Wind vibration actual measurement analysis method for large-span steel structure roof

Abstract

The invention discloses a wind vibration actual measurement analysis method for a large-span steel structure roof, and belongs to the field of steel structure engineering monitoring, and the actual measurement analysis method comprises the following specific steps: I, based on the geometric and dynamic characteristics of a large-span steel structure roof structure, optimizing the arrangement of each sensor, and constructing a multi-modal heterogeneous sensing network; iI, acquiring wind speed and wind direction time sequence data according to the multi-mode heterogeneous sensing network, establishing a reconstruction model of a non-stationary wind field, and predicting time-varying distribution of wind pressure on the surface of the structure; according to the method, high-fidelity dynamic reconstruction of the wind field is realized, the authenticity of wind load input is improved, the data quality and information diversity are improved, the prediction precision and timeliness of the wind pressure field are improved, and the scientificity and the monitoring effect of an overall layout strategy are enhanced; according to the method, the accuracy of long-term performance evaluation can be improved, environment coupling simulation of structural response is achieved, the dynamic adaptive capacity of analysis is enhanced, the structural rigidity evolution rule is depicted finely, and the robustness and local recognition capacity of the model under uncertain data are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of steel structure engineering monitoring, and particularly to a method for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof. Background Art

[0002] With the complication of urban spatial functions and the growth of the demand for large public buildings, lightweight, large-span, and flexible design schemes are widely adopted in steel roof projects of large-span space structures such as stadiums, exhibition centers, and transportation hubs. However, the problem of wind-induced vibration response under the action of wind load excitation is becoming increasingly prominent, and the structure faces safety risks such as wind-induced vibration, fatigue accumulation, and potential local damage during its service life. Traditional analysis methods mostly rely on static load assumptions and simplified models, which are difficult to reflect the time-varying dynamic effects under the real wind field and lack the support of multi-source measured data. In addition, problems such as dense sensor layout, complex data processing, and low damage identification accuracy limit the wide application of existing wind-induced vibration monitoring systems. Therefore, it is particularly important to invent a method for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof.

[0003] The existing methods for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof cannot perform high-fidelity dynamic reconstruction of the wind field, reducing the authenticity of wind load input, and the collected data is single, reducing the data quality, and the prediction accuracy and timeliness of the wind pressure field cannot be improved; in addition, the accuracy of long-term performance evaluation of the existing methods for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof is low, and the environmental coupling of the structural response cannot be simulated, reducing the dynamic adaptability of the analysis, and the robustness and local identification ability under uncertain data are poor; for this reason, we propose a method for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects existing in the prior art, and to propose a method for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof.

[0005] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for actual measurement and analysis of wind-induced vibration of a large-span steel structure roof, and the specific steps of the actual measurement and analysis method are as follows:

[0007] Ⅰ. Based on the geometric and dynamic characteristics of the large-span steel structure roof, optimize the layout of each sensor, and construct a multi-modal heterogeneous sensing network;

[0008] Ⅱ. According to the time series data of wind speed and wind direction obtained by the multi-modal heterogeneous sensing network, establish a reconstruction model of the non-stationary wind field, and predict the time-varying distribution of wind pressure on the surface of the structure;

[0009] Ⅲ. Modify the structural stiffness characteristics of the large-span steel structure roof according to the thermal-humidity sensitivity of the material, and analyze the changing trend of the wind speed in real time, and dynamically adjust the sampling frequency;

[0010] Ⅳ. Use a drone to vibrate and scan the surface of the large-span steel structure roof to generate a vibration displacement nephogram and identify the local stiffness of the large-span steel structure roof;

[0011] Ⅴ. Based on the multi-modal heterogeneous sensing network and drone acquisition, establish a simulation model of the large-span steel structure roof and evaluate the response characteristics under the safety margin and ultimate state.

[0012] As a further solution of the present invention, the specific steps of optimizing the layout of each sensor and constructing a multi-modal heterogeneous sensing network in step I are as follows:

[0013] S1.1: Establish a three-dimensional simulation model of the large-span steel structure roof according to the actual engineering drawings and BIM model, select the corresponding element types to describe each component of the large-span steel structure roof, and then set the parameters of elastic modulus, density, Poisson's ratio and cross-section properties;

[0014] S1.2: According to the constructed three-dimensional simulation model, establish a set of structural nodes, and then set the grid scale according to the complexity of the structure and the required response resolution preset, perform finite element discretization on the surface of the structure to form a discrete candidate area for layout points, and based on the set of structural nodes, establish a set of candidate points for layout points after division, calculate the sensitivity of each candidate point in the set of candidate points for layout points to the structural response, and screen out the candidate points whose sensitivity does not reach the preset threshold, perform characteristic modeling on each group of sensors, and construct a sensor response matrix according to the physical quantities and measurement accuracies of different sensors;

[0015] S1.3: Set the maximum layout quantity of each type of sensor, use integer coding to label the layout states of different candidate layout positions, randomly generate an initial population according to the preset upper limit of the layout quantity of each type of sensor, each individual in the population represents a set of sensor layout schemes, and represent different sensor layout schemes as sets of various state labels, where each bit represents the layout state of a candidate layout position;

