A load assessment method for roof steel plate structure based on ResNet

Abstract

The present invention relates to a load evaluation method for a roof steel plate structure based on ResNet. The method comprises the following steps: using Ansys finite element analysis software to perform finite element simulation on the steel plate structure of the roof, determining the strain collection position, and obtaining the strain values ​​corresponding to different loads; performing data screening and cleaning on the obtained strain values, establishing a load-strain database for roof plate mechanisms, and dividing the training set and the test set; using ResNet34 to train the training set data, and testing it on the test set to obtain the test results; and using the average relative error as an evaluation index. After completing the data preprocessing, the ResNet34 model is trained using the training set, and the model is evaluated on the test set; and then the trained ResNet model is used for load prediction of the roof steel plate structure. The present invention accurately detects the load of the roof steel plate structure, has strong practicality, and is more accurate than the traditional method of calculating the load. It can be widely used in the field of load calculation of roof steel plate structures.

Description

Technical Field

[0001] The present invention relates to the field of roof steel plate structure load calculation, and in particular to a roof steel plate structure load evaluation method based on ResNet. Background Art

[0002] In the process of building roof renovation, the steel plate roof construction method is used as the main construction process. In the actual process, the increase in roof stiffness can affect the overall stiffness of the structure and ensure the stability and safety of the structure. However, there are few domestic studies, so it is necessary to conduct finite element overall static analysis on this type of structure. Finite element analysis is a necessary means of auxiliary experimental research today, which can make up for the shortcomings of time-consuming experimental analysis, limited parameter analysis, and labor and capital consumption.

[0003] Foreign research on roof steel plates began in the middle of the last century, when plywood was first used as roof panels. In the 1980s, foreign construction industries began to use steel plates as roofs, and conducted experimental research on different roof steel plates. Through finite element and experimental analysis of the thickness and size of different roof steel plates, the bearing capacity of steel plate structures in different environments and geographical locations was obtained.

[0004] The measurement of the load borne by the steel plate is generally divided into direct method and indirect method. The direct method in engineering generally uses a force sensor, but the force sensor has certain disadvantages. The mass of the force sensor itself is large. When measuring a steel plate with a smaller mass, the force sensor will affect the measured load size. And the accuracy of the force sensor is not high, and the collected load is prone to errors. Therefore, in engineering, the indirect method is generally used to calculate the load. Use strain gauges or fiber gratings to measure the strain values ​​corresponding to different loading loads, and then calculate the calibration matrix through methods such as load calibration equations to obtain the load value. However, the load calibration equation cannot well represent nonlinear factors, because the actual steel plate measurement will be affected by temperature, humidity, etc. It will cause the strain and load to be not completely linear. Therefore, the neural network method is used to calculate the roof steel plate load. The traditional deep neural network has many parameters, a long training time, and is prone to overfitting. Under comprehensive consideration, the present invention adopts the house steel plate structure load evaluation method of ResNet34, adds the relative position information between strains, and completes the roof steel plate load calculation with higher accuracy. Summary of the invention

[0005] The present invention has two purposes: (1) to solve the problem of low calculation accuracy of roof steel plate load in traditional load calculation methods. (2) to assist in calculating the bearing capacity of roof steel plates and improve the accuracy of roof steel plate load calculation. Based on the above two purposes, the present invention provides a roof steel plate load evaluation method based on ResNet34.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A roof steel plate structure load assessment method based on ResNet34 includes the following steps:

[0008] Step 1: According to the actual structure of the roof steel plate, conduct stress analysis and establish a structural model, and establish a roof steel plate structural model in Ansys finite element analysis software;

[0009] Step 2: Perform finite element simulation in Ansys finite element simulation software, select the load loading position, and apply multiple loads of different sizes, collect the strains of all finite element structures of the steel plate structure corresponding to the load loading, and establish a load-strain database for the roof steel plate;

[0010] Step 3: Select the location of the fiber grating points during the experiment; according to the location of the experimental points, clean and filter out the corresponding strain values ​​during the simulation, and perform data preprocessing;

[0011] Establish training set, validation set and test set of the model;

[0012] Step 4: Use the training set to train the ResNet34 model, modify the two-dimensional network to a one-dimensional network, and change the input dimension to the total number of strains;

[0013] Step 5: Use the test set to test the performance of the trained ResNet34, then use the model for load evaluation of the roof steel plate structure and output the mean relative error.

