CN114894619B - Prediction method of axial compression load-strain curve of concrete-filled steel tube columns based on long short-term memory network - Google Patents

Abstract

The present invention proposes a method for predicting the axial compression load-strain curve of a steel tube concrete column based on a long short-term memory network. For design parameters, axial compression load and axial strain data, design data preprocessing and data composition methods are used for long short-term memory network offline model training; construct an axial strain data sequence, and input it and the design parameters into the trained offline model to obtain the axial compression load data output by the model; use the axial strain data and axial compression load data to restore the axial compression load-strain curve to achieve system axial compression load-strain curve prediction.

Description

Prediction method of axial compressive load-strain curve of concrete-filled steel tube columns based on long short-term memory network

Technical Field

The present invention relates to the fields of architecture and civil engineering, and in particular to a method for predicting axial compression load-strain curve of a steel tube concrete column based on a long short-term memory network.

Background technique

As a structural member that bears compression and bending, steel tube concrete columns are recognized as promising substitutes for traditional reinforced concrete and steel columns in the construction of high-rise buildings and urban viaduct bridges. Compared with traditional structural columns, steel tube concrete columns have higher bending, axial bearing capacity and seismic resistance. In addition, the mutual protection of steel tubes and filling concrete makes steel tube concrete columns have better fire resistance, impact resistance and corrosion resistance.

The axial compressive load-strain curve is one of the most basic and important features in the safety design of structural columns. Once the load-strain curve is known, key design parameters, including the elastic modulus, ultimate bearing capacity and ductility of the column, can be easily obtained. Therefore, the axial compressive load-strain curve has been the basic research target of steel tube concrete columns in the past three decades. Theoretical derivation using empirical parameters and finite element model analysis are the two main methods for calculating the load-strain curve of steel tube concrete columns. In the research based on the theoretical method, the stress-strain relationship between the outer steel tube and the inner concrete is usually assumed first, and then the load-strain curve of the column is gradually derived according to the equilibrium condition and the predetermined interaction relationship. However, with the update of materials, the applicability of this method has gradually decreased, and the theoretical method based on empirical parameters is inevitably limited to the material strength and design concept at that time.

Finite element model analysis is another method to obtain the axial load-strain curve of steel tube concrete columns. In finite element analysis, concrete and steel tubes should be simulated using appropriate unit types, usually concrete solid units and steel tube shell units, and contact models or spring models are needed to define the interaction between concrete and steel surfaces. By reasonably adopting constitutive relations and appropriate meshes, the axial compression load-strain curve of steel tube concrete columns can be obtained.

In recent years, soft computing methods and intelligent technologies have been significantly developed and widely used in civil engineering. However, the application of machine learning algorithms in the study of concrete-filled steel tube columns is still in its infancy and exploration stage. Most existing studies focus on the prediction of performance indicators (such as axial bearing capacity and stiffness) based on collected experimental databases. There is a lack of research on the use of machine learning methods to calculate the complete axial compressive load-strain curve of concrete-filled steel tube components.

Long short-term memory network is a type of recurrent neural network, which is often used to process time series. The characteristic of long short-term memory network is that it considers the long-term and short-term dependencies of data. In order to solve the above problems, it is necessary to propose a technical solution that can regard the points in the axial compression load-strain curve as a time series with long-term and short-term dependencies, and use the long short-term memory network to realize curve prediction by designing a data composition method.

Summary of the invention

In response to the above-mentioned problem of axial compressive load-strain curve prediction for steel tube concrete columns, the present invention proposes an axial compressive load-strain curve prediction method for steel tube concrete columns based on long short-term memory network, which can regard the points in the axial compressive load-strain curve as a time series with long-term and short-term dependencies, and realize curve prediction using long short-term memory network through design data construction method.

To achieve the above object, the technical solution adopted by the present invention is:

The method for predicting axial compressive load-strain curve of steel tube concrete column based on long short-term memory network is characterized by comprising the following steps:

S1: Experimental data collection;

The experimental data of axial compression load-strain curve of steel tube concrete columns of different materials and sizes are collected to establish the load-strain curve and five design parameters. The five design parameters are: steel tube diameter D, steel tube wall thickness _ts , column height H, steel strength _fs , and concrete strength _fc. The system database corresponding to the design parameters is used to construct the load-strain curve prediction training samples;

S2: data preprocessing;

For all CFST column axial compression load-strain curve experimental data, 25000με is set as the limit value for collecting axial strain. If the experimental data curve is short and does not reach the axial strain limit of 25000με, a compensation operation will be performed to extend the axial compression load-strain curve to the set axial strain limit according to the final slope of the load-strain curve. If the experimental data curve is long, a truncation operation will be performed to discard data exceeding the set limit. After the above data preprocessing, all axial load-strain curve data have a uniform axial strain length.