[0016] S1.4: Calculate the modal confidence values of each individual in the population, and check whether each individual meets the preset sensor quantity constraint. According to the modal confidence values of each individual and the constraint satisfaction situation, obtain the fitness values of each individual, adopt the roulette wheel selection method, and select multiple groups of individuals from the current population according to the fitness from high to low as the parents for the next generation of crossover;

[0017] S1.5: Randomly select two groups of individuals from the parent generation for crossover. Randomly select one or two crossover points from each of the two groups of individuals, and exchange the layout status label segments of the corresponding two groups of individuals to form two new offspring. Randomly select one or more layout positions among the offspring individuals, randomly replace the value at this position with any layout status, and at the same time detect whether the new offspring meet the preset sensor quantity constraint. If not, eliminate the corresponding offspring. After that, calculate the fitness value of the new offspring after crossover and mutation;

[0018] S1.6: If the fitness value of the new offspring is higher than that of the parent generation, retain the new offspring; otherwise, retain the parent generation. Construct a new population with the screened parent generation and new offspring, and re - perform parent selection, crossover and mutation, and population construction until the change value of the optimal fitness value of each individual in the population converges to the preset range;

[0019] S1.7: Traverse the fitness values of each group of individuals in the final population, and take the individual with the highest fitness as the optimal sensor layout scheme. Then synchronize the optimal sensor layout scheme to the three - dimensional simulation model of the long - span steel structure roof structure, and mark the sensor layout status of each layout point. Establish a multi - modal heterogeneous sensing network according to the optimal sensor layout scheme.

[0020] As a further solution of the present invention, the specific steps of establishing the reconstruction model of the non - stationary wind field in step II are as follows:

[0021] S2.1: Use the multi - modal heterogeneous sensing network to obtain long - time - series wind speed and wind direction data at multiple measuring points around the long - span steel structure roof structure. Decompose the long - time - series wind speed data into an average term and a perturbation term, and use CFD software to perform unsteady simulation on typical wind direction and speed combinations to obtain the instantaneous wind pressure and velocity field around the long - span steel structure roof structure;

[0022] S2.2: Perform joint mapping on the measured wind speed and wind direction data and the CFD results. Simulate the spatial correlation and time evolution of the wind speed field by using a time - varying covariance function to establish a non - stationary wind field model, and sample the covariance between the wind speed data collected at each measuring point at different times to establish a time - varying wind speed field with corresponding statistical characteristics;

[0023] S2.3: Perform principal component decomposition on the time - varying wind speed field generated by CFD simulation, map and obtain the time coefficients of the principal components according to the real - time collected wind speed and wind direction data, and use CFD simulation statistics to generate the corresponding principal modal spatial distribution. Based on the real - time updated time coefficients, synchronously reconstruct the non - stationary wind field model;

[0024] S2.4: Use the remaining measurement points in the measured wind speed and wind direction data that were not involved in the modeling to verify the non-stationary wind field model, and calculate the error value between the reconstructed wind speed and the measured wind speed. If the error exceeds 5%, it means that the constructed non-stationary wind field model cannot be used for wind-induced vibration load input, and the non-stationary wind field model needs to be adjusted again.

[0025] As a further solution of the present invention, the specific steps of predicting the time-varying distribution of wind pressure on the structure surface in step II are as follows:

[0026] S3.1: Extract the wind pressure distribution data on the structure surface at different times from the CFD simulation results as the target labels, and at the same time use the non-stationary wind speed field reconstruction results as the input features. Construct a complete training sample through the true labels and input features, collect multiple groups of training samples at different time steps, and perform normalization processing on each group of collected training samples;

[0027] S3.2: Establish a wind pressure prediction model based on a spatio-temporal convolutional architecture. The model includes an input layer, a spatial convolutional layer, a temporal convolutional layer, an upsampling layer, and an output layer. Divide the training samples corresponding to each time step into small batch sample data according to the set fixed time window, and randomly select a group of small batch sample data as the input data and transmit it to the wind pressure prediction model;

[0028] S3.3: The input layer of the wind pressure prediction model transmits the input data of each time step to the spatial convolutional layer. Then, multiple groups of spatial convolutional layers sequentially perform convolution operations on the input data to extract the change pattern of the wind speed vector in the spatial region and output the local spatial feature map. Then, input each generated local spatial feature map into the temporal convolutional layer. The temporal convolutional layer, based on the gating mechanism, captures the dependency relationship between the local spatial feature maps of each time step and outputs the temporal evolution information;

[0029] S3.4: After inputting the local spatial feature maps of each time step and the corresponding temporal evolution information into the upsampling layer, the upsampling layer restores the local spatial feature maps of each time step to the spatial resolution required for the wind pressure field on the structure surface through deconvolution operations, and generates the wind pressure spatial distribution representing each sample at the current prediction moment. Then, transmit the processed local spatial feature maps to the output layer. The output layer converts the local spatial feature maps of each time step into the final prediction result through linear mapping and outputs the corresponding predicted wind pressure field, where each predicted point value represents the wind pressure estimation value of the corresponding grid position on the structure surface at the current time step;

[0030] S3.5: Use the weighted mean square error function to calculate the error value between the predicted wind pressure field and the true label, and backpropagate the error value layer by layer from the output layer of the wind pressure prediction model to the input layer. At the same time, calculate the gradient value of the error value for each layer parameter, and update the network parameters through the Adam optimizer;

[0031] S3.6: After each round of training, evaluate the prediction accuracy of the updated wind pressure prediction model using the small batch of sample data that did not participate in the training, calculate the prediction error. If the prediction error range is less than 10%, the model can be used for actual prediction; otherwise, retrain the wind pressure prediction model and update the parameters. Deploy the trained wind pressure prediction model to the wind-induced vibration response analysis platform, receive the wind speed field input in real time, and automatically output the future wind pressure distribution to form a continuous loading scenario close to the actual situation.