[0014] In the above scheme, step 2, finite element analysis of the roof steel plate structure, the specific steps are as follows:

[0015] Step 21: Use Solidworks to model the roof steel plate model. The size of the steel plate model is 1000mm×600mm×5mm. Perform finite element analysis in Mechanical APDL 19.0;

[0016] The fixed part is one end of the steel plate model, with a fixed area of ​​300mm×600mm, and the upper and lower sides are fixed at the same time;

[0017] The material selected is stainless steel with an elastic modulus of 7.2×10 10Pa, Poisson's ratio is 0.33, and the size of the finite element grid is 5mm×1mm; in Ansys software, a maximum of 2898 loads can be applied, and the data for each load contains three dat files; the three dat files respectively save the partial strain values ​​and the total relative strain values ​​in the x, y, and z directions of all finite element blocks; the magnitude of the partial force and the total stress value in the x, y, and z directions; the magnitude of the applied load and the location where the load is applied;

[0018] The load range is N[100, 1000].

[0019] In the above scheme, the specific steps of step 3, data cleaning and preprocessing, are as follows:

[0020] Step 31: According to the position and coordinate values ​​of the experimental fiber Bragg grating, 270 strain values ​​under the simulation conditions are screened, and strains of other finite element parts are eliminated;

[0021] Since the range of the cleaned data is too different, the strain value of the data is Z-score standardized and the load value is normalized;

[0022] The standardized formula is as follows:

[0023]

[0024] In the above formula, μ is the mean and σ is the standard deviation;

[0025] The standard deviation is calculated as follows:

[0026]

[0027] The formula for normalization of the load is as follows:

[0028]

[0029] In the above scheme, step 4 is to train and modify the ResNet34 model. ResNet34 consists of 1×7 convolution and 4 layers, each layer contains multiple Residual Blocks; downsampling uses Conv-Downsampling downsampling.

[0030] In the above scheme, step 41: the strain data input to ResNet34 is first passed through a convolution kernel of size 1×7 and step size 2 to perform preliminary data extraction;

[0031] Then downsampling is done by a convolution kernel of size 1×3 and stride 2;

[0032] The formula for convolution downsampling is as follows:

[0033]

[0034] In the above formula, conv_out is the output dimension of downsampling, and n is the input dimension of convolution;

[0035] Since it is a one-dimensional convolution, the other dimension of the convolution input is 1;

[0036] Indicates rounding up, 3 is the size of the convolution kernel, and 2 is the step size of the convolution kernel; the formula comes from the calculation of the convolution output dimension;

[0037] If the input dimension n is an even number, set to 2m, the right side of the formula becomes After rounding up, it becomes m-1+1, and the result after convolution is m, which is the same as the result of downsampling;

[0038] If the input dimension n is an odd number, set to 2m+1, the right side of the formula becomes After rounding, it is m-1+1, and the result after convolution is still m, which is the same as the result of downsampling.

[0039] Compared with pooling downsampling, convolution downsampling has its own unique advantages. The pooling layer compresses the information after convolution, is robust to the changes of small information after convolution, and absorbs the deviation of input data. However, the disadvantage of pooling is that some data information will be lost, and more useful information may be filtered out. However, the convolution downsampling process controls the step size, which can better fuse information. When completing the downsampling operation, there will be no information loss.

[0040] In the above scheme, step 42: in order to avoid the overfitting of the model caused by the network being too deep, the disappearance of gradients, and the gradual disappearance of shallow features with convolution, etc., a connection is established between the shallow features and the deep features, the features after multiple convolutions are fused with the shallow features, and a BN (Batch normalization) layer is added during the convolution process, which perfectly solves the problems of gradient explosion and overfitting, and obtains the basic component of ResNet - Residual Block. The formula of Residual Block is as follows:

[0041] z=F(x,w i )+x

[0042] Where z and x represent the input and output vectors of the residual block, and F represents the residual part, which is generally composed of two or three convolution operations. i)+x means the residual feature is added to the mapping of the input feature, and the dimensions of the two need to be the same. Since the number of channels in the residual part will be different, a 1×1 convolution is added to the residual part to increase or decrease the dimension of the convolution result.