S3: Data composition;

Each set of design parameters and the corresponding axial compression load-strain curve is taken as a data sample; the experimental data of the axial compression load-strain curve of the steel tube concrete column after data preprocessing is evenly divided into m points according to the strain value, and each point contains the axial compression load value N and the axial strain value ε; the input data composition of the long short-term memory unit neural network is constructed using the axial strain value and the design parameters; and the output data composition of the long short-term memory unit neural network is constructed using the axial compression load value;

S4: offline model training;

After all samples are constructed through input and output data in S3, the long short-term memory unit neural network is used to perform offline model training on all samples to obtain a long short-term memory unit neural network model that meets the convergence requirements;

S5: prediction of axial compression load-strain curve;

Arrange the design parameters of the steel tube concrete column for which the axial compressive load-strain curve needs to be predicted; construct an axial strain data sequence, and take a vector with a maximum value of 25000με and m evenly distributed data values as the axial strain data sequence; use the input data composition construction method described in S3 to convert the steel tube concrete sample data for axial compressive load-strain curve prediction into an input data composition that can be used for a long short-term memory network; input the input data number into the offline model obtained in S4 to obtain the corresponding axial compressive load output; take the average value of all axial strains in each row of the input data and the axial compressive load value in the corresponding row of the output data as a point on the axial compressive load-strain plane, calculate all axial compressive load-strain points in turn to restore the complete axial compressive load-strain curve, and complete the axial compressive load-strain curve prediction.

As a preferred embodiment of the present invention:

In S3,

S31: The input data composition includes: taking k axial strains as a group, and forming the first row of a data sample with the five design parameters of the sample; then taking another group of k axial strains using a sliding window method, and forming another row of a data sample with the five design parameters; and so on until all axial strains are taken, at which time the input data composition of each data sample includes m+1-k rows and 5+k columns;

S32: The output data composition includes: taking a group of k axial compressive load values to obtain each corresponding axial compressive load value, averaging the group of compressive load values to obtain the axial compressive load corresponding to this row of input data It is used as the output data of the neural network. At this time, the output data structure of each data sample contains m+1-k rows and 1 column.

As a preferred embodiment of the present invention:

In S4,

S41: The LSTM neural network includes, in sequence: an input layer, a LSTM layer, a fully connected layer, a random drop layer, a fully connected layer, and a data regression layer; wherein the LSTM layer includes G LSTMs, i.e., the input time series length is G, and the number of input data features is C; wherein the LSTM includes:

The current moment is t-1, and the output at time t needs to be calculated through the long short-term memory unit. The input data is x _t , h _t and c _t represent the output and unit state at time t respectively; the long short-term memory unit uses the current states h _t-1 and c _t-1 to calculate the output h _t and the updated unit state c _t ; the input gate i _t , forget gate f _t , candidate unit g _t and output gate o _t are used to control the update of the long short-term memory unit. The calculation formula is as follows:

_it =σ( _{Wixt + Rit -1} + _bi )

_ft = σ(Wfxt _{+ Rfht - 1} + _bf )

g _t =σ(W _g x _t +R _g h _t-1 +b _g )

c _t = f _t ⊙ c _t-1 + i _t ⊙ g _t

o _t =σ(W _o x _t +R _o h _t-1 +b _o )

h _t = o _t ⊙ tanh(c _t )

Where _Wi , _Wf , _Wg , _Wo are input weight matrices, _Ri , _Rf , _Rg , _Ro are cyclic weight matrices, _bi , _bf , _bg , _bo are bias matrices, and the above matrices are adjusted through model training; σ is the sigmoid activation function, ⊙ is the dot multiplication operation; the output state _ht is directly used as the output of the long short-term memory unit;

S42: The offline model training described in S4 specifically includes:

The 5+k data in each row of the LSTM neural network input data described in S3 are input into each LSTM unit as the input data features described in S41, that is, the number of features C=5+k; the m+1-k rows described in S3 are respectively input into the G LSTM units of the LSTM unit layer described in S41, that is, the number of LSTM units G=m+1-k, and the backward error propagation algorithm is used to train the network.