[0032] As a further solution of the present invention, the specific steps of modifying the structural stiffness characteristics of the long-span steel structure roof according to the thermal-humidity sensitivity of the material in step III are as follows:

[0033] S4.1: Use a multi-modal heterogeneous sensing network to collect the temperature and humidity change data during the operation of the long-span steel structure roof in real time. According to the collection time, process the temperature and relative humidity into an environmental factor set that affects the stiffness change. Based on the long-term effects of thermal expansion of steel and humid environment on the structural performance, establish a thermal-humidity sensitive function of the elastic modulus to obtain the environmentally corrected material elastic modulus.

[0034] S4.2: Based on the three-dimensional simulation model of the long-span steel structure roof and the environmentally corrected material elastic modulus, convert the stiffness coefficients of each element into a dynamic stiffness matrix. Then, select the key nodes that are most sensitive to the overall stiffness change at the mid-span, support connection, and variable cross-section area of the long-span roof structure as key nodes, and extract the main diagonal elements corresponding to the structural key nodes from the dynamic stiffness matrix to construct a stiffness sequence set.

[0035] S4.3: Smooth each group of main diagonal elements in the stiffness sequence set by moving average, and at the same time use normalization processing to make the main diagonal elements of each component within a unified scale range. Then, use empirical mode decomposition to decompose each group of main diagonal elements in the stiffness sequence set into a trend term, a periodic term, and a residual term.

[0036] S4.4: Extract the quantitative indicators of the long-term change slope and inflection point position in the trend term to identify the stiffness trend. Compare and analyze the extracted stiffness trend and the periodic term with the temperature and humidity change data respectively, and then analyze the variance level and fluctuation mode corresponding to the residual term. Based on the analysis results, obtain the law of the stiffness evolution of each component of the long-span steel structure roof over time, and encapsulate it as a time-driven function.

[0037] S4.5: Based on the time-driven function, correct the nominal initial stiffness, and combine the corrected stiffness with the corresponding dynamic stiffness matrix of the three-dimensional simulation model of the current long-span steel structure roof to construct a time-dependent overall stiffness matrix and establish a time-varying dynamic equilibrium equation.

[0038] S4.6: At each time step, based on the real-time detected environmental data, the overall stiffness matrix is updated in real time, a complete simulation time axis is constructed, and at the same time, the corresponding time-varying dynamic equilibrium equation is solved to generate the structural responses of each part of the surface of the long-span steel structure roof at the current time.

[0039] As a further solution of the present invention, the specific steps for identifying the local stiffness of the long-span steel structure roof in step IV are as follows:

[0040] S5.1: The modal frequencies and mode shape information of the long-span steel structure roof at multiple measuring points are collected or simulated in real time through a multi-modal heterogeneous sensing network, and the spatial distribution of the measuring points is processed with normalized coordinates to convert the coordinates of each measuring point in the measuring point space into coordinates within the unit domain. Then, based on the collected groups of modal frequencies and mode shape information, the corresponding estimated flexibility matrix is established;

[0041] S5.2: The long-span steel structure roof is divided into multiple groups of sub-regions. According to the relative stiffness changes of each sub-region in the long-span steel structure roof analyzed in real time, a set of stiffness loss variables for each sub-region is constructed. And based on each set of stiffness loss variables and the estimated flexibility matrix, the posterior probability of each set of stiffness loss variables is calculated;

[0042] S5.3: The set of stiffness loss variables with a posterior probability lower than the preset threshold is used as the initial damage parameter, the number of iterations is set, and the normal distribution is used as the proposal distribution for sampling new samples. Then, for each iteration round, new parameters are sampled from the proposal distribution as candidate samples, and the posterior ratio of the initial damage parameter and the candidate sample is calculated;

[0043] S5.4: Based on the Metropolis-Hastings rule, the acceptance probability of the candidate sample is calculated. Then, a random number is generated. If the random number is less than the acceptance probability, the candidate sample is accepted; otherwise, the candidate sample is rejected. The selection of candidate samples and replacement are repeated until the preset number of iteration rounds is reached, and the final damage sample set is output;

[0044] S5.5: Statistical analysis is performed on the damage sample set, the maximum a posteriori estimate value of the damage parameter of each region is calculated, and at the same time, the regions with the maximum a posteriori estimate value higher than 0.2 are marked as damaged regions.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. The in-situ measurement and analysis method for wind-induced vibration of long-span steel structure roofs establishes a finite element model based on a BIM model and generates candidate areas for sensor placement. The layout scheme of multi-modal heterogeneous sensors is optimized by combining genetic algorithms with modal confidence. Then, the in-situ wind speed and direction data are collected using a sensor network, and a time-varying covariance-driven non-stationary wind field model is constructed by combining CFD unsteady simulation. After the wind speed field is decomposed by principal component analysis, time coefficients are obtained by mapping with real-time data to realize wind field reconstruction. Based on the constructed non-stationary wind speed field and CFD label data, a spatio-temporal convolutional neural network is trained to predict the surface wind pressure distribution of the structure. Through error evaluation and optimization updates, real-time high-precision prediction of the wind pressure field is carried out, providing live-level load input for wind-induced vibration response simulation, realizing high-fidelity dynamic reconstruction of the wind field, improving the authenticity of wind load input, enhancing data quality and information diversity, improving the prediction accuracy and timeliness of the wind pressure field, and enhancing the scientific nature of the overall layout strategy and monitoring effect.