[0043] In the above scheme, step 43: in ResNet34, 4 different 1ayers are composed of different numbers of residual blocks, and in the 4 1ayers, the numbers of residual blocks are 3, 4, 6, and 3 respectively;

[0044] Conv-Downsampling is used to connect any two different layers, and a fully connected layer is added at the end of the network; the softmax layer of the original ResNet34 is removed, and the final prediction result is directly output; the loss function is changed from the cross entropy loss function to the mean square error loss function, and the formula of the mean square error loss function is as follows:

[0045]

[0046] The modeling and simulation analysis of the present invention calculates the accuracy of the ResNet34 algorithm in the load evaluation of the roof steel plate structure under simulation conditions to adapt to the load calculation of the roof steel plate structure and improve the training speed and calculation accuracy of the steel plate structure load evaluation. The present invention has the following advantages:

[0047] (1) The present invention is a calculation under simulation conditions. First, the feasibility and accuracy of the ResNet34 algorithm in load calculation are verified, which saves costs and lays the foundation for the calculation of roof steel plate loads under experimental conditions.

[0048] (2) The model training speed of the present invention is fast and not prone to overfitting. Compared with the traditional load calculation method, the present invention has higher accuracy, with an average relative error of 0.0263%, which is lower than 0.231% of the load calibration equation and 0.0946% of the deep neural network.

[0049] (3) The system is implemented using the Pytorch framework, which is easy for users to expand and apply. It has certain practical application value in the promotion and expansion of roof steel plate load assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Comparison of the accuracy of different algorithms for evaluating the load on the roof steel plate structure;

[0051] Figure 2 Implementation flow chart of the present invention;

[0052] Figure 3 One-dimensional ResNet34 network structure diagram. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0054] The present invention has two purposes: (1) to solve the problem of low calculation accuracy of roof steel plate load in traditional load calculation methods. (2) to assist in calculating the bearing capacity of roof steel plates and improve the accuracy of roof steel plate load calculation. Based on the above two purposes, the present invention provides a roof steel plate load evaluation method based on ResNet34.

[0055] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0056] Step 1: According to the actual structure of the roof steel plate, perform force analysis and establish a structural model, and establish the roof steel plate structure model in Ansys finite element analysis software.

[0057] Step 2: Perform finite element simulation in Ansys finite element simulation software, select the load loading position, and apply multiple loads of different sizes. Collect the strains of all finite element structures of the steel plate structure corresponding to the load loading, and establish a load-strain database for the roof steel plate.

[0058] Step 21: Use Solidworks to model the roof steel plate model. The size of the steel plate model is 1000mm×600mm×5mm. Perform finite element analysis in Mechanical APDL 19.0. Select one end of the steel plate model as the fixed part. The fixed area is 300mm×600mm, and the upper and lower sides are fixed at the same time. The selected material is stainless steel, with an elastic modulus of 7.2×1010Pa, a Poisson's ratio of 0.33, and a finite element mesh size of 5mm×1mm. In Ansys software, up to 2898 loads can be applied, and the data for each load contains three dat files. The three dat files respectively save the partial strain values ​​and total relative strain values ​​of all finite element blocks in the x, y, and z directions; the magnitude of the component forces and the total stress values ​​in the x, y, and z directions; the magnitude of the applied load and the location of the applied load. The load range is N[100, 1000].

[0059] Step 3: Select the locations of the fiber grating points during the experiment. According to the locations of the experimental points, clean and filter out the corresponding strain values ​​during the simulation, and perform data preprocessing. Establish the training set, validation set, and test set of the model.

[0060] Step 31: According to the position of the experimental fiber Bragg grating. According to the coordinate value of the fiber Bragg grating during the experiment, 270 strain values ​​under the simulation conditions are screened, and the strain of other finite element parts is eliminated. Since the range of the data after cleaning is too large, the strain value of the data is Z-score standardized and the load value is normalized. The standardized formula is as follows:

[0061]

[0062] In the above formula, μ is the mean and σ is the standard deviation. The standard deviation is calculated as follows:

[0063]

[0064] The formula for normalization of the load is as follows:

[0065]

[0066] Moreover, the ratio of training set, validation set and test set is 7:1:2.

[0067] Step 4: Use the training set to train the ResNet34 model, modify the two-dimensional network to a one-dimensional network, and change the input dimension to the total number of strains.