As a preferred embodiment of the present invention:

In S5,

The input data composition construction method of S31 is used to transform the steel tube concrete sample data that needs to be predicted for axial compression load-strain curve into an input data composition that can be used for long short-term memory network.

Compared with the prior art, the present invention has the following beneficial effects:

The present invention uses design parameters, axial compression load and axial strain data, and designs data preprocessing and data construction methods for long short-term memory network offline model training; constructs an axial strain data sequence, and inputs it and the design parameters into the trained offline model to obtain axial compression load data output by the model; uses the axial strain data and axial compression load data to restore the axial compression load-strain curve, and realizes the prediction of the system axial compression load-strain curve.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is the curve prediction principle flow.

Figure 2 shows the input and output data structure.

Figure 3 is a long short-term memory unit layer.

Figure 4 is a long short-term memory unit.

Detailed ways

The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

The present invention proposes a method for predicting the axial compression load-strain curve of a steel tube concrete column based on a long short-term memory network. For design parameters, axial compression load and axial strain data, design data preprocessing and data composition methods are used for long short-term memory network offline model training; construct an axial strain data sequence, and input it and the design parameters into the trained offline model to obtain the axial compression load data output by the model; use the axial strain data and axial compression load data to restore the axial compression load-strain curve to achieve system axial compression load-strain curve prediction.

The curve prediction principle flow of the present invention is shown in FIG1 .

The present invention is achieved through the following technical solutions:

S1: Experimental data collection.

The experimental data of axial compression load-strain curves of steel tube concrete columns of different materials and sizes are collected, and a system database corresponding to the load-strain curve and five design parameters, namely, steel tube diameter D, steel tube wall thickness _ts , column height H, steel strength _fs , and concrete strength _fc , is established to construct a load-strain curve prediction training sample.

S2: Data preprocessing.

For all CFST column axial compression load-strain curve experimental data, 25000με is set as the limit for collecting axial strain to ensure that all load-strain curves capture the peak load point and post-peak curve characteristics. If the experimental data curve is short and does not reach the axial strain limit of 25000με, a compensation operation will be performed to extend the axial compression load-strain curve to the set axial strain limit according to the final slope of the curve. If the experimental data curve is long, a truncation operation will be performed to discard data that exceeds the set limit. After the above data preprocessing, all axial load-strain curve data have a uniform axial strain length.

S3: Data composition.

Each set of design parameters and the corresponding axial compression load-strain curve is taken as a data sample. The experimental data of the axial compression load-strain curve of the steel tube concrete column that has undergone data preprocessing is evenly divided into m points according to the strain value, and each point contains an axial compression load value N and an axial strain value ε. The input data composition of the long short-term memory unit neural network is constructed using the axial strain value and the design parameters. The output data composition of the long short-term memory unit neural network is then constructed using the axial compression load value. The input and output data composition of the present invention is shown in Figure 2, where the symbol i represents the i-th sample in the database.

S31: The input data composition includes: taking k axial strains as a group, and forming the first row of a data sample with the five design parameters of the sample; then taking another group of k axial strains using the sliding window method, and forming another row of a data sample with the five design parameters; and so on until all axial strains are taken, at which time the input data composition of each data sample includes m+1-k rows and 5+k columns.

S32: The output data composition includes: taking a group of k axial compressive load values to obtain each corresponding axial compressive load value, averaging the group of compressive load values to obtain the axial compressive load corresponding to this row of input data It is used as the output data of the neural network. At this time, the output data structure of each data sample contains m+1-k rows and 1 column.

S4: Offline model training.

After all samples are constructed through input and output data in S3, the long short-term memory unit neural network is used to perform offline model training on all samples to obtain a long short-term memory unit neural network model that meets the convergence requirements.

