traditional method of formulating a demolition scheme based on mechanical analysis or numerical simulation, the present application aims to provide a method for quickly formulating a demolition scheme for a damaged building based on a genetic algorithm or a reinforcement learning Q-learning algorithm to identify weak points during demolition, in order to solve the defects and deficiencies of the traditional calculation method.

[0006] To this end, the technical solution invented in this paper is a genetic algorithm-based method for quickly formulating a demolition scheme for a damaged building. The required process parameters include the following:

[0007] S1: Establish a data information base for the target damaged building to be demolished, mainly including the building structure type, structure plan layout, building site category, and beam column member size;

[0008] S2: Establish a structure progressive collapse analysis finite element model based on the collected information;

[0009] S3: Determine the range of components that can be removed during the collapse analysis of the demolition structure based on the collected information;

[0010] S4: Calculate the single removal effect based on the progressive collapse analysis finite element model;

[0011] S5: Select the appropriate optimization algorithm based on whether the influence of the removal sequence needs to be considered to obtain the optimal removal scheme: if the influence of the removal sequence is not considered, select the genetic algorithm to optimize the selection of the building component removal range; if the influence of the removal sequence is considered, select the Q-learning reinforcement learning algorithm to optimize the selection of the building component removal range and the removal sequence.

[0012] Preferably, the specific method of step S2 is as follows: using finite element analysis software, combining multi-scale modeling methods and the structure information in step S1, a progressive collapse analysis model of the damaged building structure is established.

[0013] Preferably, the genetic algorithm mainly includes population selection, chromosome crossover and chromosome mutation; first, encoding is carried out based on the demolition model, and an initial population is randomly generated; individual fitness is calculated, selection operation and mutation operation are carried out, and an optimization result is returned after a termination condition is reached; the individual represents each building demolition scheme, the chromosome represents the coding of each demolition scheme, the gene represents a coding element, that is, whether each column is demolished, and the survivability represents the result of each scheme in the fitness function.

[0014] Preferably, the ratio of the vertical displacement of the damaged building after each demolition to the number of demolished columns is taken as the evaluation value of the fitness function, the set parameters of each demolished column are taken as the algorithm independent variables, the set parameters of the demolition object obtained by iterative calculation of the genetic algorithm are used, and the parameter values after iteration are further used for cyclic calculation until the result meets the requirement of the final demolition effect.

[0015] Preferably, the specific method for encoding based on the demolition model is to use a binary vector for encoding. .

[0016] Preferably, the Q-learning reinforcement learning algorithm includes initializing the related parameters of the Q-Learning algorithm, selecting a polynomial decay function to control the change of the learning rate, constructing a Q table to store the state-action values, using an epsilon-greedy strategy to select an action, updating the estimated value of the corresponding state-action value in the Q table according to the current state, action, reward, next state and end flag, and searching for the action with the highest Q value in the Q table to complete the optimization and solution of the target function.

[0017] The present application has the advantages that: compared with the traditional experience method, the traditional mechanical analysis and the numerical simulation method, the present application proposes a matrix position coding mode, different positions represent different to-be-demolished components in the structure, the conversion of the demolition components and the population individual genes can be more intuitively represented, and the present application proposes a genetic algorithm or a Q-learning algorithm to solve the optimal building demolition scheme, and has the advantages of global search, high search efficiency and high accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the building demolition scheme rapid formulation method based on the genetic algorithm and the reinforcement learning of the present application;

[0019] Figure 2 is a prototype structure diagram of the building demolition scheme rapid formulation method based on the genetic algorithm and the reinforcement learning of the present application;

[0020] Figure 3 is a coding diagram of the building demolition scheme rapid formulation method based on the genetic algorithm and the reinforcement learning of the present application;

[0021] Figure 4 is a genetic algorithm flowchart;

[0022] Figure 5 is a Q-learning reinforcement learning algorithm flowchart;

[0023] Figure 6 is a simple example of genetic algorithm;

[0024] Figure 7 is a simple example of Q-learning reinforcement learning algorithm, wherein (a) matrix R; (b) Q table. DETAILED DESCRIPTION

[0025] In order to make the objects, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below with the aid of drawings and examples. The specific examples described herein are only used to explain the present application and are not intended to limit the scope of the present application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The use of the terms "including", "containing", "having" and "including" herein is used generically and is equivalent to the term "comprising" and is not intended to be limiting of the present application.