[0068] Step 41: The strain data input to ResNet34 is first extracted through a convolution kernel of size 1×7 and stride 2. Then downsampling is completed through a convolution kernel of size 1×3 and stride 2. The formula for convolution downsampling is as follows:

[0069]

[0070] In the above formula, conv_out is the output dimension of the downsampling, and n is the input dimension of the convolution. Since it is a one-dimensional convolution, the other dimension of the convolution input is 1. Indicates rounding up, 3 is the size of the convolution kernel, and 2 is the step size of the convolution kernel. The formula comes from the calculation of the convolution output dimension. If the input dimension n is an even number, set to 2m, the right side of the formula becomes After rounding up, it becomes m-1+1, and the result after convolution is m, which is the same as the result of downsampling. If the input dimension n is an odd number, set to 2m+1, the right side of the formula becomes After rounding, it is m-1+1, and the result after convolution is still m, which is the same as the result of downsampling. Compared with pooling downsampling, convolution downsampling has its own unique advantages. The pooling layer compresses the information after convolution, is robust to the changes of small information after convolution, and absorbs the deviation of input data. However, the disadvantage of pooling is that some data information will be lost, and more useful information may be filtered out. However, the convolution downsampling process controls the step size, which can better fuse information. When the same downsampling operation is completed, there will be no information loss.

[0071] Step 42: In order to avoid the overfitting of the model caused by the network being too deep, the disappearance of gradients, and the gradual disappearance of shallow features with convolution, some scholars have established a connection between shallow features and deep features, fused the features after multiple convolutions with shallow features, and added a BN (Batch normalization) layer during the convolution process, which perfectly solved the problems of gradient explosion and overfitting. The basic component of ResNet, Residual Block, is obtained. The formula of ResidualBlock is as follows:

[0072] z=F(x,w i )+x

[0073] Where z and x represent the input and output vectors of the residual block, and F represents the residual part, which is generally composed of two or three convolution operations. i )+x means the residual feature is added to the mapping of the input feature, and the dimensions of the two need to be the same. Since the number of channels in the residual part will be different, a 1×1 convolution is added to the residual part to increase or decrease the dimension of the convolution result.

[0074] Step 43: In ResNet34, 4 different layers are composed of different numbers of residual blocks. In the 4 layers, the number of residual blocks is 3, 4, 6, and 3 respectively. Any two different layers are connected using Conv-Downsampling. And a fully connected layer is added at the end of the network. The original softmax layer of ResNet34 is removed, and the final prediction result is directly output. The loss function is changed from the cross entropy loss function to the mean square error loss function. The formula of the mean square error loss function is as follows:

[0075]

[0076] Use the training set to train the ResNet34 network, and use the validation set to verify whether the network has problems such as overfitting and gradient disappearance.

[0077] Step 5: Use the test set to test the performance of the trained ResNet34, then use the model for load evaluation of the roof steel plate structure and output the mean relative error.

[0078] For the trained ResNet34 model, the training set is used to evaluate the roof steel plate load, and the average relative error is used as the evaluation index. The formula of the average relative error is as follows:

[0079]

[0080] Where M is the total number of test set samples, y i Indicates the true load value; Represents the predicted value of the neural network. The load prediction results of RetNet34 are compared with those of the traditional method.

[0081] Table 1 Comparison of detection results between the proposed method and the traditional Faster RCNN method

[0082] method Load Calibration Equations Deep Neural Networks ResNet34 Mean relative error 0.231％ 0.0946％ 0.0263％

[0083] In summary, the present invention uses Ansys finite element analysis software to perform finite element simulation on the steel plate structure of the roof, determines the strain collection position, and obtains the strain values ​​corresponding to different loads; performs data screening and cleaning on the obtained strain values, establishes a load-strain database for the roof plate structure, and divides the training set and the test set; uses ResNet34 to train the training set data, and tests it on the test set to obtain the test results; and uses the average relative error as an evaluation index. After completing the data preprocessing, the ResNet34 model is trained using the training set, and the model is evaluated on the test set; then the trained ResNet model is used for load prediction of the roof steel plate structure. The present invention accurately detects the load of the roof steel plate structure, is highly practical, and has higher accuracy than the traditional method of calculating the load.