[0047] 2. The present invention collects real-time temperature and humidity data during the operation of the structure, constructs a thermo-hygro sensitive elastic modulus model, combines three-dimensional simulation to establish a dynamic stiffness matrix, extracts the main diagonal elements of the stiffness of key nodes, performs moving average and empirical mode decomposition to identify the long-term trend and periodic fluctuations of the stiffness, and constructs a time-driven function to correct the overall stiffness to form a time-varying dynamic model. Then, based on the collected structural modal frequency and mode shape information, the flexibility matrix is calculated and sub-regions are divided. A stiffness damage variable is constructed according to the real-time stiffness change, and the MCMC algorithm is used for posterior sampling to identify the damage distribution. Finally, the potential damage area of the structure is output, which can improve the accuracy of long-term performance evaluation, realize the environmental coupling simulation of structural response, enhance the dynamic adaptability of the analysis, be able to finely depict the evolution law of structural stiffness, and enhance the robustness and local identification ability of the model under uncertain data. Description of the Drawings

[0048] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0049] Figure 1 It is a flow chart of an in-situ measurement and analysis method for wind-induced vibration of a long-span steel structure roof proposed by the present invention. Detailed Embodiments

[0050] Example 1, referring to Figure 1 A method for in-situ measurement and analysis of wind-induced vibration of a long-span steel structure roof, and the specific steps of the in-situ measurement and analysis method are as follows:

[0051] Based on the geometric and dynamic characteristics of the long-span steel structure roof, the layout of each sensor is optimized, and a multi-modal heterogeneous sensor network is constructed.

[0052] Specifically, a three-dimensional simulation model of the long-span steel structure roof is established based on the actual engineering drawings and the BIM model. The corresponding element types are selected to describe each component of the long-span steel structure roof. Then, parameters such as elastic modulus, density, Poisson's ratio, and section properties are set. According to the constructed three-dimensional simulation model, a set of structural nodes is established. Then, according to the complexity of the structure and the preset response resolution requirements, the grid scale is set, and the surface of the structure is discretized by finite elements to form a discrete set of candidate points for layout. Based on the set of structural nodes, a set of candidate points for layout after division is established. The sensitivity of each candidate point in the set of candidate points for layout to the structural response is calculated, and the candidate points whose sensitivity does not reach the preset threshold are screened out. The characteristics of each group of sensors are modeled, and according to the physical quantities and measurement accuracies of different sensors, a sensor response matrix is constructed. The maximum number of layouts for each type of sensor is set. Integer coding is used to label the layout states of different candidate layout positions. According to the preset upper limit of the number of layouts for each type of sensor, a group of initial populations is randomly generated. Each individual in the population represents a layout scheme for a group of sensors. Different sensor layout schemes are represented as sets of various state labels, where each bit represents the layout state of a candidate layout position. The modal confidence values of each individual in the population are calculated, and it is checked whether each individual meets the preset sensor number constraint. According to the modal confidence values of each individual and the constraint satisfaction situation, the fitness values of each individual are obtained. The roulette wheel selection method is adopted. According to the fitness from high to low, multiple groups of individuals are selected from the current population as the parents for the next generation of crossover. Two individuals are randomly selected from the parents for crossover. One or two crossover points are randomly selected from the two individuals respectively, and the layout state label segments of the corresponding two individuals are exchanged to form two new offspring. One or more layout positions of the new offspring individuals are randomly selected, and the value at this position is randomly replaced with any layout state. At the same time, it is detected whether the new offspring meets the preset sensor number constraint. If not, the corresponding offspring is removed. Then, the fitness values of the new offspring after crossover and mutation are calculated. If the fitness value of the new offspring is higher than that of the parents, the new offspring is retained; otherwise, the parents are retained. The selected parents and the new offspring are constructed into a new population, and the parent selection, crossover and mutation, and population construction are carried out again until the change value of the optimal fitness value of each individual in the population converges to the preset range. The fitness values of each group of individuals in the final population are traversed, and the individual with the highest fitness is used as the optimal sensor layout scheme. Then, the optimal sensor layout scheme is synchronized to the three-dimensional simulation model of the long-span steel structure roof, and the sensor layout states of each layout point are marked. A multi-modal heterogeneous sensing network is established based on the optimal sensor layout scheme.

[0053] Based on the multi-modal heterogeneous sensing network, the time series data of wind speed and wind direction are obtained, a reconstruction model of the non-stationary wind field is established, and the time-varying distribution of wind pressure on the surface of the structure is predicted.