[0027] As shown in Figure 1 A method for quickly formulating a demolition scheme of a damaged building based on a genetic algorithm, characterized in that it comprises the following steps:

[0028] Step 1: Establish a data information base of the target building to be demolished, mainly including various building structure types, structural plan layout, building site category, beam column member size, etc.

[0029] Step 2: Establish a structure progressive collapse analysis finite element model according to the collected information. A steel frame finite element model is established according to the data information base of the target building to be demolished in step one. A numerical model of the damaged building is established using the OpenSees finite element software.

[0030] As shown in Figure 2The finite element model prototype of this embodiment is a four-story steel frame structure. The steel beams and columns are simulated using Beam elements, the reinforced concrete slabs are simulated using Shell elements, and the joints are simulated using macro elements. In the design of the lateral force resisting system, the special requirements of the special flexural frame system with weakened beam flanges are followed. The flexural frames are arranged in the side spans of the structure. The other parts of the structure are designed according to the vertical force resisting system. The column and beam sections of the vertical force resisting system of the three structures are W14x90 and W24x55, respectively. The connections of the internal vertical force resisting frames are conventional single-shear plate beam-column connections. The beams and columns are made of ASTM A992 steel (the nominal yield strength is 345 MPa, and the elastic modulus is 2x10 11 N / m2)

[0031] Step 3: Determine the range of components allowed to be removed during the collapse analysis of the demolished structure.

[0032] As Figure 3 shown, all the load-bearing columns (a total of 16) of the structure are within the range of components allowed to be removed.

[0033] Step 4: Calculate the effect of single removal based on the finite element model of progressive collapse analysis.

[0034] As Figure 3 shown, all the load-bearing columns are numbered in order from low to high floors and from left to right. The specific component number to be removed is determined based on a random number. The corresponding component is removed by using the "Remove" command in OpenSees. The removed object is set in the working condition. The vertical displacement of the remaining structure is calculated as the evaluation index of the removal effect.

[0035] Step 5: Determine whether to consider the influence of removal sequence based on the actual situation to select genetic algorithm or Q-learning reinforcement learning algorithm for optimization and solution:

[0036] (1) Based on the genetic algorithm, the removal scheme of the damaged building during the removal of the components is optimized to obtain an optimal removal scheme. The removal scheme here specifically refers to the set of removal points selected each time. First, the removal model is coded, then an initial population is randomly generated, the individual fitness is calculated, then selection and mutation operations are performed, and finally the optimization result is returned after the termination condition is reached.

[0037] As Figure 3 shown, the specific method for coding based on the removal model is to use a binary vector for coding:

[0038] ;

[0039] Wherein, size represents the total number of demolition objects, here taking four-layer steel frame structure as an example, there are 4*4=16 bearing columns; all the demolition objects are sorted according to the floor and position, the 4 bearing columns of the first floor are numbered in order from left to right as 1-4, and the bearing columns of the remaining floors are numbered in order from left to right as 1-4. Indicates whether the bearing column numbered i in the demolition object is used as a demolition point in this demolition, if , it represents that the i-th column is used as a demolition point in this demolition; if , it represents that the i-th column is not used as a demolition point in this demolition; for example, the chromosome position of a certain demolition model is , it represents that the third column of the first floor and the first column of the second floor are demolished.

[0040] As shown in Figure 4 , the specific steps of the genetic algorithm are as follows:

[0041] (1) Population initialization: according to this coding mode, randomly generate demolition object information in the set range to obtain an initial population composed of several chromosomes, and appropriately stretch the population to increase individual differences and ensure convergence accuracy;

[0042] (2) Define fitness function: the fitness value is related to the demolition efficiency of single demolition, which is determined by the structure vertical displacement and the number of demolitions, the higher the ratio of structure vertical displacement to the number of demolitions, the higher the fitness; the fitness function is as follows:

[0043] (1);

[0044] Wherein, in formula (1) represents the coding information of the demolition object of single demolition, represents the structure vertical displacement distance of single demolition, represents the number of demolition columns of single demolition;

[0045] (3) Selection operation: the selection operation adopts roulette selection method, the probability of each chromosome individual being selected is proportional to the size of its fitness function value, and the optimized individual is inherited to the next generation or new individuals are generated through pairing crossover and then inherited to the next generation;

[0046] (4) Crossover operation: take each floor's four columns as a gene segment, the number of gene segments on each chromosome is equal to the number of floors, select two-point crossover method to act on the population, randomly exchange the gene segments between any positions in two chromosomes, and determine the search range in the crossover operator;

[0047] (5) Mutation operation: The mutation operator is applied to each gene segment of the chromosome according to a certain mutation probability. By adjusting the mutation probability, the convergence curve is observed and calculated to ensure the convergence speed and avoid overfitting.