[0084] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A roof steel plate structure load assessment method based on ResNet, characterized in that: It includes the following steps: Step 1: According to the actual structure of the roof steel plate, conduct stress analysis and establish a structural model, and establish a roof steel plate structural model in Ansys finite element analysis software; Step 2: Perform finite element simulation in Ansys finite element simulation software, select the load loading position, and apply multiple loads of different sizes, collect the strains of all finite element structures of the steel plate structure corresponding to the load loading, and establish a load-strain database for the roof steel plate structure; Step 3: Select the locations of the fiber grating points during the experiment, clean and filter out the corresponding strain values ​​during the simulation according to the locations of the points during the experiment, and perform data preprocessing; Establish training set, validation set and test set of the model; Step 4: Use the training set to train the ResNet34 model, modify the two-dimensional network to a one-dimensional network, and change the input dimension to the total number of strains; Step 5: Use the test set to test the performance of the trained ResNet34, then use the model for load evaluation of the roof steel plate structure and output the average relative error; In step 4, the ResNet34 model is trained and modified. ResNet34 consists of 1×7 convolution and 4 layers, each of which contains multiple Residual Blocks. Downsampling uses Conv-Downsampling. The specific steps are as follows: Step 41: The strain data input to ResNet34 is firstly passed through a convolution kernel of size 1×7 and step length 2 for preliminary data extraction; then down-sampling is completed through a convolution kernel of size 1×3 and step length 2; The formula for convolution downsampling is as follows: ； In the above formula, conv_out is the output dimension of downsampling, and n is the input dimension of convolution; Since it is a one-dimensional convolution, the other dimension of the convolution input is 1; Indicates rounding up, 3 is the size of the convolution kernel, and 2 is the step size of the convolution kernel; The formula comes from the calculation of the convolution output dimension; If the input dimension n is an even number, set to 2m, the right side of the formula becomes , after rounding up, it becomes m-1+1, and the result after convolution is m, which is the same as the result of downsampling; If the input dimension n is an odd number, set to 2m+1, the right side of the formula becomes , after rounding, it is m-1+1, and the result after convolution is still m, which is the same as the result of downsampling; Step 42: Establish a connection between shallow features and deep features, fuse the features after multiple convolutions with the shallow features, and add a BN (Batch normalization) layer during the convolution process to obtain the basic component of ResNet - Residual Block. The formula of Residual Block is as follows: ； Among them, z and x represent the input and output vectors of the residual block, and F represents the residual part, which consists of two or three convolution operations. The feature after the residual is added to the mapping of the input feature, and the dimensions of the two need to be the same; Add a 1×1 convolution to the residual part to increase or reduce the dimension of the convolution result; Step 43: In ResNet34, 4 different layers are composed of different numbers of residual blocks. In the 4 layers, the number of residual blocks is 3, 4, 6, and 3 respectively. Use Conv-Downsampling to connect any two different layers, and add a fully connected layer at the end of the network; Remove the original ResNet34 softmax layer and directly output the final prediction result; The loss function is changed from the cross entropy loss function to the mean square error loss function. The formula of the mean square error loss function is as follows: 。 2. A roof steel plate structure load assessment method based on ResNet according to claim 1, characterized in that: In step 2, the finite element analysis of the roof steel plate structure is performed in the following specific steps: Step 21: Use Solidworks to model the roof steel plate model. The size of the steel plate model is 1000mm×600mm×5mm. Perform finite element analysis in Mechanical APDL 19.

0. The fixed part is one end of the steel plate model, with a fixed area of ​​300mm×600mm, and the upper and lower sides are fixed at the same time; The material selected is stainless steel with an elastic modulus of 7.2×10 10 Pa, Poisson’s ratio is 0.33, and the size of the finite element mesh is 5 mm × 1 mm; In Ansys software, up to 2898 loads can be applied, and the data for each load contains three dat files; The three dat files respectively save the partial strain values ​​and total relative strain values ​​in the x, y, and z directions of all finite element blocks; the magnitude of the partial force and the total stress value in the x, y, and z directions; the magnitude of the applied load and the location where the load is applied; The range of the load is N [100,1000].

3. The method for evaluating the load of a roof steel plate structure based on ResNet according to claim 1, characterized in that: In step 3, the specific steps for data cleaning and preprocessing are as follows: Step 31: According to the position and coordinate values ​​of the experimental fiber Bragg grating, 270 strain values ​​under the simulation conditions are screened, and strains of other finite element parts are eliminated; The strain values ​​of the data were subjected to Z-score standardization, and the load values ​​were normalized; The standardized formula is as follows: ； In the above formula, is the mean, is the standard deviation; The standard deviation is calculated as follows: ； The formula for normalization of the load is as follows: 。

Images