[0054] Specifically, long-term wind speed and wind direction data are obtained at multiple measuring points around the long-span steel structure roof using a multi-modal heterogeneous sensing network. The long-term wind speed data is decomposed into an average term and a perturbation term. Unsteady simulations are performed on typical wind direction and speed combinations using CFD software to obtain the instantaneous wind pressure and velocity field around the long-span steel structure roof. The measured wind speed and wind direction data are jointly mapped with the CFD results. By using a time-varying covariance function to simulate the spatial correlation and time evolution of the wind speed field, a non-stationary wind field model is established. The covariance between the wind speed data collected at each measuring point at different times is sampled to establish a time-varying wind speed field with corresponding statistical characteristics. The time-varying wind speed field generated by CFD simulation is subjected to principal component decomposition, and the time coefficients of the principal components are obtained by mapping based on the real-time collected wind speed and wind direction data. The corresponding principal modal spatial distribution is generated using CFD simulation statistics. Based on the real-time updated time coefficients, the non-stationary wind field model is synchronously reconstructed. The non-stationary wind field model is verified using the remaining measuring points in the measured wind speed and wind direction data that were not involved in the modeling, and the error value between the reconstructed wind speed and the measured wind speed is calculated. If the error exceeds 5%, it indicates that the constructed non-stationary wind field model cannot be used for wind-induced vibration load input, and the non-stationary wind field model is readjusted.

[0055] Specifically, the wind pressure distribution data on the structure surface at different moments are extracted from the CFD simulation results as the target labels. At the same time, the reconstructed results of the non-stationary wind speed field are used as the input features. Complete training samples are constructed through the true labels and the input features. Multiple groups of training samples at different time steps are collected, and the collected training samples of each group are normalized. A wind pressure prediction model is established based on the spatio-temporal convolution architecture. The model includes an input layer, a spatial convolution layer, a temporal convolution layer, an upsampling layer, and an output layer. The training samples corresponding to each time step are divided into mini-batch sample data according to the set fixed time window. A group of mini-batch sample data is randomly selected as the input data and transmitted to the wind pressure prediction model. The input layer of the wind pressure prediction model transmits the input data of each time step to the spatial convolution layer. Then, multiple groups of spatial convolution layers sequentially perform convolution operations on the input data to extract the change patterns of the wind speed vector in the spatial region and output the local spatial feature maps. Then, the generated local spatial feature maps are input into the temporal convolution layer. The temporal convolution layer, based on the gating mechanism, captures the dependency relationships between the local spatial feature maps of each time step and outputs the temporal evolution information. After the local spatial feature maps of each time step and the corresponding temporal evolution information are input into the upsampling layer, the upsampling layer restores the local spatial feature maps of each time step to the spatial resolution required for the wind pressure field on the structure surface through deconvolution operations and generates the wind pressure spatial distribution representing each sample at the current prediction moment. Then, the processed local spatial feature maps are passed to the output layer. The output layer converts the local spatial feature maps of each time step into the final prediction result through linear mapping and outputs the corresponding predicted wind pressure field. The value of each prediction point represents the estimated wind pressure value at the corresponding grid position on the structure surface at the current time step. The weighted mean squared error function is used to calculate the error value between the predicted wind pressure field and the true label, and the error value is backpropagated layer by layer from the output layer of the wind pressure prediction model to the input layer. At the same time, the gradient value of the error value with respect to the parameters of each layer is calculated, and the network parameters are updated through the Adam optimizer. After each round of training, the prediction accuracy of the updated wind pressure prediction model is evaluated through the mini-batch sample data not involved in the training, and the prediction error is calculated. If the prediction error range is less than 10%, the model can be used for actual prediction; otherwise, the wind pressure prediction model is retrained and the parameters are updated. The trained wind pressure prediction model is deployed to the wind-induced vibration response analysis platform, receives the wind speed field input in real time, and automatically outputs the future wind pressure distribution to form a continuous loading scenario close to the actual situation.

[0056] Example 2, referring to Figure 1 , a field measurement and analysis method for wind-induced vibration of a large-span steel structure roof. The specific steps of the field measurement and analysis method are as follows:

[0057] Modify the structural stiffness characteristics of the large-span steel structure roof according to the thermal and moisture sensitivity of the material, and analyze the changing trend of the wind speed in real time to dynamically adjust the sampling frequency.