[0048] (6) Evolutionary operation: The optimal dismantling scheme is adopted by evolutionary search. After selection, crossover and mutation operations, the next generation of population is obtained.

[0049] Using genetic algorithms to solve building demolition schemes that do not consider the demolition order, for example... Figure 6 As shown, it is assumed that the population consists of two individuals. The implementation steps are based on... Figure 4 The flowchart shown is as follows:

[0050] Population initialization: Randomly generate two individuals for the population, with the following codes: ; ;

[0051] Crossover operation: Taking each floor as a unit, the four pillars of each floor are taken as a gene segment. The number of gene segments on each chromosome is equal to the number of floors. The two-point crossover method is applied to the population, and the gene segments between corresponding positions in the two chromosomes are randomly exchanged. Here, the gene segments corresponding to the third floor are exchanged.

[0052] Mutation operation: The mutation operator is applied to each gene segment of the chromosome according to a certain mutation probability. Here, we will not perform mutation operation for the time being.

[0053] Calculate the fitness function: Based on the individual code, the corresponding components are dismantled, and the fitness function value is calculated using the OpenSees function.

[0054] Selection operation: The selection operation adopts the roulette wheel selection method. The probability of each chromosome individual being selected is proportional to its fitness function value. Optimized individuals are inherited to the next generation or new individuals are generated through pairing and crossover and then inherited to the next generation. The optimization results are shown in Table 1.

[0055] Table 1 shows the optimization results of the genetic algorithm.

[0056] .

[0057] like Figure 5 As shown, the specific steps of Q-learning are as follows:

[0058] (1) Initialize relevant parameters: Given the exploration rate Learning rate and discount factor And set the algorithm iteration number N to initialize the state-action value function Q(s,a), and set all functions in this function to 0;

[0059] (2) Selection: adopt greedy strategy to select an action, and get a new state according to the current state and the selected action and the corresponding reward value r;

[0060] (3) Update state: when updating the state each time, update the estimated value of the corresponding state-action value in the Q table according to the current state, action, reward, and next state, and the specific update method is:

[0061] Q s , a ← Q s , a + α [ R + gamma max Q s ' , a - Q s , a ] (2) ;

[0062] wherein in formula (2), represents the Q value corresponding to the action with the maximum Q value among all possible actions in the state

[0063] (4) Optimization: find the action with the highest Q value in the Q table by searching, that is, the optimal action under the current state, which is used as the next action of the agent. The state corresponding to this action is the optimal solution state of the problem, and the optimization solution of the objective function is completed. The Q table stores the Q value of state-action pairs, which represents the cumulative reward obtained by taking a certain action from a certain state. S5 needs to search for the action with the highest Q value in the Q table under the current state; for the current state s, all possible actions a are traversed, and then the action a* with the highest Q value is found;

[0064] (5) Optimization solution: after finding the optimal action a*, it is used as the next action of the agent, and the state corresponding to a* is used as the optimal solution state of the problem. This state can make the objective function reach the minimum value. Since the Q value represents the cumulative reward obtained by taking a certain action, the action with the highest Q value will obtain the maximum cumulative reward in the long term, thereby making the objective function, that is, the optimal scheme of building demolition.

[0065] The Q-learning algorithm is used to solve the building demolition scheme considering the demolition sequence, as shown in Figure 7 , and the specific process is as follows:​

[0066] State space: the set of environment states that the agent can obtain , the same encoding as in step S6, according to whether the column at each demolition point is demolished as the state division basis, denotes the i-th result after demolition, and the total state number function is as follows:

[0067] (3);

[0068] wherein, in formula (3) denotes the total number of demolition points;

[0069] Action space: the set of environment states that the agent can obtain , according to which column at which demolition point is demolished as the action division basis, and the total number of actions is equal to ;

[0070] Reward function: Q-learning requires the agent to continuously try and error for the learning process, during which the selected action under the state is evaluated for good or bad to obtain an immediate reward value . Here, the vertical displacement of the structure is used as the reward value;

[0071] Initialization of related parameters: given the exploration rate , learning rate and discount factor , and set the number of iterations N of the algorithm, initialize the state-action value function Q(s, a), and assign all functions in this function to 0;

[0072] Action selection: select an action using the -greedy strategy, and obtain a new state and the corresponding reward value r according to the current state and the selected action. Here, state 0 (without demolition) is selected as the initial state;