S41: The LSTM neural network includes in sequence: input layer, LSTM layer, fully connected layer, random dropout layer, fully connected layer, data regression layer. The LSTM layer is shown in FIG3 , which includes G LSTMs, that is, the input time series length is G, and the number of input data features is C. The LSTM is shown in FIG4 , which includes:

The current moment is t-1, and the output at time t needs to be calculated through the long short-term memory unit. The input data is x _t , and h _t and c _t represent the output (also called hidden state) and unit state at time t, respectively. The long short-term memory unit uses the current state (h _t-1 and c _t-1 ) to calculate the output h _t and the updated unit state c _t . The input gate i _t , forget gate f _t , candidate unit g _t and output gate o _t are used to control the update of the long short-term memory unit. The calculation formula is as follows:

_it =σ( _{Wixt + Rit -1} + _bi )

_ft = σ(Wfxt _{+ Rfht - 1} + _bf )

g _t =σ(W _g x _t +R _g h _t-1 +b _g )

c _t = f _t ⊙ c _t-1 + i _t ⊙ g _t

o _t =σ(W _o x _t +R _o h _t-1 +b _o )

h _t = o _t ⊙ tanh(c _t )

Where _Wi , _Wf , _Wg , _Wo are input weight matrices, _Ri , _Rf , _Rg , _Ro are recurrent weight matrices, and _bi , _bf , _bg , _bo are bias matrices that are adjusted through model training. σ is the sigmoid activation function, and ⊙ is the dot multiplication operation. The hidden state _ht is directly used as the output of the long short-term memory unit.

S42: The offline model training described in S4 specifically includes:

The 5+k data in each row of the LSTM neural network input data described in S3 are input into each LSTM as the input data features described in S41, that is, the number of features C=5+k; the m+1-k rows described in S3 are respectively input into the G LSTM units of the LSTM layer described in S41, that is, the number of LSTM units G=m+1-k. The network is trained using a backward error propagation algorithm so that the axial compressive load predicted by the LSTM network has the smallest root mean square error with the experimental value.

S5: Prediction of axial compression load-strain curve.

Arrange the design parameters of the steel tube concrete column that needs to predict the axial compression load-strain curve. Construct the axial strain data sequence, and take the vector with a maximum value of 25000με and m evenly distributed data values as the axial strain data sequence. Use the input data composition construction method described in S31 to transform the steel tube concrete sample data that needs to predict the axial compression load-strain curve into an input data composition that can be used for the long short-term memory network. Input the input data number into the offline model obtained in S4 to obtain the corresponding axial compression load output. Take the average value of all axial strains in each row of the input data and the axial compression load value in the corresponding row of the output data as a point on the axial compression load-strain plane, calculate all axial compression load-strain points in turn to restore the complete axial compression load-strain curve, and complete the axial compression load-strain curve prediction.

The present invention uses design parameters, axial compression load and axial strain data, and designs data preprocessing and data construction methods for long short-term memory network offline model training; constructs an axial strain data sequence, and inputs it and the design parameters into the trained offline model to obtain axial compression load data output by the model; uses the axial strain data and axial compression load data to restore the axial compression load-strain curve, and realizes the prediction of the system axial compression load-strain curve.

The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims (4)

1. The method for predicting the axial compressive load-strain curve of the concrete filled steel tube column based on the long-short-term memory network is characterized by comprising the following steps of: the method comprises the following steps:

s1: collecting experimental data;

collecting experimental data of axial compression load-strain curves of steel tube concrete columns with different materials and sizes, and establishing a load-strain curve and five design parameters, wherein the five design parameters are respectively as follows: diameter D of steel pipe and wall thickness t of steel pipe _s Column height H, steel strength f _s And concrete strength f _c Constructing a load-strain curve prediction training sample through design parameters according to the corresponding system database;

s2: preprocessing data;

setting 25000 mu epsilon as a limit value for collecting axial strain for all axial pressure load-strain curve experimental data of the concrete filled steel tube column, if the experimental data curve is shorter and does not reach the axial strain limit value of 25000 mu epsilon, executing compensation operation, and extending the axial pressure load-strain curve to the set axial strain limit value according to the final slope of the load-strain curve; if the experimental data curve is longer, the cut-off operation is executed, the data exceeding the set limit value is discarded, and after the data pretreatment, all axial load-strain curve data have uniform axial strain length;

s3: data constitution;

each set of design parameters and corresponding axle load-strain curve is used as a data sample; dividing experimental data of the axial pressure load-strain curve of the steel tube concrete column subjected to data pretreatment into m points according to the strain value, wherein each point comprises an axial pressure load value N and an axial strain value epsilon; the input data of the long-term and short-term memory unit neural network is built by using the axial strain value and the design parameter; building output data of the long-term memory unit neural network by using the axle load value;

s4: training an offline model;

after all samples are formed by input and output data in the step S3, performing offline model training on all samples by adopting a long-short-period memory unit neural network to obtain a long-short-period memory unit neural network model conforming to convergence;