[0058] Specifically, a multi-modal heterogeneous sensing network is used to collect the temperature and humidity change data during the operation of the long-span steel structure roof in real time. According to the collection time, the temperature and relative humidity are processed into an environmental factor set that affects the stiffness change. Based on the long-term effects of steel thermal expansion and humid environment on the structural performance, a thermo-hygroscopic sensitivity function of the elastic modulus is established to obtain the environmentally corrected material elastic modulus. Based on the three-dimensional simulation model of the long-span steel structure roof and the environmentally corrected material elastic modulus, the stiffness coefficients of each element are converted into a dynamic stiffness matrix. Then, the nodes that are most sensitive to the overall stiffness change in the mid-span, support connection, and variable cross-section area of the long-span roof structure are selected as key nodes, and the main diagonal elements corresponding to the structural key nodes are extracted from the dynamic stiffness matrix to construct a stiffness sequence set. The main diagonal elements of each group in the stiffness sequence set are smoothed by moving average, and at the same time, normalization processing is used to make the main diagonal elements of each component within a unified scale range. Then, the empirical mode decomposition is used to decompose the main diagonal elements of each group in the stiffness sequence set into a trend term, a periodic term, and a residual term. The quantitative indicators of the long-term change slope and inflection point position in the trend term are extracted to identify the stiffness trend. The extracted stiffness trend and periodic term are respectively compared and analyzed with the temperature and humidity change data, and then the variance level and fluctuation mode corresponding to the residual term are analyzed. Based on the analysis results, the law of the stiffness evolution of each component of the long-span steel structure roof with time is obtained and encapsulated as a time-driven function. Based on the time-driven function, the nominal initial stiffness is corrected, and the corrected stiffness is combined with the corresponding dynamic stiffness matrix of the three-dimensional simulation model of the current long-span steel structure roof to construct a time-dependent overall stiffness matrix, and a time-varying dynamic equilibrium equation is established. At each time step, according to the real-time detected environmental data, the overall stiffness matrix is updated in real time, and a complete simulation time axis is constructed. At the same time, the corresponding time-varying dynamic equilibrium equation is solved to generate the structural responses of each surface of the current long-span steel structure roof.

[0059] Use a drone to vibrate and scan the surface of the long-span steel structure roof to generate a vibration displacement cloud map and identify the local stiffness of the long-span steel structure roof.

[0060] Specifically, the modal frequencies and mode shape information of the long-span steel structure roof at multiple measuring points are obtained through real-time acquisition or simulation by a multi-modal heterogeneous sensing network, and the spatial distribution of the measuring points is processed with normalized coordinates, converting the coordinates of each measuring point in the measuring point space into coordinates within the unit domain. Then, based on the collected groups of modal frequencies and mode shape information, the corresponding estimated flexibility matrix is established. The long-span steel structure roof is divided into multiple sub-regions, and according to the relative stiffness changes of each sub-region in the long-span steel structure roof analyzed in real time, a set of stiffness loss variables for each sub-region is constructed. Based on each set of stiffness loss variables and the estimated flexibility matrix, the posterior probability of each set of stiffness loss variables is calculated. The set of stiffness loss variables with a posterior probability lower than the preset threshold is used as the initial damage parameter. The number of iterations is set, and the normal distribution is used as the proposal distribution for sampling new samples. For each iteration round, new parameters are sampled from the proposal distribution as candidate samples, and the posterior ratio of the initial damage parameter and the candidate sample is calculated. Based on the Metropolis-Hastings rule, the acceptance probability of the candidate sample is calculated. Then, a random number is generated. If the random number is less than the acceptance probability, the candidate sample is accepted; otherwise, the candidate sample is rejected. The process of selecting candidate samples and replacing them is repeated until the preset number of iteration rounds is reached, and the final damage sample set is output. Statistical analysis is performed on the damage sample set to calculate the maximum a posteriori estimate value of the damage parameter for each region. At the same time, the regions with a maximum a posteriori estimate value higher than 0.2 are marked as damaged regions.

[0061] Based on the multi-modal heterogeneous sensing network and UAV acquisition, a simulation model of the long-span steel structure roof is established, and the safety margin and response characteristics under the ultimate state are evaluated.

Claims

1. A method for actual measurement and analysis of wind vibration of long-span steel structure roofs, characterized in that, The specific steps of the actual measurement and analysis method are as follows: Ⅰ. Based on the geometric and dynamic characteristics of the long-span steel structure roof, optimize the layout of each sensor and construct a multi-modal heterogeneous sensing network; Ⅱ. According to the wind speed and wind direction time series data obtained by the multi-modal heterogeneous sensing network, establish a reconstruction model of the non-stationary wind field and predict the time-varying distribution of wind pressure on the structure surface; Ⅲ. Modify the structural stiffness characteristics of the long-span steel structure roof according to the thermal-wet sensitivity of the material, and analyze the wind speed change trend in real time to dynamically adjust the sampling frequency; Ⅳ. Use a drone to scan the surface of the long-span steel structure roof to generate a vibration displacement cloud map and identify the local stiffness of the long-span steel structure roof; Ⅴ. Based on the multi-modal heterogeneous sensing network and drone acquisition, establish a simulation model of the long-span steel structure roof and evaluate the safety margin and response characteristics under the ultimate state.