[0073] State update: each time the state is updated, the estimated value of the corresponding state-action value in the Q table is updated according to the current state, action, reward, and next state. The initial state in the previous step corresponds to k possible behaviors (through k possible actions to 16 states of co-demolishing 1 column), and randomly, we select to go to state 3 (demolish a layer of the third column), and then observe the fourth row of matrix R (corresponding to state 3), which corresponds to k-1 possible behaviors: go to 15 states of co-demolishing 2 columns. Here, the learning rate is set to 0.1;

[0074] Q(0,3) = R(0,3) + 0.1*(-max{|Q(3,1)|, |Q(3,2)|, … |Q(3,15)|} = -0.18 - 0.1x2.97 = -0.477;

[0075] Optimization: By searching in the Q table, the action with the highest Q value is found, that is, the optimal action in the current state, which is taken as the next step action of the agent. The state corresponding to this action is the optimal solution state of the problem, and the optimization solution of the objective function is completed. The Q table stores the Q value of state-action pairs, which represents the cumulative reward obtained by taking a certain action from a certain state. S5 needs to search for the action with the highest Q value in the Q table under the current state; for the current state s, all possible actions a are traversed, and then the action a* with the highest Q value is found;

[0076] Optimization solution: After finding the optimal action a*, it is taken as the next step action of the agent, and the state corresponding to a* is taken as the optimal solution state of the problem. This state can make the objective function reach the minimum value. Since the Q value represents the cumulative reward obtained by taking a certain action, the action corresponding to the highest Q value will obtain the maximum cumulative reward in the long term, so that the objective function, that is, the optimal scheme of building demolition, is reached.

[0077] The above-mentioned is only a hypothetical implementation case of the present application, and is not used to limit the present application. Any modification, equivalent replacement or improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for rapidly formulating demolition plans for earthquake-damaged buildings based on genetic algorithms and reinforcement learning, characterized in that, Includes the following steps: S1: Establish a database of information on target earthquake-damaged demolished buildings, mainly including the structural type of each building, structural plan layout, building site category, and dimensions of beam and column components; S2: Establish a finite element model for progressive collapse analysis of the structure based on the collected information; S3: Determine the permissible range of components to be removed during structural collapse analysis based on the collected information; S4: Calculate the effect of a single demolition based on the finite element model of continuous collapse analysis; S5: Select the appropriate optimization algorithm based on whether the demolition order needs to be considered in the actual situation to obtain the optimal demolition scheme: if the demolition order is not considered, select the genetic algorithm to optimize the demolition range of building components; if the demolition order is considered, select the Q-learning reinforcement learning algorithm to optimize the demolition range and demolition order of building components; the genetic algorithm mainly includes population selection, chromosome crossover, and chromosome mutation; firstly, the initial population is randomly generated based on the finite element model of the continuous collapse analysis. Calculate individual fitness, perform selection and mutation operations, and return the optimization results after the termination condition is met; Each individual represents a demolition plan for a damaged building, a chromosome represents the code for each demolition plan, a gene represents the coded element (whether or not each column is demolished), and survivability represents the result of each plan in the fitness function. The ratio of the vertical displacement of the damaged building to the number of demolished columns after each demolition is used as the fitness function evaluation value. The set parameters of each demolished column are used as the algorithm's independent variables. The set parameters of the demolition objects are obtained by iterative calculation using a genetic algorithm. The iterative parameter values ​​are then used to further iterate until the result meets the final demolition effect requirements.

2. The method for rapidly formulating demolition plans for earthquake-damaged buildings based on genetic algorithms and reinforcement learning according to claim 1, characterized in that: The specific method for step S2 is as follows: Using finite element analysis software, combined with multi-scale modeling methods and the structural information in step S1, a continuous collapse analysis model of the earthquake-damaged building structure is established.

3. The method for rapidly formulating demolition plans for earthquake-damaged buildings based on genetic algorithms and reinforcement learning according to claim 1, characterized in that: The specific method for encoding based on the finite element model of the continuous collapse analysis is to use binary vector encoding. .

4. The method for rapidly formulating demolition plans for earthquake-damaged buildings based on genetic algorithms and reinforcement learning according to claim 1, characterized in that: The Q-learning reinforcement learning algorithm includes initializing the relevant parameters of the Q-Learning algorithm, using a polynomial decay function to control the change of the learning rate, constructing a Q-table to store state-action values, using an ε-greedy policy to select an action, updating the estimated value of the corresponding state-action value in the Q-table based on the current state, action, reward, next state, and end flag, and optimizing the objective function by searching in the Q-table to find the action with the highest Q value.

Images