s5: predicting an axle load-strain curve;

arranging design parameters of a concrete filled steel tube column of which the axial compression load-strain curve needs to be predicted; constructing an axial strain data sequence, and taking a vector with the maximum value of 25000 mu epsilon and containing m data values distributed evenly as the axial strain data sequence; the steel pipe concrete sample data which needs to be subjected to axle load-strain curve prediction is converted into input data constitution which can be used for a long-period and short-period memory network by utilizing the input data constitution construction method of S3; inputting the input data number into the offline model obtained in the step S4, and obtaining corresponding axle load output; taking the average value of all axial strains in each row of the input data and the axial pressure load value in the row corresponding to the output data as one point of an axial pressure load-strain plane, sequentially calculating all axial pressure load-strain points so as to restore a complete axial pressure load-strain curve, and completing the axial pressure load-strain curve prediction.

2. The method for predicting the axial compressive load-strain curve of the concrete filled steel tubular column based on the long-term memory network according to claim 1, which is characterized by comprising the following steps of:

in the step S3 of the method,

s31: the input data structure includes: taking k axial strains as a group, and forming a first row in a data sample by the k axial strains and five design parameters of the sample; then taking another group of k axial strains by adopting a sliding window method, and forming another row in one data sample by the group of k axial strains and the five design parameters; and so on until all axial strains are taken, at which time the input data for each data sample comprises m+1-k rows and 5+k columns;

s32: the output data structure includes: taking a group of k axle load values to obtain each corresponding axle load value, and averaging the group of axle load values to obtain the axle load corresponding to the line of input dataThe output data of each data sample is taken as the output data of the neural network, and the output data of each data sample comprises m+1-k rows and 1 column.

3. The method for predicting the axial compressive load-strain curve of the concrete filled steel tubular column based on the long-term memory network according to claim 1, which is characterized by comprising the following steps of:

in the step S4 of the process,

s41: the long-term and short-term memory unit neural network sequentially comprises: an input layer, a long-short-period memory unit layer, a full connection layer, a random discarding layer, a full connection layer and a data regression layer; the long-period memory unit layer comprises G long-period memory units, namely the input time sequence length is G, and the input data feature quantity is C; wherein the long-short-period memory unit comprises: at the current time of t-1, the output at the time of t needs to be calculated through a long-short-period memory unit, and the input data is x _t ，h _t And c _t Respectively representing the output and the cell state at time t; the long-short-period memory unit uses the current state h _t-1 And c _t-1 To calculate the output h _t And updated cell state c _t The method comprises the steps of carrying out a first treatment on the surface of the Using input gate i _t Forgetting door f _t Candidate unit g _t And an output gate o _t To control the update of the long-short-period memory unit, the calculation formula is as follows:

i _t ＝σ(W _i x _t +R _i h _t-1 +b _i )

f _t ＝σ(W _f x _t +R _f h _t-1 +b _f )

g _t ＝σ(W _g x _t +R _g h _t-1 +b _g )

c _t ＝f _t ⊙c _t-1 +i _t ⊙g _t

o _t ＝σ(W _o x _t +R _o h _t-1 +b _o )

h _t ＝o _t ⊙tanh(c _t )

wherein W is _i ,W _f ,W _g ,W _o Is an input weight matrix, R _i ,R _f ,R _g ,R _o Is a cyclic weight matrix, b _i ,b _f ,b _g ,b _o Is an offset matrix, and the matrix is adjusted through model training; sigma is a sigmoid activation function, and by; directly output state h _t As the output of the long-short-period memory unit;

s42: the offline model training in S4 specifically includes:

5+k data in each line of the long-period memory cell neural network input data in the step S3 are input into each long-period memory cell as the input data features in the step S41, namely the feature quantity C= 5+k; and (3) respectively inputting the m+1-k rows in the S3 into G long-short-period memory units of the long-short-period memory unit layer in the S41, namely, the number G=m+1-k of the long-short-period memory units, and training a network by adopting a backward error propagation algorithm.

4. The method for predicting the axial compressive load-strain curve of the concrete filled steel tubular column based on the long-term memory network according to claim 2, which is characterized by comprising the following steps of:

in the step S5 of the method,

the construction method of the input data composition of S31 is utilized to convert steel pipe concrete sample data which needs to be subjected to axle load-strain curve prediction into input data composition which can be used for long-period and short-period memory networks.