2. The method for actual measurement and analysis of wind-induced vibration of a long-span steel structure roof according to claim 1, characterized in that The specific steps of optimizing the layout of each sensor and constructing a multi-modal heterogeneous sensing network described in step Ⅰ are as follows: S1.1: Establish a three-dimensional simulation model of the long-span steel structure roof according to the actual engineering drawings and BIM model, select the corresponding element type to describe each component of the long-span steel structure roof, and then set the parameters of elastic modulus, density, Poisson's ratio and section properties; S1.2: According to the constructed three-dimensional simulation model, establish a set of structural nodes, and then set the grid scale according to the complexity of the structure and the required response resolution preset, perform finite element discretization on the structure surface to form a discrete candidate area for layout points, and based on the set of structural nodes, establish a set of candidate points for layout points after division, calculate the sensitivity of each candidate point in the set of candidate points for layout points to the structural response, screen out the candidate points whose sensitivity does not reach the preset threshold, perform characteristic modeling on each group of sensors, and construct a sensor response matrix according to the physical quantities and measurement accuracies of different sensors; S1.3: Set the maximum layout quantity of each type of sensor, use integer coding to label the layout states of different candidate layout positions, randomly generate an initial population according to the preset upper limit of the layout quantity of each type of sensor, each individual in the population represents a set of sensor layout schemes, and represent different sensor layout schemes as sets of various state labels, where each bit represents the layout state of a candidate layout position; S1.4: Calculate the modal confidence value of each individual in the population, and check whether each individual meets the preset sensor quantity constraint. According to the modal confidence value and constraint satisfaction situation of each individual, obtain the fitness value of each individual, and adopt the roulette wheel selection method to select multiple groups of individuals from the current population from high to low according to the fitness as the parents for the next generation of crossover; S1.5: Randomly select two groups of individuals from the parents for crossover, randomly select one or two crossover points from the two groups of individuals respectively, and exchange the layout state label segments of the corresponding two groups of individuals to form two new offspring. Randomly select one or more layout positions in the new offspring individuals, randomly replace the value of this position with any layout state, and at the same time detect whether the new offspring meets the preset sensor quantity constraint. If not, eliminate the corresponding offspring, and then calculate the fitness value of the new offspring after crossover and mutation; S1.6: If the fitness value of the new offspring is higher than that of the parent, retain the new offspring; otherwise, retain the parent. Then, construct a new population from the filtered parent and the new offspring, and repeat the operations of parent selection, crossover and mutation, and population construction until the change value of the optimal fitness value of each individual in the population converges within a preset range. S1.7: Traverse the fitness values of each group of individuals in the final population, and use the individual with the highest fitness as the optimal sensor layout scheme. Then, synchronize the optimal sensor layout scheme to the 3D simulation model of the long-span steel structure roof, mark the sensor layout status of each layout point, and establish a multi-modal heterogeneous sensing network according to the optimal sensor layout scheme.

3. The measured analysis method for wind-induced vibration of a large-span steel structure roof according to claim 2, characterized in that, The specific steps for establishing the reconstruction model of the non-stationary wind field in Step II are as follows: S2.1: Use the multi-modal heterogeneous sensing network to obtain long-time series wind speed and wind direction data at multiple measuring points around the long-span steel structure roof. Decompose the long-time series wind speed data into an average term and a perturbation term, and use CFD software to perform unsteady simulation on typical wind direction and speed combinations to obtain the instantaneous wind pressure and velocity field around the long-span steel structure roof. S2.2: Perform joint mapping on the measured wind speed and wind direction data and the CFD results. Simulate the spatial correlation and time evolution of the wind speed field by using a time-varying covariance function to establish a non-stationary wind field model, and sample the covariance between the wind speed data collected at each measuring point at different times to establish a time-varying wind speed field with corresponding statistical characteristics. S2.3: Perform principal component decomposition on the time-varying wind speed field generated by CFD simulation, map to obtain the time coefficients of the principal components according to the real-time collected wind speed and wind direction data, and use CFD simulation statistics to generate the corresponding principal modal spatial distribution. Based on the real-time updated time coefficients, synchronously reconstruct the non-stationary wind field model. S2.4: Use the remaining measuring points in the measured wind speed and wind direction data that are not involved in the modeling to verify the non-stationary wind field model, and calculate the error value between the reconstructed wind speed and the measured wind speed. If the error exceeds 5%, it means that the constructed non-stationary wind field model cannot be used for wind-induced vibration load input, and the non-stationary wind field model needs to be readjusted.

4. The measured analysis method for wind-induced vibration of a long-span steel structure roof according to claim 3, characterized in that, The specific steps for predicting the time-varying distribution of wind pressure on the structure surface in Step II are as follows: S3.1: Extract the wind pressure distribution data on the structure surface at different times from the CFD simulation results as target labels, and use the reconstruction result of the non-stationary wind speed field as input features. Then, construct a complete training sample from the real labels and input features, collect multiple groups of training samples at different time steps, and normalize each group of collected training samples. S3.2: Establish a wind pressure prediction model based on a spatio-temporal convolution architecture. The model includes an input layer, a spatial convolution layer, a temporal convolution layer, an upsampling layer, and an output layer. Divide the training samples corresponding to each time step into small batch sample data according to a set fixed time window, randomly select a group of small batch sample data as input data, and transmit it into the wind pressure prediction model. S3.3: The input layer of the wind pressure prediction model transmits the input data at each time step to the spatial convolution layer. Then, multiple groups of spatial convolution layers sequentially perform convolution operations on the input data to extract the variation patterns of the wind speed vector in the spatial region, and output local spatial feature maps. Then, the generated local spatial feature maps are input into the temporal convolution layer. Based on the gating mechanism, the temporal convolution layer captures the dependencies between the local spatial feature maps at each time step and outputs the temporal evolution information; S3.4: After inputting the local spatial feature maps at each time step and the corresponding temporal evolution information into the upsampling layer, the upsampling layer restores the local spatial feature maps at each time step to the spatial resolution required for the structural surface wind pressure field through deconvolution operations, and generates the wind pressure spatial distribution representing each sample at the current prediction moment. Then, the processed local spatial feature maps are transmitted to the output layer. The output layer converts the local spatial feature maps at each time step into the final prediction result through linear mapping and outputs the corresponding predicted wind pressure field, where the value of each prediction point represents the wind pressure estimation value of the corresponding grid position on the structural surface at the current time step; S3.5: Use the weighted mean square error function to calculate the error value between the predicted wind pressure field and the true label, and backpropagate the error value layer by layer from the output layer of the wind pressure prediction model to the input layer. At the same time, calculate the gradient value of the error value with respect to the parameters of each layer, and update the network parameters through the Adam optimizer; S3.6: After each round of training, evaluate the prediction accuracy of the updated wind pressure prediction model through a small batch of sample data that has not participated in training, and calculate the prediction error. If the prediction error range is less than 10%, the model can be used for actual prediction. Otherwise, retrain the wind pressure prediction model and update the parameters. Deploy the trained wind pressure prediction model to the wind-induced vibration response analysis platform, receive the wind speed field input in real time, and automatically output the future wind pressure distribution to form a continuous loading scenario close to the actual situation.

5. The measured analysis method for wind-induced vibration of a long-span steel structure roof according to claim 2, characterized in that, The specific steps of modifying the structural stiffness characteristics of the long-span steel structure roof according to the thermal and moisture sensitivity of the material in Step III are as follows: S4.1: Use a multi-modal heterogeneous sensing network to collect the temperature and humidity change data during the operation of the long-span steel structure roof in real time. According to the collection time, process the temperature and relative humidity into an environmental factor set that affects the stiffness change, and establish a thermal and moisture sensitivity function of the elastic modulus to obtain the environmentally corrected material elastic modulus; S4.2: Based on the three-dimensional simulation model of the long-span steel structure roof and the environmentally corrected material elastic modulus, convert the stiffness coefficients of each element into a dynamic stiffness matrix. Then, select the key nodes that are most sensitive to the overall stiffness change at the mid-span, support connection, and variable cross-section area of the long-span roof structure as the key nodes, and extract the main diagonal elements corresponding to the structural key nodes from the dynamic stiffness matrix to construct a stiffness sequence set; S4.3: Smooth each group of main diagonal elements in the stiffness sequence set through moving average. At the same time, use normalization processing to make the main diagonal elements of each component within a unified scale range. Then, use empirical mode decomposition to decompose each group of main diagonal elements in the stiffness sequence set into a trend term, a periodic term, and a residual term; S4.4: Extract various quantitative indicators of the long-term change slope and inflection point position in the trend term, identify the stiffness trend, compare and analyze the extracted stiffness trend and the periodic term with the temperature and humidity change data respectively, and then analyze the variance level and fluctuation pattern corresponding to the residual term. Based on the analysis results, obtain the law of the stiffness evolution of each component of the long-span steel structure roof over time, and encapsulate it as a time-driven function; S4.5: Based on the time-driven function, correct the nominal initial stiffness, and combine the corrected stiffness with the corresponding dynamic stiffness matrix of the three-dimensional simulation model of the current long-span steel structure roof to construct a time-dependent overall stiffness matrix and establish a time-varying dynamic equilibrium equation; S4.6: At each time step, update the overall stiffness matrix in real time according to the real-time detected environmental data, construct a complete simulation time axis, and solve the corresponding time-varying dynamic equilibrium equation simultaneously to generate the structural responses of each surface of the current long-span steel structure roof.

6. The measured analysis method for wind-induced vibration of a long-span steel structure roof according to claim 5, wherein The specific steps for identifying the local stiffness of the long-span steel structure roof described in Step IV are as follows: S5.1: Real-time collect or simulate the modal frequencies and mode shape information of the long-span steel structure roof at multiple measurement points through a multi-modal heterogeneous sensing network, perform normalized coordinate processing on the spatial distribution of the measurement points, convert the coordinates of each measurement point in the measurement point space into coordinates within the unit domain, and then establish the corresponding estimated flexibility matrix based on the collected groups of modal frequencies and mode shape information; S5.2: Divide the long-span steel structure roof into multiple groups of sub-regions, construct a set of stiffness loss variables for each sub-region according to the relative stiffness changes in each sub-region of the long-span steel structure roof analyzed in real time, and calculate the posterior probability of each set of stiffness loss variables based on each set of stiffness loss variables and the estimated flexibility matrix; S5.3: Use the set of stiffness loss variables with a posterior probability lower than the preset threshold as the initial damage parameter, set the number of iterations, use the normal distribution as the proposed distribution for sampling new samples, and then for each iteration round, sample new parameters from the proposed distribution as candidate samples and calculate the posterior ratio of the initial damage parameter to the candidate samples; S5.4: Based on the Metropolis-Hastings rule, calculate the acceptance probability of the candidate samples, and then generate a random number. If the random number is less than the acceptance probability, accept the candidate samples; otherwise, reject the candidate samples. Repeat the selection of candidate samples and replacement until the preset number of iteration rounds is reached, and output the final set of damage samples; S5.5: Conduct statistical analysis on the set of damage samples, calculate the maximum a posteriori estimate value of the damage parameter for each region, and mark the regions with a maximum a posteriori estimate value higher than 0.2 as damaged regions.

Images